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 // isOneBitSet - Return true if there is exactly one bit set in the specified
2969 static bool isOneBitSet(const ConstantInt *CI) {
2970 return CI->getValue().isPowerOf2();
2973 // isHighOnes - Return true if the constant is of the form 1+0+.
2974 // This is the same as lowones(~X).
2975 static bool isHighOnes(const ConstantInt *CI) {
2976 return (~CI->getValue() + 1).isPowerOf2();
2979 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2980 /// are carefully arranged to allow folding of expressions such as:
2982 /// (A < B) | (A > B) --> (A != B)
2984 /// Note that this is only valid if the first and second predicates have the
2985 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2987 /// Three bits are used to represent the condition, as follows:
2992 /// <=> Value Definition
2993 /// 000 0 Always false
3000 /// 111 7 Always true
3002 static unsigned getICmpCode(const ICmpInst *ICI) {
3003 switch (ICI->getPredicate()) {
3005 case ICmpInst::ICMP_UGT: return 1; // 001
3006 case ICmpInst::ICMP_SGT: return 1; // 001
3007 case ICmpInst::ICMP_EQ: return 2; // 010
3008 case ICmpInst::ICMP_UGE: return 3; // 011
3009 case ICmpInst::ICMP_SGE: return 3; // 011
3010 case ICmpInst::ICMP_ULT: return 4; // 100
3011 case ICmpInst::ICMP_SLT: return 4; // 100
3012 case ICmpInst::ICMP_NE: return 5; // 101
3013 case ICmpInst::ICMP_ULE: return 6; // 110
3014 case ICmpInst::ICMP_SLE: return 6; // 110
3017 assert(0 && "Invalid ICmp predicate!");
3022 /// getICmpValue - This is the complement of getICmpCode, which turns an
3023 /// opcode and two operands into either a constant true or false, or a brand
3024 /// new ICmp instruction. The sign is passed in to determine which kind
3025 /// of predicate to use in new icmp instructions.
3026 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3028 default: assert(0 && "Illegal ICmp code!");
3029 case 0: return ConstantInt::getFalse();
3032 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3034 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3035 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3038 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3040 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3043 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3045 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3046 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3049 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3051 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3052 case 7: return ConstantInt::getTrue();
3056 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3057 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3058 (ICmpInst::isSignedPredicate(p1) &&
3059 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3060 (ICmpInst::isSignedPredicate(p2) &&
3061 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3065 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3066 struct FoldICmpLogical {
3069 ICmpInst::Predicate pred;
3070 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3071 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3072 pred(ICI->getPredicate()) {}
3073 bool shouldApply(Value *V) const {
3074 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3075 if (PredicatesFoldable(pred, ICI->getPredicate()))
3076 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3077 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3080 Instruction *apply(Instruction &Log) const {
3081 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3082 if (ICI->getOperand(0) != LHS) {
3083 assert(ICI->getOperand(1) == LHS);
3084 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3087 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3088 unsigned LHSCode = getICmpCode(ICI);
3089 unsigned RHSCode = getICmpCode(RHSICI);
3091 switch (Log.getOpcode()) {
3092 case Instruction::And: Code = LHSCode & RHSCode; break;
3093 case Instruction::Or: Code = LHSCode | RHSCode; break;
3094 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3095 default: assert(0 && "Illegal logical opcode!"); return 0;
3098 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3099 ICmpInst::isSignedPredicate(ICI->getPredicate());
3101 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3102 if (Instruction *I = dyn_cast<Instruction>(RV))
3104 // Otherwise, it's a constant boolean value...
3105 return IC.ReplaceInstUsesWith(Log, RV);
3108 } // end anonymous namespace
3110 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3111 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3112 // guaranteed to be a binary operator.
3113 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3115 ConstantInt *AndRHS,
3116 BinaryOperator &TheAnd) {
3117 Value *X = Op->getOperand(0);
3118 Constant *Together = 0;
3120 Together = And(AndRHS, OpRHS);
3122 switch (Op->getOpcode()) {
3123 case Instruction::Xor:
3124 if (Op->hasOneUse()) {
3125 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3126 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3127 InsertNewInstBefore(And, TheAnd);
3129 return BinaryOperator::CreateXor(And, Together);
3132 case Instruction::Or:
3133 if (Together == AndRHS) // (X | C) & C --> C
3134 return ReplaceInstUsesWith(TheAnd, AndRHS);
3136 if (Op->hasOneUse() && Together != OpRHS) {
3137 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3138 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3139 InsertNewInstBefore(Or, TheAnd);
3141 return BinaryOperator::CreateAnd(Or, AndRHS);
3144 case Instruction::Add:
3145 if (Op->hasOneUse()) {
3146 // Adding a one to a single bit bit-field should be turned into an XOR
3147 // of the bit. First thing to check is to see if this AND is with a
3148 // single bit constant.
3149 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3151 // If there is only one bit set...
3152 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3153 // Ok, at this point, we know that we are masking the result of the
3154 // ADD down to exactly one bit. If the constant we are adding has
3155 // no bits set below this bit, then we can eliminate the ADD.
3156 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3158 // Check to see if any bits below the one bit set in AndRHSV are set.
3159 if ((AddRHS & (AndRHSV-1)) == 0) {
3160 // If not, the only thing that can effect the output of the AND is
3161 // the bit specified by AndRHSV. If that bit is set, the effect of
3162 // the XOR is to toggle the bit. If it is clear, then the ADD has
3164 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3165 TheAnd.setOperand(0, X);
3168 // Pull the XOR out of the AND.
3169 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3170 InsertNewInstBefore(NewAnd, TheAnd);
3171 NewAnd->takeName(Op);
3172 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3179 case Instruction::Shl: {
3180 // We know that the AND will not produce any of the bits shifted in, so if
3181 // the anded constant includes them, clear them now!
