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 std::vector<Instruction*> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass((intptr_t)&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 Instruction *commonRemTransforms(BinaryOperator &I);
176 Instruction *commonIRemTransforms(BinaryOperator &I);
177 Instruction *commonDivTransforms(BinaryOperator &I);
178 Instruction *commonIDivTransforms(BinaryOperator &I);
179 Instruction *visitUDiv(BinaryOperator &I);
180 Instruction *visitSDiv(BinaryOperator &I);
181 Instruction *visitFDiv(BinaryOperator &I);
182 Instruction *visitAnd(BinaryOperator &I);
183 Instruction *visitOr (BinaryOperator &I);
184 Instruction *visitXor(BinaryOperator &I);
185 Instruction *visitShl(BinaryOperator &I);
186 Instruction *visitAShr(BinaryOperator &I);
187 Instruction *visitLShr(BinaryOperator &I);
188 Instruction *commonShiftTransforms(BinaryOperator &I);
189 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
191 Instruction *visitFCmpInst(FCmpInst &I);
192 Instruction *visitICmpInst(ICmpInst &I);
193 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
194 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
197 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
198 ConstantInt *DivRHS);
200 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
201 ICmpInst::Predicate Cond, Instruction &I);
202 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
204 Instruction *commonCastTransforms(CastInst &CI);
205 Instruction *commonIntCastTransforms(CastInst &CI);
206 Instruction *commonPointerCastTransforms(CastInst &CI);
207 Instruction *visitTrunc(TruncInst &CI);
208 Instruction *visitZExt(ZExtInst &CI);
209 Instruction *visitSExt(SExtInst &CI);
210 Instruction *visitFPTrunc(FPTruncInst &CI);
211 Instruction *visitFPExt(CastInst &CI);
212 Instruction *visitFPToUI(FPToUIInst &FI);
213 Instruction *visitFPToSI(FPToSIInst &FI);
214 Instruction *visitUIToFP(CastInst &CI);
215 Instruction *visitSIToFP(CastInst &CI);
216 Instruction *visitPtrToInt(CastInst &CI);
217 Instruction *visitIntToPtr(IntToPtrInst &CI);
218 Instruction *visitBitCast(BitCastInst &CI);
219 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
221 Instruction *visitSelectInst(SelectInst &CI);
222 Instruction *visitCallInst(CallInst &CI);
223 Instruction *visitInvokeInst(InvokeInst &II);
224 Instruction *visitPHINode(PHINode &PN);
225 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
226 Instruction *visitAllocationInst(AllocationInst &AI);
227 Instruction *visitFreeInst(FreeInst &FI);
228 Instruction *visitLoadInst(LoadInst &LI);
229 Instruction *visitStoreInst(StoreInst &SI);
230 Instruction *visitBranchInst(BranchInst &BI);
231 Instruction *visitSwitchInst(SwitchInst &SI);
232 Instruction *visitInsertElementInst(InsertElementInst &IE);
233 Instruction *visitExtractElementInst(ExtractElementInst &EI);
234 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
235 Instruction *visitExtractValueInst(ExtractValueInst &EV);
237 // visitInstruction - Specify what to return for unhandled instructions...
238 Instruction *visitInstruction(Instruction &I) { return 0; }
241 Instruction *visitCallSite(CallSite CS);
242 bool transformConstExprCastCall(CallSite CS);
243 Instruction *transformCallThroughTrampoline(CallSite CS);
244 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
245 bool DoXform = true);
246 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
249 // InsertNewInstBefore - insert an instruction New before instruction Old
250 // in the program. Add the new instruction to the worklist.
252 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
253 assert(New && New->getParent() == 0 &&
254 "New instruction already inserted into a basic block!");
255 BasicBlock *BB = Old.getParent();
256 BB->getInstList().insert(&Old, New); // Insert inst
261 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
262 /// This also adds the cast to the worklist. Finally, this returns the
264 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
266 if (V->getType() == Ty) return V;
268 if (Constant *CV = dyn_cast<Constant>(V))
269 return ConstantExpr::getCast(opc, CV, Ty);
271 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
276 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
277 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
281 // ReplaceInstUsesWith - This method is to be used when an instruction is
282 // found to be dead, replacable with another preexisting expression. Here
283 // we add all uses of I to the worklist, replace all uses of I with the new
284 // value, then return I, so that the inst combiner will know that I was
287 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
288 AddUsersToWorkList(I); // Add all modified instrs to worklist
290 I.replaceAllUsesWith(V);
293 // If we are replacing the instruction with itself, this must be in a
294 // segment of unreachable code, so just clobber the instruction.
295 I.replaceAllUsesWith(UndefValue::get(I.getType()));
300 // UpdateValueUsesWith - This method is to be used when an value is
301 // found to be replacable with another preexisting expression or was
302 // updated. Here we add all uses of I to the worklist, replace all uses of
303 // I with the new value (unless the instruction was just updated), then
304 // return true, so that the inst combiner will know that I was modified.
306 bool UpdateValueUsesWith(Value *Old, Value *New) {
307 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
309 Old->replaceAllUsesWith(New);
310 if (Instruction *I = dyn_cast<Instruction>(Old))
312 if (Instruction *I = dyn_cast<Instruction>(New))
317 // EraseInstFromFunction - When dealing with an instruction that has side
318 // effects or produces a void value, we can't rely on DCE to delete the
319 // instruction. Instead, visit methods should return the value returned by
321 Instruction *EraseInstFromFunction(Instruction &I) {
322 assert(I.use_empty() && "Cannot erase instruction that is used!");
323 AddUsesToWorkList(I);
324 RemoveFromWorkList(&I);
326 return 0; // Don't do anything with FI
329 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
330 APInt &KnownOne, unsigned Depth = 0) const {
331 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
334 bool MaskedValueIsZero(Value *V, const APInt &Mask,
335 unsigned Depth = 0) const {
336 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
338 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
339 return llvm::ComputeNumSignBits(Op, TD, Depth);
343 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
344 /// InsertBefore instruction. This is specialized a bit to avoid inserting
345 /// casts that are known to not do anything...
347 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
348 Value *V, const Type *DestTy,
349 Instruction *InsertBefore);
351 /// SimplifyCommutative - This performs a few simplifications for
352 /// commutative operators.
353 bool SimplifyCommutative(BinaryOperator &I);
355 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
356 /// most-complex to least-complex order.
357 bool SimplifyCompare(CmpInst &I);
359 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
360 /// on the demanded bits.
361 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
362 APInt& KnownZero, APInt& KnownOne,
365 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
366 uint64_t &UndefElts, unsigned Depth = 0);
368 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
369 // PHI node as operand #0, see if we can fold the instruction into the PHI
370 // (which is only possible if all operands to the PHI are constants).
371 Instruction *FoldOpIntoPhi(Instruction &I);
373 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
374 // operator and they all are only used by the PHI, PHI together their
375 // inputs, and do the operation once, to the result of the PHI.
376 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
377 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
380 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
381 ConstantInt *AndRHS, BinaryOperator &TheAnd);
383 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
384 bool isSub, Instruction &I);
385 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
386 bool isSigned, bool Inside, Instruction &IB);
387 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
388 Instruction *MatchBSwap(BinaryOperator &I);
389 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
390 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
391 Instruction *SimplifyMemSet(MemSetInst *MI);
394 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
396 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
398 int &NumCastsRemoved);
399 unsigned GetOrEnforceKnownAlignment(Value *V,
400 unsigned PrefAlign = 0);
405 char InstCombiner::ID = 0;
406 static RegisterPass<InstCombiner>
407 X("instcombine", "Combine redundant instructions");
409 // getComplexity: Assign a complexity or rank value to LLVM Values...
410 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
411 static unsigned getComplexity(Value *V) {
412 if (isa<Instruction>(V)) {
413 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
417 if (isa<Argument>(V)) return 3;
418 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
421 // isOnlyUse - Return true if this instruction will be deleted if we stop using
423 static bool isOnlyUse(Value *V) {
424 return V->hasOneUse() || isa<Constant>(V);
427 // getPromotedType - Return the specified type promoted as it would be to pass
428 // though a va_arg area...
429 static const Type *getPromotedType(const Type *Ty) {
430 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
431 if (ITy->getBitWidth() < 32)
432 return Type::Int32Ty;
437 /// getBitCastOperand - If the specified operand is a CastInst or a constant
438 /// expression bitcast, return the operand value, otherwise return null.
439 static Value *getBitCastOperand(Value *V) {
440 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
441 return I->getOperand(0);
442 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
443 if (CE->getOpcode() == Instruction::BitCast)
444 return CE->getOperand(0);
448 /// This function is a wrapper around CastInst::isEliminableCastPair. It
449 /// simply extracts arguments and returns what that function returns.
450 static Instruction::CastOps
451 isEliminableCastPair(
452 const CastInst *CI, ///< The first cast instruction
453 unsigned opcode, ///< The opcode of the second cast instruction
454 const Type *DstTy, ///< The target type for the second cast instruction
455 TargetData *TD ///< The target data for pointer size
458 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
459 const Type *MidTy = CI->getType(); // B from above
461 // Get the opcodes of the two Cast instructions
462 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
463 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
465 return Instruction::CastOps(
466 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
467 DstTy, TD->getIntPtrType()));
470 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
471 /// in any code being generated. It does not require codegen if V is simple
472 /// enough or if the cast can be folded into other casts.
473 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
474 const Type *Ty, TargetData *TD) {
475 if (V->getType() == Ty || isa<Constant>(V)) return false;
477 // If this is another cast that can be eliminated, it isn't codegen either.
478 if (const CastInst *CI = dyn_cast<CastInst>(V))
479 if (isEliminableCastPair(CI, opcode, Ty, TD))
484 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
485 /// InsertBefore instruction. This is specialized a bit to avoid inserting
486 /// casts that are known to not do anything...
488 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
489 Value *V, const Type *DestTy,
490 Instruction *InsertBefore) {
491 if (V->getType() == DestTy) return V;
492 if (Constant *C = dyn_cast<Constant>(V))
493 return ConstantExpr::getCast(opcode, C, DestTy);
495 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
498 // SimplifyCommutative - This performs a few simplifications for commutative
501 // 1. Order operands such that they are listed from right (least complex) to
502 // left (most complex). This puts constants before unary operators before
505 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
506 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
508 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
509 bool Changed = false;
510 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
511 Changed = !I.swapOperands();
513 if (!I.isAssociative()) return Changed;
514 Instruction::BinaryOps Opcode = I.getOpcode();
515 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
516 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
517 if (isa<Constant>(I.getOperand(1))) {
518 Constant *Folded = ConstantExpr::get(I.getOpcode(),
519 cast<Constant>(I.getOperand(1)),
520 cast<Constant>(Op->getOperand(1)));
521 I.setOperand(0, Op->getOperand(0));
522 I.setOperand(1, Folded);
524 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
525 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
526 isOnlyUse(Op) && isOnlyUse(Op1)) {
527 Constant *C1 = cast<Constant>(Op->getOperand(1));
528 Constant *C2 = cast<Constant>(Op1->getOperand(1));
530 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
531 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
532 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
536 I.setOperand(0, New);
537 I.setOperand(1, Folded);
544 /// SimplifyCompare - For a CmpInst this function just orders the operands
545 /// so that theyare listed from right (least complex) to left (most complex).
546 /// This puts constants before unary operators before binary operators.
547 bool InstCombiner::SimplifyCompare(CmpInst &I) {
548 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
551 // Compare instructions are not associative so there's nothing else we can do.
555 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
556 // if the LHS is a constant zero (which is the 'negate' form).
558 static inline Value *dyn_castNegVal(Value *V) {
559 if (BinaryOperator::isNeg(V))
560 return BinaryOperator::getNegArgument(V);
562 // Constants can be considered to be negated values if they can be folded.
563 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
564 return ConstantExpr::getNeg(C);
566 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
567 if (C->getType()->getElementType()->isInteger())
568 return ConstantExpr::getNeg(C);
573 static inline Value *dyn_castNotVal(Value *V) {
574 if (BinaryOperator::isNot(V))
575 return BinaryOperator::getNotArgument(V);
577 // Constants can be considered to be not'ed values...
578 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
579 return ConstantInt::get(~C->getValue());
583 // dyn_castFoldableMul - If this value is a multiply that can be folded into
584 // other computations (because it has a constant operand), return the
585 // non-constant operand of the multiply, and set CST to point to the multiplier.
586 // Otherwise, return null.
588 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
589 if (V->hasOneUse() && V->getType()->isInteger())
590 if (Instruction *I = dyn_cast<Instruction>(V)) {
591 if (I->getOpcode() == Instruction::Mul)
592 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
593 return I->getOperand(0);
594 if (I->getOpcode() == Instruction::Shl)
595 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
596 // The multiplier is really 1 << CST.
597 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
598 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
599 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
600 return I->getOperand(0);
606 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
607 /// expression, return it.
608 static User *dyn_castGetElementPtr(Value *V) {
609 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
611 if (CE->getOpcode() == Instruction::GetElementPtr)
612 return cast<User>(V);
616 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
617 /// opcode value. Otherwise return UserOp1.
618 static unsigned getOpcode(const Value *V) {
619 if (const Instruction *I = dyn_cast<Instruction>(V))
620 return I->getOpcode();
621 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
622 return CE->getOpcode();
623 // Use UserOp1 to mean there's no opcode.
624 return Instruction::UserOp1;
627 /// AddOne - Add one to a ConstantInt
628 static ConstantInt *AddOne(ConstantInt *C) {
629 APInt Val(C->getValue());
630 return ConstantInt::get(++Val);
632 /// SubOne - Subtract one from a ConstantInt
633 static ConstantInt *SubOne(ConstantInt *C) {
634 APInt Val(C->getValue());
635 return ConstantInt::get(--Val);
637 /// Add - Add two ConstantInts together
638 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
639 return ConstantInt::get(C1->getValue() + C2->getValue());
641 /// And - Bitwise AND two ConstantInts together
642 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
643 return ConstantInt::get(C1->getValue() & C2->getValue());
645 /// Subtract - Subtract one ConstantInt from another
646 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
647 return ConstantInt::get(C1->getValue() - C2->getValue());
649 /// Multiply - Multiply two ConstantInts together
650 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
651 return ConstantInt::get(C1->getValue() * C2->getValue());
653 /// MultiplyOverflows - True if the multiply can not be expressed in an int
655 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
656 uint32_t W = C1->getBitWidth();
657 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
666 APInt MulExt = LHSExt * RHSExt;
669 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
670 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
671 return MulExt.slt(Min) || MulExt.sgt(Max);
673 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
677 /// ShrinkDemandedConstant - Check to see if the specified operand of the
678 /// specified instruction is a constant integer. If so, check to see if there
679 /// are any bits set in the constant that are not demanded. If so, shrink the
680 /// constant and return true.
681 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
683 assert(I && "No instruction?");
684 assert(OpNo < I->getNumOperands() && "Operand index too large");
686 // If the operand is not a constant integer, nothing to do.
687 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
688 if (!OpC) return false;
690 // If there are no bits set that aren't demanded, nothing to do.
691 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
692 if ((~Demanded & OpC->getValue()) == 0)
695 // This instruction is producing bits that are not demanded. Shrink the RHS.
696 Demanded &= OpC->getValue();
697 I->setOperand(OpNo, ConstantInt::get(Demanded));
701 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
702 // set of known zero and one bits, compute the maximum and minimum values that
703 // could have the specified known zero and known one bits, returning them in
705 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
706 const APInt& KnownZero,
707 const APInt& KnownOne,
708 APInt& Min, APInt& Max) {
709 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
710 assert(KnownZero.getBitWidth() == BitWidth &&
711 KnownOne.getBitWidth() == BitWidth &&
712 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
713 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
714 APInt UnknownBits = ~(KnownZero|KnownOne);
716 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
717 // bit if it is unknown.
719 Max = KnownOne|UnknownBits;
721 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
723 Max.clear(BitWidth-1);
727 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
728 // a set of known zero and one bits, compute the maximum and minimum values that
729 // could have the specified known zero and known one bits, returning them in
731 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
732 const APInt &KnownZero,
733 const APInt &KnownOne,
734 APInt &Min, APInt &Max) {
735 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
736 assert(KnownZero.getBitWidth() == BitWidth &&
737 KnownOne.getBitWidth() == BitWidth &&
738 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
739 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
740 APInt UnknownBits = ~(KnownZero|KnownOne);
742 // The minimum value is when the unknown bits are all zeros.
744 // The maximum value is when the unknown bits are all ones.
745 Max = KnownOne|UnknownBits;
748 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
749 /// value based on the demanded bits. When this function is called, it is known
750 /// that only the bits set in DemandedMask of the result of V are ever used
751 /// downstream. Consequently, depending on the mask and V, it may be possible
752 /// to replace V with a constant or one of its operands. In such cases, this
753 /// function does the replacement and returns true. In all other cases, it
754 /// returns false after analyzing the expression and setting KnownOne and known
755 /// to be one in the expression. KnownZero contains all the bits that are known
756 /// to be zero in the expression. These are provided to potentially allow the
757 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
758 /// the expression. KnownOne and KnownZero always follow the invariant that
759 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
760 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
761 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
762 /// and KnownOne must all be the same.
763 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
764 APInt& KnownZero, APInt& KnownOne,
766 assert(V != 0 && "Null pointer of Value???");
767 assert(Depth <= 6 && "Limit Search Depth");
768 uint32_t BitWidth = DemandedMask.getBitWidth();
769 const IntegerType *VTy = cast<IntegerType>(V->getType());
770 assert(VTy->getBitWidth() == BitWidth &&
771 KnownZero.getBitWidth() == BitWidth &&
772 KnownOne.getBitWidth() == BitWidth &&
773 "Value *V, DemandedMask, KnownZero and KnownOne \
774 must have same BitWidth");
775 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
776 // We know all of the bits for a constant!
777 KnownOne = CI->getValue() & DemandedMask;
778 KnownZero = ~KnownOne & DemandedMask;
784 if (!V->hasOneUse()) { // Other users may use these bits.
785 if (Depth != 0) { // Not at the root.
786 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
787 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
790 // If this is the root being simplified, allow it to have multiple uses,
791 // just set the DemandedMask to all bits.
792 DemandedMask = APInt::getAllOnesValue(BitWidth);
793 } else if (DemandedMask == 0) { // Not demanding any bits from V.
794 if (V != UndefValue::get(VTy))
795 return UpdateValueUsesWith(V, UndefValue::get(VTy));
797 } else if (Depth == 6) { // Limit search depth.
801 Instruction *I = dyn_cast<Instruction>(V);
802 if (!I) return false; // Only analyze instructions.
804 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
805 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
806 switch (I->getOpcode()) {
808 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
810 case Instruction::And:
811 // If either the LHS or the RHS are Zero, the result is zero.
812 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
813 RHSKnownZero, RHSKnownOne, Depth+1))
815 assert((RHSKnownZero & RHSKnownOne) == 0 &&
816 "Bits known to be one AND zero?");
818 // If something is known zero on the RHS, the bits aren't demanded on the
820 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
821 LHSKnownZero, LHSKnownOne, Depth+1))
823 assert((LHSKnownZero & LHSKnownOne) == 0 &&
824 "Bits known to be one AND zero?");
826 // If all of the demanded bits are known 1 on one side, return the other.
827 // These bits cannot contribute to the result of the 'and'.
828 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
829 (DemandedMask & ~LHSKnownZero))
830 return UpdateValueUsesWith(I, I->getOperand(0));
831 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
832 (DemandedMask & ~RHSKnownZero))
833 return UpdateValueUsesWith(I, I->getOperand(1));
835 // If all of the demanded bits in the inputs are known zeros, return zero.
836 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
837 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
839 // If the RHS is a constant, see if we can simplify it.
840 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
841 return UpdateValueUsesWith(I, I);
843 // Output known-1 bits are only known if set in both the LHS & RHS.
844 RHSKnownOne &= LHSKnownOne;
845 // Output known-0 are known to be clear if zero in either the LHS | RHS.
846 RHSKnownZero |= LHSKnownZero;
848 case Instruction::Or:
849 // If either the LHS or the RHS are One, the result is One.
850 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
851 RHSKnownZero, RHSKnownOne, Depth+1))
853 assert((RHSKnownZero & RHSKnownOne) == 0 &&
854 "Bits known to be one AND zero?");
855 // If something is known one on the RHS, the bits aren't demanded on the
857 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
858 LHSKnownZero, LHSKnownOne, Depth+1))
860 assert((LHSKnownZero & LHSKnownOne) == 0 &&
861 "Bits known to be one AND zero?");
863 // If all of the demanded bits are known zero on one side, return the other.
864 // These bits cannot contribute to the result of the 'or'.
865 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
866 (DemandedMask & ~LHSKnownOne))
867 return UpdateValueUsesWith(I, I->getOperand(0));
868 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
869 (DemandedMask & ~RHSKnownOne))
870 return UpdateValueUsesWith(I, I->getOperand(1));
872 // If all of the potentially set bits on one side are known to be set on
873 // the other side, just use the 'other' side.
874 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
875 (DemandedMask & (~RHSKnownZero)))
876 return UpdateValueUsesWith(I, I->getOperand(0));
877 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
878 (DemandedMask & (~LHSKnownZero)))
879 return UpdateValueUsesWith(I, I->getOperand(1));
881 // If the RHS is a constant, see if we can simplify it.
882 if (ShrinkDemandedConstant(I, 1, DemandedMask))
883 return UpdateValueUsesWith(I, I);
885 // Output known-0 bits are only known if clear in both the LHS & RHS.
886 RHSKnownZero &= LHSKnownZero;
887 // Output known-1 are known to be set if set in either the LHS | RHS.
888 RHSKnownOne |= LHSKnownOne;
890 case Instruction::Xor: {
891 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
892 RHSKnownZero, RHSKnownOne, Depth+1))
894 assert((RHSKnownZero & RHSKnownOne) == 0 &&
895 "Bits known to be one AND zero?");
896 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
897 LHSKnownZero, LHSKnownOne, Depth+1))
899 assert((LHSKnownZero & LHSKnownOne) == 0 &&
900 "Bits known to be one AND zero?");
902 // If all of the demanded bits are known zero on one side, return the other.
903 // These bits cannot contribute to the result of the 'xor'.
904 if ((DemandedMask & RHSKnownZero) == DemandedMask)
905 return UpdateValueUsesWith(I, I->getOperand(0));
906 if ((DemandedMask & LHSKnownZero) == DemandedMask)
907 return UpdateValueUsesWith(I, I->getOperand(1));
909 // Output known-0 bits are known if clear or set in both the LHS & RHS.
910 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
911 (RHSKnownOne & LHSKnownOne);
912 // Output known-1 are known to be set if set in only one of the LHS, RHS.
913 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
914 (RHSKnownOne & LHSKnownZero);
916 // If all of the demanded bits are known to be zero on one side or the
917 // other, turn this into an *inclusive* or.
918 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
919 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
921 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
923 InsertNewInstBefore(Or, *I);
924 return UpdateValueUsesWith(I, Or);
927 // If all of the demanded bits on one side are known, and all of the set
928 // bits on that side are also known to be set on the other side, turn this
929 // into an AND, as we know the bits will be cleared.
930 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
931 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
933 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
934 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
936 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
937 InsertNewInstBefore(And, *I);
938 return UpdateValueUsesWith(I, And);
942 // If the RHS is a constant, see if we can simplify it.
943 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
944 if (ShrinkDemandedConstant(I, 1, DemandedMask))
945 return UpdateValueUsesWith(I, I);
947 RHSKnownZero = KnownZeroOut;
948 RHSKnownOne = KnownOneOut;
951 case Instruction::Select:
952 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
953 RHSKnownZero, RHSKnownOne, Depth+1))
955 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
956 LHSKnownZero, LHSKnownOne, Depth+1))
958 assert((RHSKnownZero & RHSKnownOne) == 0 &&
959 "Bits known to be one AND zero?");
960 assert((LHSKnownZero & LHSKnownOne) == 0 &&
961 "Bits known to be one AND zero?");
963 // If the operands are constants, see if we can simplify them.
964 if (ShrinkDemandedConstant(I, 1, DemandedMask))
965 return UpdateValueUsesWith(I, I);
966 if (ShrinkDemandedConstant(I, 2, DemandedMask))
967 return UpdateValueUsesWith(I, I);
969 // Only known if known in both the LHS and RHS.
970 RHSKnownOne &= LHSKnownOne;
971 RHSKnownZero &= LHSKnownZero;
973 case Instruction::Trunc: {
975 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
976 DemandedMask.zext(truncBf);
977 RHSKnownZero.zext(truncBf);
978 RHSKnownOne.zext(truncBf);
979 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
980 RHSKnownZero, RHSKnownOne, Depth+1))
982 DemandedMask.trunc(BitWidth);
983 RHSKnownZero.trunc(BitWidth);
984 RHSKnownOne.trunc(BitWidth);
985 assert((RHSKnownZero & RHSKnownOne) == 0 &&
986 "Bits known to be one AND zero?");
989 case Instruction::BitCast:
990 if (!I->getOperand(0)->getType()->isInteger())
993 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
994 RHSKnownZero, RHSKnownOne, Depth+1))
996 assert((RHSKnownZero & RHSKnownOne) == 0 &&
997 "Bits known to be one AND zero?");
999 case Instruction::ZExt: {
1000 // Compute the bits in the result that are not present in the input.
1001 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1002 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1004 DemandedMask.trunc(SrcBitWidth);
1005 RHSKnownZero.trunc(SrcBitWidth);
1006 RHSKnownOne.trunc(SrcBitWidth);
1007 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1008 RHSKnownZero, RHSKnownOne, Depth+1))
1010 DemandedMask.zext(BitWidth);
1011 RHSKnownZero.zext(BitWidth);
1012 RHSKnownOne.zext(BitWidth);
1013 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1014 "Bits known to be one AND zero?");
1015 // The top bits are known to be zero.
1016 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1019 case Instruction::SExt: {
1020 // Compute the bits in the result that are not present in the input.
1021 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1022 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1024 APInt InputDemandedBits = DemandedMask &
1025 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1027 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1028 // If any of the sign extended bits are demanded, we know that the sign
1030 if ((NewBits & DemandedMask) != 0)
1031 InputDemandedBits.set(SrcBitWidth-1);
1033 InputDemandedBits.trunc(SrcBitWidth);
1034 RHSKnownZero.trunc(SrcBitWidth);
1035 RHSKnownOne.trunc(SrcBitWidth);
1036 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1037 RHSKnownZero, RHSKnownOne, Depth+1))
1039 InputDemandedBits.zext(BitWidth);
1040 RHSKnownZero.zext(BitWidth);
1041 RHSKnownOne.zext(BitWidth);
1042 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1043 "Bits known to be one AND zero?");
1045 // If the sign bit of the input is known set or clear, then we know the
1046 // top bits of the result.
1048 // If the input sign bit is known zero, or if the NewBits are not demanded
1049 // convert this into a zero extension.
1050 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1052 // Convert to ZExt cast
1053 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1054 return UpdateValueUsesWith(I, NewCast);
1055 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1056 RHSKnownOne |= NewBits;
1060 case Instruction::Add: {
1061 // Figure out what the input bits are. If the top bits of the and result
1062 // are not demanded, then the add doesn't demand them from its input
1064 uint32_t NLZ = DemandedMask.countLeadingZeros();
1066 // If there is a constant on the RHS, there are a variety of xformations
1068 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1069 // If null, this should be simplified elsewhere. Some of the xforms here
1070 // won't work if the RHS is zero.
1074 // If the top bit of the output is demanded, demand everything from the
1075 // input. Otherwise, we demand all the input bits except NLZ top bits.
1076 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1078 // Find information about known zero/one bits in the input.
1079 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1080 LHSKnownZero, LHSKnownOne, Depth+1))
1083 // If the RHS of the add has bits set that can't affect the input, reduce
1085 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1086 return UpdateValueUsesWith(I, I);
1088 // Avoid excess work.
1089 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1092 // Turn it into OR if input bits are zero.
1093 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1095 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1097 InsertNewInstBefore(Or, *I);
1098 return UpdateValueUsesWith(I, Or);
1101 // We can say something about the output known-zero and known-one bits,
1102 // depending on potential carries from the input constant and the
1103 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1104 // bits set and the RHS constant is 0x01001, then we know we have a known
1105 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1107 // To compute this, we first compute the potential carry bits. These are
1108 // the bits which may be modified. I'm not aware of a better way to do
1110 const APInt& RHSVal = RHS->getValue();
1111 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1113 // Now that we know which bits have carries, compute the known-1/0 sets.
1115 // Bits are known one if they are known zero in one operand and one in the
1116 // other, and there is no input carry.
1117 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1118 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1120 // Bits are known zero if they are known zero in both operands and there
1121 // is no input carry.
1122 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1124 // If the high-bits of this ADD are not demanded, then it does not demand
1125 // the high bits of its LHS or RHS.
1126 if (DemandedMask[BitWidth-1] == 0) {
1127 // Right fill the mask of bits for this ADD to demand the most
1128 // significant bit and all those below it.
1129 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1130 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1131 LHSKnownZero, LHSKnownOne, Depth+1))
1133 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1134 LHSKnownZero, LHSKnownOne, Depth+1))
1140 case Instruction::Sub:
1141 // If the high-bits of this SUB are not demanded, then it does not demand
1142 // the high bits of its LHS or RHS.
1143 if (DemandedMask[BitWidth-1] == 0) {
1144 // Right fill the mask of bits for this SUB to demand the most
1145 // significant bit and all those below it.
1146 uint32_t NLZ = DemandedMask.countLeadingZeros();
1147 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1148 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1149 LHSKnownZero, LHSKnownOne, Depth+1))
1151 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1152 LHSKnownZero, LHSKnownOne, Depth+1))
1155 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1156 // the known zeros and ones.
1157 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1159 case Instruction::Shl:
1160 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1161 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1162 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1163 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1164 RHSKnownZero, RHSKnownOne, Depth+1))
1166 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1167 "Bits known to be one AND zero?");
1168 RHSKnownZero <<= ShiftAmt;
1169 RHSKnownOne <<= ShiftAmt;
1170 // low bits known zero.
1172 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1175 case Instruction::LShr:
1176 // For a logical shift right
1177 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1178 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1180 // Unsigned shift right.
1181 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1182 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1183 RHSKnownZero, RHSKnownOne, Depth+1))
1185 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1186 "Bits known to be one AND zero?");
1187 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1188 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1190 // Compute the new bits that are at the top now.
1191 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1192 RHSKnownZero |= HighBits; // high bits known zero.
1196 case Instruction::AShr:
1197 // If this is an arithmetic shift right and only the low-bit is set, we can
1198 // always convert this into a logical shr, even if the shift amount is
1199 // variable. The low bit of the shift cannot be an input sign bit unless
1200 // the shift amount is >= the size of the datatype, which is undefined.
1201 if (DemandedMask == 1) {
1202 // Perform the logical shift right.
1203 Value *NewVal = BinaryOperator::CreateLShr(
1204 I->getOperand(0), I->getOperand(1), I->getName());
1205 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1206 return UpdateValueUsesWith(I, NewVal);
1209 // If the sign bit is the only bit demanded by this ashr, then there is no
1210 // need to do it, the shift doesn't change the high bit.
1211 if (DemandedMask.isSignBit())
1212 return UpdateValueUsesWith(I, I->getOperand(0));
1214 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1215 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1217 // Signed shift right.
1218 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1219 // If any of the "high bits" are demanded, we should set the sign bit as
1221 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1222 DemandedMaskIn.set(BitWidth-1);
1223 if (SimplifyDemandedBits(I->getOperand(0),
1225 RHSKnownZero, RHSKnownOne, Depth+1))
1227 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1228 "Bits known to be one AND zero?");
1229 // Compute the new bits that are at the top now.
1230 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1231 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1232 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1234 // Handle the sign bits.
1235 APInt SignBit(APInt::getSignBit(BitWidth));
1236 // Adjust to where it is now in the mask.
1237 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1239 // If the input sign bit is known to be zero, or if none of the top bits
1240 // are demanded, turn this into an unsigned shift right.
1241 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1242 (HighBits & ~DemandedMask) == HighBits) {
1243 // Perform the logical shift right.
1244 Value *NewVal = BinaryOperator::CreateLShr(
1245 I->getOperand(0), SA, I->getName());
1246 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1247 return UpdateValueUsesWith(I, NewVal);
1248 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1249 RHSKnownOne |= HighBits;
1253 case Instruction::SRem:
1254 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1255 APInt RA = Rem->getValue();
1256 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1257 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1258 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1259 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1260 LHSKnownZero, LHSKnownOne, Depth+1))
1263 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1264 LHSKnownZero |= ~LowBits;
1265 else if (LHSKnownOne[BitWidth-1])
1266 LHSKnownOne |= ~LowBits;
1268 KnownZero |= LHSKnownZero & DemandedMask;
1269 KnownOne |= LHSKnownOne & DemandedMask;
1271 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1275 case Instruction::URem: {
1276 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1277 APInt RA = Rem->getValue();
1278 if (RA.isPowerOf2()) {
1279 APInt LowBits = (RA - 1);
1280 APInt Mask2 = LowBits & DemandedMask;
1281 KnownZero |= ~LowBits & DemandedMask;
1282 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1283 KnownZero, KnownOne, Depth+1))
1286 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1291 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1292 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1293 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1294 KnownZero2, KnownOne2, Depth+1))
1297 uint32_t Leaders = KnownZero2.countLeadingOnes();
1298 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1299 KnownZero2, KnownOne2, Depth+1))
1302 Leaders = std::max(Leaders,
1303 KnownZero2.countLeadingOnes());
1304 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1307 case Instruction::Call:
1308 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1309 switch (II->getIntrinsicID()) {
1311 case Intrinsic::bswap: {
1312 // If the only bits demanded come from one byte of the bswap result,
1313 // just shift the input byte into position to eliminate the bswap.
1314 unsigned NLZ = DemandedMask.countLeadingZeros();
1315 unsigned NTZ = DemandedMask.countTrailingZeros();
1317 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1318 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1319 // have 14 leading zeros, round to 8.
1322 // If we need exactly one byte, we can do this transformation.
1323 if (BitWidth-NLZ-NTZ == 8) {
1324 unsigned ResultBit = NTZ;
1325 unsigned InputBit = BitWidth-NTZ-8;
1327 // Replace this with either a left or right shift to get the byte into
1329 Instruction *NewVal;
1330 if (InputBit > ResultBit)
1331 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1332 ConstantInt::get(I->getType(), InputBit-ResultBit));
1334 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1335 ConstantInt::get(I->getType(), ResultBit-InputBit));
1336 NewVal->takeName(I);
1337 InsertNewInstBefore(NewVal, *I);
1338 return UpdateValueUsesWith(I, NewVal);
1341 // TODO: Could compute known zero/one bits based on the input.
1346 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1350 // If the client is only demanding bits that we know, return the known
1352 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1353 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1358 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1359 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1360 /// actually used by the caller. This method analyzes which elements of the
1361 /// operand are undef and returns that information in UndefElts.
1363 /// If the information about demanded elements can be used to simplify the
1364 /// operation, the operation is simplified, then the resultant value is
1365 /// returned. This returns null if no change was made.
1366 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1367 uint64_t &UndefElts,
1369 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1370 assert(VWidth <= 64 && "Vector too wide to analyze!");
1371 uint64_t EltMask = ~0ULL >> (64-VWidth);
1372 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1373 "Invalid DemandedElts!");
1375 if (isa<UndefValue>(V)) {
1376 // If the entire vector is undefined, just return this info.
1377 UndefElts = EltMask;
1379 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1380 UndefElts = EltMask;
1381 return UndefValue::get(V->getType());
1385 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1386 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1387 Constant *Undef = UndefValue::get(EltTy);
1389 std::vector<Constant*> Elts;
1390 for (unsigned i = 0; i != VWidth; ++i)
1391 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1392 Elts.push_back(Undef);
1393 UndefElts |= (1ULL << i);
1394 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1395 Elts.push_back(Undef);
1396 UndefElts |= (1ULL << i);
1397 } else { // Otherwise, defined.