3183 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3184 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3185 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3186 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3188 if (CI->getValue() == ShlMask) {
3189 // Masking out bits that the shift already masks
3190 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3191 } else if (CI != AndRHS) { // Reducing bits set in and.
3192 TheAnd.setOperand(1, CI);
3197 case Instruction::LShr:
3199 // We know that the AND will not produce any of the bits shifted in, so if
3200 // the anded constant includes them, clear them now! This only applies to
3201 // unsigned shifts, because a signed shr may bring in set bits!
3203 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3204 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3205 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3206 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3208 if (CI->getValue() == ShrMask) {
3209 // Masking out bits that the shift already masks.
3210 return ReplaceInstUsesWith(TheAnd, Op);
3211 } else if (CI != AndRHS) {
3212 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3217 case Instruction::AShr:
3219 // See if this is shifting in some sign extension, then masking it out
3221 if (Op->hasOneUse()) {
3222 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3223 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3224 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3225 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3226 if (C == AndRHS) { // Masking out bits shifted in.
3227 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3228 // Make the argument unsigned.
3229 Value *ShVal = Op->getOperand(0);
3230 ShVal = InsertNewInstBefore(
3231 BinaryOperator::CreateLShr(ShVal, OpRHS,
3232 Op->getName()), TheAnd);
3233 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3242 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3243 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3244 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3245 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3246 /// insert new instructions.
3247 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3248 bool isSigned, bool Inside,
3250 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3251 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3252 "Lo is not <= Hi in range emission code!");
3255 if (Lo == Hi) // Trivially false.
3256 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3258 // V >= Min && V < Hi --> V < Hi
3259 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3260 ICmpInst::Predicate pred = (isSigned ?
3261 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3262 return new ICmpInst(pred, V, Hi);
3265 // Emit V-Lo <u Hi-Lo
3266 Constant *NegLo = ConstantExpr::getNeg(Lo);
3267 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3268 InsertNewInstBefore(Add, IB);
3269 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3270 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3273 if (Lo == Hi) // Trivially true.
3274 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3276 // V < Min || V >= Hi -> V > Hi-1
3277 Hi = SubOne(cast<ConstantInt>(Hi));
3278 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3279 ICmpInst::Predicate pred = (isSigned ?
3280 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3281 return new ICmpInst(pred, V, Hi);
3284 // Emit V-Lo >u Hi-1-Lo
3285 // Note that Hi has already had one subtracted from it, above.
3286 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3287 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3288 InsertNewInstBefore(Add, IB);
3289 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3290 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3293 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3294 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3295 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3296 // not, since all 1s are not contiguous.
3297 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3298 const APInt& V = Val->getValue();
3299 uint32_t BitWidth = Val->getType()->getBitWidth();
3300 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3302 // look for the first zero bit after the run of ones
3303 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3304 // look for the first non-zero bit
3305 ME = V.getActiveBits();
3309 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3310 /// where isSub determines whether the operator is a sub. If we can fold one of
3311 /// the following xforms:
3313 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3314 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3315 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3317 /// return (A +/- B).
3319 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3320 ConstantInt *Mask, bool isSub,
3322 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3323 if (!LHSI || LHSI->getNumOperands() != 2 ||
3324 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3326 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3328 switch (LHSI->getOpcode()) {
3330 case Instruction::And:
3331 if (And(N, Mask) == Mask) {
3332 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3333 if ((Mask->getValue().countLeadingZeros() +
3334 Mask->getValue().countPopulation()) ==
3335 Mask->getValue().getBitWidth())
3338 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3339 // part, we don't need any explicit masks to take them out of A. If that
3340 // is all N is, ignore it.
3341 uint32_t MB = 0, ME = 0;
3342 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3343 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3344 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3345 if (MaskedValueIsZero(RHS, Mask))
3350 case Instruction::Or:
3351 case Instruction::Xor:
3352 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3353 if ((Mask->getValue().countLeadingZeros() +
3354 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3355 && And(N, Mask)->isZero())
3362 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3364 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3365 return InsertNewInstBefore(New, I);
3368 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3369 bool Changed = SimplifyCommutative(I);
3370 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3372 if (isa<UndefValue>(Op1)) // X & undef -> 0
3373 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3377 return ReplaceInstUsesWith(I, Op1);
3379 // See if we can simplify any instructions used by the instruction whose sole
3380 // purpose is to compute bits we don't care about.
3381 if (!isa<VectorType>(I.getType())) {
3382 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3383 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3384 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3385 KnownZero, KnownOne))
3388 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3389 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3390 return ReplaceInstUsesWith(I, I.getOperand(0));
3391 } else if (isa<ConstantAggregateZero>(Op1)) {
3392 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3396 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3397 const APInt& AndRHSMask = AndRHS->getValue();
3398 APInt NotAndRHS(~AndRHSMask);
3400 // Optimize a variety of ((val OP C1) & C2) combinations...
3401 if (isa<BinaryOperator>(Op0)) {
3402 Instruction *Op0I = cast<Instruction>(Op0);
3403 Value *Op0LHS = Op0I->getOperand(0);
3404 Value *Op0RHS = Op0I->getOperand(1);
3405 switch (Op0I->getOpcode()) {
3406 case Instruction::Xor:
3407 case Instruction::Or:
3408 // If the mask is only needed on one incoming arm, push it up.
3409 if (Op0I->hasOneUse()) {
3410 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3411 // Not masking anything out for the LHS, move to RHS.
3412 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3413 Op0RHS->getName()+".masked");
3414 InsertNewInstBefore(NewRHS, I);
3415 return BinaryOperator::Create(
3416 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3418 if (!isa<Constant>(Op0RHS) &&
3419 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3420 // Not masking anything out for the RHS, move to LHS.