1398 Elts.push_back(CP->getOperand(i));
1401 // If we changed the constant, return it.
1402 Constant *NewCP = ConstantVector::get(Elts);
1403 return NewCP != CP ? NewCP : 0;
1404 } else if (isa<ConstantAggregateZero>(V)) {
1405 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1407 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1408 Constant *Zero = Constant::getNullValue(EltTy);
1409 Constant *Undef = UndefValue::get(EltTy);
1410 std::vector<Constant*> Elts;
1411 for (unsigned i = 0; i != VWidth; ++i)
1412 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1413 UndefElts = DemandedElts ^ EltMask;
1414 return ConstantVector::get(Elts);
1417 if (!V->hasOneUse()) { // Other users may use these bits.
1418 if (Depth != 0) { // Not at the root.
1419 // TODO: Just compute the UndefElts information recursively.
1423 } else if (Depth == 10) { // Limit search depth.
1427 Instruction *I = dyn_cast<Instruction>(V);
1428 if (!I) return false; // Only analyze instructions.
1430 bool MadeChange = false;
1431 uint64_t UndefElts2;
1433 switch (I->getOpcode()) {
1436 case Instruction::InsertElement: {
1437 // If this is a variable index, we don't know which element it overwrites.
1438 // demand exactly the same input as we produce.
1439 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1441 // Note that we can't propagate undef elt info, because we don't know
1442 // which elt is getting updated.
1443 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1444 UndefElts2, Depth+1);
1445 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1449 // If this is inserting an element that isn't demanded, remove this
1451 unsigned IdxNo = Idx->getZExtValue();
1452 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1453 return AddSoonDeadInstToWorklist(*I, 0);
1455 // Otherwise, the element inserted overwrites whatever was there, so the
1456 // input demanded set is simpler than the output set.
1457 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1458 DemandedElts & ~(1ULL << IdxNo),
1459 UndefElts, Depth+1);
1460 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1462 // The inserted element is defined.
1463 UndefElts |= 1ULL << IdxNo;
1466 case Instruction::BitCast: {
1467 // Vector->vector casts only.
1468 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1470 unsigned InVWidth = VTy->getNumElements();
1471 uint64_t InputDemandedElts = 0;
1474 if (VWidth == InVWidth) {
1475 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1476 // elements as are demanded of us.
1478 InputDemandedElts = DemandedElts;
1479 } else if (VWidth > InVWidth) {
1483 // If there are more elements in the result than there are in the source,
1484 // then an input element is live if any of the corresponding output
1485 // elements are live.
1486 Ratio = VWidth/InVWidth;
1487 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1488 if (DemandedElts & (1ULL << OutIdx))
1489 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1495 // If there are more elements in the source than there are in the result,
1496 // then an input element is live if the corresponding output element is
1498 Ratio = InVWidth/VWidth;
1499 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1500 if (DemandedElts & (1ULL << InIdx/Ratio))
1501 InputDemandedElts |= 1ULL << InIdx;
1504 // div/rem demand all inputs, because they don't want divide by zero.
1505 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1506 UndefElts2, Depth+1);
1508 I->setOperand(0, TmpV);
1512 UndefElts = UndefElts2;
1513 if (VWidth > InVWidth) {
1514 assert(0 && "Unimp");
1515 // If there are more elements in the result than there are in the source,
1516 // then an output element is undef if the corresponding input element is
1518 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1519 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1520 UndefElts |= 1ULL << OutIdx;
1521 } else if (VWidth < InVWidth) {
1522 assert(0 && "Unimp");
1523 // If there are more elements in the source than there are in the result,
1524 // then a result element is undef if all of the corresponding input
1525 // elements are undef.
1526 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1527 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1528 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1529 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1533 case Instruction::And:
1534 case Instruction::Or:
1535 case Instruction::Xor:
1536 case Instruction::Add:
1537 case Instruction::Sub:
1538 case Instruction::Mul:
1539 // div/rem demand all inputs, because they don't want divide by zero.
1540 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1541 UndefElts, Depth+1);
1542 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1543 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1544 UndefElts2, Depth+1);
1545 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1547 // Output elements are undefined if both are undefined. Consider things
1548 // like undef&0. The result is known zero, not undef.
1549 UndefElts &= UndefElts2;
1552 case Instruction::Call: {
1553 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1555 switch (II->getIntrinsicID()) {
1558 // Binary vector operations that work column-wise. A dest element is a
1559 // function of the corresponding input elements from the two inputs.
1560 case Intrinsic::x86_sse_sub_ss:
1561 case Intrinsic::x86_sse_mul_ss:
1562 case Intrinsic::x86_sse_min_ss:
1563 case Intrinsic::x86_sse_max_ss:
1564 case Intrinsic::x86_sse2_sub_sd:
1565 case Intrinsic::x86_sse2_mul_sd:
1566 case Intrinsic::x86_sse2_min_sd:
1567 case Intrinsic::x86_sse2_max_sd:
1568 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1569 UndefElts, Depth+1);
1570 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1571 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1572 UndefElts2, Depth+1);
1573 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1575 // If only the low elt is demanded and this is a scalarizable intrinsic,
1576 // scalarize it now.
1577 if (DemandedElts == 1) {
1578 switch (II->getIntrinsicID()) {
1580 case Intrinsic::x86_sse_sub_ss:
1581 case Intrinsic::x86_sse_mul_ss:
1582 case Intrinsic::x86_sse2_sub_sd:
1583 case Intrinsic::x86_sse2_mul_sd:
1584 // TODO: Lower MIN/MAX/ABS/etc
1585 Value *LHS = II->getOperand(1);
1586 Value *RHS = II->getOperand(2);
1587 // Extract the element as scalars.
1588 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1589 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1591 switch (II->getIntrinsicID()) {
1592 default: assert(0 && "Case stmts out of sync!");
1593 case Intrinsic::x86_sse_sub_ss:
1594 case Intrinsic::x86_sse2_sub_sd:
1595 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1596 II->getName()), *II);
1598 case Intrinsic::x86_sse_mul_ss:
1599 case Intrinsic::x86_sse2_mul_sd:
1600 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1601 II->getName()), *II);
1606 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1608 InsertNewInstBefore(New, *II);
1609 AddSoonDeadInstToWorklist(*II, 0);
1614 // Output elements are undefined if both are undefined. Consider things
1615 // like undef&0. The result is known zero, not undef.
1616 UndefElts &= UndefElts2;
1622 return MadeChange ? I : 0;
1626 /// AssociativeOpt - Perform an optimization on an associative operator. This
1627 /// function is designed to check a chain of associative operators for a
1628 /// potential to apply a certain optimization. Since the optimization may be
1629 /// applicable if the expression was reassociated, this checks the chain, then
1630 /// reassociates the expression as necessary to expose the optimization
1631 /// opportunity. This makes use of a special Functor, which must define
1632 /// 'shouldApply' and 'apply' methods.
1634 template<typename Functor>
1635 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1636 unsigned Opcode = Root.getOpcode();
1637 Value *LHS = Root.getOperand(0);
1639 // Quick check, see if the immediate LHS matches...
1640 if (F.shouldApply(LHS))
1641 return F.apply(Root);
1643 // Otherwise, if the LHS is not of the same opcode as the root, return.
1644 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1645 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1646 // Should we apply this transform to the RHS?
1647 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1649 // If not to the RHS, check to see if we should apply to the LHS...
1650 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1651 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1655 // If the functor wants to apply the optimization to the RHS of LHSI,
1656 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1658 // Now all of the instructions are in the current basic block, go ahead
1659 // and perform the reassociation.
1660 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1662 // First move the selected RHS to the LHS of the root...
1663 Root.setOperand(0, LHSI->getOperand(1));
1665 // Make what used to be the LHS of the root be the user of the root...
1666 Value *ExtraOperand = TmpLHSI->getOperand(1);
1667 if (&Root == TmpLHSI) {
1668 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1671 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1672 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1673 BasicBlock::iterator ARI = &Root; ++ARI;
1674 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1677 // Now propagate the ExtraOperand down the chain of instructions until we
1679 while (TmpLHSI != LHSI) {
1680 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1681 // Move the instruction to immediately before the chain we are
1682 // constructing to avoid breaking dominance properties.
1683 NextLHSI->moveBefore(ARI);
1686 Value *NextOp = NextLHSI->getOperand(1);
1687 NextLHSI->setOperand(1, ExtraOperand);
1689 ExtraOperand = NextOp;
1692 // Now that the instructions are reassociated, have the functor perform
1693 // the transformation...
1694 return F.apply(Root);
1697 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1704 // AddRHS - Implements: X + X --> X << 1
1707 AddRHS(Value *rhs) : RHS(rhs) {}
1708 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1709 Instruction *apply(BinaryOperator &Add) const {
1710 return BinaryOperator::CreateShl(Add.getOperand(0),
1711 ConstantInt::get(Add.getType(), 1));
1715 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1717 struct AddMaskingAnd {
1719 AddMaskingAnd(Constant *c) : C2(c) {}
1720 bool shouldApply(Value *LHS) const {
1722 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1723 ConstantExpr::getAnd(C1, C2)->isNullValue();
1725 Instruction *apply(BinaryOperator &Add) const {
1726 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1732 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1734 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1735 if (Constant *SOC = dyn_cast<Constant>(SO))
1736 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1738 return IC->InsertNewInstBefore(CastInst::Create(
1739 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1742 // Figure out if the constant is the left or the right argument.
1743 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1744 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1746 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1748 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1749 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1752 Value *Op0 = SO, *Op1 = ConstOperand;
1754 std::swap(Op0, Op1);
1756 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1757 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1758 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1759 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1760 SO->getName()+".cmp");
1762 assert(0 && "Unknown binary instruction type!");
1765 return IC->InsertNewInstBefore(New, I);
1768 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1769 // constant as the other operand, try to fold the binary operator into the
1770 // select arguments. This also works for Cast instructions, which obviously do
1771 // not have a second operand.
1772 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1774 // Don't modify shared select instructions
1775 if (!SI->hasOneUse()) return 0;
1776 Value *TV = SI->getOperand(1);
1777 Value *FV = SI->getOperand(2);
1779 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1780 // Bool selects with constant operands can be folded to logical ops.
1781 if (SI->getType() == Type::Int1Ty) return 0;
1783 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1784 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1786 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1793 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1794 /// node as operand #0, see if we can fold the instruction into the PHI (which
1795 /// is only possible if all operands to the PHI are constants).
1796 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1797 PHINode *PN = cast<PHINode>(I.getOperand(0));
1798 unsigned NumPHIValues = PN->getNumIncomingValues();
1799 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1801 // Check to see if all of the operands of the PHI are constants. If there is
1802 // one non-constant value, remember the BB it is. If there is more than one
1803 // or if *it* is a PHI, bail out.
1804 BasicBlock *NonConstBB = 0;
1805 for (unsigned i = 0; i != NumPHIValues; ++i)
1806 if (!isa<Constant>(PN->getIncomingValue(i))) {
1807 if (NonConstBB) return 0; // More than one non-const value.
1808 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1809 NonConstBB = PN->getIncomingBlock(i);
1811 // If the incoming non-constant value is in I's block, we have an infinite
1813 if (NonConstBB == I.getParent())
1817 // If there is exactly one non-constant value, we can insert a copy of the
1818 // operation in that block. However, if this is a critical edge, we would be
1819 // inserting the computation one some other paths (e.g. inside a loop). Only
1820 // do this if the pred block is unconditionally branching into the phi block.
1822 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1823 if (!BI || !BI->isUnconditional()) return 0;
1826 // Okay, we can do the transformation: create the new PHI node.
1827 PHINode *NewPN = PHINode::Create(I.getType(), "");
1828 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1829 InsertNewInstBefore(NewPN, *PN);
1830 NewPN->takeName(PN);
1832 // Next, add all of the operands to the PHI.
1833 if (I.getNumOperands() == 2) {
1834 Constant *C = cast<Constant>(I.getOperand(1));
1835 for (unsigned i = 0; i != NumPHIValues; ++i) {
1837 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1838 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1839 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1841 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1843 assert(PN->getIncomingBlock(i) == NonConstBB);
1844 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1845 InV = BinaryOperator::Create(BO->getOpcode(),
1846 PN->getIncomingValue(i), C, "phitmp",
1847 NonConstBB->getTerminator());
1848 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1849 InV = CmpInst::Create(CI->getOpcode(),
1851 PN->getIncomingValue(i), C, "phitmp",
1852 NonConstBB->getTerminator());
1854 assert(0 && "Unknown binop!");
1856 AddToWorkList(cast<Instruction>(InV));
1858 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1861 CastInst *CI = cast<CastInst>(&I);
1862 const Type *RetTy = CI->getType();
1863 for (unsigned i = 0; i != NumPHIValues; ++i) {
1865 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1866 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1868 assert(PN->getIncomingBlock(i) == NonConstBB);
1869 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1870 I.getType(), "phitmp",
1871 NonConstBB->getTerminator());
1872 AddToWorkList(cast<Instruction>(InV));
1874 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1877 return ReplaceInstUsesWith(I, NewPN);
1881 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1882 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1883 /// This basically requires proving that the add in the original type would not
1884 /// overflow to change the sign bit or have a carry out.
1885 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1886 // There are different heuristics we can use for this. Here are some simple
1889 // Add has the property that adding any two 2's complement numbers can only
1890 // have one carry bit which can change a sign. As such, if LHS and RHS each
1891 // have at least two sign bits, we know that the addition of the two values will
1892 // sign extend fine.
1893 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1897 // If one of the operands only has one non-zero bit, and if the other operand
1898 // has a known-zero bit in a more significant place than it (not including the
1899 // sign bit) the ripple may go up to and fill the zero, but won't change the
1900 // sign. For example, (X & ~4) + 1.
1908 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1909 bool Changed = SimplifyCommutative(I);
1910 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1912 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1913 // X + undef -> undef
1914 if (isa<UndefValue>(RHS))
1915 return ReplaceInstUsesWith(I, RHS);
1918 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1919 if (RHSC->isNullValue())
1920 return ReplaceInstUsesWith(I, LHS);
1921 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1922 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1923 (I.getType())->getValueAPF()))
1924 return ReplaceInstUsesWith(I, LHS);
1927 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1928 // X + (signbit) --> X ^ signbit
1929 const APInt& Val = CI->getValue();
1930 uint32_t BitWidth = Val.getBitWidth();
1931 if (Val == APInt::getSignBit(BitWidth))
1932 return BinaryOperator::CreateXor(LHS, RHS);
1934 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1935 // (X & 254)+1 -> (X&254)|1
1936 if (!isa<VectorType>(I.getType())) {
1937 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1938 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1939 KnownZero, KnownOne))
1944 if (isa<PHINode>(LHS))
1945 if (Instruction *NV = FoldOpIntoPhi(I))
1948 ConstantInt *XorRHS = 0;
1950 if (isa<ConstantInt>(RHSC) &&
1951 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1952 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1953 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1955 uint32_t Size = TySizeBits / 2;
1956 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1957 APInt CFF80Val(-C0080Val);
1959 if (TySizeBits > Size) {
1960 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1961 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1962 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1963 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1964 // This is a sign extend if the top bits are known zero.
1965 if (!MaskedValueIsZero(XorLHS,
1966 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1967 Size = 0; // Not a sign ext, but can't be any others either.
1972 C0080Val = APIntOps::lshr(C0080Val, Size);
1973 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1974 } while (Size >= 1);
1976 // FIXME: This shouldn't be necessary. When the backends can handle types
1977 // with funny bit widths then this switch statement should be removed. It
1978 // is just here to get the size of the "middle" type back up to something
1979 // that the back ends can handle.
1980 const Type *MiddleType = 0;
1983 case 32: MiddleType = Type::Int32Ty; break;
1984 case 16: MiddleType = Type::Int16Ty; break;
1985 case 8: MiddleType = Type::Int8Ty; break;
1988 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1989 InsertNewInstBefore(NewTrunc, I);
1990 return new SExtInst(NewTrunc, I.getType(), I.getName());
1995 if (I.getType() == Type::Int1Ty)
1996 return BinaryOperator::CreateXor(LHS, RHS);
1999 if (I.getType()->isInteger()) {
2000 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2002 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2003 if (RHSI->getOpcode() == Instruction::Sub)
2004 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2005 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2007 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2008 if (LHSI->getOpcode() == Instruction::Sub)
2009 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2010 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2015 // -A + -B --> -(A + B)
2016 if (Value *LHSV = dyn_castNegVal(LHS)) {
2017 if (LHS->getType()->isIntOrIntVector()) {
2018 if (Value *RHSV = dyn_castNegVal(RHS)) {
2019 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2020 InsertNewInstBefore(NewAdd, I);
2021 return BinaryOperator::CreateNeg(NewAdd);
2025 return BinaryOperator::CreateSub(RHS, LHSV);
2029 if (!isa<Constant>(RHS))
2030 if (Value *V = dyn_castNegVal(RHS))
2031 return BinaryOperator::CreateSub(LHS, V);
2035 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2036 if (X == RHS) // X*C + X --> X * (C+1)
2037 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2039 // X*C1 + X*C2 --> X * (C1+C2)
2041 if (X == dyn_castFoldableMul(RHS, C1))
2042 return BinaryOperator::CreateMul(X, Add(C1, C2));
2045 // X + X*C --> X * (C+1)
2046 if (dyn_castFoldableMul(RHS, C2) == LHS)
2047 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2049 // X + ~X --> -1 since ~X = -X-1
2050 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2051 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2054 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2055 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2056 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2059 // A+B --> A|B iff A and B have no bits set in common.
2060 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2061 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2062 APInt LHSKnownOne(IT->getBitWidth(), 0);
2063 APInt LHSKnownZero(IT->getBitWidth(), 0);
2064 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2065 if (LHSKnownZero != 0) {
2066 APInt RHSKnownOne(IT->getBitWidth(), 0);
2067 APInt RHSKnownZero(IT->getBitWidth(), 0);
2068 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2070 // No bits in common -> bitwise or.
2071 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2072 return BinaryOperator::CreateOr(LHS, RHS);
2076 // W*X + Y*Z --> W * (X+Z) iff W == Y
2077 if (I.getType()->isIntOrIntVector()) {
2078 Value *W, *X, *Y, *Z;
2079 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2080 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2084 } else if (Y == X) {
2086 } else if (X == Z) {
2093 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2094 LHS->getName()), I);
2095 return BinaryOperator::CreateMul(W, NewAdd);
2100 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2102 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2103 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2105 // (X & FF00) + xx00 -> (X+xx00) & FF00
2106 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2107 Constant *Anded = And(CRHS, C2);
2108 if (Anded == CRHS) {
2109 // See if all bits from the first bit set in the Add RHS up are included
2110 // in the mask. First, get the rightmost bit.
2111 const APInt& AddRHSV = CRHS->getValue();
2113 // Form a mask of all bits from the lowest bit added through the top.
2114 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2116 // See if the and mask includes all of these bits.
2117 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2119 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2120 // Okay, the xform is safe. Insert the new add pronto.
2121 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2122 LHS->getName()), I);
2123 return BinaryOperator::CreateAnd(NewAdd, C2);
2128 // Try to fold constant add into select arguments.
2129 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2130 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2134 // add (cast *A to intptrtype) B ->
2135 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2137 CastInst *CI = dyn_cast<CastInst>(LHS);
2140 CI = dyn_cast<CastInst>(RHS);
2143 if (CI && CI->getType()->isSized() &&
2144 (CI->getType()->getPrimitiveSizeInBits() ==
2145 TD->getIntPtrType()->getPrimitiveSizeInBits())
2146 && isa<PointerType>(CI->getOperand(0)->getType())) {
2148 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2149 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2150 PointerType::get(Type::Int8Ty, AS), I);
2151 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2152 return new PtrToIntInst(I2, CI->getType());
2156 // add (select X 0 (sub n A)) A --> select X A n
2158 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2161 SI = dyn_cast<SelectInst>(RHS);
2164 if (SI && SI->hasOneUse()) {
2165 Value *TV = SI->getTrueValue();
2166 Value *FV = SI->getFalseValue();
2169 // Can we fold the add into the argument of the select?
2170 // We check both true and false select arguments for a matching subtract.
2171 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2172 A == Other) // Fold the add into the true select value.
2173 return SelectInst::Create(SI->getCondition(), N, A);
2174 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2175 A == Other) // Fold the add into the false select value.
2176 return SelectInst::Create(SI->getCondition(), A, N);
2180 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2181 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2182 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2183 return ReplaceInstUsesWith(I, LHS);
2185 // Check for (add (sext x), y), see if we can merge this into an
2186 // integer add followed by a sext.
2187 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2188 // (add (sext x), cst) --> (sext (add x, cst'))
2189 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2191 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2192 if (LHSConv->hasOneUse() &&
2193 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2194 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2195 // Insert the new, smaller add.
2196 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2198 InsertNewInstBefore(NewAdd, I);
2199 return new SExtInst(NewAdd, I.getType());
2203 // (add (sext x), (sext y)) --> (sext (add int x, y))
2204 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2205 // Only do this if x/y have the same type, if at last one of them has a
2206 // single use (so we don't increase the number of sexts), and if the
2207 // integer add will not overflow.
2208 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2209 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2210 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2211 RHSConv->getOperand(0))) {
2212 // Insert the new integer add.
2213 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2214 RHSConv->getOperand(0),
2216 InsertNewInstBefore(NewAdd, I);
2217 return new SExtInst(NewAdd, I.getType());
2222 // Check for (add double (sitofp x), y), see if we can merge this into an
2223 // integer add followed by a promotion.
2224 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2225 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2226 // ... if the constant fits in the integer value. This is useful for things
2227 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2228 // requires a constant pool load, and generally allows the add to be better
2230 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2232 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2233 if (LHSConv->hasOneUse() &&
2234 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2235 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2236 // Insert the new integer add.
2237 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2239 InsertNewInstBefore(NewAdd, I);
2240 return new SIToFPInst(NewAdd, I.getType());
2244 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2245 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2246 // Only do this if x/y have the same type, if at last one of them has a
2247 // single use (so we don't increase the number of int->fp conversions),
2248 // and if the integer add will not overflow.
2249 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2250 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2251 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2252 RHSConv->getOperand(0))) {
2253 // Insert the new integer add.
2254 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2255 RHSConv->getOperand(0),
2257 InsertNewInstBefore(NewAdd, I);
2258 return new SIToFPInst(NewAdd, I.getType());
2263 return Changed ? &I : 0;
2266 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2267 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2269 if (Op0 == Op1) // sub X, X -> 0
2270 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2272 // If this is a 'B = x-(-A)', change to B = x+A...
2273 if (Value *V = dyn_castNegVal(Op1))
2274 return BinaryOperator::CreateAdd(Op0, V);
2276 if (isa<UndefValue>(Op0))
2277 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2278 if (isa<UndefValue>(Op1))
2279 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2281 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2282 // Replace (-1 - A) with (~A)...
2283 if (C->isAllOnesValue())
2284 return BinaryOperator::CreateNot(Op1);
2286 // C - ~X == X + (1+C)
2288 if (match(Op1, m_Not(m_Value(X))))
2289 return BinaryOperator::CreateAdd(X, AddOne(C));
2291 // -(X >>u 31) -> (X >>s 31)
2292 // -(X >>s 31) -> (X >>u 31)
2294 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2295 if (SI->getOpcode() == Instruction::LShr) {
2296 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2297 // Check to see if we are shifting out everything but the sign bit.
2298 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2299 SI->getType()->getPrimitiveSizeInBits()-1) {
2300 // Ok, the transformation is safe. Insert AShr.
2301 return BinaryOperator::Create(Instruction::AShr,
2302 SI->getOperand(0), CU, SI->getName());
2306 else if (SI->getOpcode() == Instruction::AShr) {
2307 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2308 // Check to see if we are shifting out everything but the sign bit.
2309 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2310 SI->getType()->getPrimitiveSizeInBits()-1) {
2311 // Ok, the transformation is safe. Insert LShr.
2312 return BinaryOperator::CreateLShr(
2313 SI->getOperand(0), CU, SI->getName());
2320 // Try to fold constant sub into select arguments.
2321 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2322 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2325 if (isa<PHINode>(Op0))
2326 if (Instruction *NV = FoldOpIntoPhi(I))
2330 if (I.getType() == Type::Int1Ty)
2331 return BinaryOperator::CreateXor(Op0, Op1);
2333 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2334 if (Op1I->getOpcode() == Instruction::Add &&
2335 !Op0->getType()->isFPOrFPVector()) {
2336 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2337 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2338 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2339 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2340 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2341 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2342 // C1-(X+C2) --> (C1-C2)-X
2343 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2344 Op1I->getOperand(0));
2348 if (Op1I->hasOneUse()) {
2349 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2350 // is not used by anyone else...
2352 if (Op1I->getOpcode() == Instruction::Sub &&
2353 !Op1I->getType()->isFPOrFPVector()) {
2354 // Swap the two operands of the subexpr...
2355 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2356 Op1I->setOperand(0, IIOp1);
2357 Op1I->setOperand(1, IIOp0);
2359 // Create the new top level add instruction...
2360 return BinaryOperator::CreateAdd(Op0, Op1);
2363 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2365 if (Op1I->getOpcode() == Instruction::And &&
2366 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2367 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2370 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2371 return BinaryOperator::CreateAnd(Op0, NewNot);
2374 // 0 - (X sdiv C) -> (X sdiv -C)
2375 if (Op1I->getOpcode() == Instruction::SDiv)
2376 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2378 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2379 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2380 ConstantExpr::getNeg(DivRHS));
2382 // X - X*C --> X * (1-C)
2383 ConstantInt *C2 = 0;
2384 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2385 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2386 return BinaryOperator::CreateMul(Op0, CP1);
2389 // X - ((X / Y) * Y) --> X % Y
2390 if (Op1I->getOpcode() == Instruction::Mul)
2391 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2392 if (Op0 == I->getOperand(0) &&
2393 Op1I->getOperand(1) == I->getOperand(1)) {
2394 if (I->getOpcode() == Instruction::SDiv)
2395 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2396 if (I->getOpcode() == Instruction::UDiv)
2397 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2402 if (!Op0->getType()->isFPOrFPVector())
2403 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2404 if (Op0I->getOpcode() == Instruction::Add) {
2405 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2406 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2407 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2408 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2409 } else if (Op0I->getOpcode() == Instruction::Sub) {
2410 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2411 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2416 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2417 if (X == Op1) // X*C - X --> X * (C-1)
2418 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2420 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2421 if (X == dyn_castFoldableMul(Op1, C2))
2422 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2427 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2428 /// comparison only checks the sign bit. If it only checks the sign bit, set
2429 /// TrueIfSigned if the result of the comparison is true when the input value is
2431 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2432 bool &TrueIfSigned) {
2434 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2435 TrueIfSigned = true;
2436 return RHS->isZero();
2437 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2438 TrueIfSigned = true;
2439 return RHS->isAllOnesValue();
2440 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2441 TrueIfSigned = false;
2442 return RHS->isAllOnesValue();
2443 case ICmpInst::ICMP_UGT:
2444 // True if LHS u> RHS and RHS == high-bit-mask - 1
2445 TrueIfSigned = true;
2446 return RHS->getValue() ==
2447 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2448 case ICmpInst::ICMP_UGE:
2449 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2450 TrueIfSigned = true;
2451 return RHS->getValue().isSignBit();
2457 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2458 bool Changed = SimplifyCommutative(I);
2459 Value *Op0 = I.getOperand(0);
2461 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2462 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2464 // Simplify mul instructions with a constant RHS...
2465 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2466 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2468 // ((X << C1)*C2) == (X * (C2 << C1))
2469 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2470 if (SI->getOpcode() == Instruction::Shl)
2471 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2472 return BinaryOperator::CreateMul(SI->getOperand(0),
2473 ConstantExpr::getShl(CI, ShOp));
2476 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2477 if (CI->equalsInt(1)) // X * 1 == X
2478 return ReplaceInstUsesWith(I, Op0);
2479 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2480 return BinaryOperator::CreateNeg(Op0, I.getName());
2482 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2483 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2484 return BinaryOperator::CreateShl(Op0,
2485 ConstantInt::get(Op0->getType(), Val.logBase2()));
2487 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2488 if (Op1F->isNullValue())
2489 return ReplaceInstUsesWith(I, Op1);
2491 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2492 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2493 // We need a better interface for long double here.
2494 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2495 if (Op1F->isExactlyValue(1.0))
2496 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2499 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2500 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2501 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2502 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2503 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2505 InsertNewInstBefore(Add, I);
2506 Value *C1C2 = ConstantExpr::getMul(Op1,
2507 cast<Constant>(Op0I->getOperand(1)));
2508 return BinaryOperator::CreateAdd(Add, C1C2);
2512 // Try to fold constant mul into select arguments.
2513 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2514 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2517 if (isa<PHINode>(Op0))
2518 if (Instruction *NV = FoldOpIntoPhi(I))
2522 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2523 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2524 return BinaryOperator::CreateMul(Op0v, Op1v);
2526 if (I.getType() == Type::Int1Ty)
2527 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2529 // If one of the operands of the multiply is a cast from a boolean value, then
2530 // we know the bool is either zero or one, so this is a 'masking' multiply.
2531 // See if we can simplify things based on how the boolean was originally
2533 CastInst *BoolCast = 0;
2534 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2535 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2538 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2539 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2542 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2543 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2544 const Type *SCOpTy = SCIOp0->getType();
2547 // If the icmp is true iff the sign bit of X is set, then convert this
2548 // multiply into a shift/and combination.
2549 if (isa<ConstantInt>(SCIOp1) &&
2550 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2552 // Shift the X value right to turn it into "all signbits".
2553 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2554 SCOpTy->getPrimitiveSizeInBits()-1);
2556 InsertNewInstBefore(
2557 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2558 BoolCast->getOperand(0)->getName()+
2561 // If the multiply type is not the same as the source type, sign extend
2562 // or truncate to the multiply type.
2563 if (I.getType() != V->getType()) {
2564 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2565 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2566 Instruction::CastOps opcode =
2567 (SrcBits == DstBits ? Instruction::BitCast :
2568 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2569 V = InsertCastBefore(opcode, V, I.getType(), I);
2572 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2573 return BinaryOperator::CreateAnd(V, OtherOp);
2578 return Changed ? &I : 0;
2581 /// This function implements the transforms on div instructions that work
2582 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2583 /// used by the visitors to those instructions.
2584 /// @brief Transforms common to all three div instructions
2585 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2586 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2588 // undef / X -> 0 for integer.
2589 // undef / X -> undef for FP (the undef could be a snan).
2590 if (isa<UndefValue>(Op0)) {
2591 if (Op0->getType()->isFPOrFPVector())
2592 return ReplaceInstUsesWith(I, Op0);
2593 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2596 // X / undef -> undef
2597 if (isa<UndefValue>(Op1))
2598 return ReplaceInstUsesWith(I, Op1);
2600 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2601 // This does not apply for fdiv.
2602 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2603 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2604 // the same basic block, then we replace the select with Y, and the
2605 // condition of the select with false (if the cond value is in the same BB).
2606 // If the select has uses other than the div, this allows them to be
2607 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2608 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2609 if (ST->isNullValue()) {
2610 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2611 if (CondI && CondI->getParent() == I.getParent())
2612 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2613 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2614 I.setOperand(1, SI->getOperand(2));
2616 UpdateValueUsesWith(SI, SI->getOperand(2));
2620 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2621 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2622 if (ST->isNullValue()) {
2623 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2624 if (CondI && CondI->getParent() == I.getParent())
2625 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2626 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2627 I.setOperand(1, SI->getOperand(1));
2629 UpdateValueUsesWith(SI, SI->getOperand(1));
2637 /// This function implements the transforms common to both integer division
2638 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2639 /// division instructions.
2640 /// @brief Common integer divide transforms
2641 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2642 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2644 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2646 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2647 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2648 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2649 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2652 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2653 return ReplaceInstUsesWith(I, CI);
2656 if (Instruction *Common = commonDivTransforms(I))
2659 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2661 if (RHS->equalsInt(1))
2662 return ReplaceInstUsesWith(I, Op0);
2664 // (X / C1) / C2 -> X / (C1*C2)
2665 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2666 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2667 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2668 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2669 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2671 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2672 Multiply(RHS, LHSRHS));
2675 if (!RHS->isZero()) { // avoid X udiv 0
2676 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2677 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2679 if (isa<PHINode>(Op0))
2680 if (Instruction *NV = FoldOpIntoPhi(I))
2685 // 0 / X == 0, we don't need to preserve faults!
2686 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2687 if (LHS->equalsInt(0))
2688 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2690 // It can't be division by zero, hence it must be division by one.
2691 if (I.getType() == Type::Int1Ty)
2692 return ReplaceInstUsesWith(I, Op0);
2697 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2698 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2700 // Handle the integer div common cases
2701 if (Instruction *Common = commonIDivTransforms(I))
2704 // X udiv C^2 -> X >> C
2705 // Check to see if this is an unsigned division with an exact power of 2,
2706 // if so, convert to a right shift.
2707 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2708 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2709 return BinaryOperator::CreateLShr(Op0,
2710 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2713 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2714 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2715 if (RHSI->getOpcode() == Instruction::Shl &&
2716 isa<ConstantInt>(RHSI->getOperand(0))) {
2717 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2718 if (C1.isPowerOf2()) {
2719 Value *N = RHSI->getOperand(1);
2720 const Type *NTy = N->getType();
2721 if (uint32_t C2 = C1.logBase2()) {
2722 Constant *C2V = ConstantInt::get(NTy, C2);
2723 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2725 return BinaryOperator::CreateLShr(Op0, N);
2730 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2731 // where C1&C2 are powers of two.
2732 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2733 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2734 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2735 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2736 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2737 // Compute the shift amounts
2738 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2739 // Construct the "on true" case of the select
2740 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2741 Instruction *TSI = BinaryOperator::CreateLShr(
2742 Op0, TC, SI->getName()+".t");
2743 TSI = InsertNewInstBefore(TSI, I);
2745 // Construct the "on false" case of the select
2746 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2747 Instruction *FSI = BinaryOperator::CreateLShr(
2748 Op0, FC, SI->getName()+".f");
2749 FSI = InsertNewInstBefore(FSI, I);
2751 // construct the select instruction and return it.
2752 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2758 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2759 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2761 // Handle the integer div common cases
2762 if (Instruction *Common = commonIDivTransforms(I))
2765 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2767 if (RHS->isAllOnesValue())
2768 return BinaryOperator::CreateNeg(Op0);
2771 if (Value *LHSNeg = dyn_castNegVal(Op0))
2772 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2775 // If the sign bits of both operands are zero (i.e. we can prove they are
2776 // unsigned inputs), turn this into a udiv.
2777 if (I.getType()->isInteger()) {
2778 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2779 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2780 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2781 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2788 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2789 return commonDivTransforms(I);
2792 /// This function implements the transforms on rem instructions that work
2793 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2794 /// is used by the visitors to those instructions.
2795 /// @brief Transforms common to all three rem instructions
2796 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2797 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2799 // 0 % X == 0 for integer, we don't need to preserve faults!
2800 if (Constant *LHS = dyn_cast<Constant>(Op0))
2801 if (LHS->isNullValue())
2802 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2804 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2805 if (I.getType()->isFPOrFPVector())
2806 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2807 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2809 if (isa<UndefValue>(Op1))
2810 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2812 // Handle cases involving: rem X, (select Cond, Y, Z)
2813 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2814 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2815 // the same basic block, then we replace the select with Y, and the
2816 // condition of the select with false (if the cond value is in the same
2817 // BB). If the select has uses other than the div, this allows them to be
2819 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2820 if (ST->isNullValue()) {
2821 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2822 if (CondI && CondI->getParent() == I.getParent())
2823 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2824 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2825 I.setOperand(1, SI->getOperand(2));
2827 UpdateValueUsesWith(SI, SI->getOperand(2));
2830 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2831 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2832 if (ST->isNullValue()) {
2833 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2834 if (CondI && CondI->getParent() == I.getParent())
2835 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2836 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2837 I.setOperand(1, SI->getOperand(1));
2839 UpdateValueUsesWith(SI, SI->getOperand(1));
2847 /// This function implements the transforms common to both integer remainder
2848 /// instructions (urem and srem). It is called by the visitors to those integer
2849 /// remainder instructions.
2850 /// @brief Common integer remainder transforms
2851 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2852 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2854 if (Instruction *common = commonRemTransforms(I))
2857 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2858 // X % 0 == undef, we don't need to preserve faults!
2859 if (RHS->equalsInt(0))
2860 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2862 if (RHS->equalsInt(1)) // X % 1 == 0
2863 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2865 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2866 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2867 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2869 } else if (isa<PHINode>(Op0I)) {
2870 if (Instruction *NV = FoldOpIntoPhi(I))
2874 // See if we can fold away this rem instruction.
2875 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2876 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2877 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2878 KnownZero, KnownOne))
2886 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2887 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2889 if (Instruction *common = commonIRemTransforms(I))
2892 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2893 // X urem C^2 -> X and C
2894 // Check to see if this is an unsigned remainder with an exact power of 2,
2895 // if so, convert to a bitwise and.
2896 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2897 if (C->getValue().isPowerOf2())
2898 return BinaryOperator::CreateAnd(Op0, SubOne(C));
2901 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2902 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2903 if (RHSI->getOpcode() == Instruction::Shl &&
2904 isa<ConstantInt>(RHSI->getOperand(0))) {
2905 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2906 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2907 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
2909 return BinaryOperator::CreateAnd(Op0, Add);
2914 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2915 // where C1&C2 are powers of two.
2916 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2917 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2918 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2919 // STO == 0 and SFO == 0 handled above.
2920 if ((STO->getValue().isPowerOf2()) &&
2921 (SFO->getValue().isPowerOf2())) {
2922 Value *TrueAnd = InsertNewInstBefore(
2923 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2924 Value *FalseAnd = InsertNewInstBefore(
2925 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2926 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
2934 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2935 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2937 // Handle the integer rem common cases
2938 if (Instruction *common = commonIRemTransforms(I))
2941 if (Value *RHSNeg = dyn_castNegVal(Op1))
2942 if (!isa<ConstantInt>(RHSNeg) ||
2943 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2945 AddUsesToWorkList(I);
2946 I.setOperand(1, RHSNeg);
2950 // If the sign bits of both operands are zero (i.e. we can prove they are
2951 // unsigned inputs), turn this into a urem.
2952 if (I.getType()->isInteger()) {
2953 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2954 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2955 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2956 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
2963 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2964 return commonRemTransforms(I);
2967 // isMaxValueMinusOne - return true if this is Max-1
2968 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2969 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2971 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2972 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2975 // isMinValuePlusOne - return true if this is Min+1
2976 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2978 return C->getValue() == 1; // unsigned
2980 // Calculate 1111111111000000000000
2981 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2982 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2985 // isOneBitSet - Return true if there is exactly one bit set in the specified
2987 static bool isOneBitSet(const ConstantInt *CI) {
2988 return CI->getValue().isPowerOf2();
2991 // isHighOnes - Return true if the constant is of the form 1+0+.
2992 // This is the same as lowones(~X).
2993 static bool isHighOnes(const ConstantInt *CI) {
2994 return (~CI->getValue() + 1).isPowerOf2();
2997 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2998 /// are carefully arranged to allow folding of expressions such as:
3000 /// (A < B) | (A > B) --> (A != B)
3002 /// Note that this is only valid if the first and second predicates have the
3003 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3005 /// Three bits are used to represent the condition, as follows:
3010 /// <=> Value Definition
3011 /// 000 0 Always false
3018 /// 111 7 Always true
3020 static unsigned getICmpCode(const ICmpInst *ICI) {
3021 switch (ICI->getPredicate()) {
3023 case ICmpInst::ICMP_UGT: return 1; // 001
3024 case ICmpInst::ICMP_SGT: return 1; // 001
3025 case ICmpInst::ICMP_EQ: return 2; // 010
3026 case ICmpInst::ICMP_UGE: return 3; // 011
3027 case ICmpInst::ICMP_SGE: return 3; // 011
3028 case ICmpInst::ICMP_ULT: return 4; // 100
3029 case ICmpInst::ICMP_SLT: return 4; // 100
3030 case ICmpInst::ICMP_NE: return 5; // 101
3031 case ICmpInst::ICMP_ULE: return 6; // 110
3032 case ICmpInst::ICMP_SLE: return 6; // 110
3035 assert(0 && "Invalid ICmp predicate!");
3040 /// getICmpValue - This is the complement of getICmpCode, which turns an
3041 /// opcode and two operands into either a constant true or false, or a brand
3042 /// new ICmp instruction. The sign is passed in to determine which kind
3043 /// of predicate to use in new icmp instructions.
3044 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3046 default: assert(0 && "Illegal ICmp code!");
3047 case 0: return ConstantInt::getFalse();
3050 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3052 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3053 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3056 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3058 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3061 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3063 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3064 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3067 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3069 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3070 case 7: return ConstantInt::getTrue();
3074 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3075 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3076 (ICmpInst::isSignedPredicate(p1) &&
3077 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3078 (ICmpInst::isSignedPredicate(p2) &&
3079 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3083 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3084 struct FoldICmpLogical {
3087 ICmpInst::Predicate pred;
3088 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3089 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3090 pred(ICI->getPredicate()) {}
3091 bool shouldApply(Value *V) const {
3092 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3093 if (PredicatesFoldable(pred, ICI->getPredicate()))
3094 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3095 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3098 Instruction *apply(Instruction &Log) const {
3099 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3100 if (ICI->getOperand(0) != LHS) {
3101 assert(ICI->getOperand(1) == LHS);
3102 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3105 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3106 unsigned LHSCode = getICmpCode(ICI);
3107 unsigned RHSCode = getICmpCode(RHSICI);
3109 switch (Log.getOpcode()) {
3110 case Instruction::And: Code = LHSCode & RHSCode; break;
3111 case Instruction::Or: Code = LHSCode | RHSCode; break;
3112 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3113 default: assert(0 && "Illegal logical opcode!"); return 0;
3116 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3117 ICmpInst::isSignedPredicate(ICI->getPredicate());
3119 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3120 if (Instruction *I = dyn_cast<Instruction>(RV))
3122 // Otherwise, it's a constant boolean value...
3123 return IC.ReplaceInstUsesWith(Log, RV);
3126 } // end anonymous namespace
3128 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3129 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3130 // guaranteed to be a binary operator.
3131 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3133 ConstantInt *AndRHS,
3134 BinaryOperator &TheAnd) {
3135 Value *X = Op->getOperand(0);
3136 Constant *Together = 0;
3138 Together = And(AndRHS, OpRHS);
3140 switch (Op->getOpcode()) {
3141 case Instruction::Xor:
3142 if (Op->hasOneUse()) {
3143 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3144 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3145 InsertNewInstBefore(And, TheAnd);
3147 return BinaryOperator::CreateXor(And, Together);
3150 case Instruction::Or:
3151 if (Together == AndRHS) // (X | C) & C --> C
3152 return ReplaceInstUsesWith(TheAnd, AndRHS);
3154 if (Op->hasOneUse() && Together != OpRHS) {
3155 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3156 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3157 InsertNewInstBefore(Or, TheAnd);
3159 return BinaryOperator::CreateAnd(Or, AndRHS);
3162 case Instruction::Add:
3163 if (Op->hasOneUse()) {
3164 // Adding a one to a single bit bit-field should be turned into an XOR
3165 // of the bit. First thing to check is to see if this AND is with a
3166 // single bit constant.
3167 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3169 // If there is only one bit set...
3170 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3171 // Ok, at this point, we know that we are masking the result of the
3172 // ADD down to exactly one bit. If the constant we are adding has
3173 // no bits set below this bit, then we can eliminate the ADD.
3174 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3176 // Check to see if any bits below the one bit set in AndRHSV are set.
3177 if ((AddRHS & (AndRHSV-1)) == 0) {
3178 // If not, the only thing that can effect the output of the AND is
3179 // the bit specified by AndRHSV. If that bit is set, the effect of
3180 // the XOR is to toggle the bit. If it is clear, then the ADD has
3182 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3183 TheAnd.setOperand(0, X);
3186 // Pull the XOR out of the AND.
3187 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3188 InsertNewInstBefore(NewAnd, TheAnd);
3189 NewAnd->takeName(Op);
3190 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3197 case Instruction::Shl: {
3198 // We know that the AND will not produce any of the bits shifted in, so if
3199 // the anded constant includes them, clear them now!
3201 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3202 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3203 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3204 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3206 if (CI->getValue() == ShlMask) {
3207 // Masking out bits that the shift already masks
3208 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3209 } else if (CI != AndRHS) { // Reducing bits set in and.
3210 TheAnd.setOperand(1, CI);
3215 case Instruction::LShr:
3217 // We know that the AND will not produce any of the bits shifted in, so if
3218 // the anded constant includes them, clear them now! This only applies to
3219 // unsigned shifts, because a signed shr may bring in set bits!
3221 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3222 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3223 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3224 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3226 if (CI->getValue() == ShrMask) {
3227 // Masking out bits that the shift already masks.
3228 return ReplaceInstUsesWith(TheAnd, Op);
3229 } else if (CI != AndRHS) {
3230 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3235 case Instruction::AShr:
3237 // See if this is shifting in some sign extension, then masking it out
3239 if (Op->hasOneUse()) {
3240 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3241 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3242 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3243 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3244 if (C == AndRHS) { // Masking out bits shifted in.
3245 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3246 // Make the argument unsigned.
3247 Value *ShVal = Op->getOperand(0);
3248 ShVal = InsertNewInstBefore(
3249 BinaryOperator::CreateLShr(ShVal, OpRHS,
3250 Op->getName()), TheAnd);
3251 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3260 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3261 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3262 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3263 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3264 /// insert new instructions.
3265 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3266 bool isSigned, bool Inside,
3268 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3269 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3270 "Lo is not <= Hi in range emission code!");
3273 if (Lo == Hi) // Trivially false.
3274 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3276 // V >= Min && V < Hi --> V < Hi
3277 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3278 ICmpInst::Predicate pred = (isSigned ?
3279 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3280 return new ICmpInst(pred, V, Hi);
3283 // Emit V-Lo <u Hi-Lo
3284 Constant *NegLo = ConstantExpr::getNeg(Lo);
3285 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3286 InsertNewInstBefore(Add, IB);
3287 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3288 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3291 if (Lo == Hi) // Trivially true.
3292 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3294 // V < Min || V >= Hi -> V > Hi-1
3295 Hi = SubOne(cast<ConstantInt>(Hi));
3296 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3297 ICmpInst::Predicate pred = (isSigned ?
3298 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3299 return new ICmpInst(pred, V, Hi);
3302 // Emit V-Lo >u Hi-1-Lo
3303 // Note that Hi has already had one subtracted from it, above.
3304 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3305 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3306 InsertNewInstBefore(Add, IB);
3307 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3308 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3311 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3312 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3313 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3314 // not, since all 1s are not contiguous.
3315 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3316 const APInt& V = Val->getValue();
3317 uint32_t BitWidth = Val->getType()->getBitWidth();
3318 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3320 // look for the first zero bit after the run of ones
3321 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3322 // look for the first non-zero bit
3323 ME = V.getActiveBits();
3327 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3328 /// where isSub determines whether the operator is a sub. If we can fold one of
3329 /// the following xforms:
3331 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3332 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3333 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3335 /// return (A +/- B).
3337 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3338 ConstantInt *Mask, bool isSub,
3340 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3341 if (!LHSI || LHSI->getNumOperands() != 2 ||
3342 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3344 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3346 switch (LHSI->getOpcode()) {
3348 case Instruction::And:
3349 if (And(N, Mask) == Mask) {
3350 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3351 if ((Mask->getValue().countLeadingZeros() +
3352 Mask->getValue().countPopulation()) ==
3353 Mask->getValue().getBitWidth())
3356 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3357 // part, we don't need any explicit masks to take them out of A. If that
3358 // is all N is, ignore it.
3359 uint32_t MB = 0, ME = 0;
3360 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3361 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3362 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3363 if (MaskedValueIsZero(RHS, Mask))
3368 case Instruction::Or:
3369 case Instruction::Xor:
3370 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3371 if ((Mask->getValue().countLeadingZeros() +
3372 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3373 && And(N, Mask)->isZero())
3380 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3382 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3383 return InsertNewInstBefore(New, I);
3386 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3387 bool Changed = SimplifyCommutative(I);
3388 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3390 if (isa<UndefValue>(Op1)) // X & undef -> 0
3391 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3395 return ReplaceInstUsesWith(I, Op1);
3397 // See if we can simplify any instructions used by the instruction whose sole
3398 // purpose is to compute bits we don't care about.
3399 if (!isa<VectorType>(I.getType())) {
3400 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3401 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3402 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3403 KnownZero, KnownOne))
3406 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3407 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3408 return ReplaceInstUsesWith(I, I.getOperand(0));
3409 } else if (isa<ConstantAggregateZero>(Op1)) {
3410 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3414 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3415 const APInt& AndRHSMask = AndRHS->getValue();
3416 APInt NotAndRHS(~AndRHSMask);
3418 // Optimize a variety of ((val OP C1) & C2) combinations...
3419 if (isa<BinaryOperator>(Op0)) {
3420 Instruction *Op0I = cast<Instruction>(Op0);
3421 Value *Op0LHS = Op0I->getOperand(0);
3422 Value *Op0RHS = Op0I->getOperand(1);
3423 switch (Op0I->getOpcode()) {
3424 case Instruction::Xor:
3425 case Instruction::Or:
3426 // If the mask is only needed on one incoming arm, push it up.
3427 if (Op0I->hasOneUse()) {
3428 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3429 // Not masking anything out for the LHS, move to RHS.
3430 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3431 Op0RHS->getName()+".masked");
3432 InsertNewInstBefore(NewRHS, I);
3433 return BinaryOperator::Create(
3434 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3436 if (!isa<Constant>(Op0RHS) &&
3437 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3438 // Not masking anything out for the RHS, move to LHS.
3439 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3440 Op0LHS->getName()+".masked");
3441 InsertNewInstBefore(NewLHS, I);
3442 return BinaryOperator::Create(
3443 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3448 case Instruction::Add:
3449 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3450 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3451 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3452 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3453 return BinaryOperator::CreateAnd(V, AndRHS);
3454 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3455 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3458 case Instruction::Sub:
3459 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3460 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3461 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3462 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3463 return BinaryOperator::CreateAnd(V, AndRHS);
3465 // (A - N) & AndRHS -> -N & AndRHS where A & AndRHS == 0
3466 if (Op0I->hasOneUse() && MaskedValueIsZero(Op0LHS, AndRHSMask)) {
3467 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3468 if (!A || !A->isZero()) {
3469 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3470 InsertNewInstBefore(NewNeg, I);
3471 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3477 case Instruction::Shl:
3478 case Instruction::LShr:
3479 // (1 << x) & 1 --> zext(x == 0)
3480 // (1 >> x) & 1 --> zext(x == 0)
3481 if (AndRHSMask.getLimitedValue() == 1 && Op0LHS == AndRHS) {
3482 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3483 Constant::getNullValue(I.getType()));
3484 InsertNewInstBefore(NewICmp, I);
3485 return new ZExtInst(NewICmp, I.getType());
3490 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3491 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3493 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3494 // If this is an integer truncation or change from signed-to-unsigned, and
3495 // if the source is an and/or with immediate, transform it. This
3496 // frequently occurs for bitfield accesses.
3497 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3498 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3499 CastOp->getNumOperands() == 2)
3500 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3501 if (CastOp->getOpcode() == Instruction::And) {
3502 // Change: and (cast (and X, C1) to T), C2
3503 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3504 // This will fold the two constants together, which may allow
3505 // other simplifications.
3506 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3507 CastOp->getOperand(0), I.getType(),
3508 CastOp->getName()+".shrunk");
3509 NewCast = InsertNewInstBefore(NewCast, I);
3510 // trunc_or_bitcast(C1)&C2
3511 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3512 C3 = ConstantExpr::getAnd(C3, AndRHS);
3513 return BinaryOperator::CreateAnd(NewCast, C3);
3514 } else if (CastOp->getOpcode() == Instruction::Or) {
3515 // Change: and (cast (or X, C1) to T), C2
3516 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3517 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3518 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3519 return ReplaceInstUsesWith(I, AndRHS);
3525 // Try to fold constant and into select arguments.
3526 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3527 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3529 if (isa<PHINode>(Op0))
3530 if (Instruction *NV = FoldOpIntoPhi(I))
3534 Value *Op0NotVal = dyn_castNotVal(Op0);
3535 Value *Op1NotVal = dyn_castNotVal(Op1);
3537 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3538 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3540 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3541 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3542 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3543 I.getName()+".demorgan");
3544 InsertNewInstBefore(Or, I);
3545 return BinaryOperator::CreateNot(Or);
3549 Value *A = 0, *B = 0, *C = 0, *D = 0;
3550 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3551 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3552 return ReplaceInstUsesWith(I, Op1);
3554 // (A|B) & ~(A&B) -> A^B
3555 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3556 if ((A == C && B == D) || (A == D && B == C))
3557 return BinaryOperator::CreateXor(A, B);
3561 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3562 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3563 return ReplaceInstUsesWith(I, Op0);
3565 // ~(A&B) & (A|B) -> A^B
3566 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3567 if ((A == C && B == D) || (A == D && B == C))
3568 return BinaryOperator::CreateXor(A, B);
3572 if (Op0->hasOneUse() &&
3573 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3574 if (A == Op1) { // (A^B)&A -> A&(A^B)
3575 I.swapOperands(); // Simplify below
3576 std::swap(Op0, Op1);
3577 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3578 cast<BinaryOperator>(Op0)->swapOperands();
3579 I.swapOperands(); // Simplify below
3580 std::swap(Op0, Op1);
3583 if (Op1->hasOneUse() &&
3584 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3585 if (B == Op0) { // B&(A^B) -> B&(B^A)
3586 cast<BinaryOperator>(Op1)->swapOperands();
3589 if (A == Op0) { // A&(A^B) -> A & ~B
3590 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3591 InsertNewInstBefore(NotB, I);
3592 return BinaryOperator::CreateAnd(A, NotB);
3598 { // (icmp ugt/ult A, C) & (icmp B, C) --> (icmp (A|B), C)
3599 // where C is a power of 2
3601 ConstantInt *C1, *C2;
3602 ICmpInst::Predicate LHSCC, RHSCC;
3603 if (match(&I, m_And(m_ICmp(LHSCC, m_Value(A), m_ConstantInt(C1)),
3604 m_ICmp(RHSCC, m_Value(B), m_ConstantInt(C2)))))
3605 if (C1 == C2 && LHSCC == RHSCC && C1->getValue().isPowerOf2() &&
3606 (LHSCC == ICmpInst::ICMP_ULT || LHSCC == ICmpInst::ICMP_UGT)) {
3607 Instruction *NewOr = BinaryOperator::CreateOr(A, B);
3608 InsertNewInstBefore(NewOr, I);
3609 return new ICmpInst(LHSCC, NewOr, C1);
3613 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3614 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3615 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3618 Value *LHSVal, *RHSVal;
3619 ConstantInt *LHSCst, *RHSCst;
3620 ICmpInst::Predicate LHSCC, RHSCC;
3621 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3622 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3623 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3624 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3625 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3626 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3627 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3628 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3630 // Don't try to fold ICMP_SLT + ICMP_ULT.
3631 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3632 ICmpInst::isSignedPredicate(LHSCC) ==
3633 ICmpInst::isSignedPredicate(RHSCC))) {
3634 // Ensure that the larger constant is on the RHS.
3635 ICmpInst::Predicate GT;
3636 if (ICmpInst::isSignedPredicate(LHSCC) ||
3637 (ICmpInst::isEquality(LHSCC) &&
3638 ICmpInst::isSignedPredicate(RHSCC)))
3639 GT = ICmpInst::ICMP_SGT;
3641 GT = ICmpInst::ICMP_UGT;
3643 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3644 ICmpInst *LHS = cast<ICmpInst>(Op0);
3645 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3646 std::swap(LHS, RHS);
3647 std::swap(LHSCst, RHSCst);
3648 std::swap(LHSCC, RHSCC);
3651 // At this point, we know we have have two icmp instructions
3652 // comparing a value against two constants and and'ing the result
3653 // together. Because of the above check, we know that we only have
3654 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3655 // (from the FoldICmpLogical check above), that the two constants
3656 // are not equal and that the larger constant is on the RHS
3657 assert(LHSCst != RHSCst && "Compares not folded above?");
3660 default: assert(0 && "Unknown integer condition code!");
3661 case ICmpInst::ICMP_EQ:
3663 default: assert(0 && "Unknown integer condition code!");
3664 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3665 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3666 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3667 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3668 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3669 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3670 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3671 return ReplaceInstUsesWith(I, LHS);
3673 case ICmpInst::ICMP_NE:
3675 default: assert(0 && "Unknown integer condition code!");
3676 case ICmpInst::ICMP_ULT:
3677 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3678 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3679 break; // (X != 13 & X u< 15) -> no change
3680 case ICmpInst::ICMP_SLT:
3681 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3682 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3683 break; // (X != 13 & X s< 15) -> no change
3684 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3685 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3686 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3687 return ReplaceInstUsesWith(I, RHS);
3688 case ICmpInst::ICMP_NE:
3689 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3690 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3691 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
3692 LHSVal->getName()+".off");
3693 InsertNewInstBefore(Add, I);
3694 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3695 ConstantInt::get(Add->getType(), 1));
3697 break; // (X != 13 & X != 15) -> no change
3700 case ICmpInst::ICMP_ULT:
3702 default: assert(0 && "Unknown integer condition code!");
3703 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3704 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3705 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3706 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3708 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3709 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3710 return ReplaceInstUsesWith(I, LHS);
3711 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3715 case ICmpInst::ICMP_SLT:
3717 default: assert(0 && "Unknown integer condition code!");
3718 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3719 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3720 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3721 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3723 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3724 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3725 return ReplaceInstUsesWith(I, LHS);
3726 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3730 case ICmpInst::ICMP_UGT:
3732 default: assert(0 && "Unknown integer condition code!");
3733 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3734 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3735 return ReplaceInstUsesWith(I, RHS);
3736 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3738 case ICmpInst::ICMP_NE:
3739 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3740 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3741 break; // (X u> 13 & X != 15) -> no change
3742 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3743 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3745 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3749 case ICmpInst::ICMP_SGT:
3751 default: assert(0 && "Unknown integer condition code!");
3752 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3753 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3754 return ReplaceInstUsesWith(I, RHS);
3755 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3757 case ICmpInst::ICMP_NE:
3758 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3759 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3760 break; // (X s> 13 & X != 15) -> no change
3761 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3762 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3764 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3772 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3773 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3774 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3775 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3776 const Type *SrcTy = Op0C->getOperand(0)->getType();
3777 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3778 // Only do this if the casts both really cause code to be generated.
3779 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3781 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3783 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3784 Op1C->getOperand(0),
3786 InsertNewInstBefore(NewOp, I);
3787 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3791 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3792 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3793 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3794 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3795 SI0->getOperand(1) == SI1->getOperand(1) &&
3796 (SI0->hasOneUse() || SI1->hasOneUse())) {
3797 Instruction *NewOp =
3798 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3800 SI0->getName()), I);
3801 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3802 SI1->getOperand(1));
3806 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3807 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3808 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3809 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3810 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3811 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3812 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3813 // If either of the constants are nans, then the whole thing returns
3815 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3816 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3817 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3818 RHS->getOperand(0));
3823 return Changed ? &I : 0;
3826 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3827 /// in the result. If it does, and if the specified byte hasn't been filled in
3828 /// yet, fill it in and return false.
3829 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3830 Instruction *I = dyn_cast<Instruction>(V);
3831 if (I == 0) return true;
3833 // If this is an or instruction, it is an inner node of the bswap.
3834 if (I->getOpcode() == Instruction::Or)
3835 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3836 CollectBSwapParts(I->getOperand(1), ByteValues);
3838 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3839 // If this is a shift by a constant int, and it is "24", then its operand
3840 // defines a byte. We only handle unsigned types here.
3841 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3842 // Not shifting the entire input by N-1 bytes?
3843 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3844 8*(ByteValues.size()-1))
3848 if (I->getOpcode() == Instruction::Shl) {
3849 // X << 24 defines the top byte with the lowest of the input bytes.
3850 DestNo = ByteValues.size()-1;
3852 // X >>u 24 defines the low byte with the highest of the input bytes.
3856 // If the destination byte value is already defined, the values are or'd
3857 // together, which isn't a bswap (unless it's an or of the same bits).
3858 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3860 ByteValues[DestNo] = I->getOperand(0);
3864 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3866 Value *Shift = 0, *ShiftLHS = 0;
3867 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3868 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3869 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3871 Instruction *SI = cast<Instruction>(Shift);
3873 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3874 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3875 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3878 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3880 if (AndAmt->getValue().getActiveBits() > 64)
3882 uint64_t AndAmtVal = AndAmt->getZExtValue();
3883 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3884 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3886 // Unknown mask for bswap.
3887 if (DestByte == ByteValues.size()) return true;
3889 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3891 if (SI->getOpcode() == Instruction::Shl)
3892 SrcByte = DestByte - ShiftBytes;
3894 SrcByte = DestByte + ShiftBytes;
3896 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3897 if (SrcByte != ByteValues.size()-DestByte-1)
3900 // If the destination byte value is already defined, the values are or'd
3901 // together, which isn't a bswap (unless it's an or of the same bits).
3902 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3904 ByteValues[DestByte] = SI->getOperand(0);
3908 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3909 /// If so, insert the new bswap intrinsic and return it.
3910 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3911 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3912 if (!ITy || ITy->getBitWidth() % 16)
3913 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3915 /// ByteValues - For each byte of the result, we keep track of which value
3916 /// defines each byte.
3917 SmallVector<Value*, 8> ByteValues;
3918 ByteValues.resize(ITy->getBitWidth()/8);
3920 // Try to find all the pieces corresponding to the bswap.
3921 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3922 CollectBSwapParts(I.getOperand(1), ByteValues))
3925 // Check to see if all of the bytes come from the same value.
3926 Value *V = ByteValues[0];
3927 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3929 // Check to make sure that all of the bytes come from the same value.
3930 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3931 if (ByteValues[i] != V)
3933 const Type *Tys[] = { ITy };
3934 Module *M = I.getParent()->getParent()->getParent();
3935 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3936 return CallInst::Create(F, V);
3940 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3941 bool Changed = SimplifyCommutative(I);
3942 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3944 if (isa<UndefValue>(Op1)) // X | undef -> -1
3945 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3949 return ReplaceInstUsesWith(I, Op0);
3951 // See if we can simplify any instructions used by the instruction whose sole
3952 // purpose is to compute bits we don't care about.
3953 if (!isa<VectorType>(I.getType())) {
3954 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3955 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3956 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3957 KnownZero, KnownOne))
3959 } else if (isa<ConstantAggregateZero>(Op1)) {
3960 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3961 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3962 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3963 return ReplaceInstUsesWith(I, I.getOperand(1));
3969 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3970 ConstantInt *C1 = 0; Value *X = 0;
3971 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3972 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3973 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3974 InsertNewInstBefore(Or, I);
3976 return BinaryOperator::CreateAnd(Or,
3977 ConstantInt::get(RHS->getValue() | C1->getValue()));
3980 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3981 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3982 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3983 InsertNewInstBefore(Or, I);
3985 return BinaryOperator::CreateXor(Or,
3986 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3989 // Try to fold constant and into select arguments.
3990 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3991 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3993 if (isa<PHINode>(Op0))
3994 if (Instruction *NV = FoldOpIntoPhi(I))
3998 Value *A = 0, *B = 0;
3999 ConstantInt *C1 = 0, *C2 = 0;
4001 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4002 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4003 return ReplaceInstUsesWith(I, Op1);
4004 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4005 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4006 return ReplaceInstUsesWith(I, Op0);
4008 // (A | B) | C and A | (B | C) -> bswap if possible.
4009 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4010 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4011 match(Op1, m_Or(m_Value(), m_Value())) ||
4012 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4013 match(Op1, m_Shift(m_Value(), m_Value())))) {
4014 if (Instruction *BSwap = MatchBSwap(I))
4018 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4019 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4020 MaskedValueIsZero(Op1, C1->getValue())) {
4021 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4022 InsertNewInstBefore(NOr, I);
4024 return BinaryOperator::CreateXor(NOr, C1);
4027 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4028 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4029 MaskedValueIsZero(Op0, C1->getValue())) {
4030 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4031 InsertNewInstBefore(NOr, I);
4033 return BinaryOperator::CreateXor(NOr, C1);
4037 Value *C = 0, *D = 0;
4038 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4039 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4040 Value *V1 = 0, *V2 = 0, *V3 = 0;
4041 C1 = dyn_cast<ConstantInt>(C);
4042 C2 = dyn_cast<ConstantInt>(D);
4043 if (C1 && C2) { // (A & C1)|(B & C2)
4044 // If we have: ((V + N) & C1) | (V & C2)
4045 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4046 // replace with V+N.
4047 if (C1->getValue() == ~C2->getValue()) {
4048 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4049 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4050 // Add commutes, try both ways.
4051 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4052 return ReplaceInstUsesWith(I, A);
4053 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4054 return ReplaceInstUsesWith(I, A);
4056 // Or commutes, try both ways.
4057 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4058 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4059 // Add commutes, try both ways.
4060 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4061 return ReplaceInstUsesWith(I, B);
4062 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4063 return ReplaceInstUsesWith(I, B);
4066 V1 = 0; V2 = 0; V3 = 0;
4069 // Check to see if we have any common things being and'ed. If so, find the
4070 // terms for V1 & (V2|V3).
4071 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4072 if (A == B) // (A & C)|(A & D) == A & (C|D)
4073 V1 = A, V2 = C, V3 = D;
4074 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4075 V1 = A, V2 = B, V3 = C;
4076 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4077 V1 = C, V2 = A, V3 = D;
4078 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4079 V1 = C, V2 = A, V3 = B;
4083 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4084 return BinaryOperator::CreateAnd(V1, Or);
4089 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4090 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4091 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4092 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4093 SI0->getOperand(1) == SI1->getOperand(1) &&
4094 (SI0->hasOneUse() || SI1->hasOneUse())) {
4095 Instruction *NewOp =
4096 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4098 SI0->getName()), I);
4099 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4100 SI1->getOperand(1));
4104 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4105 if (A == Op1) // ~A | A == -1
4106 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4110 // Note, A is still live here!
4111 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4113 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4115 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4116 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4117 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4118 I.getName()+".demorgan"), I);
4119 return BinaryOperator::CreateNot(And);
4123 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4124 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4125 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4128 Value *LHSVal, *RHSVal;
4129 ConstantInt *LHSCst, *RHSCst;
4130 ICmpInst::Predicate LHSCC, RHSCC;
4131 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4132 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4133 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4134 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4135 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4136 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4137 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4138 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4139 // We can't fold (ugt x, C) | (sgt x, C2).
4140 PredicatesFoldable(LHSCC, RHSCC)) {
4141 // Ensure that the larger constant is on the RHS.
4142 ICmpInst *LHS = cast<ICmpInst>(Op0);
4144 if (ICmpInst::isSignedPredicate(LHSCC))
4145 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4147 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4150 std::swap(LHS, RHS);
4151 std::swap(LHSCst, RHSCst);
4152 std::swap(LHSCC, RHSCC);
4155 // At this point, we know we have have two icmp instructions
4156 // comparing a value against two constants and or'ing the result
4157 // together. Because of the above check, we know that we only have
4158 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4159 // FoldICmpLogical check above), that the two constants are not
4161 assert(LHSCst != RHSCst && "Compares not folded above?");
4164 default: assert(0 && "Unknown integer condition code!");
4165 case ICmpInst::ICMP_EQ:
4167 default: assert(0 && "Unknown integer condition code!");
4168 case ICmpInst::ICMP_EQ:
4169 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4170 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4171 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4172 LHSVal->getName()+".off");
4173 InsertNewInstBefore(Add, I);
4174 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4175 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4177 break; // (X == 13 | X == 15) -> no change
4178 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4179 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4181 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4182 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4183 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4184 return ReplaceInstUsesWith(I, RHS);
4187 case ICmpInst::ICMP_NE:
4189 default: assert(0 && "Unknown integer condition code!");
4190 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4191 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4192 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4193 return ReplaceInstUsesWith(I, LHS);
4194 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4195 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4196 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4197 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4200 case ICmpInst::ICMP_ULT:
4202 default: assert(0 && "Unknown integer condition code!");
4203 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4205 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4206 // If RHSCst is [us]MAXINT, it is always false. Not handling
4207 // this can cause overflow.
4208 if (RHSCst->isMaxValue(false))
4209 return ReplaceInstUsesWith(I, LHS);
4210 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4212 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4214 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4215 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4216 return ReplaceInstUsesWith(I, RHS);
4217 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4221 case ICmpInst::ICMP_SLT:
4223 default: assert(0 && "Unknown integer condition code!");
4224 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4226 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4227 // If RHSCst is [us]MAXINT, it is always false. Not handling
4228 // this can cause overflow.
4229 if (RHSCst->isMaxValue(true))
4230 return ReplaceInstUsesWith(I, LHS);
4231 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4233 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4235 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4236 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4237 return ReplaceInstUsesWith(I, RHS);
4238 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4242 case ICmpInst::ICMP_UGT:
4244 default: assert(0 && "Unknown integer condition code!");
4245 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4246 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4247 return ReplaceInstUsesWith(I, LHS);
4248 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4250 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4251 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4252 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4253 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4257 case ICmpInst::ICMP_SGT:
4259 default: assert(0 && "Unknown integer condition code!");
4260 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4261 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4262 return ReplaceInstUsesWith(I, LHS);
4263 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4265 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4266 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4267 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4268 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4276 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4277 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4278 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4279 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4280 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4281 !isa<ICmpInst>(Op1C->getOperand(0))) {
4282 const Type *SrcTy = Op0C->getOperand(0)->getType();
4283 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4284 // Only do this if the casts both really cause code to be
4286 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4288 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4290 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4291 Op1C->getOperand(0),
4293 InsertNewInstBefore(NewOp, I);
4294 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4301 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4302 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4303 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4304 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4305 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4306 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4307 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4308 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4309 // If either of the constants are nans, then the whole thing returns
4311 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4312 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4314 // Otherwise, no need to compare the two constants, compare the
4316 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4317 RHS->getOperand(0));
4322 return Changed ? &I : 0;
4327 // XorSelf - Implements: X ^ X --> 0
4330 XorSelf(Value *rhs) : RHS(rhs) {}
4331 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4332 Instruction *apply(BinaryOperator &Xor) const {
4339 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4340 bool Changed = SimplifyCommutative(I);
4341 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4343 if (isa<UndefValue>(Op1)) {
4344 if (isa<UndefValue>(Op0))
4345 // Handle undef ^ undef -> 0 special case. This is a common
4347 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4348 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4351 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4352 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4353 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4354 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4357 // See if we can simplify any instructions used by the instruction whose sole
4358 // purpose is to compute bits we don't care about.
4359 if (!isa<VectorType>(I.getType())) {
4360 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4361 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4362 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4363 KnownZero, KnownOne))
4365 } else if (isa<ConstantAggregateZero>(Op1)) {
4366 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4369 // Is this a ~ operation?