3421 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3422 Op0LHS->getName()+".masked");
3423 InsertNewInstBefore(NewLHS, I);
3424 return BinaryOperator::Create(
3425 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3430 case Instruction::Add:
3431 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3432 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3433 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3434 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3435 return BinaryOperator::CreateAnd(V, AndRHS);
3436 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3437 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3440 case Instruction::Sub:
3441 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3442 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3443 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3444 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3445 return BinaryOperator::CreateAnd(V, AndRHS);
3447 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3448 // has 1's for all bits that the subtraction with A might affect.
3449 if (Op0I->hasOneUse()) {
3450 uint32_t BitWidth = AndRHSMask.getBitWidth();
3451 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3452 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3454 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3455 if (!(A && A->isZero()) && // avoid infinite recursion.
3456 MaskedValueIsZero(Op0LHS, Mask)) {
3457 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3458 InsertNewInstBefore(NewNeg, I);
3459 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3464 case Instruction::Shl:
3465 case Instruction::LShr:
3466 // (1 << x) & 1 --> zext(x == 0)
3467 // (1 >> x) & 1 --> zext(x == 0)
3468 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3469 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3470 Constant::getNullValue(I.getType()));
3471 InsertNewInstBefore(NewICmp, I);
3472 return new ZExtInst(NewICmp, I.getType());
3477 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3478 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3480 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3481 // If this is an integer truncation or change from signed-to-unsigned, and
3482 // if the source is an and/or with immediate, transform it. This
3483 // frequently occurs for bitfield accesses.
3484 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3485 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3486 CastOp->getNumOperands() == 2)
3487 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3488 if (CastOp->getOpcode() == Instruction::And) {
3489 // Change: and (cast (and X, C1) to T), C2
3490 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3491 // This will fold the two constants together, which may allow
3492 // other simplifications.
3493 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3494 CastOp->getOperand(0), I.getType(),
3495 CastOp->getName()+".shrunk");
3496 NewCast = InsertNewInstBefore(NewCast, I);
3497 // trunc_or_bitcast(C1)&C2
3498 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3499 C3 = ConstantExpr::getAnd(C3, AndRHS);
3500 return BinaryOperator::CreateAnd(NewCast, C3);
3501 } else if (CastOp->getOpcode() == Instruction::Or) {
3502 // Change: and (cast (or X, C1) to T), C2
3503 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3504 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3505 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3506 return ReplaceInstUsesWith(I, AndRHS);
3512 // Try to fold constant and into select arguments.
3513 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3514 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3516 if (isa<PHINode>(Op0))
3517 if (Instruction *NV = FoldOpIntoPhi(I))
3521 Value *Op0NotVal = dyn_castNotVal(Op0);
3522 Value *Op1NotVal = dyn_castNotVal(Op1);
3524 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3525 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3527 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3528 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3529 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3530 I.getName()+".demorgan");
3531 InsertNewInstBefore(Or, I);
3532 return BinaryOperator::CreateNot(Or);
3536 Value *A = 0, *B = 0, *C = 0, *D = 0;
3537 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3538 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3539 return ReplaceInstUsesWith(I, Op1);
3541 // (A|B) & ~(A&B) -> A^B
3542 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3543 if ((A == C && B == D) || (A == D && B == C))
3544 return BinaryOperator::CreateXor(A, B);
3548 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3549 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3550 return ReplaceInstUsesWith(I, Op0);
3552 // ~(A&B) & (A|B) -> A^B
3553 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3554 if ((A == C && B == D) || (A == D && B == C))
3555 return BinaryOperator::CreateXor(A, B);
3559 if (Op0->hasOneUse() &&
3560 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3561 if (A == Op1) { // (A^B)&A -> A&(A^B)
3562 I.swapOperands(); // Simplify below
3563 std::swap(Op0, Op1);
3564 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3565 cast<BinaryOperator>(Op0)->swapOperands();
3566 I.swapOperands(); // Simplify below
3567 std::swap(Op0, Op1);
3570 if (Op1->hasOneUse() &&
3571 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3572 if (B == Op0) { // B&(A^B) -> B&(B^A)
3573 cast<BinaryOperator>(Op1)->swapOperands();
3576 if (A == Op0) { // A&(A^B) -> A & ~B
3577 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3578 InsertNewInstBefore(NotB, I);
3579 return BinaryOperator::CreateAnd(A, NotB);
3585 { // (icmp ugt/ult A, C) & (icmp B, C) --> (icmp (A|B), C)
3586 // where C is a power of 2
3588 ConstantInt *C1, *C2;
3589 ICmpInst::Predicate LHSCC, RHSCC;
3590 if (match(&I, m_And(m_ICmp(LHSCC, m_Value(A), m_ConstantInt(C1)),
3591 m_ICmp(RHSCC, m_Value(B), m_ConstantInt(C2)))))
3592 if (C1 == C2 && LHSCC == RHSCC && C1->getValue().isPowerOf2() &&
3593 (LHSCC == ICmpInst::ICMP_ULT || LHSCC == ICmpInst::ICMP_UGT)) {
3594 Instruction *NewOr = BinaryOperator::CreateOr(A, B);
3595 InsertNewInstBefore(NewOr, I);
3596 return new ICmpInst(LHSCC, NewOr, C1);
3600 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3601 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3602 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3605 Value *LHSVal, *RHSVal;
3606 ConstantInt *LHSCst, *RHSCst;
3607 ICmpInst::Predicate LHSCC, RHSCC;
3608 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3609 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3610 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3611 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3612 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3613 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3614 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3615 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3617 // Don't try to fold ICMP_SLT + ICMP_ULT.
3618 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3619 ICmpInst::isSignedPredicate(LHSCC) ==
3620 ICmpInst::isSignedPredicate(RHSCC))) {
3621 // Ensure that the larger constant is on the RHS.