4370 if (Value *NotOp = dyn_castNotVal(&I)) {
4371 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4372 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4373 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4374 if (Op0I->getOpcode() == Instruction::And ||
4375 Op0I->getOpcode() == Instruction::Or) {
4376 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4377 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4379 BinaryOperator::CreateNot(Op0I->getOperand(1),
4380 Op0I->getOperand(1)->getName()+".not");
4381 InsertNewInstBefore(NotY, I);
4382 if (Op0I->getOpcode() == Instruction::And)
4383 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4385 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4392 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4393 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4394 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4395 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4396 return new ICmpInst(ICI->getInversePredicate(),
4397 ICI->getOperand(0), ICI->getOperand(1));
4399 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4400 return new FCmpInst(FCI->getInversePredicate(),
4401 FCI->getOperand(0), FCI->getOperand(1));
4404 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4405 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4406 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4407 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4408 Instruction::CastOps Opcode = Op0C->getOpcode();
4409 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4410 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4411 Op0C->getDestTy())) {
4412 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4413 CI->getOpcode(), CI->getInversePredicate(),
4414 CI->getOperand(0), CI->getOperand(1)), I);
4415 NewCI->takeName(CI);
4416 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4423 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4424 // ~(c-X) == X-c-1 == X+(-c-1)
4425 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4426 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4427 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4428 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4429 ConstantInt::get(I.getType(), 1));
4430 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4433 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4434 if (Op0I->getOpcode() == Instruction::Add) {
4435 // ~(X-c) --> (-c-1)-X
4436 if (RHS->isAllOnesValue()) {
4437 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4438 return BinaryOperator::CreateSub(
4439 ConstantExpr::getSub(NegOp0CI,
4440 ConstantInt::get(I.getType(), 1)),
4441 Op0I->getOperand(0));
4442 } else if (RHS->getValue().isSignBit()) {
4443 // (X + C) ^ signbit -> (X + C + signbit)
4444 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4445 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4448 } else if (Op0I->getOpcode() == Instruction::Or) {
4449 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4450 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4451 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4452 // Anything in both C1 and C2 is known to be zero, remove it from
4454 Constant *CommonBits = And(Op0CI, RHS);
4455 NewRHS = ConstantExpr::getAnd(NewRHS,
4456 ConstantExpr::getNot(CommonBits));
4457 AddToWorkList(Op0I);
4458 I.setOperand(0, Op0I->getOperand(0));
4459 I.setOperand(1, NewRHS);
4466 // Try to fold constant and into select arguments.
4467 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4468 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4470 if (isa<PHINode>(Op0))
4471 if (Instruction *NV = FoldOpIntoPhi(I))
4475 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4477 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4479 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4481 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4484 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4487 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4488 if (A == Op0) { // B^(B|A) == (A|B)^B
4489 Op1I->swapOperands();
4491 std::swap(Op0, Op1);
4492 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4493 I.swapOperands(); // Simplified below.
4494 std::swap(Op0, Op1);
4496 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4497 if (Op0 == A) // A^(A^B) == B
4498 return ReplaceInstUsesWith(I, B);
4499 else if (Op0 == B) // A^(B^A) == B
4500 return ReplaceInstUsesWith(I, A);
4501 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4502 if (A == Op0) { // A^(A&B) -> A^(B&A)
4503 Op1I->swapOperands();
4506 if (B == Op0) { // A^(B&A) -> (B&A)^A
4507 I.swapOperands(); // Simplified below.
4508 std::swap(Op0, Op1);
4513 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4516 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4517 if (A == Op1) // (B|A)^B == (A|B)^B
4519 if (B == Op1) { // (A|B)^B == A & ~B
4521 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4522 return BinaryOperator::CreateAnd(A, NotB);
4524 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4525 if (Op1 == A) // (A^B)^A == B
4526 return ReplaceInstUsesWith(I, B);
4527 else if (Op1 == B) // (B^A)^A == B
4528 return ReplaceInstUsesWith(I, A);
4529 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4530 if (A == Op1) // (A&B)^A -> (B&A)^A
4532 if (B == Op1 && // (B&A)^A == ~B & A
4533 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4535 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4536 return BinaryOperator::CreateAnd(N, Op1);
4541 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4542 if (Op0I && Op1I && Op0I->isShift() &&
4543 Op0I->getOpcode() == Op1I->getOpcode() &&
4544 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4545 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4546 Instruction *NewOp =
4547 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4548 Op1I->getOperand(0),
4549 Op0I->getName()), I);
4550 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4551 Op1I->getOperand(1));
4555 Value *A, *B, *C, *D;
4556 // (A & B)^(A | B) -> A ^ B
4557 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4558 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4559 if ((A == C && B == D) || (A == D && B == C))
4560 return BinaryOperator::CreateXor(A, B);
4562 // (A | B)^(A & B) -> A ^ B
4563 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4564 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4565 if ((A == C && B == D) || (A == D && B == C))
4566 return BinaryOperator::CreateXor(A, B);
4570 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4571 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4572 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4573 // (X & Y)^(X & Y) -> (Y^Z) & X
4574 Value *X = 0, *Y = 0, *Z = 0;
4576 X = A, Y = B, Z = D;
4578 X = A, Y = B, Z = C;
4580 X = B, Y = A, Z = D;
4582 X = B, Y = A, Z = C;
4585 Instruction *NewOp =
4586 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4587 return BinaryOperator::CreateAnd(NewOp, X);
4592 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4593 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4594 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4597 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4598 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4599 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4600 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4601 const Type *SrcTy = Op0C->getOperand(0)->getType();
4602 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4603 // Only do this if the casts both really cause code to be generated.
4604 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4606 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4608 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4609 Op1C->getOperand(0),
4611 InsertNewInstBefore(NewOp, I);
4612 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4617 return Changed ? &I : 0;
4620 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4621 /// overflowed for this type.
4622 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4623 ConstantInt *In2, bool IsSigned = false) {
4624 Result = cast<ConstantInt>(Add(In1, In2));
4627 if (In2->getValue().isNegative())
4628 return Result->getValue().sgt(In1->getValue());
4630 return Result->getValue().slt(In1->getValue());
4632 return Result->getValue().ult(In1->getValue());
4635 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4636 /// code necessary to compute the offset from the base pointer (without adding
4637 /// in the base pointer). Return the result as a signed integer of intptr size.
4638 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4639 TargetData &TD = IC.getTargetData();
4640 gep_type_iterator GTI = gep_type_begin(GEP);
4641 const Type *IntPtrTy = TD.getIntPtrType();
4642 Value *Result = Constant::getNullValue(IntPtrTy);
4644 // Build a mask for high order bits.
4645 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4646 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4648 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
4651 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4652 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4653 if (OpC->isZero()) continue;
4655 // Handle a struct index, which adds its field offset to the pointer.
4656 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4657 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4659 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4660 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4662 Result = IC.InsertNewInstBefore(
4663 BinaryOperator::CreateAdd(Result,
4664 ConstantInt::get(IntPtrTy, Size),
4665 GEP->getName()+".offs"), I);
4669 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4670 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4671 Scale = ConstantExpr::getMul(OC, Scale);
4672 if (Constant *RC = dyn_cast<Constant>(Result))
4673 Result = ConstantExpr::getAdd(RC, Scale);
4675 // Emit an add instruction.
4676 Result = IC.InsertNewInstBefore(
4677 BinaryOperator::CreateAdd(Result, Scale,
4678 GEP->getName()+".offs"), I);
4682 // Convert to correct type.
4683 if (Op->getType() != IntPtrTy) {
4684 if (Constant *OpC = dyn_cast<Constant>(Op))
4685 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4687 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4688 Op->getName()+".c"), I);
4691 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4692 if (Constant *OpC = dyn_cast<Constant>(Op))
4693 Op = ConstantExpr::getMul(OpC, Scale);
4694 else // We'll let instcombine(mul) convert this to a shl if possible.
4695 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
4696 GEP->getName()+".idx"), I);
4699 // Emit an add instruction.
4700 if (isa<Constant>(Op) && isa<Constant>(Result))
4701 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4702 cast<Constant>(Result));
4704 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
4705 GEP->getName()+".offs"), I);
4711 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4712 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4713 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4714 /// complex, and scales are involved. The above expression would also be legal
4715 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4716 /// later form is less amenable to optimization though, and we are allowed to
4717 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4719 /// If we can't emit an optimized form for this expression, this returns null.
4721 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4723 TargetData &TD = IC.getTargetData();
4724 gep_type_iterator GTI = gep_type_begin(GEP);
4726 // Check to see if this gep only has a single variable index. If so, and if
4727 // any constant indices are a multiple of its scale, then we can compute this
4728 // in terms of the scale of the variable index. For example, if the GEP
4729 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4730 // because the expression will cross zero at the same point.
4731 unsigned i, e = GEP->getNumOperands();
4733 for (i = 1; i != e; ++i, ++GTI) {
4734 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4735 // Compute the aggregate offset of constant indices.
4736 if (CI->isZero()) continue;
4738 // Handle a struct index, which adds its field offset to the pointer.
4739 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4740 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4742 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4743 Offset += Size*CI->getSExtValue();
4746 // Found our variable index.
4751 // If there are no variable indices, we must have a constant offset, just
4752 // evaluate it the general way.
4753 if (i == e) return 0;
4755 Value *VariableIdx = GEP->getOperand(i);
4756 // Determine the scale factor of the variable element. For example, this is
4757 // 4 if the variable index is into an array of i32.
4758 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4760 // Verify that there are no other variable indices. If so, emit the hard way.
4761 for (++i, ++GTI; i != e; ++i, ++GTI) {
4762 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4765 // Compute the aggregate offset of constant indices.
4766 if (CI->isZero()) continue;
4768 // Handle a struct index, which adds its field offset to the pointer.
4769 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4770 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4772 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4773 Offset += Size*CI->getSExtValue();
4777 // Okay, we know we have a single variable index, which must be a
4778 // pointer/array/vector index. If there is no offset, life is simple, return
4780 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4782 // Cast to intptrty in case a truncation occurs. If an extension is needed,
4783 // we don't need to bother extending: the extension won't affect where the
4784 // computation crosses zero.
4785 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
4786 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
4787 VariableIdx->getNameStart(), &I);
4791 // Otherwise, there is an index. The computation we will do will be modulo
4792 // the pointer size, so get it.
4793 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4795 Offset &= PtrSizeMask;
4796 VariableScale &= PtrSizeMask;
4798 // To do this transformation, any constant index must be a multiple of the
4799 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
4800 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
4801 // multiple of the variable scale.
4802 int64_t NewOffs = Offset / (int64_t)VariableScale;
4803 if (Offset != NewOffs*(int64_t)VariableScale)
4806 // Okay, we can do this evaluation. Start by converting the index to intptr.
4807 const Type *IntPtrTy = TD.getIntPtrType();
4808 if (VariableIdx->getType() != IntPtrTy)
4809 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
4811 VariableIdx->getNameStart(), &I);
4812 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
4813 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
4817 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4818 /// else. At this point we know that the GEP is on the LHS of the comparison.
4819 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4820 ICmpInst::Predicate Cond,
4822 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4824 // Look through bitcasts.
4825 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
4826 RHS = BCI->getOperand(0);
4828 Value *PtrBase = GEPLHS->getOperand(0);
4829 if (PtrBase == RHS) {
4830 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4831 // This transformation (ignoring the base and scales) is valid because we
4832 // know pointers can't overflow. See if we can output an optimized form.
4833 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
4835 // If not, synthesize the offset the hard way.
4837 Offset = EmitGEPOffset(GEPLHS, I, *this);
4838 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4839 Constant::getNullValue(Offset->getType()));
4840 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4841 // If the base pointers are different, but the indices are the same, just
4842 // compare the base pointer.
4843 if (PtrBase != GEPRHS->getOperand(0)) {
4844 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4845 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4846 GEPRHS->getOperand(0)->getType();
4848 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4849 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4850 IndicesTheSame = false;
4854 // If all indices are the same, just compare the base pointers.
4856 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4857 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4859 // Otherwise, the base pointers are different and the indices are
4860 // different, bail out.
4864 // If one of the GEPs has all zero indices, recurse.
4865 bool AllZeros = true;
4866 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4867 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4868 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4873 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4874 ICmpInst::getSwappedPredicate(Cond), I);
4876 // If the other GEP has all zero indices, recurse.
4878 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4879 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4880 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4885 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4887 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4888 // If the GEPs only differ by one index, compare it.
4889 unsigned NumDifferences = 0; // Keep track of # differences.
4890 unsigned DiffOperand = 0; // The operand that differs.
4891 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4892 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4893 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4894 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4895 // Irreconcilable differences.
4899 if (NumDifferences++) break;
4904 if (NumDifferences == 0) // SAME GEP?
4905 return ReplaceInstUsesWith(I, // No comparison is needed here.
4906 ConstantInt::get(Type::Int1Ty,
4907 ICmpInst::isTrueWhenEqual(Cond)));
4909 else if (NumDifferences == 1) {
4910 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4911 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4912 // Make sure we do a signed comparison here.
4913 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4917 // Only lower this if the icmp is the only user of the GEP or if we expect
4918 // the result to fold to a constant!
4919 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4920 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4921 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4922 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4923 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4924 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4930 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
4932 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
4935 if (!isa<ConstantFP>(RHSC)) return 0;
4936 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4938 // Get the width of the mantissa. We don't want to hack on conversions that
4939 // might lose information from the integer, e.g. "i64 -> float"
4940 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4941 if (MantissaWidth == -1) return 0; // Unknown.
4943 // Check to see that the input is converted from an integer type that is small
4944 // enough that preserves all bits. TODO: check here for "known" sign bits.
4945 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4946 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
4948 // If this is a uitofp instruction, we need an extra bit to hold the sign.
4949 if (isa<UIToFPInst>(LHSI))
4952 // If the conversion would lose info, don't hack on this.
4953 if ((int)InputSize > MantissaWidth)
4956 // Otherwise, we can potentially simplify the comparison. We know that it
4957 // will always come through as an integer value and we know the constant is
4958 // not a NAN (it would have been previously simplified).
4959 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4961 ICmpInst::Predicate Pred;
4962 switch (I.getPredicate()) {
4963 default: assert(0 && "Unexpected predicate!");
4964 case FCmpInst::FCMP_UEQ:
4965 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
4966 case FCmpInst::FCMP_UGT:
4967 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
4968 case FCmpInst::FCMP_UGE:
4969 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
4970 case FCmpInst::FCMP_ULT:
4971 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
4972 case FCmpInst::FCMP_ULE:
4973 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
4974 case FCmpInst::FCMP_UNE:
4975 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
4976 case FCmpInst::FCMP_ORD:
4977 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4978 case FCmpInst::FCMP_UNO:
4979 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4982 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4984 // Now we know that the APFloat is a normal number, zero or inf.
4986 // See if the FP constant is too large for the integer. For example,
4987 // comparing an i8 to 300.0.
4988 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
4990 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4991 // and large values.
4992 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
4993 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4994 APFloat::rmNearestTiesToEven);
4995 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4996 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4997 Pred == ICmpInst::ICMP_SLE)
4998 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4999 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5002 // See if the RHS value is < SignedMin.
5003 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5004 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5005 APFloat::rmNearestTiesToEven);
5006 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5007 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5008 Pred == ICmpInst::ICMP_SGE)
5009 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5010 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5013 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
5014 // it may still be fractional. See if it is fractional by casting the FP
5015 // value to the integer value and back, checking for equality. Don't do this
5016 // for zero, because -0.0 is not fractional.
5017 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5018 if (!RHS.isZero() &&
5019 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5020 // If we had a comparison against a fractional value, we have to adjust
5021 // the compare predicate and sometimes the value. RHSC is rounded towards
5022 // zero at this point.
5024 default: assert(0 && "Unexpected integer comparison!");
5025 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5026 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5027 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5028 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5029 case ICmpInst::ICMP_SLE:
5030 // (float)int <= 4.4 --> int <= 4
5031 // (float)int <= -4.4 --> int < -4
5032 if (RHS.isNegative())
5033 Pred = ICmpInst::ICMP_SLT;
5035 case ICmpInst::ICMP_SLT:
5036 // (float)int < -4.4 --> int < -4
5037 // (float)int < 4.4 --> int <= 4
5038 if (!RHS.isNegative())
5039 Pred = ICmpInst::ICMP_SLE;
5041 case ICmpInst::ICMP_SGT:
5042 // (float)int > 4.4 --> int > 4
5043 // (float)int > -4.4 --> int >= -4
5044 if (RHS.isNegative())
5045 Pred = ICmpInst::ICMP_SGE;
5047 case ICmpInst::ICMP_SGE:
5048 // (float)int >= -4.4 --> int >= -4
5049 // (float)int >= 4.4 --> int > 4
5050 if (!RHS.isNegative())
5051 Pred = ICmpInst::ICMP_SGT;
5056 // Lower this FP comparison into an appropriate integer version of the
5058 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5061 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5062 bool Changed = SimplifyCompare(I);
5063 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5065 // Fold trivial predicates.
5066 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5067 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5068 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5069 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5071 // Simplify 'fcmp pred X, X'
5073 switch (I.getPredicate()) {
5074 default: assert(0 && "Unknown predicate!");
5075 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5076 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5077 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5078 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5079 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5080 case FCmpInst::FCMP_OLT: // True if ordered and less than
5081 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5082 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5084 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5085 case FCmpInst::FCMP_ULT: // True if unordered or less than
5086 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5087 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5088 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5089 I.setPredicate(FCmpInst::FCMP_UNO);
5090 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5093 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5094 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5095 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5096 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5097 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5098 I.setPredicate(FCmpInst::FCMP_ORD);
5099 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5104 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5105 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5107 // Handle fcmp with constant RHS
5108 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5109 // If the constant is a nan, see if we can fold the comparison based on it.
5110 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5111 if (CFP->getValueAPF().isNaN()) {
5112 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5113 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5114 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5115 "Comparison must be either ordered or unordered!");
5116 // True if unordered.
5117 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5121 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5122 switch (LHSI->getOpcode()) {
5123 case Instruction::PHI:
5124 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5125 // block. If in the same block, we're encouraging jump threading. If
5126 // not, we are just pessimizing the code by making an i1 phi.
5127 if (LHSI->getParent() == I.getParent())
5128 if (Instruction *NV = FoldOpIntoPhi(I))
5131 case Instruction::SIToFP:
5132 case Instruction::UIToFP:
5133 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5136 case Instruction::Select:
5137 // If either operand of the select is a constant, we can fold the
5138 // comparison into the select arms, which will cause one to be
5139 // constant folded and the select turned into a bitwise or.
5140 Value *Op1 = 0, *Op2 = 0;
5141 if (LHSI->hasOneUse()) {
5142 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5143 // Fold the known value into the constant operand.
5144 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5145 // Insert a new FCmp of the other select operand.
5146 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5147 LHSI->getOperand(2), RHSC,
5149 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5150 // Fold the known value into the constant operand.
5151 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5152 // Insert a new FCmp of the other select operand.
5153 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5154 LHSI->getOperand(1), RHSC,
5160 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5165 return Changed ? &I : 0;
5168 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5169 bool Changed = SimplifyCompare(I);
5170 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5171 const Type *Ty = Op0->getType();
5175 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5176 I.isTrueWhenEqual()));
5178 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5179 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5181 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5182 // addresses never equal each other! We already know that Op0 != Op1.
5183 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5184 isa<ConstantPointerNull>(Op0)) &&
5185 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5186 isa<ConstantPointerNull>(Op1)))
5187 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5188 !I.isTrueWhenEqual()));
5190 // icmp's with boolean values can always be turned into bitwise operations
5191 if (Ty == Type::Int1Ty) {
5192 switch (I.getPredicate()) {
5193 default: assert(0 && "Invalid icmp instruction!");
5194 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5195 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5196 InsertNewInstBefore(Xor, I);
5197 return BinaryOperator::CreateNot(Xor);
5199 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5200 return BinaryOperator::CreateXor(Op0, Op1);
5202 case ICmpInst::ICMP_UGT:
5203 case ICmpInst::ICMP_SGT:
5204 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5206 case ICmpInst::ICMP_ULT:
5207 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5208 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5209 InsertNewInstBefore(Not, I);
5210 return BinaryOperator::CreateAnd(Not, Op1);
5212 case ICmpInst::ICMP_UGE:
5213 case ICmpInst::ICMP_SGE:
5214 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5216 case ICmpInst::ICMP_ULE:
5217 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5218 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5219 InsertNewInstBefore(Not, I);
5220 return BinaryOperator::CreateOr(Not, Op1);
5225 // See if we are doing a comparison between a constant and an instruction that
5226 // can be folded into the comparison.
5227 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5230 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5231 if (I.isEquality() && CI->isNullValue() &&
5232 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5233 // (icmp cond A B) if cond is equality
5234 return new ICmpInst(I.getPredicate(), A, B);
5237 switch (I.getPredicate()) {
5239 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5240 if (CI->isMinValue(false))
5241 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5242 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5243 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5244 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5245 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5246 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5247 if (CI->isMinValue(true))
5248 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5249 ConstantInt::getAllOnesValue(Op0->getType()));
5253 case ICmpInst::ICMP_SLT:
5254 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5255 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5256 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5257 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5258 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5259 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5262 case ICmpInst::ICMP_UGT:
5263 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5264 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5265 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5266 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5267 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5268 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5270 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5271 if (CI->isMaxValue(true))
5272 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5273 ConstantInt::getNullValue(Op0->getType()));
5276 case ICmpInst::ICMP_SGT:
5277 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5278 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5279 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5280 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5281 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5282 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5285 case ICmpInst::ICMP_ULE:
5286 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5287 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5288 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5289 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5290 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5291 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5294 case ICmpInst::ICMP_SLE:
5295 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5296 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5297 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5298 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5299 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5300 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5303 case ICmpInst::ICMP_UGE:
5304 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5305 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5306 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5307 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5308 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5309 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5312 case ICmpInst::ICMP_SGE:
5313 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5314 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5315 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5316 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5317 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5318 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5322 // If we still have a icmp le or icmp ge instruction, turn it into the
5323 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5324 // already been handled above, this requires little checking.
5326 switch (I.getPredicate()) {
5328 case ICmpInst::ICMP_ULE:
5329 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5330 case ICmpInst::ICMP_SLE:
5331 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5332 case ICmpInst::ICMP_UGE:
5333 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5334 case ICmpInst::ICMP_SGE:
5335 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5338 // See if we can fold the comparison based on bits known to be zero or one
5339 // in the input. If this comparison is a normal comparison, it demands all
5340 // bits, if it is a sign bit comparison, it only demands the sign bit.
5343 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5345 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5346 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5347 if (SimplifyDemandedBits(Op0,
5348 isSignBit ? APInt::getSignBit(BitWidth)
5349 : APInt::getAllOnesValue(BitWidth),
5350 KnownZero, KnownOne, 0))
5353 // Given the known and unknown bits, compute a range that the LHS could be
5355 if ((KnownOne | KnownZero) != 0) {
5356 // Compute the Min, Max and RHS values based on the known bits. For the
5357 // EQ and NE we use unsigned values.
5358 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5359 const APInt& RHSVal = CI->getValue();
5360 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5361 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5364 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5367 switch (I.getPredicate()) { // LE/GE have been folded already.
5368 default: assert(0 && "Unknown icmp opcode!");
5369 case ICmpInst::ICMP_EQ:
5370 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5371 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5373 case ICmpInst::ICMP_NE:
5374 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5375 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5377 case ICmpInst::ICMP_ULT:
5378 if (Max.ult(RHSVal))
5379 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5380 if (Min.uge(RHSVal))
5381 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5383 case ICmpInst::ICMP_UGT:
5384 if (Min.ugt(RHSVal))
5385 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5386 if (Max.ule(RHSVal))
5387 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5389 case ICmpInst::ICMP_SLT:
5390 if (Max.slt(RHSVal))
5391 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5392 if (Min.sgt(RHSVal))
5393 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5395 case ICmpInst::ICMP_SGT:
5396 if (Min.sgt(RHSVal))
5397 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5398 if (Max.sle(RHSVal))
5399 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5404 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5405 // instruction, see if that instruction also has constants so that the
5406 // instruction can be folded into the icmp
5407 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5408 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5412 // Handle icmp with constant (but not simple integer constant) RHS
5413 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5414 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5415 switch (LHSI->getOpcode()) {
5416 case Instruction::GetElementPtr:
5417 if (RHSC->isNullValue()) {
5418 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5419 bool isAllZeros = true;
5420 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5421 if (!isa<Constant>(LHSI->getOperand(i)) ||
5422 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5427 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5428 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5432 case Instruction::PHI:
5433 // Only fold icmp into the PHI if the phi and fcmp are in the same
5434 // block. If in the same block, we're encouraging jump threading. If
5435 // not, we are just pessimizing the code by making an i1 phi.
5436 if (LHSI->getParent() == I.getParent())
5437 if (Instruction *NV = FoldOpIntoPhi(I))
5440 case Instruction::Select: {
5441 // If either operand of the select is a constant, we can fold the
5442 // comparison into the select arms, which will cause one to be
5443 // constant folded and the select turned into a bitwise or.
5444 Value *Op1 = 0, *Op2 = 0;
5445 if (LHSI->hasOneUse()) {
5446 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5447 // Fold the known value into the constant operand.
5448 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5449 // Insert a new ICmp of the other select operand.
5450 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5451 LHSI->getOperand(2), RHSC,
5453 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5454 // Fold the known value into the constant operand.
5455 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5456 // Insert a new ICmp of the other select operand.
5457 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5458 LHSI->getOperand(1), RHSC,
5464 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5467 case Instruction::Malloc:
5468 // If we have (malloc != null), and if the malloc has a single use, we
5469 // can assume it is successful and remove the malloc.
5470 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5471 AddToWorkList(LHSI);
5472 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5473 !I.isTrueWhenEqual()));
5479 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5480 if (User *GEP = dyn_castGetElementPtr(Op0))
5481 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5483 if (User *GEP = dyn_castGetElementPtr(Op1))
5484 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5485 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5488 // Test to see if the operands of the icmp are casted versions of other
5489 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5491 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5492 if (isa<PointerType>(Op0->getType()) &&
5493 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5494 // We keep moving the cast from the left operand over to the right
5495 // operand, where it can often be eliminated completely.
5496 Op0 = CI->getOperand(0);
5498 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5499 // so eliminate it as well.
5500 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5501 Op1 = CI2->getOperand(0);
5503 // If Op1 is a constant, we can fold the cast into the constant.
5504 if (Op0->getType() != Op1->getType()) {
5505 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5506 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5508 // Otherwise, cast the RHS right before the icmp
5509 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5512 return new ICmpInst(I.getPredicate(), Op0, Op1);
5516 if (isa<CastInst>(Op0)) {
5517 // Handle the special case of: icmp (cast bool to X), <cst>
5518 // This comes up when you have code like
5521 // For generality, we handle any zero-extension of any operand comparison
5522 // with a constant or another cast from the same type.
5523 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5524 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5528 // ~x < ~y --> y < x
5530 if (match(Op0, m_Not(m_Value(A))) &&
5531 match(Op1, m_Not(m_Value(B))))
5532 return new ICmpInst(I.getPredicate(), B, A);
5535 if (I.isEquality()) {
5536 Value *A, *B, *C, *D;
5538 // -x == -y --> x == y
5539 if (match(Op0, m_Neg(m_Value(A))) &&
5540 match(Op1, m_Neg(m_Value(B))))
5541 return new ICmpInst(I.getPredicate(), A, B);
5543 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5544 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5545 Value *OtherVal = A == Op1 ? B : A;
5546 return new ICmpInst(I.getPredicate(), OtherVal,
5547 Constant::getNullValue(A->getType()));
5550 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5551 // A^c1 == C^c2 --> A == C^(c1^c2)
5552 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5553 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5554 if (Op1->hasOneUse()) {
5555 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5556 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5557 return new ICmpInst(I.getPredicate(), A,
5558 InsertNewInstBefore(Xor, I));
5561 // A^B == A^D -> B == D
5562 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5563 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5564 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5565 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5569 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5570 (A == Op0 || B == Op0)) {
5571 // A == (A^B) -> B == 0
5572 Value *OtherVal = A == Op0 ? B : A;
5573 return new ICmpInst(I.getPredicate(), OtherVal,
5574 Constant::getNullValue(A->getType()));
5576 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5577 // (A-B) == A -> B == 0
5578 return new ICmpInst(I.getPredicate(), B,
5579 Constant::getNullValue(B->getType()));
5581 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5582 // A == (A-B) -> B == 0
5583 return new ICmpInst(I.getPredicate(), B,
5584 Constant::getNullValue(B->getType()));
5587 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5588 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5589 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5590 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5591 Value *X = 0, *Y = 0, *Z = 0;
5594 X = B; Y = D; Z = A;
5595 } else if (A == D) {
5596 X = B; Y = C; Z = A;
5597 } else if (B == C) {
5598 X = A; Y = D; Z = B;
5599 } else if (B == D) {
5600 X = A; Y = C; Z = B;
5603 if (X) { // Build (X^Y) & Z
5604 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
5605 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
5606 I.setOperand(0, Op1);
5607 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5612 return Changed ? &I : 0;
5616 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5617 /// and CmpRHS are both known to be integer constants.
5618 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5619 ConstantInt *DivRHS) {
5620 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5621 const APInt &CmpRHSV = CmpRHS->getValue();
5623 // FIXME: If the operand types don't match the type of the divide
5624 // then don't attempt this transform. The code below doesn't have the
5625 // logic to deal with a signed divide and an unsigned compare (and
5626 // vice versa). This is because (x /s C1) <s C2 produces different
5627 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5628 // (x /u C1) <u C2. Simply casting the operands and result won't
5629 // work. :( The if statement below tests that condition and bails
5631 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5632 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5634 if (DivRHS->isZero())
5635 return 0; // The ProdOV computation fails on divide by zero.
5637 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5638 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5639 // C2 (CI). By solving for X we can turn this into a range check
5640 // instead of computing a divide.
5641 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5643 // Determine if the product overflows by seeing if the product is
5644 // not equal to the divide. Make sure we do the same kind of divide
5645 // as in the LHS instruction that we're folding.
5646 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5647 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5649 // Get the ICmp opcode
5650 ICmpInst::Predicate Pred = ICI.getPredicate();
5652 // Figure out the interval that is being checked. For example, a comparison
5653 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5654 // Compute this interval based on the constants involved and the signedness of
5655 // the compare/divide. This computes a half-open interval, keeping track of
5656 // whether either value in the interval overflows. After analysis each
5657 // overflow variable is set to 0 if it's corresponding bound variable is valid
5658 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5659 int LoOverflow = 0, HiOverflow = 0;
5660 ConstantInt *LoBound = 0, *HiBound = 0;
5663 if (!DivIsSigned) { // udiv
5664 // e.g. X/5 op 3 --> [15, 20)
5666 HiOverflow = LoOverflow = ProdOV;
5668 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5669 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5670 if (CmpRHSV == 0) { // (X / pos) op 0
5671 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5672 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5674 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5675 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5676 HiOverflow = LoOverflow = ProdOV;
5678 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5679 } else { // (X / pos) op neg
5680 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5681 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5682 LoOverflow = AddWithOverflow(LoBound, Prod,
5683 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5684 HiBound = AddOne(Prod);
5685 HiOverflow = ProdOV ? -1 : 0;
5687 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5688 if (CmpRHSV == 0) { // (X / neg) op 0
5689 // e.g. X/-5 op 0 --> [-4, 5)
5690 LoBound = AddOne(DivRHS);
5691 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5692 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5693 HiOverflow = 1; // [INTMIN+1, overflow)
5694 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5696 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5697 // e.g. X/-5 op 3 --> [-19, -14)
5698 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5700 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5701 HiBound = AddOne(Prod);
5702 } else { // (X / neg) op neg
5703 // e.g. X/-5 op -3 --> [15, 20)
5705 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5706 HiBound = Subtract(Prod, DivRHS);
5709 // Dividing by a negative swaps the condition. LT <-> GT
5710 Pred = ICmpInst::getSwappedPredicate(Pred);
5713 Value *X = DivI->getOperand(0);
5715 default: assert(0 && "Unhandled icmp opcode!");
5716 case ICmpInst::ICMP_EQ:
5717 if (LoOverflow && HiOverflow)
5718 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5719 else if (HiOverflow)
5720 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5721 ICmpInst::ICMP_UGE, X, LoBound);
5722 else if (LoOverflow)
5723 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5724 ICmpInst::ICMP_ULT, X, HiBound);
5726 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5727 case ICmpInst::ICMP_NE:
5728 if (LoOverflow && HiOverflow)
5729 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5730 else if (HiOverflow)
5731 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5732 ICmpInst::ICMP_ULT, X, LoBound);
5733 else if (LoOverflow)
5734 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5735 ICmpInst::ICMP_UGE, X, HiBound);
5737 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5738 case ICmpInst::ICMP_ULT:
5739 case ICmpInst::ICMP_SLT:
5740 if (LoOverflow == +1) // Low bound is greater than input range.
5741 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5742 if (LoOverflow == -1) // Low bound is less than input range.
5743 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5744 return new ICmpInst(Pred, X, LoBound);
5745 case ICmpInst::ICMP_UGT:
5746 case ICmpInst::ICMP_SGT:
5747 if (HiOverflow == +1) // High bound greater than input range.
5748 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5749 else if (HiOverflow == -1) // High bound less than input range.
5750 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5751 if (Pred == ICmpInst::ICMP_UGT)
5752 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5754 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5759 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5761 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5764 const APInt &RHSV = RHS->getValue();
5766 switch (LHSI->getOpcode()) {
5767 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5768 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5769 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5771 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5772 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5773 Value *CompareVal = LHSI->getOperand(0);
5775 // If the sign bit of the XorCST is not set, there is no change to
5776 // the operation, just stop using the Xor.
5777 if (!XorCST->getValue().isNegative()) {
5778 ICI.setOperand(0, CompareVal);
5779 AddToWorkList(LHSI);
5783 // Was the old condition true if the operand is positive?
5784 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5786 // If so, the new one isn't.
5787 isTrueIfPositive ^= true;
5789 if (isTrueIfPositive)
5790 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5792 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5796 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5797 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5798 LHSI->getOperand(0)->hasOneUse()) {
5799 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5801 // If the LHS is an AND of a truncating cast, we can widen the
5802 // and/compare to be the input width without changing the value
5803 // produced, eliminating a cast.
5804 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5805 // We can do this transformation if either the AND constant does not
5806 // have its sign bit set or if it is an equality comparison.
5807 // Extending a relational comparison when we're checking the sign
5808 // bit would not work.
5809 if (Cast->hasOneUse() &&
5810 (ICI.isEquality() ||
5811 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5813 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5814 APInt NewCST = AndCST->getValue();
5815 NewCST.zext(BitWidth);
5817 NewCI.zext(BitWidth);
5818 Instruction *NewAnd =
5819 BinaryOperator::CreateAnd(Cast->getOperand(0),
5820 ConstantInt::get(NewCST),LHSI->getName());
5821 InsertNewInstBefore(NewAnd, ICI);
5822 return new ICmpInst(ICI.getPredicate(), NewAnd,
5823 ConstantInt::get(NewCI));
5827 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5828 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5829 // happens a LOT in code produced by the C front-end, for bitfield
5831 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5832 if (Shift && !Shift->isShift())
5836 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5837 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5838 const Type *AndTy = AndCST->getType(); // Type of the and.
5840 // We can fold this as long as we can't shift unknown bits
5841 // into the mask. This can only happen with signed shift
5842 // rights, as they sign-extend.
5844 bool CanFold = Shift->isLogicalShift();
5846 // To test for the bad case of the signed shr, see if any
5847 // of the bits shifted in could be tested after the mask.
5848 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5849 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5851 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5852 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5853 AndCST->getValue()) == 0)
5859 if (Shift->getOpcode() == Instruction::Shl)
5860 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5862 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5864 // Check to see if we are shifting out any of the bits being
5866 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5867 // If we shifted bits out, the fold is not going to work out.
5868 // As a special case, check to see if this means that the
5869 // result is always true or false now.
5870 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5871 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5872 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5873 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5875 ICI.setOperand(1, NewCst);
5876 Constant *NewAndCST;
5877 if (Shift->getOpcode() == Instruction::Shl)
5878 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5880 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5881 LHSI->setOperand(1, NewAndCST);
5882 LHSI->setOperand(0, Shift->getOperand(0));
5883 AddToWorkList(Shift); // Shift is dead.
5884 AddUsesToWorkList(ICI);
5890 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5891 // preferable because it allows the C<<Y expression to be hoisted out
5892 // of a loop if Y is invariant and X is not.
5893 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5894 ICI.isEquality() && !Shift->isArithmeticShift() &&
5895 isa<Instruction>(Shift->getOperand(0))) {
5898 if (Shift->getOpcode() == Instruction::LShr) {
5899 NS = BinaryOperator::CreateShl(AndCST,
5900 Shift->getOperand(1), "tmp");
5902 // Insert a logical shift.
5903 NS = BinaryOperator::CreateLShr(AndCST,
5904 Shift->getOperand(1), "tmp");
5906 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5908 // Compute X & (C << Y).
5909 Instruction *NewAnd =
5910 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
5911 InsertNewInstBefore(NewAnd, ICI);
5913 ICI.setOperand(0, NewAnd);
5919 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5920 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5923 uint32_t TypeBits = RHSV.getBitWidth();
5925 // Check that the shift amount is in range. If not, don't perform
5926 // undefined shifts. When the shift is visited it will be
5928 if (ShAmt->uge(TypeBits))
5931 if (ICI.isEquality()) {
5932 // If we are comparing against bits always shifted out, the
5933 // comparison cannot succeed.
5935 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5936 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5937 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5938 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5939 return ReplaceInstUsesWith(ICI, Cst);
5942 if (LHSI->hasOneUse()) {
5943 // Otherwise strength reduce the shift into an and.
5944 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5946 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5949 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5950 Mask, LHSI->getName()+".mask");
5951 Value *And = InsertNewInstBefore(AndI, ICI);
5952 return new ICmpInst(ICI.getPredicate(), And,
5953 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5957 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5958 bool TrueIfSigned = false;
5959 if (LHSI->hasOneUse() &&
5960 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5961 // (X << 31) <s 0 --> (X&1) != 0
5962 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5963 (TypeBits-ShAmt->getZExtValue()-1));
5965 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5966 Mask, LHSI->getName()+".mask");
5967 Value *And = InsertNewInstBefore(AndI, ICI);
5969 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5970 And, Constant::getNullValue(And->getType()));
5975 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5976 case Instruction::AShr: {
5977 // Only handle equality comparisons of shift-by-constant.
5978 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5979 if (!ShAmt || !ICI.isEquality()) break;
5981 // Check that the shift amount is in range. If not, don't perform
5982 // undefined shifts. When the shift is visited it will be
5984 uint32_t TypeBits = RHSV.getBitWidth();
5985 if (ShAmt->uge(TypeBits))
5988 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5990 // If we are comparing against bits always shifted out, the
5991 // comparison cannot succeed.
5992 APInt Comp = RHSV << ShAmtVal;
5993 if (LHSI->getOpcode() == Instruction::LShr)
5994 Comp = Comp.lshr(ShAmtVal);
5996 Comp = Comp.ashr(ShAmtVal);
5998 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5999 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6000 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6001 return ReplaceInstUsesWith(ICI, Cst);
6004 // Otherwise, check to see if the bits shifted out are known to be zero.
6005 // If so, we can compare against the unshifted value:
6006 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6007 if (LHSI->hasOneUse() &&
6008 MaskedValueIsZero(LHSI->getOperand(0),
6009 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6010 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6011 ConstantExpr::getShl(RHS, ShAmt));
6014 if (LHSI->hasOneUse()) {
6015 // Otherwise strength reduce the shift into an and.
6016 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6017 Constant *Mask = ConstantInt::get(Val);
6020 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6021 Mask, LHSI->getName()+".mask");
6022 Value *And = InsertNewInstBefore(AndI, ICI);
6023 return new ICmpInst(ICI.getPredicate(), And,
6024 ConstantExpr::getShl(RHS, ShAmt));
6029 case Instruction::SDiv:
6030 case Instruction::UDiv:
6031 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6032 // Fold this div into the comparison, producing a range check.
6033 // Determine, based on the divide type, what the range is being
6034 // checked. If there is an overflow on the low or high side, remember
6035 // it, otherwise compute the range [low, hi) bounding the new value.
6036 // See: InsertRangeTest above for the kinds of replacements possible.
6037 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6038 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6043 case Instruction::Add:
6044 // Fold: icmp pred (add, X, C1), C2
6046 if (!ICI.isEquality()) {
6047 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6049 const APInt &LHSV = LHSC->getValue();
6051 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6054 if (ICI.isSignedPredicate()) {
6055 if (CR.getLower().isSignBit()) {
6056 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6057 ConstantInt::get(CR.getUpper()));
6058 } else if (CR.getUpper().isSignBit()) {
6059 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6060 ConstantInt::get(CR.getLower()));
6063 if (CR.getLower().isMinValue()) {
6064 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6065 ConstantInt::get(CR.getUpper()));
6066 } else if (CR.getUpper().isMinValue()) {
6067 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6068 ConstantInt::get(CR.getLower()));
6075 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6076 if (ICI.isEquality()) {
6077 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6079 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6080 // the second operand is a constant, simplify a bit.
6081 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6082 switch (BO->getOpcode()) {
6083 case Instruction::SRem:
6084 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6085 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6086 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6087 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6088 Instruction *NewRem =
6089 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6091 InsertNewInstBefore(NewRem, ICI);
6092 return new ICmpInst(ICI.getPredicate(), NewRem,
6093 Constant::getNullValue(BO->getType()));
6097 case Instruction::Add:
6098 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6099 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6100 if (BO->hasOneUse())
6101 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6102 Subtract(RHS, BOp1C));
6103 } else if (RHSV == 0) {
6104 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6105 // efficiently invertible, or if the add has just this one use.
6106 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6108 if (Value *NegVal = dyn_castNegVal(BOp1))
6109 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6110 else if (Value *NegVal = dyn_castNegVal(BOp0))
6111 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6112 else if (BO->hasOneUse()) {
6113 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6114 InsertNewInstBefore(Neg, ICI);
6116 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6120 case Instruction::Xor:
6121 // For the xor case, we can xor two constants together, eliminating
6122 // the explicit xor.
6123 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6124 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6125 ConstantExpr::getXor(RHS, BOC));
6128 case Instruction::Sub:
6129 // Replace (([sub|xor] A, B) != 0) with (A != B)
6131 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6135 case Instruction::Or:
6136 // If bits are being or'd in that are not present in the constant we
6137 // are comparing against, then the comparison could never succeed!
6138 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6139 Constant *NotCI = ConstantExpr::getNot(RHS);
6140 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6141 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6146 case Instruction::And:
6147 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6148 // If bits are being compared against that are and'd out, then the
6149 // comparison can never succeed!
6150 if ((RHSV & ~BOC->getValue()) != 0)
6151 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6154 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6155 if (RHS == BOC && RHSV.isPowerOf2())
6156 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6157 ICmpInst::ICMP_NE, LHSI,
6158 Constant::getNullValue(RHS->getType()));
6160 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6161 if (BOC->getValue().isSignBit()) {
6162 Value *X = BO->getOperand(0);
6163 Constant *Zero = Constant::getNullValue(X->getType());
6164 ICmpInst::Predicate pred = isICMP_NE ?
6165 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6166 return new ICmpInst(pred, X, Zero);
6169 // ((X & ~7) == 0) --> X < 8
6170 if (RHSV == 0 && isHighOnes(BOC)) {
6171 Value *X = BO->getOperand(0);
6172 Constant *NegX = ConstantExpr::getNeg(BOC);
6173 ICmpInst::Predicate pred = isICMP_NE ?
6174 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6175 return new ICmpInst(pred, X, NegX);
6180 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6181 // Handle icmp {eq|ne} <intrinsic>, intcst.
6182 if (II->getIntrinsicID() == Intrinsic::bswap) {
6184 ICI.setOperand(0, II->getOperand(1));
6185 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6189 } else { // Not a ICMP_EQ/ICMP_NE
6190 // If the LHS is a cast from an integral value of the same size,
6191 // then since we know the RHS is a constant, try to simlify.
6192 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6193 Value *CastOp = Cast->getOperand(0);
6194 const Type *SrcTy = CastOp->getType();
6195 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6196 if (SrcTy->isInteger() &&
6197 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6198 // If this is an unsigned comparison, try to make the comparison use
6199 // smaller constant values.
6200 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6201 // X u< 128 => X s> -1
6202 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6203 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6204 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6205 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6206 // X u> 127 => X s< 0
6207 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6208 Constant::getNullValue(SrcTy));
6216 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6217 /// We only handle extending casts so far.
6219 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6220 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6221 Value *LHSCIOp = LHSCI->getOperand(0);
6222 const Type *SrcTy = LHSCIOp->getType();
6223 const Type *DestTy = LHSCI->getType();
6226 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6227 // integer type is the same size as the pointer type.
6228 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6229 getTargetData().getPointerSizeInBits() ==
6230 cast<IntegerType>(DestTy)->getBitWidth()) {
6232 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6233 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6234 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6235 RHSOp = RHSC->getOperand(0);
6236 // If the pointer types don't match, insert a bitcast.
6237 if (LHSCIOp->getType() != RHSOp->getType())
6238 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6242 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6245 // The code below only handles extension cast instructions, so far.
6247 if (LHSCI->getOpcode() != Instruction::ZExt &&
6248 LHSCI->getOpcode() != Instruction::SExt)
6251 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6252 bool isSignedCmp = ICI.isSignedPredicate();
6254 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6255 // Not an extension from the same type?
6256 RHSCIOp = CI->getOperand(0);
6257 if (RHSCIOp->getType() != LHSCIOp->getType())
6260 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6261 // and the other is a zext), then we can't handle this.
6262 if (CI->getOpcode() != LHSCI->getOpcode())
6265 // Deal with equality cases early.
6266 if (ICI.isEquality())
6267 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6269 // A signed comparison of sign extended values simplifies into a
6270 // signed comparison.
6271 if (isSignedCmp && isSignedExt)
6272 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6274 // The other three cases all fold into an unsigned comparison.
6275 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6278 // If we aren't dealing with a constant on the RHS, exit early
6279 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6283 // Compute the constant that would happen if we truncated to SrcTy then
6284 // reextended to DestTy.
6285 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6286 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6288 // If the re-extended constant didn't change...
6290 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6291 // For example, we might have:
6292 // %A = sext short %X to uint
6293 // %B = icmp ugt uint %A, 1330
6294 // It is incorrect to transform this into
6295 // %B = icmp ugt short %X, 1330
6296 // because %A may have negative value.
6298 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6299 // OR operation is EQ/NE.
6300 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6301 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6306 // The re-extended constant changed so the constant cannot be represented
6307 // in the shorter type. Consequently, we cannot emit a simple comparison.
6309 // First, handle some easy cases. We know the result cannot be equal at this
6310 // point so handle the ICI.isEquality() cases
6311 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6312 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6313 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6314 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6316 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6317 // should have been folded away previously and not enter in here.
6320 // We're performing a signed comparison.
6321 if (cast<ConstantInt>(CI)->getValue().isNegative())
6322 Result = ConstantInt::getFalse(); // X < (small) --> false
6324 Result = ConstantInt::getTrue(); // X < (large) --> true
6326 // We're performing an unsigned comparison.
6328 // We're performing an unsigned comp with a sign extended value.
6329 // This is true if the input is >= 0. [aka >s -1]
6330 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6331 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6332 NegOne, ICI.getName()), ICI);
6334 // Unsigned extend & unsigned compare -> always true.
6335 Result = ConstantInt::getTrue();
6339 // Finally, return the value computed.
6340 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6341 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6342 return ReplaceInstUsesWith(ICI, Result);
6344 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6345 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6346 "ICmp should be folded!");
6347 if (Constant *CI = dyn_cast<Constant>(Result))
6348 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6350 return BinaryOperator::CreateNot(Result);
6354 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6355 return commonShiftTransforms(I);
6358 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6359 return commonShiftTransforms(I);
6362 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6363 if (Instruction *R = commonShiftTransforms(I))
6366 Value *Op0 = I.getOperand(0);
6368 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6369 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6370 if (CSI->isAllOnesValue())
6371 return ReplaceInstUsesWith(I, CSI);
6373 // See if we can turn a signed shr into an unsigned shr.
6374 if (MaskedValueIsZero(Op0,
6375 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6376 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6381 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6382 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6383 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6385 // shl X, 0 == X and shr X, 0 == X
6386 // shl 0, X == 0 and shr 0, X == 0
6387 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6388 Op0 == Constant::getNullValue(Op0->getType()))
6389 return ReplaceInstUsesWith(I, Op0);
6391 if (isa<UndefValue>(Op0)) {
6392 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6393 return ReplaceInstUsesWith(I, Op0);
6394 else // undef << X -> 0, undef >>u X -> 0
6395 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6397 if (isa<UndefValue>(Op1)) {
6398 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6399 return ReplaceInstUsesWith(I, Op0);
6400 else // X << undef, X >>u undef -> 0
6401 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6404 // Try to fold constant and into select arguments.
6405 if (isa<Constant>(Op0))
6406 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6407 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6410 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6411 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6416 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6417 BinaryOperator &I) {
6418 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6420 // See if we can simplify any instructions used by the instruction whose sole
6421 // purpose is to compute bits we don't care about.
6422 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6423 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6424 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6425 KnownZero, KnownOne))
6428 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6429 // of a signed value.
6431 if (Op1->uge(TypeBits)) {
6432 if (I.getOpcode() != Instruction::AShr)
6433 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6435 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6440 // ((X*C1) << C2) == (X * (C1 << C2))
6441 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6442 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6443 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6444 return BinaryOperator::CreateMul(BO->getOperand(0),
6445 ConstantExpr::getShl(BOOp, Op1));
6447 // Try to fold constant and into select arguments.
6448 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6449 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6451 if (isa<PHINode>(Op0))
6452 if (Instruction *NV = FoldOpIntoPhi(I))
6455 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6456 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6457 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6458 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6459 // place. Don't try to do this transformation in this case. Also, we
6460 // require that the input operand is a shift-by-constant so that we have
6461 // confidence that the shifts will get folded together. We could do this
6462 // xform in more cases, but it is unlikely to be profitable.
6463 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6464 isa<ConstantInt>(TrOp->getOperand(1))) {
6465 // Okay, we'll do this xform. Make the shift of shift.
6466 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6467 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6469 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6471 // For logical shifts, the truncation has the effect of making the high
6472 // part of the register be zeros. Emulate this by inserting an AND to
6473 // clear the top bits as needed. This 'and' will usually be zapped by
6474 // other xforms later if dead.
6475 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6476 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6477 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6479 // The mask we constructed says what the trunc would do if occurring
6480 // between the shifts. We want to know the effect *after* the second
6481 // shift. We know that it is a logical shift by a constant, so adjust the
6482 // mask as appropriate.
6483 if (I.getOpcode() == Instruction::Shl)
6484 MaskV <<= Op1->getZExtValue();
6486 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6487 MaskV = MaskV.lshr(Op1->getZExtValue());
6490 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6492 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6494 // Return the value truncated to the interesting size.
6495 return new TruncInst(And, I.getType());
6499 if (Op0->hasOneUse()) {
6500 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6501 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6504 switch (Op0BO->getOpcode()) {
6506 case Instruction::Add:
6507 case Instruction::And:
6508 case Instruction::Or:
6509 case Instruction::Xor: {
6510 // These operators commute.
6511 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6512 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6513 match(Op0BO->getOperand(1),
6514 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6515 Instruction *YS = BinaryOperator::CreateShl(
6516 Op0BO->getOperand(0), Op1,
6518 InsertNewInstBefore(YS, I); // (Y << C)
6520 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6521 Op0BO->getOperand(1)->getName());
6522 InsertNewInstBefore(X, I); // (X + (Y << C))
6523 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6524 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6525 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6528 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6529 Value *Op0BOOp1 = Op0BO->getOperand(1);
6530 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6532 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6533 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6535 Instruction *YS = BinaryOperator::CreateShl(
6536 Op0BO->getOperand(0), Op1,
6538 InsertNewInstBefore(YS, I); // (Y << C)
6540 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6541 V1->getName()+".mask");
6542 InsertNewInstBefore(XM, I); // X & (CC << C)
6544 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6549 case Instruction::Sub: {
6550 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6551 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6552 match(Op0BO->getOperand(0),
6553 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6554 Instruction *YS = BinaryOperator::CreateShl(
6555 Op0BO->getOperand(1), Op1,
6557 InsertNewInstBefore(YS, I); // (Y << C)
6559 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6560 Op0BO->getOperand(0)->getName());
6561 InsertNewInstBefore(X, I); // (X + (Y << C))
6562 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6563 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6564 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6567 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6568 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6569 match(Op0BO->getOperand(0),
6570 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6571 m_ConstantInt(CC))) && V2 == Op1 &&
6572 cast<BinaryOperator>(Op0BO->getOperand(0))
6573 ->getOperand(0)->hasOneUse()) {
6574 Instruction *YS = BinaryOperator::CreateShl(
6575 Op0BO->getOperand(1), Op1,
6577 InsertNewInstBefore(YS, I); // (Y << C)
6579 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6580 V1->getName()+".mask");
6581 InsertNewInstBefore(XM, I); // X & (CC << C)
6583 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
6591 // If the operand is an bitwise operator with a constant RHS, and the
6592 // shift is the only use, we can pull it out of the shift.
6593 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6594 bool isValid = true; // Valid only for And, Or, Xor
6595 bool highBitSet = false; // Transform if high bit of constant set?
6597 switch (Op0BO->getOpcode()) {
6598 default: isValid = false; break; // Do not perform transform!
6599 case Instruction::Add:
6600 isValid = isLeftShift;
6602 case Instruction::Or:
6603 case Instruction::Xor:
6606 case Instruction::And:
6611 // If this is a signed shift right, and the high bit is modified
6612 // by the logical operation, do not perform the transformation.
6613 // The highBitSet boolean indicates the value of the high bit of
6614 // the constant which would cause it to be modified for this
6617 if (isValid && I.getOpcode() == Instruction::AShr)
6618 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6621 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6623 Instruction *NewShift =
6624 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6625 InsertNewInstBefore(NewShift, I);
6626 NewShift->takeName(Op0BO);
6628 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
6635 // Find out if this is a shift of a shift by a constant.
6636 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6637 if (ShiftOp && !ShiftOp->isShift())
6640 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6641 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6642 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6643 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6644 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6645 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6646 Value *X = ShiftOp->getOperand(0);
6648 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6649 if (AmtSum > TypeBits)
6652 const IntegerType *Ty = cast<IntegerType>(I.getType());
6654 // Check for (X << c1) << c2 and (X >> c1) >> c2
6655 if (I.getOpcode() == ShiftOp->getOpcode()) {
6656 return BinaryOperator::Create(I.getOpcode(), X,
6657 ConstantInt::get(Ty, AmtSum));
6658 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6659 I.getOpcode() == Instruction::AShr) {
6660 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6661 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
6662 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6663 I.getOpcode() == Instruction::LShr) {
6664 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6665 Instruction *Shift =
6666 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
6667 InsertNewInstBefore(Shift, I);
6669 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6670 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6673 // Okay, if we get here, one shift must be left, and the other shift must be
6674 // right. See if the amounts are equal.
6675 if (ShiftAmt1 == ShiftAmt2) {
6676 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6677 if (I.getOpcode() == Instruction::Shl) {
6678 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6679 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6681 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6682 if (I.getOpcode() == Instruction::LShr) {
6683 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6684 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6686 // We can simplify ((X << C) >>s C) into a trunc + sext.
6687 // NOTE: we could do this for any C, but that would make 'unusual' integer
6688 // types. For now, just stick to ones well-supported by the code
6690 const Type *SExtType = 0;
6691 switch (Ty->getBitWidth() - ShiftAmt1) {
6698 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6703 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6704 InsertNewInstBefore(NewTrunc, I);
6705 return new SExtInst(NewTrunc, Ty);
6707 // Otherwise, we can't handle it yet.
6708 } else if (ShiftAmt1 < ShiftAmt2) {
6709 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6711 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6712 if (I.getOpcode() == Instruction::Shl) {
6713 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6714 ShiftOp->getOpcode() == Instruction::AShr);
6715 Instruction *Shift =
6716 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6717 InsertNewInstBefore(Shift, I);
6719 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6720 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6723 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6724 if (I.getOpcode() == Instruction::LShr) {
6725 assert(ShiftOp->getOpcode() == Instruction::Shl);
6726 Instruction *Shift =
6727 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
6728 InsertNewInstBefore(Shift, I);
6730 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6731 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6734 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6736 assert(ShiftAmt2 < ShiftAmt1);
6737 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6739 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6740 if (I.getOpcode() == Instruction::Shl) {
6741 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6742 ShiftOp->getOpcode() == Instruction::AShr);
6743 Instruction *Shift =
6744 BinaryOperator::Create(ShiftOp->getOpcode(), X,
6745 ConstantInt::get(Ty, ShiftDiff));
6746 InsertNewInstBefore(Shift, I);
6748 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6749 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6752 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6753 if (I.getOpcode() == Instruction::LShr) {
6754 assert(ShiftOp->getOpcode() == Instruction::Shl);
6755 Instruction *Shift =
6756 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6757 InsertNewInstBefore(Shift, I);
6759 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6760 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6763 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6770 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6771 /// expression. If so, decompose it, returning some value X, such that Val is
6774 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6776 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6777 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6778 Offset = CI->getZExtValue();
6780 return ConstantInt::get(Type::Int32Ty, 0);
6781 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6782 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6783 if (I->getOpcode() == Instruction::Shl) {
6784 // This is a value scaled by '1 << the shift amt'.
6785 Scale = 1U << RHS->getZExtValue();
6787 return I->getOperand(0);
6788 } else if (I->getOpcode() == Instruction::Mul) {
6789 // This value is scaled by 'RHS'.
6790 Scale = RHS->getZExtValue();
6792 return I->getOperand(0);
6793 } else if (I->getOpcode() == Instruction::Add) {
6794 // We have X+C. Check to see if we really have (X*C2)+C1,
6795 // where C1 is divisible by C2.
6798 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6799 Offset += RHS->getZExtValue();
6806 // Otherwise, we can't look past this.
6813 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6814 /// try to eliminate the cast by moving the type information into the alloc.
6815 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6816 AllocationInst &AI) {
6817 const PointerType *PTy = cast<PointerType>(CI.getType());
6819 // Remove any uses of AI that are dead.
6820 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6822 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6823 Instruction *User = cast<Instruction>(*UI++);
6824 if (isInstructionTriviallyDead(User)) {
6825 while (UI != E && *UI == User)
6826 ++UI; // If this instruction uses AI more than once, don't break UI.
6829 DOUT << "IC: DCE: " << *User;
6830 EraseInstFromFunction(*User);
6834 // Get the type really allocated and the type casted to.
6835 const Type *AllocElTy = AI.getAllocatedType();
6836 const Type *CastElTy = PTy->getElementType();
6837 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6839 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6840 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6841 if (CastElTyAlign < AllocElTyAlign) return 0;
6843 // If the allocation has multiple uses, only promote it if we are strictly
6844 // increasing the alignment of the resultant allocation. If we keep it the
6845 // same, we open the door to infinite loops of various kinds.
6846 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6848 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6849 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6850 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6852 // See if we can satisfy the modulus by pulling a scale out of the array
6854 unsigned ArraySizeScale;
6856 Value *NumElements = // See if the array size is a decomposable linear expr.
6857 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6859 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6861 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6862 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6864 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6869 // If the allocation size is constant, form a constant mul expression
6870 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6871 if (isa<ConstantInt>(NumElements))
6872 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6873 // otherwise multiply the amount and the number of elements
6874 else if (Scale != 1) {
6875 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
6876 Amt = InsertNewInstBefore(Tmp, AI);
6880 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6881 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6882 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
6883 Amt = InsertNewInstBefore(Tmp, AI);
6886 AllocationInst *New;
6887 if (isa<MallocInst>(AI))
6888 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6890 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6891 InsertNewInstBefore(New, AI);
6894 // If the allocation has multiple uses, insert a cast and change all things
6895 // that used it to use the new cast. This will also hack on CI, but it will
6897 if (!AI.hasOneUse()) {
6898 AddUsesToWorkList(AI);
6899 // New is the allocation instruction, pointer typed. AI is the original
6900 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6901 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6902 InsertNewInstBefore(NewCast, AI);
6903 AI.replaceAllUsesWith(NewCast);
6905 return ReplaceInstUsesWith(CI, New);
6908 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6909 /// and return it as type Ty without inserting any new casts and without
6910 /// changing the computed value. This is used by code that tries to decide
6911 /// whether promoting or shrinking integer operations to wider or smaller types
6912 /// will allow us to eliminate a truncate or extend.
6914 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6915 /// extension operation if Ty is larger.
6917 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
6918 /// should return true if trunc(V) can be computed by computing V in the smaller
6919 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
6920 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
6921 /// efficiently truncated.
6923 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
6924 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
6925 /// the final result.
6926 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6928 int &NumCastsRemoved) {
6929 // We can always evaluate constants in another type.
6930 if (isa<ConstantInt>(V))
6933 Instruction *I = dyn_cast<Instruction>(V);
6934 if (!I) return false;
6936 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6938 // If this is an extension or truncate, we can often eliminate it.
6939 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6940 // If this is a cast from the destination type, we can trivially eliminate
6941 // it, and this will remove a cast overall.
6942 if (I->getOperand(0)->getType() == Ty) {
6943 // If the first operand is itself a cast, and is eliminable, do not count
6944 // this as an eliminable cast. We would prefer to eliminate those two
6946 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
6952 // We can't extend or shrink something that has multiple uses: doing so would
6953 // require duplicating the instruction in general, which isn't profitable.
6954 if (!I->hasOneUse()) return false;
6956 switch (I->getOpcode()) {
6957 case Instruction::Add:
6958 case Instruction::Sub:
6959 case Instruction::Mul:
6960 case Instruction::And:
6961 case Instruction::Or:
6962 case Instruction::Xor:
6963 // These operators can all arbitrarily be extended or truncated.
6964 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6966 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6969 case Instruction::Shl:
6970 // If we are truncating the result of this SHL, and if it's a shift of a
6971 // constant amount, we can always perform a SHL in a smaller type.
6972 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6973 uint32_t BitWidth = Ty->getBitWidth();
6974 if (BitWidth < OrigTy->getBitWidth() &&
6975 CI->getLimitedValue(BitWidth) < BitWidth)
6976 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6980 case Instruction::LShr:
6981 // If this is a truncate of a logical shr, we can truncate it to a smaller
6982 // lshr iff we know that the bits we would otherwise be shifting in are
6984 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6985 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6986 uint32_t BitWidth = Ty->getBitWidth();
6987 if (BitWidth < OrigBitWidth &&
6988 MaskedValueIsZero(I->getOperand(0),
6989 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6990 CI->getLimitedValue(BitWidth) < BitWidth) {
6991 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6996 case Instruction::ZExt:
6997 case Instruction::SExt:
6998 case Instruction::Trunc:
6999 // If this is the same kind of case as our original (e.g. zext+zext), we
7000 // can safely replace it. Note that replacing it does not reduce the number
7001 // of casts in the input.
7002 if (I->getOpcode() == CastOpc)
7005 case Instruction::Select: {
7006 SelectInst *SI = cast<SelectInst>(I);
7007 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7009 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7012 case Instruction::PHI: {
7013 // We can change a phi if we can change all operands.
7014 PHINode *PN = cast<PHINode>(I);
7015 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7016 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7022 // TODO: Can handle more cases here.
7029 /// EvaluateInDifferentType - Given an expression that
7030 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7031 /// evaluate the expression.
7032 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7034 if (Constant *C = dyn_cast<Constant>(V))
7035 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7037 // Otherwise, it must be an instruction.
7038 Instruction *I = cast<Instruction>(V);
7039 Instruction *Res = 0;
7040 switch (I->getOpcode()) {
7041 case Instruction::Add:
7042 case Instruction::Sub:
7043 case Instruction::Mul:
7044 case Instruction::And:
7045 case Instruction::Or:
7046 case Instruction::Xor:
7047 case Instruction::AShr:
7048 case Instruction::LShr:
7049 case Instruction::Shl: {
7050 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7051 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7052 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7056 case Instruction::Trunc:
7057 case Instruction::ZExt:
7058 case Instruction::SExt:
7059 // If the source type of the cast is the type we're trying for then we can
7060 // just return the source. There's no need to insert it because it is not
7062 if (I->getOperand(0)->getType() == Ty)
7063 return I->getOperand(0);
7065 // Otherwise, must be the same type of cast, so just reinsert a new one.
7066 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7069 case Instruction::Select: {
7070 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7071 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7072 Res = SelectInst::Create(I->getOperand(0), True, False);
7075 case Instruction::PHI: {
7076 PHINode *OPN = cast<PHINode>(I);
7077 PHINode *NPN = PHINode::Create(Ty);
7078 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7079 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7080 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7086 // TODO: Can handle more cases here.
7087 assert(0 && "Unreachable!");
7092 return InsertNewInstBefore(Res, *I);
7095 /// @brief Implement the transforms common to all CastInst visitors.
7096 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7097 Value *Src = CI.getOperand(0);
7099 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7100 // eliminate it now.
7101 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7102 if (Instruction::CastOps opc =
7103 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7104 // The first cast (CSrc) is eliminable so we need to fix up or replace
7105 // the second cast (CI). CSrc will then have a good chance of being dead.
7106 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7110 // If we are casting a select then fold the cast into the select
7111 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7112 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7115 // If we are casting a PHI then fold the cast into the PHI
7116 if (isa<PHINode>(Src))
7117 if (Instruction *NV = FoldOpIntoPhi(CI))
7123 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7124 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7125 Value *Src = CI.getOperand(0);
7127 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7128 // If casting the result of a getelementptr instruction with no offset, turn
7129 // this into a cast of the original pointer!
7130 if (GEP->hasAllZeroIndices()) {
7131 // Changing the cast operand is usually not a good idea but it is safe
7132 // here because the pointer operand is being replaced with another
7133 // pointer operand so the opcode doesn't need to change.
7135 CI.setOperand(0, GEP->getOperand(0));
7139 // If the GEP has a single use, and the base pointer is a bitcast, and the
7140 // GEP computes a constant offset, see if we can convert these three
7141 // instructions into fewer. This typically happens with unions and other
7142 // non-type-safe code.
7143 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7144 if (GEP->hasAllConstantIndices()) {
7145 // We are guaranteed to get a constant from EmitGEPOffset.
7146 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7147 int64_t Offset = OffsetV->getSExtValue();
7149 // Get the base pointer input of the bitcast, and the type it points to.
7150 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7151 const Type *GEPIdxTy =
7152 cast<PointerType>(OrigBase->getType())->getElementType();
7153 if (GEPIdxTy->isSized()) {
7154 SmallVector<Value*, 8> NewIndices;
7156 // Start with the index over the outer type. Note that the type size
7157 // might be zero (even if the offset isn't zero) if the indexed type
7158 // is something like [0 x {int, int}]
7159 const Type *IntPtrTy = TD->getIntPtrType();
7160 int64_t FirstIdx = 0;
7161 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7162 FirstIdx = Offset/TySize;
7165 // Handle silly modulus not returning values values [0..TySize).
7169 assert(Offset >= 0);
7171 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7174 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7176 // Index into the types. If we fail, set OrigBase to null.
7178 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7179 const StructLayout *SL = TD->getStructLayout(STy);
7180 if (Offset < (int64_t)SL->getSizeInBytes()) {
7181 unsigned Elt = SL->getElementContainingOffset(Offset);
7182 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7184 Offset -= SL->getElementOffset(Elt);
7185 GEPIdxTy = STy->getElementType(Elt);
7187 // Otherwise, we can't index into this, bail out.
7191 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7192 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7193 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7194 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7197 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7199 GEPIdxTy = STy->getElementType();
7201 // Otherwise, we can't index into this, bail out.
7207 // If we were able to index down into an element, create the GEP
7208 // and bitcast the result. This eliminates one bitcast, potentially
7210 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7212 NewIndices.end(), "");
7213 InsertNewInstBefore(NGEP, CI);
7214 NGEP->takeName(GEP);
7216 if (isa<BitCastInst>(CI))
7217 return new BitCastInst(NGEP, CI.getType());
7218 assert(isa<PtrToIntInst>(CI));
7219 return new PtrToIntInst(NGEP, CI.getType());
7226 return commonCastTransforms(CI);
7231 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7232 /// integer types. This function implements the common transforms for all those
7234 /// @brief Implement the transforms common to CastInst with integer operands
7235 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7236 if (Instruction *Result = commonCastTransforms(CI))
7239 Value *Src = CI.getOperand(0);
7240 const Type *SrcTy = Src->getType();
7241 const Type *DestTy = CI.getType();
7242 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7243 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7245 // See if we can simplify any instructions used by the LHS whose sole
7246 // purpose is to compute bits we don't care about.
7247 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7248 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7249 KnownZero, KnownOne))
7252 // If the source isn't an instruction or has more than one use then we
7253 // can't do anything more.
7254 Instruction *SrcI = dyn_cast<Instruction>(Src);
7255 if (!SrcI || !Src->hasOneUse())
7258 // Attempt to propagate the cast into the instruction for int->int casts.
7259 int NumCastsRemoved = 0;
7260 if (!isa<BitCastInst>(CI) &&
7261 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7262 CI.getOpcode(), NumCastsRemoved)) {
7263 // If this cast is a truncate, evaluting in a different type always
7264 // eliminates the cast, so it is always a win. If this is a zero-extension,
7265 // we need to do an AND to maintain the clear top-part of the computation,
7266 // so we require that the input have eliminated at least one cast. If this
7267 // is a sign extension, we insert two new casts (to do the extension) so we
7268 // require that two casts have been eliminated.
7270 switch (CI.getOpcode()) {
7272 // All the others use floating point so we shouldn't actually
7273 // get here because of the check above.
7274 assert(0 && "Unknown cast type");
7275 case Instruction::Trunc:
7278 case Instruction::ZExt:
7279 DoXForm = NumCastsRemoved >= 1;
7281 case Instruction::SExt:
7282 DoXForm = NumCastsRemoved >= 2;
7287 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7288 CI.getOpcode() == Instruction::SExt);
7289 assert(Res->getType() == DestTy);
7290 switch (CI.getOpcode()) {
7291 default: assert(0 && "Unknown cast type!");
7292 case Instruction::Trunc:
7293 case Instruction::BitCast:
7294 // Just replace this cast with the result.
7295 return ReplaceInstUsesWith(CI, Res);
7296 case Instruction::ZExt: {
7297 // We need to emit an AND to clear the high bits.
7298 assert(SrcBitSize < DestBitSize && "Not a zext?");
7299 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7301 return BinaryOperator::CreateAnd(Res, C);
7303 case Instruction::SExt:
7304 // We need to emit a cast to truncate, then a cast to sext.
7305 return CastInst::Create(Instruction::SExt,
7306 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7312 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7313 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7315 switch (SrcI->getOpcode()) {
7316 case Instruction::Add:
7317 case Instruction::Mul:
7318 case Instruction::And:
7319 case Instruction::Or:
7320 case Instruction::Xor:
7321 // If we are discarding information, rewrite.
7322 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7323 // Don't insert two casts if they cannot be eliminated. We allow
7324 // two casts to be inserted if the sizes are the same. This could
7325 // only be converting signedness, which is a noop.