3622 ICmpInst::Predicate GT;
3623 if (ICmpInst::isSignedPredicate(LHSCC) ||
3624 (ICmpInst::isEquality(LHSCC) &&
3625 ICmpInst::isSignedPredicate(RHSCC)))
3626 GT = ICmpInst::ICMP_SGT;
3628 GT = ICmpInst::ICMP_UGT;
3630 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3631 ICmpInst *LHS = cast<ICmpInst>(Op0);
3632 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3633 std::swap(LHS, RHS);
3634 std::swap(LHSCst, RHSCst);
3635 std::swap(LHSCC, RHSCC);
3638 // At this point, we know we have have two icmp instructions
3639 // comparing a value against two constants and and'ing the result
3640 // together. Because of the above check, we know that we only have
3641 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3642 // (from the FoldICmpLogical check above), that the two constants
3643 // are not equal and that the larger constant is on the RHS
3644 assert(LHSCst != RHSCst && "Compares not folded above?");
3647 default: assert(0 && "Unknown integer condition code!");
3648 case ICmpInst::ICMP_EQ:
3650 default: assert(0 && "Unknown integer condition code!");
3651 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3652 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3653 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3654 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3655 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3656 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3657 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3658 return ReplaceInstUsesWith(I, LHS);
3660 case ICmpInst::ICMP_NE:
3662 default: assert(0 && "Unknown integer condition code!");
3663 case ICmpInst::ICMP_ULT:
3664 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3665 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3666 break; // (X != 13 & X u< 15) -> no change
3667 case ICmpInst::ICMP_SLT:
3668 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3669 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3670 break; // (X != 13 & X s< 15) -> no change
3671 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3672 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3673 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3674 return ReplaceInstUsesWith(I, RHS);
3675 case ICmpInst::ICMP_NE:
3676 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3677 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3678 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
3679 LHSVal->getName()+".off");
3680 InsertNewInstBefore(Add, I);
3681 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3682 ConstantInt::get(Add->getType(), 1));
3684 break; // (X != 13 & X != 15) -> no change
3687 case ICmpInst::ICMP_ULT:
3689 default: assert(0 && "Unknown integer condition code!");
3690 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3691 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3692 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3693 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3695 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3696 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3697 return ReplaceInstUsesWith(I, LHS);
3698 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3702 case ICmpInst::ICMP_SLT:
3704 default: assert(0 && "Unknown integer condition code!");
3705 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3706 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3707 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3708 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3710 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3711 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3712 return ReplaceInstUsesWith(I, LHS);
3713 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3717 case ICmpInst::ICMP_UGT:
3719 default: assert(0 && "Unknown integer condition code!");
3720 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3721 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3722 return ReplaceInstUsesWith(I, RHS);
3723 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3725 case ICmpInst::ICMP_NE:
3726 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3727 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3728 break; // (X u> 13 & X != 15) -> no change
3729 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3730 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3732 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3736 case ICmpInst::ICMP_SGT:
3738 default: assert(0 && "Unknown integer condition code!");
3739 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3740 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3741 return ReplaceInstUsesWith(I, RHS);
3742 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3744 case ICmpInst::ICMP_NE:
3745 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3746 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3747 break; // (X s> 13 & X != 15) -> no change
3748 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3749 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3751 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3759 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3760 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3761 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3762 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3763 const Type *SrcTy = Op0C->getOperand(0)->getType();
3764 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3765 // Only do this if the casts both really cause code to be generated.
3766 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3768 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3770 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3771 Op1C->getOperand(0),
3773 InsertNewInstBefore(NewOp, I);
3774 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3778 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3779 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3780 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3781 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3782 SI0->getOperand(1) == SI1->getOperand(1) &&
3783 (SI0->hasOneUse() || SI1->hasOneUse())) {
3784 Instruction *NewOp =
3785 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3787 SI0->getName()), I);
3788 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3789 SI1->getOperand(1));
3793 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3794 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3795 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3796 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3797 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3798 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3799 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3800 // If either of the constants are nans, then the whole thing returns
3802 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3803 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3804 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3805 RHS->getOperand(0));
3810 return Changed ? &I : 0;
3813 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3814 /// in the result. If it does, and if the specified byte hasn't been filled in
3815 /// yet, fill it in and return false.
3816 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3817 Instruction *I = dyn_cast<Instruction>(V);
3818 if (I == 0) return true;
3820 // If this is an or instruction, it is an inner node of the bswap.
3821 if (I->getOpcode() == Instruction::Or)
3822 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3823 CollectBSwapParts(I->getOperand(1), ByteValues);
3825 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3826 // If this is a shift by a constant int, and it is "24", then its operand
3827 // defines a byte. We only handle unsigned types here.
3828 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3829 // Not shifting the entire input by N-1 bytes?
3830 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3831 8*(ByteValues.size()-1))
3835 if (I->getOpcode() == Instruction::Shl) {
3836 // X << 24 defines the top byte with the lowest of the input bytes.
3837 DestNo = ByteValues.size()-1;
3839 // X >>u 24 defines the low byte with the highest of the input bytes.
3843 // If the destination byte value is already defined, the values are or'd
3844 // together, which isn't a bswap (unless it's an or of the same bits).
3845 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3847 ByteValues[DestNo] = I->getOperand(0);
3851 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3853 Value *Shift = 0, *ShiftLHS = 0;
3854 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3855 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3856 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3858 Instruction *SI = cast<Instruction>(Shift);
3860 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3861 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3862 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3865 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3867 if (AndAmt->getValue().getActiveBits() > 64)
3869 uint64_t AndAmtVal = AndAmt->getZExtValue();
3870 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3871 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3873 // Unknown mask for bswap.
3874 if (DestByte == ByteValues.size()) return true;
3876 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3878 if (SI->getOpcode() == Instruction::Shl)
3879 SrcByte = DestByte - ShiftBytes;
3881 SrcByte = DestByte + ShiftBytes;
3883 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3884 if (SrcByte != ByteValues.size()-DestByte-1)
3887 // If the destination byte value is already defined, the values are or'd
3888 // together, which isn't a bswap (unless it's an or of the same bits).