7326 if (DestBitSize == SrcBitSize ||
7327 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7328 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7329 Instruction::CastOps opcode = CI.getOpcode();
7330 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7331 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7332 return BinaryOperator::Create(
7333 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7337 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7338 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7339 SrcI->getOpcode() == Instruction::Xor &&
7340 Op1 == ConstantInt::getTrue() &&
7341 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7342 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7343 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7346 case Instruction::SDiv:
7347 case Instruction::UDiv:
7348 case Instruction::SRem:
7349 case Instruction::URem:
7350 // If we are just changing the sign, rewrite.
7351 if (DestBitSize == SrcBitSize) {
7352 // Don't insert two casts if they cannot be eliminated. We allow
7353 // two casts to be inserted if the sizes are the same. This could
7354 // only be converting signedness, which is a noop.
7355 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7356 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7357 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7359 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7361 return BinaryOperator::Create(
7362 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7367 case Instruction::Shl:
7368 // Allow changing the sign of the source operand. Do not allow
7369 // changing the size of the shift, UNLESS the shift amount is a
7370 // constant. We must not change variable sized shifts to a smaller
7371 // size, because it is undefined to shift more bits out than exist
7373 if (DestBitSize == SrcBitSize ||
7374 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7375 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7376 Instruction::BitCast : Instruction::Trunc);
7377 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7378 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7379 return BinaryOperator::CreateShl(Op0c, Op1c);
7382 case Instruction::AShr:
7383 // If this is a signed shr, and if all bits shifted in are about to be
7384 // truncated off, turn it into an unsigned shr to allow greater
7386 if (DestBitSize < SrcBitSize &&
7387 isa<ConstantInt>(Op1)) {
7388 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7389 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7390 // Insert the new logical shift right.
7391 return BinaryOperator::CreateLShr(Op0, Op1);
7399 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7400 if (Instruction *Result = commonIntCastTransforms(CI))
7403 Value *Src = CI.getOperand(0);
7404 const Type *Ty = CI.getType();
7405 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7406 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7408 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7409 switch (SrcI->getOpcode()) {
7411 case Instruction::LShr:
7412 // We can shrink lshr to something smaller if we know the bits shifted in
7413 // are already zeros.
7414 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7415 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7417 // Get a mask for the bits shifting in.
7418 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7419 Value* SrcIOp0 = SrcI->getOperand(0);
7420 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7421 if (ShAmt >= DestBitWidth) // All zeros.
7422 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7424 // Okay, we can shrink this. Truncate the input, then return a new
7426 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7427 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7429 return BinaryOperator::CreateLShr(V1, V2);
7431 } else { // This is a variable shr.
7433 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7434 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7435 // loop-invariant and CSE'd.
7436 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7437 Value *One = ConstantInt::get(SrcI->getType(), 1);
7439 Value *V = InsertNewInstBefore(
7440 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7442 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7443 SrcI->getOperand(0),
7445 Value *Zero = Constant::getNullValue(V->getType());
7446 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7456 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7457 /// in order to eliminate the icmp.
7458 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7460 // If we are just checking for a icmp eq of a single bit and zext'ing it
7461 // to an integer, then shift the bit to the appropriate place and then
7462 // cast to integer to avoid the comparison.
7463 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7464 const APInt &Op1CV = Op1C->getValue();
7466 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7467 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7468 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7469 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7470 if (!DoXform) return ICI;
7472 Value *In = ICI->getOperand(0);
7473 Value *Sh = ConstantInt::get(In->getType(),
7474 In->getType()->getPrimitiveSizeInBits()-1);
7475 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7476 In->getName()+".lobit"),
7478 if (In->getType() != CI.getType())
7479 In = CastInst::CreateIntegerCast(In, CI.getType(),
7480 false/*ZExt*/, "tmp", &CI);
7482 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7483 Constant *One = ConstantInt::get(In->getType(), 1);
7484 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7485 In->getName()+".not"),
7489 return ReplaceInstUsesWith(CI, In);
7494 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7495 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7496 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7497 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7498 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7499 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7500 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7501 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7502 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7503 // This only works for EQ and NE
7504 ICI->isEquality()) {
7505 // If Op1C some other power of two, convert:
7506 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7507 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7508 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7509 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7511 APInt KnownZeroMask(~KnownZero);
7512 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7513 if (!DoXform) return ICI;
7515 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7516 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7517 // (X&4) == 2 --> false
7518 // (X&4) != 2 --> true
7519 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7520 Res = ConstantExpr::getZExt(Res, CI.getType());
7521 return ReplaceInstUsesWith(CI, Res);
7524 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7525 Value *In = ICI->getOperand(0);
7527 // Perform a logical shr by shiftamt.
7528 // Insert the shift to put the result in the low bit.
7529 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7530 ConstantInt::get(In->getType(), ShiftAmt),
7531 In->getName()+".lobit"), CI);
7534 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7535 Constant *One = ConstantInt::get(In->getType(), 1);
7536 In = BinaryOperator::CreateXor(In, One, "tmp");
7537 InsertNewInstBefore(cast<Instruction>(In), CI);
7540 if (CI.getType() == In->getType())
7541 return ReplaceInstUsesWith(CI, In);
7543 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7551 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7552 // If one of the common conversion will work ..
7553 if (Instruction *Result = commonIntCastTransforms(CI))
7556 Value *Src = CI.getOperand(0);
7558 // If this is a cast of a cast
7559 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7560 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7561 // types and if the sizes are just right we can convert this into a logical
7562 // 'and' which will be much cheaper than the pair of casts.
7563 if (isa<TruncInst>(CSrc)) {
7564 // Get the sizes of the types involved
7565 Value *A = CSrc->getOperand(0);
7566 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7567 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7568 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7569 // If we're actually extending zero bits and the trunc is a no-op
7570 if (MidSize < DstSize && SrcSize == DstSize) {
7571 // Replace both of the casts with an And of the type mask.
7572 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7573 Constant *AndConst = ConstantInt::get(AndValue);
7575 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
7576 // Unfortunately, if the type changed, we need to cast it back.
7577 if (And->getType() != CI.getType()) {
7578 And->setName(CSrc->getName()+".mask");
7579 InsertNewInstBefore(And, CI);
7580 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
7587 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7588 return transformZExtICmp(ICI, CI);
7590 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7591 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7592 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7593 // of the (zext icmp) will be transformed.
7594 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7595 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7596 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7597 (transformZExtICmp(LHS, CI, false) ||
7598 transformZExtICmp(RHS, CI, false))) {
7599 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7600 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7601 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
7608 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7609 if (Instruction *I = commonIntCastTransforms(CI))
7612 Value *Src = CI.getOperand(0);
7614 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7615 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7616 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7617 // If we are just checking for a icmp eq of a single bit and zext'ing it
7618 // to an integer, then shift the bit to the appropriate place and then
7619 // cast to integer to avoid the comparison.
7620 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7621 const APInt &Op1CV = Op1C->getValue();
7623 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7624 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7625 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7626 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7627 Value *In = ICI->getOperand(0);
7628 Value *Sh = ConstantInt::get(In->getType(),
7629 In->getType()->getPrimitiveSizeInBits()-1);
7630 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
7631 In->getName()+".lobit"),
7633 if (In->getType() != CI.getType())
7634 In = CastInst::CreateIntegerCast(In, CI.getType(),
7635 true/*SExt*/, "tmp", &CI);
7637 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7638 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
7639 In->getName()+".not"), CI);
7641 return ReplaceInstUsesWith(CI, In);
7646 // See if the value being truncated is already sign extended. If so, just
7647 // eliminate the trunc/sext pair.
7648 if (getOpcode(Src) == Instruction::Trunc) {
7649 Value *Op = cast<User>(Src)->getOperand(0);
7650 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
7651 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
7652 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
7653 unsigned NumSignBits = ComputeNumSignBits(Op);
7655 if (OpBits == DestBits) {
7656 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7657 // bits, it is already ready.
7658 if (NumSignBits > DestBits-MidBits)
7659 return ReplaceInstUsesWith(CI, Op);
7660 } else if (OpBits < DestBits) {
7661 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7662 // bits, just sext from i32.
7663 if (NumSignBits > OpBits-MidBits)
7664 return new SExtInst(Op, CI.getType(), "tmp");
7666 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7667 // bits, just truncate to i32.
7668 if (NumSignBits > OpBits-MidBits)
7669 return new TruncInst(Op, CI.getType(), "tmp");
7676 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7677 /// in the specified FP type without changing its value.
7678 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7679 APFloat F = CFP->getValueAPF();
7680 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7681 return ConstantFP::get(F);
7685 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7686 /// through it until we get the source value.
7687 static Value *LookThroughFPExtensions(Value *V) {
7688 if (Instruction *I = dyn_cast<Instruction>(V))
7689 if (I->getOpcode() == Instruction::FPExt)
7690 return LookThroughFPExtensions(I->getOperand(0));
7692 // If this value is a constant, return the constant in the smallest FP type
7693 // that can accurately represent it. This allows us to turn
7694 // (float)((double)X+2.0) into x+2.0f.
7695 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7696 if (CFP->getType() == Type::PPC_FP128Ty)
7697 return V; // No constant folding of this.
7698 // See if the value can be truncated to float and then reextended.
7699 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7701 if (CFP->getType() == Type::DoubleTy)
7702 return V; // Won't shrink.
7703 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7705 // Don't try to shrink to various long double types.
7711 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7712 if (Instruction *I = commonCastTransforms(CI))
7715 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7716 // smaller than the destination type, we can eliminate the truncate by doing
7717 // the add as the smaller type. This applies to add/sub/mul/div as well as
7718 // many builtins (sqrt, etc).
7719 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7720 if (OpI && OpI->hasOneUse()) {
7721 switch (OpI->getOpcode()) {
7723 case Instruction::Add:
7724 case Instruction::Sub:
7725 case Instruction::Mul:
7726 case Instruction::FDiv:
7727 case Instruction::FRem:
7728 const Type *SrcTy = OpI->getType();
7729 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7730 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7731 if (LHSTrunc->getType() != SrcTy &&
7732 RHSTrunc->getType() != SrcTy) {
7733 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7734 // If the source types were both smaller than the destination type of
7735 // the cast, do this xform.
7736 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7737 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7738 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7740 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7742 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7751 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7752 return commonCastTransforms(CI);
7755 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
7756 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
7757 // mantissa to accurately represent all values of X. For example, do not
7758 // do this with i64->float->i64.
7759 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
7760 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7761 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
7762 SrcI->getType()->getFPMantissaWidth())
7763 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7765 return commonCastTransforms(FI);
7768 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
7769 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
7770 // mantissa to accurately represent all values of X. For example, do not
7771 // do this with i64->float->i64.
7772 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
7773 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7774 (int)FI.getType()->getPrimitiveSizeInBits() <=
7775 SrcI->getType()->getFPMantissaWidth())
7776 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7778 return commonCastTransforms(FI);
7781 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7782 return commonCastTransforms(CI);
7785 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7786 return commonCastTransforms(CI);
7789 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7790 return commonPointerCastTransforms(CI);
7793 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7794 if (Instruction *I = commonCastTransforms(CI))
7797 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7798 if (!DestPointee->isSized()) return 0;
7800 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7803 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7804 m_ConstantInt(Cst)))) {
7805 // If the source and destination operands have the same type, see if this
7806 // is a single-index GEP.
7807 if (X->getType() == CI.getType()) {
7808 // Get the size of the pointee type.
7809 uint64_t Size = TD->getABITypeSize(DestPointee);
7811 // Convert the constant to intptr type.
7812 APInt Offset = Cst->getValue();
7813 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7815 // If Offset is evenly divisible by Size, we can do this xform.
7816 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7817 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7818 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7821 // TODO: Could handle other cases, e.g. where add is indexing into field of
7823 } else if (CI.getOperand(0)->hasOneUse() &&
7824 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7825 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7826 // "inttoptr+GEP" instead of "add+intptr".
7828 // Get the size of the pointee type.
7829 uint64_t Size = TD->getABITypeSize(DestPointee);
7831 // Convert the constant to intptr type.
7832 APInt Offset = Cst->getValue();
7833 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7835 // If Offset is evenly divisible by Size, we can do this xform.
7836 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7837 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7839 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7841 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7847 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7848 // If the operands are integer typed then apply the integer transforms,
7849 // otherwise just apply the common ones.
7850 Value *Src = CI.getOperand(0);
7851 const Type *SrcTy = Src->getType();
7852 const Type *DestTy = CI.getType();
7854 if (SrcTy->isInteger() && DestTy->isInteger()) {
7855 if (Instruction *Result = commonIntCastTransforms(CI))
7857 } else if (isa<PointerType>(SrcTy)) {
7858 if (Instruction *I = commonPointerCastTransforms(CI))
7861 if (Instruction *Result = commonCastTransforms(CI))
7866 // Get rid of casts from one type to the same type. These are useless and can
7867 // be replaced by the operand.
7868 if (DestTy == Src->getType())
7869 return ReplaceInstUsesWith(CI, Src);
7871 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7872 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7873 const Type *DstElTy = DstPTy->getElementType();
7874 const Type *SrcElTy = SrcPTy->getElementType();
7876 // If the address spaces don't match, don't eliminate the bitcast, which is
7877 // required for changing types.
7878 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7881 // If we are casting a malloc or alloca to a pointer to a type of the same
7882 // size, rewrite the allocation instruction to allocate the "right" type.
7883 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7884 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7887 // If the source and destination are pointers, and this cast is equivalent
7888 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7889 // This can enhance SROA and other transforms that want type-safe pointers.
7890 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7891 unsigned NumZeros = 0;
7892 while (SrcElTy != DstElTy &&
7893 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7894 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7895 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7899 // If we found a path from the src to dest, create the getelementptr now.
7900 if (SrcElTy == DstElTy) {
7901 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7902 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7903 ((Instruction*) NULL));
7907 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7908 if (SVI->hasOneUse()) {
7909 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7910 // a bitconvert to a vector with the same # elts.
7911 if (isa<VectorType>(DestTy) &&
7912 cast<VectorType>(DestTy)->getNumElements() ==
7913 SVI->getType()->getNumElements()) {
7915 // If either of the operands is a cast from CI.getType(), then
7916 // evaluating the shuffle in the casted destination's type will allow
7917 // us to eliminate at least one cast.
7918 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7919 Tmp->getOperand(0)->getType() == DestTy) ||
7920 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7921 Tmp->getOperand(0)->getType() == DestTy)) {
7922 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7923 SVI->getOperand(0), DestTy, &CI);
7924 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7925 SVI->getOperand(1), DestTy, &CI);
7926 // Return a new shuffle vector. Use the same element ID's, as we
7927 // know the vector types match #elts.
7928 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7936 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7938 /// %D = select %cond, %C, %A
7940 /// %C = select %cond, %B, 0
7943 /// Assuming that the specified instruction is an operand to the select, return
7944 /// a bitmask indicating which operands of this instruction are foldable if they
7945 /// equal the other incoming value of the select.
7947 static unsigned GetSelectFoldableOperands(Instruction *I) {
7948 switch (I->getOpcode()) {
7949 case Instruction::Add:
7950 case Instruction::Mul:
7951 case Instruction::And:
7952 case Instruction::Or:
7953 case Instruction::Xor:
7954 return 3; // Can fold through either operand.
7955 case Instruction::Sub: // Can only fold on the amount subtracted.
7956 case Instruction::Shl: // Can only fold on the shift amount.
7957 case Instruction::LShr:
7958 case Instruction::AShr:
7961 return 0; // Cannot fold
7965 /// GetSelectFoldableConstant - For the same transformation as the previous
7966 /// function, return the identity constant that goes into the select.
7967 static Constant *GetSelectFoldableConstant(Instruction *I) {
7968 switch (I->getOpcode()) {
7969 default: assert(0 && "This cannot happen!"); abort();
7970 case Instruction::Add:
7971 case Instruction::Sub:
7972 case Instruction::Or:
7973 case Instruction::Xor:
7974 case Instruction::Shl:
7975 case Instruction::LShr:
7976 case Instruction::AShr:
7977 return Constant::getNullValue(I->getType());
7978 case Instruction::And:
7979 return Constant::getAllOnesValue(I->getType());
7980 case Instruction::Mul:
7981 return ConstantInt::get(I->getType(), 1);
7985 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7986 /// have the same opcode and only one use each. Try to simplify this.
7987 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7989 if (TI->getNumOperands() == 1) {
7990 // If this is a non-volatile load or a cast from the same type,
7993 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7996 return 0; // unknown unary op.
7999 // Fold this by inserting a select from the input values.
8000 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8001 FI->getOperand(0), SI.getName()+".v");
8002 InsertNewInstBefore(NewSI, SI);
8003 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8007 // Only handle binary operators here.
8008 if (!isa<BinaryOperator>(TI))
8011 // Figure out if the operations have any operands in common.
8012 Value *MatchOp, *OtherOpT, *OtherOpF;
8014 if (TI->getOperand(0) == FI->getOperand(0)) {
8015 MatchOp = TI->getOperand(0);
8016 OtherOpT = TI->getOperand(1);
8017 OtherOpF = FI->getOperand(1);
8018 MatchIsOpZero = true;
8019 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8020 MatchOp = TI->getOperand(1);
8021 OtherOpT = TI->getOperand(0);
8022 OtherOpF = FI->getOperand(0);
8023 MatchIsOpZero = false;
8024 } else if (!TI->isCommutative()) {
8026 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8027 MatchOp = TI->getOperand(0);
8028 OtherOpT = TI->getOperand(1);
8029 OtherOpF = FI->getOperand(0);
8030 MatchIsOpZero = true;
8031 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8032 MatchOp = TI->getOperand(1);
8033 OtherOpT = TI->getOperand(0);
8034 OtherOpF = FI->getOperand(1);
8035 MatchIsOpZero = true;
8040 // If we reach here, they do have operations in common.
8041 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8042 OtherOpF, SI.getName()+".v");
8043 InsertNewInstBefore(NewSI, SI);
8045 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8047 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8049 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8051 assert(0 && "Shouldn't get here");
8055 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8056 Value *CondVal = SI.getCondition();
8057 Value *TrueVal = SI.getTrueValue();
8058 Value *FalseVal = SI.getFalseValue();
8060 // select true, X, Y -> X
8061 // select false, X, Y -> Y
8062 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8063 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8065 // select C, X, X -> X
8066 if (TrueVal == FalseVal)
8067 return ReplaceInstUsesWith(SI, TrueVal);
8069 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8070 return ReplaceInstUsesWith(SI, FalseVal);
8071 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8072 return ReplaceInstUsesWith(SI, TrueVal);
8073 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8074 if (isa<Constant>(TrueVal))
8075 return ReplaceInstUsesWith(SI, TrueVal);
8077 return ReplaceInstUsesWith(SI, FalseVal);
8080 if (SI.getType() == Type::Int1Ty) {
8081 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8082 if (C->getZExtValue()) {
8083 // Change: A = select B, true, C --> A = or B, C
8084 return BinaryOperator::CreateOr(CondVal, FalseVal);
8086 // Change: A = select B, false, C --> A = and !B, C
8088 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8089 "not."+CondVal->getName()), SI);
8090 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8092 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8093 if (C->getZExtValue() == false) {
8094 // Change: A = select B, C, false --> A = and B, C
8095 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8097 // Change: A = select B, C, true --> A = or !B, C
8099 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8100 "not."+CondVal->getName()), SI);
8101 return BinaryOperator::CreateOr(NotCond, TrueVal);
8105 // select a, b, a -> a&b
8106 // select a, a, b -> a|b
8107 if (CondVal == TrueVal)
8108 return BinaryOperator::CreateOr(CondVal, FalseVal);
8109 else if (CondVal == FalseVal)
8110 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8113 // Selecting between two integer constants?
8114 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8115 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8116 // select C, 1, 0 -> zext C to int
8117 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8118 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8119 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8120 // select C, 0, 1 -> zext !C to int
8122 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8123 "not."+CondVal->getName()), SI);
8124 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8127 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8129 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8131 // (x <s 0) ? -1 : 0 -> ashr x, 31
8132 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8133 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8134 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8135 // The comparison constant and the result are not neccessarily the
8136 // same width. Make an all-ones value by inserting a AShr.
8137 Value *X = IC->getOperand(0);
8138 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8139 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8140 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8142 InsertNewInstBefore(SRA, SI);
8144 // Finally, convert to the type of the select RHS. We figure out
8145 // if this requires a SExt, Trunc or BitCast based on the sizes.
8146 Instruction::CastOps opc = Instruction::BitCast;
8147 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8148 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8149 if (SRASize < SISize)
8150 opc = Instruction::SExt;
8151 else if (SRASize > SISize)
8152 opc = Instruction::Trunc;
8153 return CastInst::Create(opc, SRA, SI.getType());
8158 // If one of the constants is zero (we know they can't both be) and we
8159 // have an icmp instruction with zero, and we have an 'and' with the
8160 // non-constant value, eliminate this whole mess. This corresponds to
8161 // cases like this: ((X & 27) ? 27 : 0)
8162 if (TrueValC->isZero() || FalseValC->isZero())
8163 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8164 cast<Constant>(IC->getOperand(1))->isNullValue())
8165 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8166 if (ICA->getOpcode() == Instruction::And &&
8167 isa<ConstantInt>(ICA->getOperand(1)) &&
8168 (ICA->getOperand(1) == TrueValC ||
8169 ICA->getOperand(1) == FalseValC) &&
8170 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8171 // Okay, now we know that everything is set up, we just don't
8172 // know whether we have a icmp_ne or icmp_eq and whether the
8173 // true or false val is the zero.
8174 bool ShouldNotVal = !TrueValC->isZero();
8175 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8178 V = InsertNewInstBefore(BinaryOperator::Create(
8179 Instruction::Xor, V, ICA->getOperand(1)), SI);
8180 return ReplaceInstUsesWith(SI, V);
8185 // See if we are selecting two values based on a comparison of the two values.
8186 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8187 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8188 // Transform (X == Y) ? X : Y -> Y
8189 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8190 // This is not safe in general for floating point:
8191 // consider X== -0, Y== +0.
8192 // It becomes safe if either operand is a nonzero constant.
8193 ConstantFP *CFPt, *CFPf;
8194 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8195 !CFPt->getValueAPF().isZero()) ||
8196 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8197 !CFPf->getValueAPF().isZero()))
8198 return ReplaceInstUsesWith(SI, FalseVal);
8200 // Transform (X != Y) ? X : Y -> X
8201 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8202 return ReplaceInstUsesWith(SI, TrueVal);
8203 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8205 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8206 // Transform (X == Y) ? Y : X -> X
8207 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8208 // This is not safe in general for floating point:
8209 // consider X== -0, Y== +0.
8210 // It becomes safe if either operand is a nonzero constant.
8211 ConstantFP *CFPt, *CFPf;
8212 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8213 !CFPt->getValueAPF().isZero()) ||
8214 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8215 !CFPf->getValueAPF().isZero()))
8216 return ReplaceInstUsesWith(SI, FalseVal);
8218 // Transform (X != Y) ? Y : X -> Y
8219 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8220 return ReplaceInstUsesWith(SI, TrueVal);
8221 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8225 // See if we are selecting two values based on a comparison of the two values.
8226 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8227 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8228 // Transform (X == Y) ? X : Y -> Y
8229 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8230 return ReplaceInstUsesWith(SI, FalseVal);
8231 // Transform (X != Y) ? X : Y -> X
8232 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8233 return ReplaceInstUsesWith(SI, TrueVal);
8234 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8236 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8237 // Transform (X == Y) ? Y : X -> X
8238 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8239 return ReplaceInstUsesWith(SI, FalseVal);
8240 // Transform (X != Y) ? Y : X -> Y
8241 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8242 return ReplaceInstUsesWith(SI, TrueVal);
8243 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8247 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8248 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8249 if (TI->hasOneUse() && FI->hasOneUse()) {
8250 Instruction *AddOp = 0, *SubOp = 0;
8252 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8253 if (TI->getOpcode() == FI->getOpcode())
8254 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8257 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8258 // even legal for FP.
8259 if (TI->getOpcode() == Instruction::Sub &&
8260 FI->getOpcode() == Instruction::Add) {
8261 AddOp = FI; SubOp = TI;
8262 } else if (FI->getOpcode() == Instruction::Sub &&
8263 TI->getOpcode() == Instruction::Add) {
8264 AddOp = TI; SubOp = FI;
8268 Value *OtherAddOp = 0;
8269 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8270 OtherAddOp = AddOp->getOperand(1);
8271 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8272 OtherAddOp = AddOp->getOperand(0);
8276 // So at this point we know we have (Y -> OtherAddOp):
8277 // select C, (add X, Y), (sub X, Z)
8278 Value *NegVal; // Compute -Z
8279 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8280 NegVal = ConstantExpr::getNeg(C);
8282 NegVal = InsertNewInstBefore(
8283 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8286 Value *NewTrueOp = OtherAddOp;
8287 Value *NewFalseOp = NegVal;
8289 std::swap(NewTrueOp, NewFalseOp);
8290 Instruction *NewSel =
8291 SelectInst::Create(CondVal, NewTrueOp,
8292 NewFalseOp, SI.getName() + ".p");
8294 NewSel = InsertNewInstBefore(NewSel, SI);
8295 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8300 // See if we can fold the select into one of our operands.
8301 if (SI.getType()->isInteger()) {
8302 // See the comment above GetSelectFoldableOperands for a description of the
8303 // transformation we are doing here.
8304 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8305 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8306 !isa<Constant>(FalseVal))
8307 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8308 unsigned OpToFold = 0;
8309 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8311 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8316 Constant *C = GetSelectFoldableConstant(TVI);
8317 Instruction *NewSel =
8318 SelectInst::Create(SI.getCondition(),
8319 TVI->getOperand(2-OpToFold), C);
8320 InsertNewInstBefore(NewSel, SI);
8321 NewSel->takeName(TVI);
8322 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8323 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8325 assert(0 && "Unknown instruction!!");
8330 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8331 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8332 !isa<Constant>(TrueVal))
8333 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8334 unsigned OpToFold = 0;
8335 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8337 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8342 Constant *C = GetSelectFoldableConstant(FVI);
8343 Instruction *NewSel =
8344 SelectInst::Create(SI.getCondition(), C,
8345 FVI->getOperand(2-OpToFold));
8346 InsertNewInstBefore(NewSel, SI);
8347 NewSel->takeName(FVI);
8348 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8349 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8351 assert(0 && "Unknown instruction!!");
8356 if (BinaryOperator::isNot(CondVal)) {
8357 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8358 SI.setOperand(1, FalseVal);
8359 SI.setOperand(2, TrueVal);
8366 /// EnforceKnownAlignment - If the specified pointer points to an object that
8367 /// we control, modify the object's alignment to PrefAlign. This isn't
8368 /// often possible though. If alignment is important, a more reliable approach
8369 /// is to simply align all global variables and allocation instructions to
8370 /// their preferred alignment from the beginning.
8372 static unsigned EnforceKnownAlignment(Value *V,
8373 unsigned Align, unsigned PrefAlign) {
8375 User *U = dyn_cast<User>(V);
8376 if (!U) return Align;
8378 switch (getOpcode(U)) {
8380 case Instruction::BitCast:
8381 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8382 case Instruction::GetElementPtr: {
8383 // If all indexes are zero, it is just the alignment of the base pointer.
8384 bool AllZeroOperands = true;
8385 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8386 if (!isa<Constant>(*i) ||
8387 !cast<Constant>(*i)->isNullValue()) {
8388 AllZeroOperands = false;
8392 if (AllZeroOperands) {
8393 // Treat this like a bitcast.
8394 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8400 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8401 // If there is a large requested alignment and we can, bump up the alignment
8403 if (!GV->isDeclaration()) {
8404 GV->setAlignment(PrefAlign);
8407 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8408 // If there is a requested alignment and if this is an alloca, round up. We
8409 // don't do this for malloc, because some systems can't respect the request.
8410 if (isa<AllocaInst>(AI)) {
8411 AI->setAlignment(PrefAlign);
8419 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8420 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8421 /// and it is more than the alignment of the ultimate object, see if we can
8422 /// increase the alignment of the ultimate object, making this check succeed.
8423 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8424 unsigned PrefAlign) {
8425 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8426 sizeof(PrefAlign) * CHAR_BIT;
8427 APInt Mask = APInt::getAllOnesValue(BitWidth);
8428 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8429 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8430 unsigned TrailZ = KnownZero.countTrailingOnes();
8431 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8433 if (PrefAlign > Align)
8434 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8436 // We don't need to make any adjustment.
8440 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8441 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8442 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8443 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8444 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8446 if (CopyAlign < MinAlign) {
8447 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8451 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8453 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8454 if (MemOpLength == 0) return 0;
8456 // Source and destination pointer types are always "i8*" for intrinsic. See
8457 // if the size is something we can handle with a single primitive load/store.
8458 // A single load+store correctly handles overlapping memory in the memmove
8460 unsigned Size = MemOpLength->getZExtValue();
8461 if (Size == 0) return MI; // Delete this mem transfer.
8463 if (Size > 8 || (Size&(Size-1)))
8464 return 0; // If not 1/2/4/8 bytes, exit.
8466 // Use an integer load+store unless we can find something better.
8467 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8469 // Memcpy forces the use of i8* for the source and destination. That means
8470 // that if you're using memcpy to move one double around, you'll get a cast
8471 // from double* to i8*. We'd much rather use a double load+store rather than
8472 // an i64 load+store, here because this improves the odds that the source or
8473 // dest address will be promotable. See if we can find a better type than the
8474 // integer datatype.
8475 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8476 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8477 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8478 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8479 // down through these levels if so.
8480 while (!SrcETy->isSingleValueType()) {
8481 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8482 if (STy->getNumElements() == 1)
8483 SrcETy = STy->getElementType(0);
8486 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8487 if (ATy->getNumElements() == 1)
8488 SrcETy = ATy->getElementType();
8495 if (SrcETy->isSingleValueType())
8496 NewPtrTy = PointerType::getUnqual(SrcETy);
8501 // If the memcpy/memmove provides better alignment info than we can
8503 SrcAlign = std::max(SrcAlign, CopyAlign);
8504 DstAlign = std::max(DstAlign, CopyAlign);
8506 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8507 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8508 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8509 InsertNewInstBefore(L, *MI);
8510 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8512 // Set the size of the copy to 0, it will be deleted on the next iteration.
8513 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8517 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8518 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8519 if (MI->getAlignment()->getZExtValue() < Alignment) {
8520 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8524 // Extract the length and alignment and fill if they are constant.
8525 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8526 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8527 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8529 uint64_t Len = LenC->getZExtValue();
8530 Alignment = MI->getAlignment()->getZExtValue();
8532 // If the length is zero, this is a no-op
8533 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8535 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8536 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8537 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8539 Value *Dest = MI->getDest();
8540 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8542 // Alignment 0 is identity for alignment 1 for memset, but not store.
8543 if (Alignment == 0) Alignment = 1;
8545 // Extract the fill value and store.
8546 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8547 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8550 // Set the size of the copy to 0, it will be deleted on the next iteration.
8551 MI->setLength(Constant::getNullValue(LenC->getType()));
8559 /// visitCallInst - CallInst simplification. This mostly only handles folding
8560 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8561 /// the heavy lifting.
8563 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8564 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8565 if (!II) return visitCallSite(&CI);
8567 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8569 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8570 bool Changed = false;
8572 // memmove/cpy/set of zero bytes is a noop.
8573 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8574 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8576 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8577 if (CI->getZExtValue() == 1) {
8578 // Replace the instruction with just byte operations. We would
8579 // transform other cases to loads/stores, but we don't know if
8580 // alignment is sufficient.
8584 // If we have a memmove and the source operation is a constant global,
8585 // then the source and dest pointers can't alias, so we can change this
8586 // into a call to memcpy.
8587 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8588 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8589 if (GVSrc->isConstant()) {
8590 Module *M = CI.getParent()->getParent()->getParent();
8591 Intrinsic::ID MemCpyID;
8592 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8593 MemCpyID = Intrinsic::memcpy_i32;
8595 MemCpyID = Intrinsic::memcpy_i64;
8596 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8600 // memmove(x,x,size) -> noop.
8601 if (MMI->getSource() == MMI->getDest())
8602 return EraseInstFromFunction(CI);
8605 // If we can determine a pointer alignment that is bigger than currently
8606 // set, update the alignment.
8607 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8608 if (Instruction *I = SimplifyMemTransfer(MI))
8610 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8611 if (Instruction *I = SimplifyMemSet(MSI))
8615 if (Changed) return II;
8618 switch (II->getIntrinsicID()) {
8620 case Intrinsic::bswap:
8621 // bswap(bswap(x)) -> x
8622 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
8623 if (Operand->getIntrinsicID() == Intrinsic::bswap)
8624 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
8626 case Intrinsic::ppc_altivec_lvx:
8627 case Intrinsic::ppc_altivec_lvxl:
8628 case Intrinsic::x86_sse_loadu_ps:
8629 case Intrinsic::x86_sse2_loadu_pd:
8630 case Intrinsic::x86_sse2_loadu_dq:
8631 // Turn PPC lvx -> load if the pointer is known aligned.
8632 // Turn X86 loadups -> load if the pointer is known aligned.
8633 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8634 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8635 PointerType::getUnqual(II->getType()),
8637 return new LoadInst(Ptr);
8640 case Intrinsic::ppc_altivec_stvx:
8641 case Intrinsic::ppc_altivec_stvxl:
8642 // Turn stvx -> store if the pointer is known aligned.
8643 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8644 const Type *OpPtrTy =
8645 PointerType::getUnqual(II->getOperand(1)->getType());
8646 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8647 return new StoreInst(II->getOperand(1), Ptr);
8650 case Intrinsic::x86_sse_storeu_ps:
8651 case Intrinsic::x86_sse2_storeu_pd:
8652 case Intrinsic::x86_sse2_storeu_dq:
8653 case Intrinsic::x86_sse2_storel_dq:
8654 // Turn X86 storeu -> store if the pointer is known aligned.
8655 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8656 const Type *OpPtrTy =
8657 PointerType::getUnqual(II->getOperand(2)->getType());
8658 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8659 return new StoreInst(II->getOperand(2), Ptr);
8663 case Intrinsic::x86_sse_cvttss2si: {
8664 // These intrinsics only demands the 0th element of its input vector. If
8665 // we can simplify the input based on that, do so now.
8667 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8669 II->setOperand(1, V);
8675 case Intrinsic::ppc_altivec_vperm:
8676 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8677 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8678 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8680 // Check that all of the elements are integer constants or undefs.
8681 bool AllEltsOk = true;
8682 for (unsigned i = 0; i != 16; ++i) {
8683 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8684 !isa<UndefValue>(Mask->getOperand(i))) {
8691 // Cast the input vectors to byte vectors.
8692 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8693 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8694 Value *Result = UndefValue::get(Op0->getType());
8696 // Only extract each element once.
8697 Value *ExtractedElts[32];
8698 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8700 for (unsigned i = 0; i != 16; ++i) {
8701 if (isa<UndefValue>(Mask->getOperand(i)))
8703 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8704 Idx &= 31; // Match the hardware behavior.
8706 if (ExtractedElts[Idx] == 0) {
8708 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8709 InsertNewInstBefore(Elt, CI);
8710 ExtractedElts[Idx] = Elt;
8713 // Insert this value into the result vector.
8714 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
8716 InsertNewInstBefore(cast<Instruction>(Result), CI);
8718 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
8723 case Intrinsic::stackrestore: {
8724 // If the save is right next to the restore, remove the restore. This can
8725 // happen when variable allocas are DCE'd.
8726 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8727 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8728 BasicBlock::iterator BI = SS;
8730 return EraseInstFromFunction(CI);
8734 // Scan down this block to see if there is another stack restore in the
8735 // same block without an intervening call/alloca.
8736 BasicBlock::iterator BI = II;
8737 TerminatorInst *TI = II->getParent()->getTerminator();
8738 bool CannotRemove = false;
8739 for (++BI; &*BI != TI; ++BI) {
8740 if (isa<AllocaInst>(BI)) {
8741 CannotRemove = true;
8744 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
8745 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
8746 // If there is a stackrestore below this one, remove this one.
8747 if (II->getIntrinsicID() == Intrinsic::stackrestore)
8748 return EraseInstFromFunction(CI);
8749 // Otherwise, ignore the intrinsic.
8751 // If we found a non-intrinsic call, we can't remove the stack
8753 CannotRemove = true;
8759 // If the stack restore is in a return/unwind block and if there are no
8760 // allocas or calls between the restore and the return, nuke the restore.
8761 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8762 return EraseInstFromFunction(CI);
8767 return visitCallSite(II);
8770 // InvokeInst simplification
8772 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8773 return visitCallSite(&II);
8776 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8777 /// passed through the varargs area, we can eliminate the use of the cast.
8778 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8779 const CastInst * const CI,
8780 const TargetData * const TD,
8782 if (!CI->isLosslessCast())
8785 // The size of ByVal arguments is derived from the type, so we
8786 // can't change to a type with a different size. If the size were
8787 // passed explicitly we could avoid this check.
8788 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
8792 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
8793 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
8794 if (!SrcTy->isSized() || !DstTy->isSized())
8796 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
8801 // visitCallSite - Improvements for call and invoke instructions.
8803 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8804 bool Changed = false;
8806 // If the callee is a constexpr cast of a function, attempt to move the cast
8807 // to the arguments of the call/invoke.
8808 if (transformConstExprCastCall(CS)) return 0;
8810 Value *Callee = CS.getCalledValue();
8812 if (Function *CalleeF = dyn_cast<Function>(Callee))
8813 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8814 Instruction *OldCall = CS.getInstruction();
8815 // If the call and callee calling conventions don't match, this call must
8816 // be unreachable, as the call is undefined.
8817 new StoreInst(ConstantInt::getTrue(),
8818 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8820 if (!OldCall->use_empty())
8821 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8822 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8823 return EraseInstFromFunction(*OldCall);
8827 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8828 // This instruction is not reachable, just remove it. We insert a store to
8829 // undef so that we know that this code is not reachable, despite the fact
8830 // that we can't modify the CFG here.
8831 new StoreInst(ConstantInt::getTrue(),
8832 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8833 CS.getInstruction());
8835 if (!CS.getInstruction()->use_empty())
8836 CS.getInstruction()->
8837 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8839 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8840 // Don't break the CFG, insert a dummy cond branch.
8841 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8842 ConstantInt::getTrue(), II);
8844 return EraseInstFromFunction(*CS.getInstruction());
8847 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8848 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8849 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8850 return transformCallThroughTrampoline(CS);
8852 const PointerType *PTy = cast<PointerType>(Callee->getType());
8853 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8854 if (FTy->isVarArg()) {
8855 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
8856 // See if we can optimize any arguments passed through the varargs area of
8858 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8859 E = CS.arg_end(); I != E; ++I, ++ix) {
8860 CastInst *CI = dyn_cast<CastInst>(*I);
8861 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
8862 *I = CI->getOperand(0);
8868 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8869 // Inline asm calls cannot throw - mark them 'nounwind'.
8870 CS.setDoesNotThrow();
8874 return Changed ? CS.getInstruction() : 0;
8877 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8878 // attempt to move the cast to the arguments of the call/invoke.
8880 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8881 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8882 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8883 if (CE->getOpcode() != Instruction::BitCast ||
8884 !isa<Function>(CE->getOperand(0)))
8886 Function *Callee = cast<Function>(CE->getOperand(0));
8887 Instruction *Caller = CS.getInstruction();
8888 const PAListPtr &CallerPAL = CS.getParamAttrs();
8890 // Okay, this is a cast from a function to a different type. Unless doing so
8891 // would cause a type conversion of one of our arguments, change this call to
8892 // be a direct call with arguments casted to the appropriate types.
8894 const FunctionType *FT = Callee->getFunctionType();
8895 const Type *OldRetTy = Caller->getType();
8896 const Type *NewRetTy = FT->getReturnType();
8898 if (isa<StructType>(NewRetTy))
8899 return false; // TODO: Handle multiple return values.
8901 // Check to see if we are changing the return type...
8902 if (OldRetTy != NewRetTy) {
8903 if (Callee->isDeclaration() &&
8904 // Conversion is ok if changing from one pointer type to another or from
8905 // a pointer to an integer of the same size.
8906 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
8907 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
8908 return false; // Cannot transform this return value.
8910 if (!Caller->use_empty() &&
8911 // void -> non-void is handled specially
8912 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
8913 return false; // Cannot transform this return value.
8915 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8916 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8917 if (RAttrs & ParamAttr::typeIncompatible(NewRetTy))
8918 return false; // Attribute not compatible with transformed value.
8921 // If the callsite is an invoke instruction, and the return value is used by
8922 // a PHI node in a successor, we cannot change the return type of the call
8923 // because there is no place to put the cast instruction (without breaking
8924 // the critical edge). Bail out in this case.
8925 if (!Caller->use_empty())
8926 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8927 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8929 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8930 if (PN->getParent() == II->getNormalDest() ||
8931 PN->getParent() == II->getUnwindDest())
8935 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8936 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8938 CallSite::arg_iterator AI = CS.arg_begin();
8939 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8940 const Type *ParamTy = FT->getParamType(i);
8941 const Type *ActTy = (*AI)->getType();
8943 if (!CastInst::isCastable(ActTy, ParamTy))
8944 return false; // Cannot transform this parameter value.
8946 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8947 return false; // Attribute not compatible with transformed value.
8949 // Converting from one pointer type to another or between a pointer and an
8950 // integer of the same size is safe even if we do not have a body.
8951 bool isConvertible = ActTy == ParamTy ||
8952 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
8953 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
8954 if (Callee->isDeclaration() && !isConvertible) return false;
8957 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8958 Callee->isDeclaration())
8959 return false; // Do not delete arguments unless we have a function body.
8961 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
8962 !CallerPAL.isEmpty())
8963 // In this case we have more arguments than the new function type, but we
8964 // won't be dropping them. Check that these extra arguments have attributes
8965 // that are compatible with being a vararg call argument.
8966 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
8967 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
8969 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
8970 if (PAttrs & ParamAttr::VarArgsIncompatible)
8974 // Okay, we decided that this is a safe thing to do: go ahead and start
8975 // inserting cast instructions as necessary...
8976 std::vector<Value*> Args;
8977 Args.reserve(NumActualArgs);
8978 SmallVector<ParamAttrsWithIndex, 8> attrVec;
8979 attrVec.reserve(NumCommonArgs);
8981 // Get any return attributes.
8982 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8984 // If the return value is not being used, the type may not be compatible
8985 // with the existing attributes. Wipe out any problematic attributes.
8986 RAttrs &= ~ParamAttr::typeIncompatible(NewRetTy);
8988 // Add the new return attributes.
8990 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8992 AI = CS.arg_begin();
8993 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8994 const Type *ParamTy = FT->getParamType(i);
8995 if ((*AI)->getType() == ParamTy) {
8996 Args.push_back(*AI);
8998 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8999 false, ParamTy, false);
9000 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9001 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9004 // Add any parameter attributes.
9005 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9006 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9009 // If the function takes more arguments than the call was taking, add them
9011 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9012 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9014 // If we are removing arguments to the function, emit an obnoxious warning...
9015 if (FT->getNumParams() < NumActualArgs) {
9016 if (!FT->isVarArg()) {
9017 cerr << "WARNING: While resolving call to function '"
9018 << Callee->getName() << "' arguments were dropped!\n";
9020 // Add all of the arguments in their promoted form to the arg list...
9021 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9022 const Type *PTy = getPromotedType((*AI)->getType());
9023 if (PTy != (*AI)->getType()) {
9024 // Must promote to pass through va_arg area!
9025 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9027 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9028 InsertNewInstBefore(Cast, *Caller);
9029 Args.push_back(Cast);
9031 Args.push_back(*AI);
9034 // Add any parameter attributes.
9035 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9036 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9041 if (NewRetTy == Type::VoidTy)
9042 Caller->setName(""); // Void type should not have a name.
9044 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9047 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9048 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9049 Args.begin(), Args.end(),
9050 Caller->getName(), Caller);
9051 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9052 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9054 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9055 Caller->getName(), Caller);
9056 CallInst *CI = cast<CallInst>(Caller);
9057 if (CI->isTailCall())
9058 cast<CallInst>(NC)->setTailCall();
9059 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9060 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9063 // Insert a cast of the return type as necessary.
9065 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9066 if (NV->getType() != Type::VoidTy) {
9067 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9069 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9071 // If this is an invoke instruction, we should insert it after the first
9072 // non-phi, instruction in the normal successor block.
9073 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9074 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9075 InsertNewInstBefore(NC, *I);
9077 // Otherwise, it's a call, just insert cast right after the call instr
9078 InsertNewInstBefore(NC, *Caller);
9080 AddUsersToWorkList(*Caller);
9082 NV = UndefValue::get(Caller->getType());
9086 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9087 Caller->replaceAllUsesWith(NV);
9088 Caller->eraseFromParent();
9089 RemoveFromWorkList(Caller);
9093 // transformCallThroughTrampoline - Turn a call to a function created by the
9094 // init_trampoline intrinsic into a direct call to the underlying function.
9096 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9097 Value *Callee = CS.getCalledValue();
9098 const PointerType *PTy = cast<PointerType>(Callee->getType());
9099 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9100 const PAListPtr &Attrs = CS.getParamAttrs();
9102 // If the call already has the 'nest' attribute somewhere then give up -
9103 // otherwise 'nest' would occur twice after splicing in the chain.
9104 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9107 IntrinsicInst *Tramp =
9108 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9110 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9111 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9112 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9114 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9115 if (!NestAttrs.isEmpty()) {
9116 unsigned NestIdx = 1;
9117 const Type *NestTy = 0;
9118 ParameterAttributes NestAttr = ParamAttr::None;
9120 // Look for a parameter marked with the 'nest' attribute.
9121 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9122 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9123 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9124 // Record the parameter type and any other attributes.
9126 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9131 Instruction *Caller = CS.getInstruction();
9132 std::vector<Value*> NewArgs;
9133 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9135 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9136 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9138 // Insert the nest argument into the call argument list, which may
9139 // mean appending it. Likewise for attributes.
9141 // Add any function result attributes.
9142 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9143 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9147 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9149 if (Idx == NestIdx) {
9150 // Add the chain argument and attributes.
9151 Value *NestVal = Tramp->getOperand(3);
9152 if (NestVal->getType() != NestTy)
9153 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9154 NewArgs.push_back(NestVal);
9155 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9161 // Add the original argument and attributes.
9162 NewArgs.push_back(*I);
9163 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9165 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9171 // The trampoline may have been bitcast to a bogus type (FTy).
9172 // Handle this by synthesizing a new function type, equal to FTy
9173 // with the chain parameter inserted.
9175 std::vector<const Type*> NewTypes;
9176 NewTypes.reserve(FTy->getNumParams()+1);
9178 // Insert the chain's type into the list of parameter types, which may
9179 // mean appending it.
9182 FunctionType::param_iterator I = FTy->param_begin(),
9183 E = FTy->param_end();
9187 // Add the chain's type.
9188 NewTypes.push_back(NestTy);
9193 // Add the original type.
9194 NewTypes.push_back(*I);
9200 // Replace the trampoline call with a direct call. Let the generic
9201 // code sort out any function type mismatches.
9202 FunctionType *NewFTy =
9203 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9204 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9205 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9206 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9208 Instruction *NewCaller;
9209 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9210 NewCaller = InvokeInst::Create(NewCallee,
9211 II->getNormalDest(), II->getUnwindDest(),
9212 NewArgs.begin(), NewArgs.end(),
9213 Caller->getName(), Caller);
9214 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9215 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9217 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9218 Caller->getName(), Caller);
9219 if (cast<CallInst>(Caller)->isTailCall())
9220 cast<CallInst>(NewCaller)->setTailCall();
9221 cast<CallInst>(NewCaller)->
9222 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9223 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9225 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9226 Caller->replaceAllUsesWith(NewCaller);
9227 Caller->eraseFromParent();
9228 RemoveFromWorkList(Caller);
9233 // Replace the trampoline call with a direct call. Since there is no 'nest'
9234 // parameter, there is no need to adjust the argument list. Let the generic
9235 // code sort out any function type mismatches.
9236 Constant *NewCallee =
9237 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9238 CS.setCalledFunction(NewCallee);
9239 return CS.getInstruction();
9242 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9243 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9244 /// and a single binop.
9245 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9246 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9247 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9248 isa<CmpInst>(FirstInst));
9249 unsigned Opc = FirstInst->getOpcode();
9250 Value *LHSVal = FirstInst->getOperand(0);
9251 Value *RHSVal = FirstInst->getOperand(1);
9253 const Type *LHSType = LHSVal->getType();
9254 const Type *RHSType = RHSVal->getType();
9256 // Scan to see if all operands are the same opcode, all have one use, and all
9257 // kill their operands (i.e. the operands have one use).
9258 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9259 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9260 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9261 // Verify type of the LHS matches so we don't fold cmp's of different
9262 // types or GEP's with different index types.
9263 I->getOperand(0)->getType() != LHSType ||
9264 I->getOperand(1)->getType() != RHSType)
9267 // If they are CmpInst instructions, check their predicates
9268 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9269 if (cast<CmpInst>(I)->getPredicate() !=
9270 cast<CmpInst>(FirstInst)->getPredicate())
9273 // Keep track of which operand needs a phi node.
9274 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9275 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9278 // Otherwise, this is safe to transform, determine if it is profitable.
9280 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9281 // Indexes are often folded into load/store instructions, so we don't want to
9282 // hide them behind a phi.
9283 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9286 Value *InLHS = FirstInst->getOperand(0);
9287 Value *InRHS = FirstInst->getOperand(1);
9288 PHINode *NewLHS = 0, *NewRHS = 0;
9290 NewLHS = PHINode::Create(LHSType,
9291 FirstInst->getOperand(0)->getName() + ".pn");
9292 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9293 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9294 InsertNewInstBefore(NewLHS, PN);
9299 NewRHS = PHINode::Create(RHSType,
9300 FirstInst->getOperand(1)->getName() + ".pn");
9301 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9302 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9303 InsertNewInstBefore(NewRHS, PN);
9307 // Add all operands to the new PHIs.
9308 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9310 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9311 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9314 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9315 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9319 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9320 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9321 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9322 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9325 assert(isa<GetElementPtrInst>(FirstInst));
9326 return GetElementPtrInst::Create(LHSVal, RHSVal);
9330 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9331 /// of the block that defines it. This means that it must be obvious the value
9332 /// of the load is not changed from the point of the load to the end of the
9335 /// Finally, it is safe, but not profitable, to sink a load targetting a
9336 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9338 static bool isSafeToSinkLoad(LoadInst *L) {
9339 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9341 for (++BBI; BBI != E; ++BBI)
9342 if (BBI->mayWriteToMemory())
9345 // Check for non-address taken alloca. If not address-taken already, it isn't
9346 // profitable to do this xform.
9347 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9348 bool isAddressTaken = false;
9349 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9351 if (isa<LoadInst>(UI)) continue;
9352 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9353 // If storing TO the alloca, then the address isn't taken.
9354 if (SI->getOperand(1) == AI) continue;
9356 isAddressTaken = true;
9360 if (!isAddressTaken)
9368 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9369 // operator and they all are only used by the PHI, PHI together their
9370 // inputs, and do the operation once, to the result of the PHI.
9371 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9372 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9374 // Scan the instruction, looking for input operations that can be folded away.
9375 // If all input operands to the phi are the same instruction (e.g. a cast from
9376 // the same type or "+42") we can pull the operation through the PHI, reducing
9377 // code size and simplifying code.
9378 Constant *ConstantOp = 0;
9379 const Type *CastSrcTy = 0;
9380 bool isVolatile = false;
9381 if (isa<CastInst>(FirstInst)) {
9382 CastSrcTy = FirstInst->getOperand(0)->getType();
9383 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9384 // Can fold binop, compare or shift here if the RHS is a constant,
9385 // otherwise call FoldPHIArgBinOpIntoPHI.
9386 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9387 if (ConstantOp == 0)
9388 return FoldPHIArgBinOpIntoPHI(PN);
9389 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9390 isVolatile = LI->isVolatile();
9391 // We can't sink the load if the loaded value could be modified between the
9392 // load and the PHI.
9393 if (LI->getParent() != PN.getIncomingBlock(0) ||
9394 !isSafeToSinkLoad(LI))
9397 // If the PHI is of volatile loads and the load block has multiple
9398 // successors, sinking it would remove a load of the volatile value from
9399 // the path through the other successor.
9401 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9404 } else if (isa<GetElementPtrInst>(FirstInst)) {
9405 if (FirstInst->getNumOperands() == 2)
9406 return FoldPHIArgBinOpIntoPHI(PN);
9407 // Can't handle general GEPs yet.
9410 return 0; // Cannot fold this operation.
9413 // Check to see if all arguments are the same operation.
9414 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9415 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9416 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9417 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9420 if (I->getOperand(0)->getType() != CastSrcTy)
9421 return 0; // Cast operation must match.
9422 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9423 // We can't sink the load if the loaded value could be modified between
9424 // the load and the PHI.
9425 if (LI->isVolatile() != isVolatile ||
9426 LI->getParent() != PN.getIncomingBlock(i) ||
9427 !isSafeToSinkLoad(LI))
9430 // If the PHI is of volatile loads and the load block has multiple
9431 // successors, sinking it would remove a load of the volatile value from
9432 // the path through the other successor.
9434 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9438 } else if (I->getOperand(1) != ConstantOp) {
9443 // Okay, they are all the same operation. Create a new PHI node of the
9444 // correct type, and PHI together all of the LHS's of the instructions.
9445 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9446 PN.getName()+".in");
9447 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9449 Value *InVal = FirstInst->getOperand(0);
9450 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9452 // Add all operands to the new PHI.
9453 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9454 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9455 if (NewInVal != InVal)
9457 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9462 // The new PHI unions all of the same values together. This is really
9463 // common, so we handle it intelligently here for compile-time speed.
9467 InsertNewInstBefore(NewPN, PN);
9471 // Insert and return the new operation.
9472 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9473 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9474 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9475 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9476 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9477 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9478 PhiVal, ConstantOp);
9479 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9481 // If this was a volatile load that we are merging, make sure to loop through
9482 // and mark all the input loads as non-volatile. If we don't do this, we will
9483 // insert a new volatile load and the old ones will not be deletable.
9485 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9486 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9488 return new LoadInst(PhiVal, "", isVolatile);
9491 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9493 static bool DeadPHICycle(PHINode *PN,
9494 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9495 if (PN->use_empty()) return true;
9496 if (!PN->hasOneUse()) return false;
9498 // Remember this node, and if we find the cycle, return.
9499 if (!PotentiallyDeadPHIs.insert(PN))
9502 // Don't scan crazily complex things.
9503 if (PotentiallyDeadPHIs.size() == 16)
9506 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9507 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9512 /// PHIsEqualValue - Return true if this phi node is always equal to
9513 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9514 /// z = some value; x = phi (y, z); y = phi (x, z)
9515 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9516 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9517 // See if we already saw this PHI node.
9518 if (!ValueEqualPHIs.insert(PN))
9521 // Don't scan crazily complex things.
9522 if (ValueEqualPHIs.size() == 16)
9525 // Scan the operands to see if they are either phi nodes or are equal to
9527 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9528 Value *Op = PN->getIncomingValue(i);
9529 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9530 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9532 } else if (Op != NonPhiInVal)
9540 // PHINode simplification
9542 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9543 // If LCSSA is around, don't mess with Phi nodes
9544 if (MustPreserveLCSSA) return 0;
9546 if (Value *V = PN.hasConstantValue())
9547 return ReplaceInstUsesWith(PN, V);
9549 // If all PHI operands are the same operation, pull them through the PHI,
9550 // reducing code size.
9551 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9552 PN.getIncomingValue(0)->hasOneUse())
9553 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9556 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9557 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9558 // PHI)... break the cycle.
9559 if (PN.hasOneUse()) {
9560 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9561 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9562 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9563 PotentiallyDeadPHIs.insert(&PN);
9564 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9565 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9568 // If this phi has a single use, and if that use just computes a value for
9569 // the next iteration of a loop, delete the phi. This occurs with unused
9570 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9571 // common case here is good because the only other things that catch this
9572 // are induction variable analysis (sometimes) and ADCE, which is only run
9574 if (PHIUser->hasOneUse() &&
9575 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9576 PHIUser->use_back() == &PN) {
9577 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9581 // We sometimes end up with phi cycles that non-obviously end up being the
9582 // same value, for example:
9583 // z = some value; x = phi (y, z); y = phi (x, z)
9584 // where the phi nodes don't necessarily need to be in the same block. Do a
9585 // quick check to see if the PHI node only contains a single non-phi value, if
9586 // so, scan to see if the phi cycle is actually equal to that value.
9588 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9589 // Scan for the first non-phi operand.
9590 while (InValNo != NumOperandVals &&
9591 isa<PHINode>(PN.getIncomingValue(InValNo)))
9594 if (InValNo != NumOperandVals) {
9595 Value *NonPhiInVal = PN.getOperand(InValNo);
9597 // Scan the rest of the operands to see if there are any conflicts, if so
9598 // there is no need to recursively scan other phis.
9599 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9600 Value *OpVal = PN.getIncomingValue(InValNo);
9601 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9605 // If we scanned over all operands, then we have one unique value plus
9606 // phi values. Scan PHI nodes to see if they all merge in each other or
9608 if (InValNo == NumOperandVals) {
9609 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9610 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9611 return ReplaceInstUsesWith(PN, NonPhiInVal);
9618 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9619 Instruction *InsertPoint,
9621 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9622 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9623 // We must cast correctly to the pointer type. Ensure that we
9624 // sign extend the integer value if it is smaller as this is
9625 // used for address computation.
9626 Instruction::CastOps opcode =
9627 (VTySize < PtrSize ? Instruction::SExt :
9628 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9629 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9633 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9634 Value *PtrOp = GEP.getOperand(0);
9635 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9636 // If so, eliminate the noop.
9637 if (GEP.getNumOperands() == 1)
9638 return ReplaceInstUsesWith(GEP, PtrOp);
9640 if (isa<UndefValue>(GEP.getOperand(0)))
9641 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9643 bool HasZeroPointerIndex = false;
9644 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9645 HasZeroPointerIndex = C->isNullValue();
9647 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9648 return ReplaceInstUsesWith(GEP, PtrOp);
9650 // Eliminate unneeded casts for indices.
9651 bool MadeChange = false;
9653 gep_type_iterator GTI = gep_type_begin(GEP);
9654 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
9655 i != e; ++i, ++GTI) {
9656 if (isa<SequentialType>(*GTI)) {
9657 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
9658 if (CI->getOpcode() == Instruction::ZExt ||
9659 CI->getOpcode() == Instruction::SExt) {
9660 const Type *SrcTy = CI->getOperand(0)->getType();
9661 // We can eliminate a cast from i32 to i64 iff the target
9662 // is a 32-bit pointer target.
9663 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9665 *i = CI->getOperand(0);
9669 // If we are using a wider index than needed for this platform, shrink it
9670 // to what we need. If the incoming value needs a cast instruction,
9671 // insert it. This explicit cast can make subsequent optimizations more
9674 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9675 if (Constant *C = dyn_cast<Constant>(Op)) {
9676 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
9679 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9687 if (MadeChange) return &GEP;
9689 // If this GEP instruction doesn't move the pointer, and if the input operand
9690 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9691 // real input to the dest type.
9692 if (GEP.hasAllZeroIndices()) {
9693 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9694 // If the bitcast is of an allocation, and the allocation will be
9695 // converted to match the type of the cast, don't touch this.
9696 if (isa<AllocationInst>(BCI->getOperand(0))) {
9697 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9698 if (Instruction *I = visitBitCast(*BCI)) {
9701 BCI->getParent()->getInstList().insert(BCI, I);
9702 ReplaceInstUsesWith(*BCI, I);
9707 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9711 // Combine Indices - If the source pointer to this getelementptr instruction
9712 // is a getelementptr instruction, combine the indices of the two
9713 // getelementptr instructions into a single instruction.
9715 SmallVector<Value*, 8> SrcGEPOperands;
9716 if (User *Src = dyn_castGetElementPtr(PtrOp))
9717 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9719 if (!SrcGEPOperands.empty()) {
9720 // Note that if our source is a gep chain itself that we wait for that
9721 // chain to be resolved before we perform this transformation. This
9722 // avoids us creating a TON of code in some cases.
9724 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9725 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9726 return 0; // Wait until our source is folded to completion.
9728 SmallVector<Value*, 8> Indices;
9730 // Find out whether the last index in the source GEP is a sequential idx.
9731 bool EndsWithSequential = false;
9732 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9733 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9734 EndsWithSequential = !isa<StructType>(*I);
9736 // Can we combine the two pointer arithmetics offsets?
9737 if (EndsWithSequential) {
9738 // Replace: gep (gep %P, long B), long A, ...
9739 // With: T = long A+B; gep %P, T, ...
9741 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9742 if (SO1 == Constant::getNullValue(SO1->getType())) {
9744 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9747 // If they aren't the same type, convert both to an integer of the
9748 // target's pointer size.
9749 if (SO1->getType() != GO1->getType()) {
9750 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9751 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9752 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9753 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9755 unsigned PS = TD->getPointerSizeInBits();
9756 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9757 // Convert GO1 to SO1's type.
9758 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9760 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9761 // Convert SO1 to GO1's type.
9762 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9764 const Type *PT = TD->getIntPtrType();
9765 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9766 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9770 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9771 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9773 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
9774 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9778 // Recycle the GEP we already have if possible.
9779 if (SrcGEPOperands.size() == 2) {
9780 GEP.setOperand(0, SrcGEPOperands[0]);
9781 GEP.setOperand(1, Sum);
9784 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9785 SrcGEPOperands.end()-1);
9786 Indices.push_back(Sum);
9787 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9789 } else if (isa<Constant>(*GEP.idx_begin()) &&
9790 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9791 SrcGEPOperands.size() != 1) {
9792 // Otherwise we can do the fold if the first index of the GEP is a zero
9793 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9794 SrcGEPOperands.end());
9795 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9798 if (!Indices.empty())
9799 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9800 Indices.end(), GEP.getName());
9802 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9803 // GEP of global variable. If all of the indices for this GEP are
9804 // constants, we can promote this to a constexpr instead of an instruction.
9806 // Scan for nonconstants...
9807 SmallVector<Constant*, 8> Indices;
9808 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9809 for (; I != E && isa<Constant>(*I); ++I)
9810 Indices.push_back(cast<Constant>(*I));
9812 if (I == E) { // If they are all constants...
9813 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9814 &Indices[0],Indices.size());
9816 // Replace all uses of the GEP with the new constexpr...
9817 return ReplaceInstUsesWith(GEP, CE);
9819 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9820 if (!isa<PointerType>(X->getType())) {
9821 // Not interesting. Source pointer must be a cast from pointer.
9822 } else if (HasZeroPointerIndex) {
9823 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9824 // into : GEP [10 x i8]* X, i32 0, ...
9826 // This occurs when the program declares an array extern like "int X[];"
9828 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9829 const PointerType *XTy = cast<PointerType>(X->getType());
9830 if (const ArrayType *XATy =
9831 dyn_cast<ArrayType>(XTy->getElementType()))
9832 if (const ArrayType *CATy =
9833 dyn_cast<ArrayType>(CPTy->getElementType()))
9834 if (CATy->getElementType() == XATy->getElementType()) {
9835 // At this point, we know that the cast source type is a pointer
9836 // to an array of the same type as the destination pointer
9837 // array. Because the array type is never stepped over (there
9838 // is a leading zero) we can fold the cast into this GEP.
9839 GEP.setOperand(0, X);
9842 } else if (GEP.getNumOperands() == 2) {
9843 // Transform things like:
9844 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9845 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9846 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9847 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9848 if (isa<ArrayType>(SrcElTy) &&
9849 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9850 TD->getABITypeSize(ResElTy)) {
9852 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9853 Idx[1] = GEP.getOperand(1);
9854 Value *V = InsertNewInstBefore(
9855 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9856 // V and GEP are both pointer types --> BitCast
9857 return new BitCastInst(V, GEP.getType());
9860 // Transform things like:
9861 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9862 // (where tmp = 8*tmp2) into:
9863 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9865 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9866 uint64_t ArrayEltSize =
9867 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9869 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9870 // allow either a mul, shift, or constant here.
9872 ConstantInt *Scale = 0;
9873 if (ArrayEltSize == 1) {
9874 NewIdx = GEP.getOperand(1);
9875 Scale = ConstantInt::get(NewIdx->getType(), 1);
9876 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9877 NewIdx = ConstantInt::get(CI->getType(), 1);
9879 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9880 if (Inst->getOpcode() == Instruction::Shl &&
9881 isa<ConstantInt>(Inst->getOperand(1))) {
9882 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9883 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9884 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9885 NewIdx = Inst->getOperand(0);
9886 } else if (Inst->getOpcode() == Instruction::Mul &&
9887 isa<ConstantInt>(Inst->getOperand(1))) {
9888 Scale = cast<ConstantInt>(Inst->getOperand(1));
9889 NewIdx = Inst->getOperand(0);
9893 // If the index will be to exactly the right offset with the scale taken
9894 // out, perform the transformation. Note, we don't know whether Scale is
9895 // signed or not. We'll use unsigned version of division/modulo
9896 // operation after making sure Scale doesn't have the sign bit set.
9897 if (Scale && Scale->getSExtValue() >= 0LL &&
9898 Scale->getZExtValue() % ArrayEltSize == 0) {
9899 Scale = ConstantInt::get(Scale->getType(),
9900 Scale->getZExtValue() / ArrayEltSize);
9901 if (Scale->getZExtValue() != 1) {
9902 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9904 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
9905 NewIdx = InsertNewInstBefore(Sc, GEP);
9908 // Insert the new GEP instruction.
9910 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9912 Instruction *NewGEP =
9913 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9914 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9915 // The NewGEP must be pointer typed, so must the old one -> BitCast
9916 return new BitCastInst(NewGEP, GEP.getType());
9925 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9926 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9927 if (AI.isArrayAllocation()) { // Check C != 1
9928 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9930 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9931 AllocationInst *New = 0;
9933 // Create and insert the replacement instruction...
9934 if (isa<MallocInst>(AI))
9935 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9937 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9938 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9941 InsertNewInstBefore(New, AI);
9943 // Scan to the end of the allocation instructions, to skip over a block of
9944 // allocas if possible...
9946 BasicBlock::iterator It = New;
9947 while (isa<AllocationInst>(*It)) ++It;
9949 // Now that I is pointing to the first non-allocation-inst in the block,
9950 // insert our getelementptr instruction...
9952 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9956 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9957 New->getName()+".sub", It);
9959 // Now make everything use the getelementptr instead of the original
9961 return ReplaceInstUsesWith(AI, V);
9962 } else if (isa<UndefValue>(AI.getArraySize())) {
9963 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9967 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9968 // Note that we only do this for alloca's, because malloc should allocate and
9969 // return a unique pointer, even for a zero byte allocation.
9970 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9971 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9972 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9977 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9978 Value *Op = FI.getOperand(0);
9980 // free undef -> unreachable.
9981 if (isa<UndefValue>(Op)) {
9982 // Insert a new store to null because we cannot modify the CFG here.
9983 new StoreInst(ConstantInt::getTrue(),
9984 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9985 return EraseInstFromFunction(FI);
9988 // If we have 'free null' delete the instruction. This can happen in stl code
9989 // when lots of inlining happens.
9990 if (isa<ConstantPointerNull>(Op))
9991 return EraseInstFromFunction(FI);
9993 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9994 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9995 FI.setOperand(0, CI->getOperand(0));
9999 // Change free (gep X, 0,0,0,0) into free(X)
10000 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10001 if (GEPI->hasAllZeroIndices()) {
10002 AddToWorkList(GEPI);
10003 FI.setOperand(0, GEPI->getOperand(0));
10008 // Change free(malloc) into nothing, if the malloc has a single use.
10009 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10010 if (MI->hasOneUse()) {
10011 EraseInstFromFunction(FI);
10012 return EraseInstFromFunction(*MI);
10019 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10020 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10021 const TargetData *TD) {
10022 User *CI = cast<User>(LI.getOperand(0));
10023 Value *CastOp = CI->getOperand(0);
10025 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10026 // Instead of loading constant c string, use corresponding integer value
10027 // directly if string length is small enough.
10029 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10030 unsigned len = Str.length();
10031 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10032 unsigned numBits = Ty->getPrimitiveSizeInBits();
10033 // Replace LI with immediate integer store.
10034 if ((numBits >> 3) == len + 1) {
10035 APInt StrVal(numBits, 0);
10036 APInt SingleChar(numBits, 0);
10037 if (TD->isLittleEndian()) {
10038 for (signed i = len-1; i >= 0; i--) {
10039 SingleChar = (uint64_t) Str[i];
10040 StrVal = (StrVal << 8) | SingleChar;
10043 for (unsigned i = 0; i < len; i++) {
10044 SingleChar = (uint64_t) Str[i];
10045 StrVal = (StrVal << 8) | SingleChar;
10047 // Append NULL at the end.
10049 StrVal = (StrVal << 8) | SingleChar;
10051 Value *NL = ConstantInt::get(StrVal);
10052 return IC.ReplaceInstUsesWith(LI, NL);
10057 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10058 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10059 const Type *SrcPTy = SrcTy->getElementType();
10061 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10062 isa<VectorType>(DestPTy)) {
10063 // If the source is an array, the code below will not succeed. Check to
10064 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10066 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10067 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10068 if (ASrcTy->getNumElements() != 0) {
10070 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10071 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10072 SrcTy = cast<PointerType>(CastOp->getType());
10073 SrcPTy = SrcTy->getElementType();
10076 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10077 isa<VectorType>(SrcPTy)) &&
10078 // Do not allow turning this into a load of an integer, which is then
10079 // casted to a pointer, this pessimizes pointer analysis a lot.
10080 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10081 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10082 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10084 // Okay, we are casting from one integer or pointer type to another of
10085 // the same size. Instead of casting the pointer before the load, cast
10086 // the result of the loaded value.
10087 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10089 LI.isVolatile()),LI);
10090 // Now cast the result of the load.
10091 return new BitCastInst(NewLoad, LI.getType());
10098 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10099 /// from this value cannot trap. If it is not obviously safe to load from the
10100 /// specified pointer, we do a quick local scan of the basic block containing
10101 /// ScanFrom, to determine if the address is already accessed.
10102 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10103 // If it is an alloca it is always safe to load from.
10104 if (isa<AllocaInst>(V)) return true;
10106 // If it is a global variable it is mostly safe to load from.
10107 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10108 // Don't try to evaluate aliases. External weak GV can be null.
10109 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10111 // Otherwise, be a little bit agressive by scanning the local block where we
10112 // want to check to see if the pointer is already being loaded or stored
10113 // from/to. If so, the previous load or store would have already trapped,
10114 // so there is no harm doing an extra load (also, CSE will later eliminate
10115 // the load entirely).
10116 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10121 // If we see a free or a call (which might do a free) the pointer could be
10123 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10126 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10127 if (LI->getOperand(0) == V) return true;
10128 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10129 if (SI->getOperand(1) == V) return true;
10136 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10137 /// until we find the underlying object a pointer is referring to or something
10138 /// we don't understand. Note that the returned pointer may be offset from the
10139 /// input, because we ignore GEP indices.