3889 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3891 ByteValues[DestByte] = SI->getOperand(0);
3895 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3896 /// If so, insert the new bswap intrinsic and return it.
3897 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3898 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3899 if (!ITy || ITy->getBitWidth() % 16)
3900 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3902 /// ByteValues - For each byte of the result, we keep track of which value
3903 /// defines each byte.
3904 SmallVector<Value*, 8> ByteValues;
3905 ByteValues.resize(ITy->getBitWidth()/8);
3907 // Try to find all the pieces corresponding to the bswap.
3908 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3909 CollectBSwapParts(I.getOperand(1), ByteValues))
3912 // Check to see if all of the bytes come from the same value.
3913 Value *V = ByteValues[0];
3914 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3916 // Check to make sure that all of the bytes come from the same value.
3917 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3918 if (ByteValues[i] != V)
3920 const Type *Tys[] = { ITy };
3921 Module *M = I.getParent()->getParent()->getParent();
3922 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3923 return CallInst::Create(F, V);
3927 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3928 bool Changed = SimplifyCommutative(I);
3929 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3931 if (isa<UndefValue>(Op1)) // X | undef -> -1
3932 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3936 return ReplaceInstUsesWith(I, Op0);
3938 // See if we can simplify any instructions used by the instruction whose sole
3939 // purpose is to compute bits we don't care about.
3940 if (!isa<VectorType>(I.getType())) {
3941 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3942 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3943 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3944 KnownZero, KnownOne))
3946 } else if (isa<ConstantAggregateZero>(Op1)) {
3947 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3948 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3949 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3950 return ReplaceInstUsesWith(I, I.getOperand(1));
3956 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3957 ConstantInt *C1 = 0; Value *X = 0;
3958 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3959 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3960 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3961 InsertNewInstBefore(Or, I);
3963 return BinaryOperator::CreateAnd(Or,
3964 ConstantInt::get(RHS->getValue() | C1->getValue()));
3967 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3968 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3969 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3970 InsertNewInstBefore(Or, I);
3972 return BinaryOperator::CreateXor(Or,
3973 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3976 // Try to fold constant and into select arguments.
3977 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3978 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3980 if (isa<PHINode>(Op0))
3981 if (Instruction *NV = FoldOpIntoPhi(I))
3985 Value *A = 0, *B = 0;
3986 ConstantInt *C1 = 0, *C2 = 0;
3988 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3989 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3990 return ReplaceInstUsesWith(I, Op1);
3991 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3992 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3993 return ReplaceInstUsesWith(I, Op0);
3995 // (A | B) | C and A | (B | C) -> bswap if possible.
3996 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3997 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3998 match(Op1, m_Or(m_Value(), m_Value())) ||
3999 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4000 match(Op1, m_Shift(m_Value(), m_Value())))) {
4001 if (Instruction *BSwap = MatchBSwap(I))
4005 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4006 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4007 MaskedValueIsZero(Op1, C1->getValue())) {
4008 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4009 InsertNewInstBefore(NOr, I);
4011 return BinaryOperator::CreateXor(NOr, C1);
4014 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4015 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4016 MaskedValueIsZero(Op0, C1->getValue())) {
4017 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4018 InsertNewInstBefore(NOr, I);
4020 return BinaryOperator::CreateXor(NOr, C1);
4024 Value *C = 0, *D = 0;
4025 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4026 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4027 Value *V1 = 0, *V2 = 0, *V3 = 0;
4028 C1 = dyn_cast<ConstantInt>(C);
4029 C2 = dyn_cast<ConstantInt>(D);
4030 if (C1 && C2) { // (A & C1)|(B & C2)
4031 // If we have: ((V + N) & C1) | (V & C2)
4032 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4033 // replace with V+N.
4034 if (C1->getValue() == ~C2->getValue()) {
4035 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4036 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4037 // Add commutes, try both ways.
4038 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4039 return ReplaceInstUsesWith(I, A);
4040 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4041 return ReplaceInstUsesWith(I, A);
4043 // Or commutes, try both ways.
4044 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4045 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4046 // Add commutes, try both ways.
4047 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4048 return ReplaceInstUsesWith(I, B);
4049 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4050 return ReplaceInstUsesWith(I, B);
4053 V1 = 0; V2 = 0; V3 = 0;
4056 // Check to see if we have any common things being and'ed. If so, find the
4057 // terms for V1 & (V2|V3).
4058 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4059 if (A == B) // (A & C)|(A & D) == A & (C|D)
4060 V1 = A, V2 = C, V3 = D;
4061 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4062 V1 = A, V2 = B, V3 = C;
4063 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4064 V1 = C, V2 = A, V3 = D;
4065 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4066 V1 = C, V2 = A, V3 = B;
4070 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4071 return BinaryOperator::CreateAnd(V1, Or);
4076 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4077 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4078 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4079 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4080 SI0->getOperand(1) == SI1->getOperand(1) &&
4081 (SI0->hasOneUse() || SI1->hasOneUse())) {
4082 Instruction *NewOp =
4083 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4085 SI0->getName()), I);
4086 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4087 SI1->getOperand(1));
4091 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4092 if (A == Op1) // ~A | A == -1
4093 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4097 // Note, A is still live here!
4098 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4100 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4102 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4103 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4104 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4105 I.getName()+".demorgan"), I);
4106 return BinaryOperator::CreateNot(And);
4110 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4111 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4112 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4115 Value *LHSVal, *RHSVal;
4116 ConstantInt *LHSCst, *RHSCst;
4117 ICmpInst::Predicate LHSCC, RHSCC;
4118 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4119 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4120 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4121 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4122 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4123 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4124 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4125 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4126 // We can't fold (ugt x, C) | (sgt x, C2).
4127 PredicatesFoldable(LHSCC, RHSCC)) {
4128 // Ensure that the larger constant is on the RHS.
4129 ICmpInst *LHS = cast<ICmpInst>(Op0);
4131 if (ICmpInst::isSignedPredicate(LHSCC))
4132 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4134 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4137 std::swap(LHS, RHS);
4138 std::swap(LHSCst, RHSCst);
4139 std::swap(LHSCC, RHSCC);
4142 // At this point, we know we have have two icmp instructions
4143 // comparing a value against two constants and or'ing the result
4144 // together. Because of the above check, we know that we only have
4145 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4146 // FoldICmpLogical check above), that the two constants are not
4148 assert(LHSCst != RHSCst && "Compares not folded above?");
4151 default: assert(0 && "Unknown integer condition code!");
4152 case ICmpInst::ICMP_EQ:
4154 default: assert(0 && "Unknown integer condition code!");
4155 case ICmpInst::ICMP_EQ:
4156 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4157 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4158 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4159 LHSVal->getName()+".off");
4160 InsertNewInstBefore(Add, I);
4161 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4162 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4164 break; // (X == 13 | X == 15) -> no change
4165 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4166 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4168 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4169 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4170 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4171 return ReplaceInstUsesWith(I, RHS);
4174 case ICmpInst::ICMP_NE:
4176 default: assert(0 && "Unknown integer condition code!");
4177 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4178 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4179 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4180 return ReplaceInstUsesWith(I, LHS);
4181 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4182 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4183 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4184 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4187 case ICmpInst::ICMP_ULT:
4189 default: assert(0 && "Unknown integer condition code!");
4190 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4192 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4193 // If RHSCst is [us]MAXINT, it is always false. Not handling
4194 // this can cause overflow.
4195 if (RHSCst->isMaxValue(false))
4196 return ReplaceInstUsesWith(I, LHS);
4197 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4199 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4201 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4202 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4203 return ReplaceInstUsesWith(I, RHS);
4204 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4208 case ICmpInst::ICMP_SLT:
4210 default: assert(0 && "Unknown integer condition code!");
4211 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4213 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4214 // If RHSCst is [us]MAXINT, it is always false. Not handling
4215 // this can cause overflow.
4216 if (RHSCst->isMaxValue(true))
4217 return ReplaceInstUsesWith(I, LHS);
4218 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4220 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4222 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4223 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4224 return ReplaceInstUsesWith(I, RHS);
4225 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4229 case ICmpInst::ICMP_UGT:
4231 default: assert(0 && "Unknown integer condition code!");
4232 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4233 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4234 return ReplaceInstUsesWith(I, LHS);
4235 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4237 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4238 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4239 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4240 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4244 case ICmpInst::ICMP_SGT:
4246 default: assert(0 && "Unknown integer condition code!");
4247 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4248 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4249 return ReplaceInstUsesWith(I, LHS);
4250 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4252 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4253 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4254 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4255 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4263 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4264 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4265 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4266 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4267 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4268 !isa<ICmpInst>(Op1C->getOperand(0))) {
4269 const Type *SrcTy = Op0C->getOperand(0)->getType();
4270 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4271 // Only do this if the casts both really cause code to be
4273 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4275 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4277 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4278 Op1C->getOperand(0),
4280 InsertNewInstBefore(NewOp, I);
4281 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4288 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4289 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4290 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4291 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4292 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4293 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4294 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4295 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4296 // If either of the constants are nans, then the whole thing returns
4298 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4299 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4301 // Otherwise, no need to compare the two constants, compare the
4303 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4304 RHS->getOperand(0));
4309 return Changed ? &I : 0;
4314 // XorSelf - Implements: X ^ X --> 0
4317 XorSelf(Value *rhs) : RHS(rhs) {}
4318 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4319 Instruction *apply(BinaryOperator &Xor) const {
4326 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4327 bool Changed = SimplifyCommutative(I);
4328 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4330 if (isa<UndefValue>(Op1)) {
4331 if (isa<UndefValue>(Op0))
4332 // Handle undef ^ undef -> 0 special case. This is a common
4334 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4335 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4338 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4339 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4340 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4341 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4344 // See if we can simplify any instructions used by the instruction whose sole
4345 // purpose is to compute bits we don't care about.
4346 if (!isa<VectorType>(I.getType())) {
4347 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4348 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4349 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4350 KnownZero, KnownOne))
4352 } else if (isa<ConstantAggregateZero>(Op1)) {
4353 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4356 // Is this a ~ operation?
4357 if (Value *NotOp = dyn_castNotVal(&I)) {
4358 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4359 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4360 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4361 if (Op0I->getOpcode() == Instruction::And ||
4362 Op0I->getOpcode() == Instruction::Or) {
4363 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4364 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4366 BinaryOperator::CreateNot(Op0I->getOperand(1),
4367 Op0I->getOperand(1)->getName()+".not");
4368 InsertNewInstBefore(NotY, I);
4369 if (Op0I->getOpcode() == Instruction::And)
4370 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4372 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4379 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4380 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4381 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4382 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4383 return new ICmpInst(ICI->getInversePredicate(),
4384 ICI->getOperand(0), ICI->getOperand(1));
4386 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4387 return new FCmpInst(FCI->getInversePredicate(),
4388 FCI->getOperand(0), FCI->getOperand(1));
4391 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4392 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4393 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4394 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4395 Instruction::CastOps Opcode = Op0C->getOpcode();
4396 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4397 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4398 Op0C->getDestTy())) {
4399 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4400 CI->getOpcode(), CI->getInversePredicate(),
4401 CI->getOperand(0), CI->getOperand(1)), I);
4402 NewCI->takeName(CI);
4403 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4410 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4411 // ~(c-X) == X-c-1 == X+(-c-1)
4412 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4413 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4414 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4415 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4416 ConstantInt::get(I.getType(), 1));
4417 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4420 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4421 if (Op0I->getOpcode() == Instruction::Add) {
4422 // ~(X-c) --> (-c-1)-X
4423 if (RHS->isAllOnesValue()) {
4424 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4425 return BinaryOperator::CreateSub(
4426 ConstantExpr::getSub(NegOp0CI,
4427 ConstantInt::get(I.getType(), 1)),
4428 Op0I->getOperand(0));
4429 } else if (RHS->getValue().isSignBit()) {
4430 // (X + C) ^ signbit -> (X + C + signbit)
4431 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4432 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4435 } else if (Op0I->getOpcode() == Instruction::Or) {
4436 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4437 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4438 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4439 // Anything in both C1 and C2 is known to be zero, remove it from
4441 Constant *CommonBits = And(Op0CI, RHS);
4442 NewRHS = ConstantExpr::getAnd(NewRHS,
4443 ConstantExpr::getNot(CommonBits));
4444 AddToWorkList(Op0I);
4445 I.setOperand(0, Op0I->getOperand(0));
4446 I.setOperand(1, NewRHS);
4453 // Try to fold constant and into select arguments.