10140 static Value *GetUnderlyingObject(Value *Ptr) {
10142 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10143 if (CE->getOpcode() == Instruction::BitCast ||
10144 CE->getOpcode() == Instruction::GetElementPtr)
10145 Ptr = CE->getOperand(0);
10148 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10149 Ptr = BCI->getOperand(0);
10150 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10151 Ptr = GEP->getOperand(0);
10158 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10159 Value *Op = LI.getOperand(0);
10161 // Attempt to improve the alignment.
10162 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10164 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10165 LI.getAlignment()))
10166 LI.setAlignment(KnownAlign);
10168 // load (cast X) --> cast (load X) iff safe
10169 if (isa<CastInst>(Op))
10170 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10173 // None of the following transforms are legal for volatile loads.
10174 if (LI.isVolatile()) return 0;
10176 if (&LI.getParent()->front() != &LI) {
10177 BasicBlock::iterator BBI = &LI; --BBI;
10178 // If the instruction immediately before this is a store to the same
10179 // address, do a simple form of store->load forwarding.
10180 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10181 if (SI->getOperand(1) == LI.getOperand(0))
10182 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10183 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10184 if (LIB->getOperand(0) == LI.getOperand(0))
10185 return ReplaceInstUsesWith(LI, LIB);
10188 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10189 const Value *GEPI0 = GEPI->getOperand(0);
10190 // TODO: Consider a target hook for valid address spaces for this xform.
10191 if (isa<ConstantPointerNull>(GEPI0) &&
10192 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10193 // Insert a new store to null instruction before the load to indicate
10194 // that this code is not reachable. We do this instead of inserting
10195 // an unreachable instruction directly because we cannot modify the
10197 new StoreInst(UndefValue::get(LI.getType()),
10198 Constant::getNullValue(Op->getType()), &LI);
10199 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10203 if (Constant *C = dyn_cast<Constant>(Op)) {
10204 // load null/undef -> undef
10205 // TODO: Consider a target hook for valid address spaces for this xform.
10206 if (isa<UndefValue>(C) || (C->isNullValue() &&
10207 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10208 // Insert a new store to null instruction before the load to indicate that
10209 // this code is not reachable. We do this instead of inserting an
10210 // unreachable instruction directly because we cannot modify the CFG.
10211 new StoreInst(UndefValue::get(LI.getType()),
10212 Constant::getNullValue(Op->getType()), &LI);
10213 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10216 // Instcombine load (constant global) into the value loaded.
10217 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10218 if (GV->isConstant() && !GV->isDeclaration())
10219 return ReplaceInstUsesWith(LI, GV->getInitializer());
10221 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10222 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10223 if (CE->getOpcode() == Instruction::GetElementPtr) {
10224 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10225 if (GV->isConstant() && !GV->isDeclaration())
10227 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10228 return ReplaceInstUsesWith(LI, V);
10229 if (CE->getOperand(0)->isNullValue()) {
10230 // Insert a new store to null instruction before the load to indicate
10231 // that this code is not reachable. We do this instead of inserting
10232 // an unreachable instruction directly because we cannot modify the
10234 new StoreInst(UndefValue::get(LI.getType()),
10235 Constant::getNullValue(Op->getType()), &LI);
10236 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10239 } else if (CE->isCast()) {
10240 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10246 // If this load comes from anywhere in a constant global, and if the global
10247 // is all undef or zero, we know what it loads.
10248 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10249 if (GV->isConstant() && GV->hasInitializer()) {
10250 if (GV->getInitializer()->isNullValue())
10251 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10252 else if (isa<UndefValue>(GV->getInitializer()))
10253 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10257 if (Op->hasOneUse()) {
10258 // Change select and PHI nodes to select values instead of addresses: this
10259 // helps alias analysis out a lot, allows many others simplifications, and
10260 // exposes redundancy in the code.
10262 // Note that we cannot do the transformation unless we know that the
10263 // introduced loads cannot trap! Something like this is valid as long as
10264 // the condition is always false: load (select bool %C, int* null, int* %G),
10265 // but it would not be valid if we transformed it to load from null
10266 // unconditionally.
10268 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10269 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10270 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10271 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10272 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10273 SI->getOperand(1)->getName()+".val"), LI);
10274 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10275 SI->getOperand(2)->getName()+".val"), LI);
10276 return SelectInst::Create(SI->getCondition(), V1, V2);
10279 // load (select (cond, null, P)) -> load P
10280 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10281 if (C->isNullValue()) {
10282 LI.setOperand(0, SI->getOperand(2));
10286 // load (select (cond, P, null)) -> load P
10287 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10288 if (C->isNullValue()) {
10289 LI.setOperand(0, SI->getOperand(1));
10297 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10299 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10300 User *CI = cast<User>(SI.getOperand(1));
10301 Value *CastOp = CI->getOperand(0);
10303 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10304 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10305 const Type *SrcPTy = SrcTy->getElementType();
10307 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10308 // If the source is an array, the code below will not succeed. Check to
10309 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10311 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10312 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10313 if (ASrcTy->getNumElements() != 0) {
10315 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10316 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10317 SrcTy = cast<PointerType>(CastOp->getType());
10318 SrcPTy = SrcTy->getElementType();
10321 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10322 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10323 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10325 // Okay, we are casting from one integer or pointer type to another of
10326 // the same size. Instead of casting the pointer before
10327 // the store, cast the value to be stored.
10329 Value *SIOp0 = SI.getOperand(0);
10330 Instruction::CastOps opcode = Instruction::BitCast;
10331 const Type* CastSrcTy = SIOp0->getType();
10332 const Type* CastDstTy = SrcPTy;
10333 if (isa<PointerType>(CastDstTy)) {
10334 if (CastSrcTy->isInteger())
10335 opcode = Instruction::IntToPtr;
10336 } else if (isa<IntegerType>(CastDstTy)) {
10337 if (isa<PointerType>(SIOp0->getType()))
10338 opcode = Instruction::PtrToInt;
10340 if (Constant *C = dyn_cast<Constant>(SIOp0))
10341 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10343 NewCast = IC.InsertNewInstBefore(
10344 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10346 return new StoreInst(NewCast, CastOp);
10353 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10354 Value *Val = SI.getOperand(0);
10355 Value *Ptr = SI.getOperand(1);
10357 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10358 EraseInstFromFunction(SI);
10363 // If the RHS is an alloca with a single use, zapify the store, making the
10365 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10366 if (isa<AllocaInst>(Ptr)) {
10367 EraseInstFromFunction(SI);
10372 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10373 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10374 GEP->getOperand(0)->hasOneUse()) {
10375 EraseInstFromFunction(SI);
10381 // Attempt to improve the alignment.
10382 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10384 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10385 SI.getAlignment()))
10386 SI.setAlignment(KnownAlign);
10388 // Do really simple DSE, to catch cases where there are several consequtive
10389 // stores to the same location, separated by a few arithmetic operations. This
10390 // situation often occurs with bitfield accesses.
10391 BasicBlock::iterator BBI = &SI;
10392 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10396 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10397 // Prev store isn't volatile, and stores to the same location?
10398 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10401 EraseInstFromFunction(*PrevSI);
10407 // If this is a load, we have to stop. However, if the loaded value is from
10408 // the pointer we're loading and is producing the pointer we're storing,
10409 // then *this* store is dead (X = load P; store X -> P).
10410 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10411 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10412 EraseInstFromFunction(SI);
10416 // Otherwise, this is a load from some other location. Stores before it
10417 // may not be dead.
10421 // Don't skip over loads or things that can modify memory.
10422 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10427 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10429 // store X, null -> turns into 'unreachable' in SimplifyCFG
10430 if (isa<ConstantPointerNull>(Ptr)) {
10431 if (!isa<UndefValue>(Val)) {
10432 SI.setOperand(0, UndefValue::get(Val->getType()));
10433 if (Instruction *U = dyn_cast<Instruction>(Val))
10434 AddToWorkList(U); // Dropped a use.
10437 return 0; // Do not modify these!
10440 // store undef, Ptr -> noop
10441 if (isa<UndefValue>(Val)) {
10442 EraseInstFromFunction(SI);
10447 // If the pointer destination is a cast, see if we can fold the cast into the
10449 if (isa<CastInst>(Ptr))
10450 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10452 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10454 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10458 // If this store is the last instruction in the basic block, and if the block
10459 // ends with an unconditional branch, try to move it to the successor block.
10461 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10462 if (BI->isUnconditional())
10463 if (SimplifyStoreAtEndOfBlock(SI))
10464 return 0; // xform done!
10469 /// SimplifyStoreAtEndOfBlock - Turn things like:
10470 /// if () { *P = v1; } else { *P = v2 }
10471 /// into a phi node with a store in the successor.
10473 /// Simplify things like:
10474 /// *P = v1; if () { *P = v2; }
10475 /// into a phi node with a store in the successor.
10477 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10478 BasicBlock *StoreBB = SI.getParent();
10480 // Check to see if the successor block has exactly two incoming edges. If
10481 // so, see if the other predecessor contains a store to the same location.
10482 // if so, insert a PHI node (if needed) and move the stores down.
10483 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10485 // Determine whether Dest has exactly two predecessors and, if so, compute
10486 // the other predecessor.
10487 pred_iterator PI = pred_begin(DestBB);
10488 BasicBlock *OtherBB = 0;
10489 if (*PI != StoreBB)
10492 if (PI == pred_end(DestBB))
10495 if (*PI != StoreBB) {
10500 if (++PI != pred_end(DestBB))
10503 // Bail out if all the relevant blocks aren't distinct (this can happen,
10504 // for example, if SI is in an infinite loop)
10505 if (StoreBB == DestBB || OtherBB == DestBB)
10508 // Verify that the other block ends in a branch and is not otherwise empty.
10509 BasicBlock::iterator BBI = OtherBB->getTerminator();
10510 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10511 if (!OtherBr || BBI == OtherBB->begin())
10514 // If the other block ends in an unconditional branch, check for the 'if then
10515 // else' case. there is an instruction before the branch.
10516 StoreInst *OtherStore = 0;
10517 if (OtherBr->isUnconditional()) {
10518 // If this isn't a store, or isn't a store to the same location, bail out.
10520 OtherStore = dyn_cast<StoreInst>(BBI);
10521 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10524 // Otherwise, the other block ended with a conditional branch. If one of the
10525 // destinations is StoreBB, then we have the if/then case.
10526 if (OtherBr->getSuccessor(0) != StoreBB &&
10527 OtherBr->getSuccessor(1) != StoreBB)
10530 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10531 // if/then triangle. See if there is a store to the same ptr as SI that
10532 // lives in OtherBB.
10534 // Check to see if we find the matching store.
10535 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10536 if (OtherStore->getOperand(1) != SI.getOperand(1))
10540 // If we find something that may be using or overwriting the stored
10541 // value, or if we run out of instructions, we can't do the xform.
10542 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
10543 BBI == OtherBB->begin())
10547 // In order to eliminate the store in OtherBr, we have to
10548 // make sure nothing reads or overwrites the stored value in
10550 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10551 // FIXME: This should really be AA driven.
10552 if (I->mayReadFromMemory() || I->mayWriteToMemory())
10557 // Insert a PHI node now if we need it.
10558 Value *MergedVal = OtherStore->getOperand(0);
10559 if (MergedVal != SI.getOperand(0)) {
10560 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10561 PN->reserveOperandSpace(2);
10562 PN->addIncoming(SI.getOperand(0), SI.getParent());
10563 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10564 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10567 // Advance to a place where it is safe to insert the new store and
10569 BBI = DestBB->getFirstNonPHI();
10570 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10571 OtherStore->isVolatile()), *BBI);
10573 // Nuke the old stores.
10574 EraseInstFromFunction(SI);
10575 EraseInstFromFunction(*OtherStore);
10581 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10582 // Change br (not X), label True, label False to: br X, label False, True
10584 BasicBlock *TrueDest;
10585 BasicBlock *FalseDest;
10586 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10587 !isa<Constant>(X)) {
10588 // Swap Destinations and condition...
10589 BI.setCondition(X);
10590 BI.setSuccessor(0, FalseDest);
10591 BI.setSuccessor(1, TrueDest);
10595 // Cannonicalize fcmp_one -> fcmp_oeq
10596 FCmpInst::Predicate FPred; Value *Y;
10597 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10598 TrueDest, FalseDest)))
10599 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10600 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10601 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10602 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10603 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10604 NewSCC->takeName(I);
10605 // Swap Destinations and condition...
10606 BI.setCondition(NewSCC);
10607 BI.setSuccessor(0, FalseDest);
10608 BI.setSuccessor(1, TrueDest);
10609 RemoveFromWorkList(I);
10610 I->eraseFromParent();
10611 AddToWorkList(NewSCC);
10615 // Cannonicalize icmp_ne -> icmp_eq
10616 ICmpInst::Predicate IPred;
10617 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10618 TrueDest, FalseDest)))
10619 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10620 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10621 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10622 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10623 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10624 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10625 NewSCC->takeName(I);
10626 // Swap Destinations and condition...
10627 BI.setCondition(NewSCC);
10628 BI.setSuccessor(0, FalseDest);
10629 BI.setSuccessor(1, TrueDest);
10630 RemoveFromWorkList(I);
10631 I->eraseFromParent();;
10632 AddToWorkList(NewSCC);
10639 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10640 Value *Cond = SI.getCondition();
10641 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10642 if (I->getOpcode() == Instruction::Add)
10643 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10644 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10645 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10646 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10648 SI.setOperand(0, I->getOperand(0));
10656 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
10657 // See if we are trying to extract a known value. If so, use that instead.
10658 if (Value *Elt = FindInsertedValue(EV.getOperand(0), EV.idx_begin(),
10659 EV.idx_end(), &EV))
10660 return ReplaceInstUsesWith(EV, Elt);
10666 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10667 /// is to leave as a vector operation.
10668 static bool CheapToScalarize(Value *V, bool isConstant) {
10669 if (isa<ConstantAggregateZero>(V))
10671 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10672 if (isConstant) return true;
10673 // If all elts are the same, we can extract.
10674 Constant *Op0 = C->getOperand(0);
10675 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10676 if (C->getOperand(i) != Op0)
10680 Instruction *I = dyn_cast<Instruction>(V);
10681 if (!I) return false;
10683 // Insert element gets simplified to the inserted element or is deleted if
10684 // this is constant idx extract element and its a constant idx insertelt.
10685 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10686 isa<ConstantInt>(I->getOperand(2)))
10688 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10690 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10691 if (BO->hasOneUse() &&
10692 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10693 CheapToScalarize(BO->getOperand(1), isConstant)))
10695 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10696 if (CI->hasOneUse() &&
10697 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10698 CheapToScalarize(CI->getOperand(1), isConstant)))
10704 /// Read and decode a shufflevector mask.
10706 /// It turns undef elements into values that are larger than the number of
10707 /// elements in the input.
10708 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10709 unsigned NElts = SVI->getType()->getNumElements();
10710 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10711 return std::vector<unsigned>(NElts, 0);
10712 if (isa<UndefValue>(SVI->getOperand(2)))
10713 return std::vector<unsigned>(NElts, 2*NElts);
10715 std::vector<unsigned> Result;
10716 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10717 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
10718 if (isa<UndefValue>(*i))
10719 Result.push_back(NElts*2); // undef -> 8
10721 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
10725 /// FindScalarElement - Given a vector and an element number, see if the scalar
10726 /// value is already around as a register, for example if it were inserted then
10727 /// extracted from the vector.
10728 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10729 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10730 const VectorType *PTy = cast<VectorType>(V->getType());
10731 unsigned Width = PTy->getNumElements();
10732 if (EltNo >= Width) // Out of range access.
10733 return UndefValue::get(PTy->getElementType());
10735 if (isa<UndefValue>(V))
10736 return UndefValue::get(PTy->getElementType());
10737 else if (isa<ConstantAggregateZero>(V))
10738 return Constant::getNullValue(PTy->getElementType());
10739 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10740 return CP->getOperand(EltNo);
10741 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10742 // If this is an insert to a variable element, we don't know what it is.
10743 if (!isa<ConstantInt>(III->getOperand(2)))
10745 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10747 // If this is an insert to the element we are looking for, return the
10749 if (EltNo == IIElt)
10750 return III->getOperand(1);
10752 // Otherwise, the insertelement doesn't modify the value, recurse on its
10754 return FindScalarElement(III->getOperand(0), EltNo);
10755 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10756 unsigned InEl = getShuffleMask(SVI)[EltNo];
10758 return FindScalarElement(SVI->getOperand(0), InEl);
10759 else if (InEl < Width*2)
10760 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10762 return UndefValue::get(PTy->getElementType());
10765 // Otherwise, we don't know.
10769 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10770 // If vector val is undef, replace extract with scalar undef.
10771 if (isa<UndefValue>(EI.getOperand(0)))
10772 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10774 // If vector val is constant 0, replace extract with scalar 0.
10775 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10776 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10778 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10779 // If vector val is constant with all elements the same, replace EI with
10780 // that element. When the elements are not identical, we cannot replace yet
10781 // (we do that below, but only when the index is constant).
10782 Constant *op0 = C->getOperand(0);
10783 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10784 if (C->getOperand(i) != op0) {
10789 return ReplaceInstUsesWith(EI, op0);
10792 // If extracting a specified index from the vector, see if we can recursively
10793 // find a previously computed scalar that was inserted into the vector.
10794 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10795 unsigned IndexVal = IdxC->getZExtValue();
10796 unsigned VectorWidth =
10797 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10799 // If this is extracting an invalid index, turn this into undef, to avoid
10800 // crashing the code below.
10801 if (IndexVal >= VectorWidth)
10802 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10804 // This instruction only demands the single element from the input vector.
10805 // If the input vector has a single use, simplify it based on this use
10807 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10808 uint64_t UndefElts;
10809 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10812 EI.setOperand(0, V);
10817 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10818 return ReplaceInstUsesWith(EI, Elt);
10820 // If the this extractelement is directly using a bitcast from a vector of
10821 // the same number of elements, see if we can find the source element from
10822 // it. In this case, we will end up needing to bitcast the scalars.
10823 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10824 if (const VectorType *VT =
10825 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10826 if (VT->getNumElements() == VectorWidth)
10827 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10828 return new BitCastInst(Elt, EI.getType());
10832 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10833 if (I->hasOneUse()) {
10834 // Push extractelement into predecessor operation if legal and
10835 // profitable to do so
10836 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10837 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10838 if (CheapToScalarize(BO, isConstantElt)) {
10839 ExtractElementInst *newEI0 =
10840 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10841 EI.getName()+".lhs");
10842 ExtractElementInst *newEI1 =
10843 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10844 EI.getName()+".rhs");
10845 InsertNewInstBefore(newEI0, EI);
10846 InsertNewInstBefore(newEI1, EI);
10847 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
10849 } else if (isa<LoadInst>(I)) {
10851 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10852 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10853 PointerType::get(EI.getType(), AS),EI);
10854 GetElementPtrInst *GEP =
10855 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
10856 InsertNewInstBefore(GEP, EI);
10857 return new LoadInst(GEP);
10860 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10861 // Extracting the inserted element?
10862 if (IE->getOperand(2) == EI.getOperand(1))
10863 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10864 // If the inserted and extracted elements are constants, they must not
10865 // be the same value, extract from the pre-inserted value instead.
10866 if (isa<Constant>(IE->getOperand(2)) &&
10867 isa<Constant>(EI.getOperand(1))) {
10868 AddUsesToWorkList(EI);
10869 EI.setOperand(0, IE->getOperand(0));
10872 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10873 // If this is extracting an element from a shufflevector, figure out where
10874 // it came from and extract from the appropriate input element instead.
10875 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10876 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10878 if (SrcIdx < SVI->getType()->getNumElements())
10879 Src = SVI->getOperand(0);
10880 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10881 SrcIdx -= SVI->getType()->getNumElements();
10882 Src = SVI->getOperand(1);
10884 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10886 return new ExtractElementInst(Src, SrcIdx);
10893 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10894 /// elements from either LHS or RHS, return the shuffle mask and true.
10895 /// Otherwise, return false.
10896 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10897 std::vector<Constant*> &Mask) {
10898 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10899 "Invalid CollectSingleShuffleElements");
10900 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10902 if (isa<UndefValue>(V)) {
10903 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10905 } else if (V == LHS) {
10906 for (unsigned i = 0; i != NumElts; ++i)
10907 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10909 } else if (V == RHS) {
10910 for (unsigned i = 0; i != NumElts; ++i)
10911 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10913 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10914 // If this is an insert of an extract from some other vector, include it.
10915 Value *VecOp = IEI->getOperand(0);
10916 Value *ScalarOp = IEI->getOperand(1);
10917 Value *IdxOp = IEI->getOperand(2);
10919 if (!isa<ConstantInt>(IdxOp))
10921 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10923 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10924 // Okay, we can handle this if the vector we are insertinting into is
10925 // transitively ok.
10926 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10927 // If so, update the mask to reflect the inserted undef.
10928 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10931 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10932 if (isa<ConstantInt>(EI->getOperand(1)) &&
10933 EI->getOperand(0)->getType() == V->getType()) {
10934 unsigned ExtractedIdx =
10935 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10937 // This must be extracting from either LHS or RHS.
10938 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10939 // Okay, we can handle this if the vector we are insertinting into is
10940 // transitively ok.
10941 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10942 // If so, update the mask to reflect the inserted value.
10943 if (EI->getOperand(0) == LHS) {
10944 Mask[InsertedIdx & (NumElts-1)] =
10945 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10947 assert(EI->getOperand(0) == RHS);
10948 Mask[InsertedIdx & (NumElts-1)] =
10949 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10958 // TODO: Handle shufflevector here!
10963 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10964 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10965 /// that computes V and the LHS value of the shuffle.
10966 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10968 assert(isa<VectorType>(V->getType()) &&
10969 (RHS == 0 || V->getType() == RHS->getType()) &&
10970 "Invalid shuffle!");
10971 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10973 if (isa<UndefValue>(V)) {
10974 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10976 } else if (isa<ConstantAggregateZero>(V)) {
10977 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10979 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10980 // If this is an insert of an extract from some other vector, include it.
10981 Value *VecOp = IEI->getOperand(0);
10982 Value *ScalarOp = IEI->getOperand(1);
10983 Value *IdxOp = IEI->getOperand(2);
10985 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10986 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10987 EI->getOperand(0)->getType() == V->getType()) {
10988 unsigned ExtractedIdx =
10989 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10990 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10992 // Either the extracted from or inserted into vector must be RHSVec,
10993 // otherwise we'd end up with a shuffle of three inputs.
10994 if (EI->getOperand(0) == RHS || RHS == 0) {
10995 RHS = EI->getOperand(0);
10996 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10997 Mask[InsertedIdx & (NumElts-1)] =
10998 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11002 if (VecOp == RHS) {
11003 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11004 // Everything but the extracted element is replaced with the RHS.
11005 for (unsigned i = 0; i != NumElts; ++i) {
11006 if (i != InsertedIdx)
11007 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11012 // If this insertelement is a chain that comes from exactly these two
11013 // vectors, return the vector and the effective shuffle.
11014 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11015 return EI->getOperand(0);
11020 // TODO: Handle shufflevector here!
11022 // Otherwise, can't do anything fancy. Return an identity vector.
11023 for (unsigned i = 0; i != NumElts; ++i)
11024 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11028 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11029 Value *VecOp = IE.getOperand(0);
11030 Value *ScalarOp = IE.getOperand(1);
11031 Value *IdxOp = IE.getOperand(2);
11033 // Inserting an undef or into an undefined place, remove this.
11034 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11035 ReplaceInstUsesWith(IE, VecOp);
11037 // If the inserted element was extracted from some other vector, and if the
11038 // indexes are constant, try to turn this into a shufflevector operation.
11039 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11040 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11041 EI->getOperand(0)->getType() == IE.getType()) {
11042 unsigned NumVectorElts = IE.getType()->getNumElements();
11043 unsigned ExtractedIdx =
11044 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11045 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11047 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11048 return ReplaceInstUsesWith(IE, VecOp);
11050 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11051 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11053 // If we are extracting a value from a vector, then inserting it right
11054 // back into the same place, just use the input vector.
11055 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11056 return ReplaceInstUsesWith(IE, VecOp);
11058 // We could theoretically do this for ANY input. However, doing so could
11059 // turn chains of insertelement instructions into a chain of shufflevector
11060 // instructions, and right now we do not merge shufflevectors. As such,
11061 // only do this in a situation where it is clear that there is benefit.
11062 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11063 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11064 // the values of VecOp, except then one read from EIOp0.
11065 // Build a new shuffle mask.
11066 std::vector<Constant*> Mask;
11067 if (isa<UndefValue>(VecOp))
11068 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11070 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11071 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11074 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11075 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11076 ConstantVector::get(Mask));
11079 // If this insertelement isn't used by some other insertelement, turn it
11080 // (and any insertelements it points to), into one big shuffle.
11081 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11082 std::vector<Constant*> Mask;
11084 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11085 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11086 // We now have a shuffle of LHS, RHS, Mask.
11087 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11096 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11097 Value *LHS = SVI.getOperand(0);
11098 Value *RHS = SVI.getOperand(1);
11099 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11101 bool MadeChange = false;
11103 // Undefined shuffle mask -> undefined value.
11104 if (isa<UndefValue>(SVI.getOperand(2)))
11105 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11107 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11108 // the undef, change them to undefs.
11109 if (isa<UndefValue>(SVI.getOperand(1))) {
11110 // Scan to see if there are any references to the RHS. If so, replace them
11111 // with undef element refs and set MadeChange to true.
11112 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11113 if (Mask[i] >= e && Mask[i] != 2*e) {
11120 // Remap any references to RHS to use LHS.
11121 std::vector<Constant*> Elts;
11122 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11123 if (Mask[i] == 2*e)
11124 Elts.push_back(UndefValue::get(Type::Int32Ty));
11126 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11128 SVI.setOperand(2, ConstantVector::get(Elts));
11132 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11133 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11134 if (LHS == RHS || isa<UndefValue>(LHS)) {
11135 if (isa<UndefValue>(LHS) && LHS == RHS) {
11136 // shuffle(undef,undef,mask) -> undef.
11137 return ReplaceInstUsesWith(SVI, LHS);
11140 // Remap any references to RHS to use LHS.
11141 std::vector<Constant*> Elts;
11142 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11143 if (Mask[i] >= 2*e)
11144 Elts.push_back(UndefValue::get(Type::Int32Ty));
11146 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11147 (Mask[i] < e && isa<UndefValue>(LHS)))
11148 Mask[i] = 2*e; // Turn into undef.
11150 Mask[i] &= (e-1); // Force to LHS.
11151 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11154 SVI.setOperand(0, SVI.getOperand(1));
11155 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11156 SVI.setOperand(2, ConstantVector::get(Elts));
11157 LHS = SVI.getOperand(0);
11158 RHS = SVI.getOperand(1);
11162 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11163 bool isLHSID = true, isRHSID = true;
11165 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11166 if (Mask[i] >= e*2) continue; // Ignore undef values.
11167 // Is this an identity shuffle of the LHS value?
11168 isLHSID &= (Mask[i] == i);
11170 // Is this an identity shuffle of the RHS value?
11171 isRHSID &= (Mask[i]-e == i);
11174 // Eliminate identity shuffles.
11175 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11176 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11178 // If the LHS is a shufflevector itself, see if we can combine it with this
11179 // one without producing an unusual shuffle. Here we are really conservative:
11180 // we are absolutely afraid of producing a shuffle mask not in the input
11181 // program, because the code gen may not be smart enough to turn a merged
11182 // shuffle into two specific shuffles: it may produce worse code. As such,
11183 // we only merge two shuffles if the result is one of the two input shuffle
11184 // masks. In this case, merging the shuffles just removes one instruction,
11185 // which we know is safe. This is good for things like turning:
11186 // (splat(splat)) -> splat.
11187 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11188 if (isa<UndefValue>(RHS)) {
11189 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11191 std::vector<unsigned> NewMask;
11192 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11193 if (Mask[i] >= 2*e)
11194 NewMask.push_back(2*e);
11196 NewMask.push_back(LHSMask[Mask[i]]);
11198 // If the result mask is equal to the src shuffle or this shuffle mask, do
11199 // the replacement.
11200 if (NewMask == LHSMask || NewMask == Mask) {
11201 std::vector<Constant*> Elts;
11202 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11203 if (NewMask[i] >= e*2) {
11204 Elts.push_back(UndefValue::get(Type::Int32Ty));
11206 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11209 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11210 LHSSVI->getOperand(1),
11211 ConstantVector::get(Elts));
11216 return MadeChange ? &SVI : 0;
11222 /// TryToSinkInstruction - Try to move the specified instruction from its
11223 /// current block into the beginning of DestBlock, which can only happen if it's
11224 /// safe to move the instruction past all of the instructions between it and the
11225 /// end of its block.
11226 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11227 assert(I->hasOneUse() && "Invariants didn't hold!");
11229 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11230 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11233 // Do not sink alloca instructions out of the entry block.
11234 if (isa<AllocaInst>(I) && I->getParent() ==
11235 &DestBlock->getParent()->getEntryBlock())
11238 // We can only sink load instructions if there is nothing between the load and
11239 // the end of block that could change the value.
11240 if (I->mayReadFromMemory()) {
11241 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11243 if (Scan->mayWriteToMemory())
11247 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11249 I->moveBefore(InsertPos);
11255 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11256 /// all reachable code to the worklist.
11258 /// This has a couple of tricks to make the code faster and more powerful. In
11259 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11260 /// them to the worklist (this significantly speeds up instcombine on code where
11261 /// many instructions are dead or constant). Additionally, if we find a branch
11262 /// whose condition is a known constant, we only visit the reachable successors.
11264 static void AddReachableCodeToWorklist(BasicBlock *BB,
11265 SmallPtrSet<BasicBlock*, 64> &Visited,
11267 const TargetData *TD) {
11268 std::vector<BasicBlock*> Worklist;
11269 Worklist.push_back(BB);
11271 while (!Worklist.empty()) {
11272 BB = Worklist.back();
11273 Worklist.pop_back();
11275 // We have now visited this block! If we've already been here, ignore it.
11276 if (!Visited.insert(BB)) continue;
11278 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11279 Instruction *Inst = BBI++;
11281 // DCE instruction if trivially dead.
11282 if (isInstructionTriviallyDead(Inst)) {
11284 DOUT << "IC: DCE: " << *Inst;
11285 Inst->eraseFromParent();
11289 // ConstantProp instruction if trivially constant.
11290 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11291 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11292 Inst->replaceAllUsesWith(C);
11294 Inst->eraseFromParent();
11298 IC.AddToWorkList(Inst);
11301 // Recursively visit successors. If this is a branch or switch on a
11302 // constant, only visit the reachable successor.
11303 TerminatorInst *TI = BB->getTerminator();
11304 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11305 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11306 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11307 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11308 Worklist.push_back(ReachableBB);
11311 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11312 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11313 // See if this is an explicit destination.
11314 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11315 if (SI->getCaseValue(i) == Cond) {
11316 BasicBlock *ReachableBB = SI->getSuccessor(i);
11317 Worklist.push_back(ReachableBB);
11321 // Otherwise it is the default destination.
11322 Worklist.push_back(SI->getSuccessor(0));
11327 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11328 Worklist.push_back(TI->getSuccessor(i));
11332 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11333 bool Changed = false;
11334 TD = &getAnalysis<TargetData>();
11336 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11337 << F.getNameStr() << "\n");
11340 // Do a depth-first traversal of the function, populate the worklist with
11341 // the reachable instructions. Ignore blocks that are not reachable. Keep
11342 // track of which blocks we visit.
11343 SmallPtrSet<BasicBlock*, 64> Visited;
11344 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11346 // Do a quick scan over the function. If we find any blocks that are
11347 // unreachable, remove any instructions inside of them. This prevents
11348 // the instcombine code from having to deal with some bad special cases.
11349 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11350 if (!Visited.count(BB)) {
11351 Instruction *Term = BB->getTerminator();
11352 while (Term != BB->begin()) { // Remove instrs bottom-up
11353 BasicBlock::iterator I = Term; --I;
11355 DOUT << "IC: DCE: " << *I;
11358 if (!I->use_empty())
11359 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11360 I->eraseFromParent();
11365 while (!Worklist.empty()) {
11366 Instruction *I = RemoveOneFromWorkList();
11367 if (I == 0) continue; // skip null values.
11369 // Check to see if we can DCE the instruction.
11370 if (isInstructionTriviallyDead(I)) {
11371 // Add operands to the worklist.
11372 if (I->getNumOperands() < 4)
11373 AddUsesToWorkList(*I);
11376 DOUT << "IC: DCE: " << *I;
11378 I->eraseFromParent();
11379 RemoveFromWorkList(I);
11383 // Instruction isn't dead, see if we can constant propagate it.
11384 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11385 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11387 // Add operands to the worklist.
11388 AddUsesToWorkList(*I);
11389 ReplaceInstUsesWith(*I, C);
11392 I->eraseFromParent();
11393 RemoveFromWorkList(I);
11397 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
11398 // See if we can constant fold its operands.
11399 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
11400 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
11401 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
11407 // See if we can trivially sink this instruction to a successor basic block.
11408 // FIXME: Remove GetResultInst test when first class support for aggregates
11410 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11411 BasicBlock *BB = I->getParent();
11412 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11413 if (UserParent != BB) {
11414 bool UserIsSuccessor = false;
11415 // See if the user is one of our successors.
11416 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11417 if (*SI == UserParent) {
11418 UserIsSuccessor = true;
11422 // If the user is one of our immediate successors, and if that successor
11423 // only has us as a predecessors (we'd have to split the critical edge
11424 // otherwise), we can keep going.
11425 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11426 next(pred_begin(UserParent)) == pred_end(UserParent))
11427 // Okay, the CFG is simple enough, try to sink this instruction.
11428 Changed |= TryToSinkInstruction(I, UserParent);
11432 // Now that we have an instruction, try combining it to simplify it...
11436 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11437 if (Instruction *Result = visit(*I)) {
11439 // Should we replace the old instruction with a new one?
11441 DOUT << "IC: Old = " << *I
11442 << " New = " << *Result;
11444 // Everything uses the new instruction now.
11445 I->replaceAllUsesWith(Result);
11447 // Push the new instruction and any users onto the worklist.
11448 AddToWorkList(Result);
11449 AddUsersToWorkList(*Result);
11451 // Move the name to the new instruction first.
11452 Result->takeName(I);
11454 // Insert the new instruction into the basic block...
11455 BasicBlock *InstParent = I->getParent();
11456 BasicBlock::iterator InsertPos = I;
11458 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11459 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11462 InstParent->getInstList().insert(InsertPos, Result);
11464 // Make sure that we reprocess all operands now that we reduced their
11466 AddUsesToWorkList(*I);
11468 // Instructions can end up on the worklist more than once. Make sure
11469 // we do not process an instruction that has been deleted.
11470 RemoveFromWorkList(I);
11472 // Erase the old instruction.
11473 InstParent->getInstList().erase(I);
11476 DOUT << "IC: Mod = " << OrigI
11477 << " New = " << *I;
11480 // If the instruction was modified, it's possible that it is now dead.
11481 // if so, remove it.
11482 if (isInstructionTriviallyDead(I)) {
11483 // Make sure we process all operands now that we are reducing their
11485 AddUsesToWorkList(*I);
11487 // Instructions may end up in the worklist more than once. Erase all
11488 // occurrences of this instruction.
11489 RemoveFromWorkList(I);
11490 I->eraseFromParent();
11493 AddUsersToWorkList(*I);
11500 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11502 // Do an explicit clear, this shrinks the map if needed.
11503 WorklistMap.clear();
11508 bool InstCombiner::runOnFunction(Function &F) {
11509 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11511 bool EverMadeChange = false;
11513 // Iterate while there is work to do.
11514 unsigned Iteration = 0;
11515 while (DoOneIteration(F, Iteration++))
11516 EverMadeChange = true;
11517 return EverMadeChange;
11520 FunctionPass *llvm::createInstructionCombiningPass() {
11521 return new InstCombiner();