4454 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4455 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4457 if (isa<PHINode>(Op0))
4458 if (Instruction *NV = FoldOpIntoPhi(I))
4462 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4464 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4466 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4468 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4471 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4474 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4475 if (A == Op0) { // B^(B|A) == (A|B)^B
4476 Op1I->swapOperands();
4478 std::swap(Op0, Op1);
4479 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4480 I.swapOperands(); // Simplified below.
4481 std::swap(Op0, Op1);
4483 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4484 if (Op0 == A) // A^(A^B) == B
4485 return ReplaceInstUsesWith(I, B);
4486 else if (Op0 == B) // A^(B^A) == B
4487 return ReplaceInstUsesWith(I, A);
4488 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4489 if (A == Op0) { // A^(A&B) -> A^(B&A)
4490 Op1I->swapOperands();
4493 if (B == Op0) { // A^(B&A) -> (B&A)^A
4494 I.swapOperands(); // Simplified below.
4495 std::swap(Op0, Op1);
4500 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4503 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4504 if (A == Op1) // (B|A)^B == (A|B)^B
4506 if (B == Op1) { // (A|B)^B == A & ~B
4508 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4509 return BinaryOperator::CreateAnd(A, NotB);
4511 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4512 if (Op1 == A) // (A^B)^A == B
4513 return ReplaceInstUsesWith(I, B);
4514 else if (Op1 == B) // (B^A)^A == B
4515 return ReplaceInstUsesWith(I, A);
4516 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4517 if (A == Op1) // (A&B)^A -> (B&A)^A
4519 if (B == Op1 && // (B&A)^A == ~B & A
4520 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4522 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4523 return BinaryOperator::CreateAnd(N, Op1);
4528 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4529 if (Op0I && Op1I && Op0I->isShift() &&
4530 Op0I->getOpcode() == Op1I->getOpcode() &&
4531 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4532 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4533 Instruction *NewOp =
4534 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4535 Op1I->getOperand(0),
4536 Op0I->getName()), I);
4537 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4538 Op1I->getOperand(1));
4542 Value *A, *B, *C, *D;
4543 // (A & B)^(A | B) -> A ^ B
4544 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4545 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4546 if ((A == C && B == D) || (A == D && B == C))
4547 return BinaryOperator::CreateXor(A, B);
4549 // (A | B)^(A & B) -> A ^ B
4550 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4551 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4552 if ((A == C && B == D) || (A == D && B == C))
4553 return BinaryOperator::CreateXor(A, B);
4557 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4558 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4559 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4560 // (X & Y)^(X & Y) -> (Y^Z) & X
4561 Value *X = 0, *Y = 0, *Z = 0;
4563 X = A, Y = B, Z = D;
4565 X = A, Y = B, Z = C;
4567 X = B, Y = A, Z = D;
4569 X = B, Y = A, Z = C;
4572 Instruction *NewOp =
4573 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4574 return BinaryOperator::CreateAnd(NewOp, X);
4579 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4580 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4581 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4584 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4585 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4586 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4587 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4588 const Type *SrcTy = Op0C->getOperand(0)->getType();
4589 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4590 // Only do this if the casts both really cause code to be generated.
4591 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4593 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4595 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4596 Op1C->getOperand(0),
4598 InsertNewInstBefore(NewOp, I);
4599 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4604 return Changed ? &I : 0;
4607 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4608 /// overflowed for this type.
4609 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4610 ConstantInt *In2, bool IsSigned = false) {
4611 Result = cast<ConstantInt>(Add(In1, In2));
4614 if (In2->getValue().isNegative())
4615 return Result->getValue().sgt(In1->getValue());
4617 return Result->getValue().slt(In1->getValue());
4619 return Result->getValue().ult(In1->getValue());
4622 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4623 /// code necessary to compute the offset from the base pointer (without adding
4624 /// in the base pointer). Return the result as a signed integer of intptr size.
4625 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4626 TargetData &TD = IC.getTargetData();
4627 gep_type_iterator GTI = gep_type_begin(GEP);
4628 const Type *IntPtrTy = TD.getIntPtrType();
4629 Value *Result = Constant::getNullValue(IntPtrTy);
4631 // Build a mask for high order bits.
4632 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4633 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4635 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
4638 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4639 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4640 if (OpC->isZero()) continue;
4642 // Handle a struct index, which adds its field offset to the pointer.
4643 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4644 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4646 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4647 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4649 Result = IC.InsertNewInstBefore(
4650 BinaryOperator::CreateAdd(Result,
4651 ConstantInt::get(IntPtrTy, Size),
4652 GEP->getName()+".offs"), I);
4656 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4657 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4658 Scale = ConstantExpr::getMul(OC, Scale);
4659 if (Constant *RC = dyn_cast<Constant>(Result))
4660 Result = ConstantExpr::getAdd(RC, Scale);
4662 // Emit an add instruction.
4663 Result = IC.InsertNewInstBefore(
4664 BinaryOperator::CreateAdd(Result, Scale,
4665 GEP->getName()+".offs"), I);
4669 // Convert to correct type.
4670 if (Op->getType() != IntPtrTy) {
4671 if (Constant *OpC = dyn_cast<Constant>(Op))
4672 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4674 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4675 Op->getName()+".c"), I);
4678 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4679 if (Constant *OpC = dyn_cast<Constant>(Op))
4680 Op = ConstantExpr::getMul(OpC, Scale);
4681 else // We'll let instcombine(mul) convert this to a shl if possible.
4682 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
4683 GEP->getName()+".idx"), I);
4686 // Emit an add instruction.
4687 if (isa<Constant>(Op) && isa<Constant>(Result))
4688 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4689 cast<Constant>(Result));
4691 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
4692 GEP->getName()+".offs"), I);
4698 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4699 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4700 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4701 /// complex, and scales are involved. The above expression would also be legal
4702 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4703 /// later form is less amenable to optimization though, and we are allowed to
4704 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4706 /// If we can't emit an optimized form for this expression, this returns null.
4708 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4710 TargetData &TD = IC.getTargetData();
4711 gep_type_iterator GTI = gep_type_begin(GEP);
4713 // Check to see if this gep only has a single variable index. If so, and if
4714 // any constant indices are a multiple of its scale, then we can compute this
4715 // in terms of the scale of the variable index. For example, if the GEP
4716 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4717 // because the expression will cross zero at the same point.
4718 unsigned i, e = GEP->getNumOperands();
4720 for (i = 1; i != e; ++i, ++GTI) {
4721 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4722 // Compute the aggregate offset of constant indices.
4723 if (CI->isZero()) continue;
4725 // Handle a struct index, which adds its field offset to the pointer.
4726 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4727 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4729 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4730 Offset += Size*CI->getSExtValue();
4733 // Found our variable index.
4738 // If there are no variable indices, we must have a constant offset, just
4739 // evaluate it the general way.
4740 if (i == e) return 0;
4742 Value *VariableIdx = GEP->getOperand(i);
4743 // Determine the scale factor of the variable element. For example, this is
4744 // 4 if the variable index is into an array of i32.
4745 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4747 // Verify that there are no other variable indices. If so, emit the hard way.
4748 for (++i, ++GTI; i != e; ++i, ++GTI) {
4749 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4752 // Compute the aggregate offset of constant indices.
4753 if (CI->isZero()) continue;
4755 // Handle a struct index, which adds its field offset to the pointer.
4756 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4757 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4759 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4760 Offset += Size*CI->getSExtValue();
4764 // Okay, we know we have a single variable index, which must be a
4765 // pointer/array/vector index. If there is no offset, life is simple, return
4767 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4769 // Cast to intptrty in case a truncation occurs. If an extension is needed,
4770 // we don't need to bother extending: the extension won't affect where the
4771 // computation crosses zero.
4772 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
4773 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
4774 VariableIdx->getNameStart(), &I);
4778 // Otherwise, there is an index. The computation we will do will be modulo
4779 // the pointer size, so get it.
4780 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4782 Offset &= PtrSizeMask;
4783 VariableScale &= PtrSizeMask;
4785 // To do this transformation, any constant index must be a multiple of the
4786 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
4787 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
4788 // multiple of the variable scale.
4789 int64_t NewOffs = Offset / (int64_t)VariableScale;
4790 if (Offset != NewOffs*(int64_t)VariableScale)
4793 // Okay, we can do this evaluation. Start by converting the index to intptr.
4794 const Type *IntPtrTy = TD.getIntPtrType();
4795 if (VariableIdx->getType() != IntPtrTy)
4796 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
4798 VariableIdx->getNameStart(), &I);
4799 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
4800 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
4804 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4805 /// else. At this point we know that the GEP is on the LHS of the comparison.
4806 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4807 ICmpInst::Predicate Cond,
4809 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4811 // Look through bitcasts.
4812 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
4813 RHS = BCI->getOperand(0);
4815 Value *PtrBase = GEPLHS->getOperand(0);
4816 if (PtrBase == RHS) {
4817 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4818 // This transformation (ignoring the base and scales) is valid because we
4819 // know pointers can't overflow. See if we can output an optimized form.
4820 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
4822 // If not, synthesize the offset the hard way.
4824 Offset = EmitGEPOffset(GEPLHS, I, *this);
4825 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4826 Constant::getNullValue(Offset->getType()));
4827 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4828 // If the base pointers are different, but the indices are the same, just
4829 // compare the base pointer.
4830 if (PtrBase != GEPRHS->getOperand(0)) {
4831 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4832 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4833 GEPRHS->getOperand(0)->getType();
4835 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4836 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4837 IndicesTheSame = false;
4841 // If all indices are the same, just compare the base pointers.
4843 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4844 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4846 // Otherwise, the base pointers are different and the indices are
4847 // different, bail out.
4851 // If one of the GEPs has all zero indices, recurse.
4852 bool AllZeros = true;
4853 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4854 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4855 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4860 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4861 ICmpInst::getSwappedPredicate(Cond), I);
4863 // If the other GEP has all zero indices, recurse.
4865 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4866 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4867 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4872 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4874 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4875 // If the GEPs only differ by one index, compare it.
4876 unsigned NumDifferences = 0; // Keep track of # differences.
4877 unsigned DiffOperand = 0; // The operand that differs.
4878 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4879 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4880 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4881 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4882 // Irreconcilable differences.
4886 if (NumDifferences++) break;
4891 if (NumDifferences == 0) // SAME GEP?
4892 return ReplaceInstUsesWith(I, // No comparison is needed here.
4893 ConstantInt::get(Type::Int1Ty,
4894 ICmpInst::isTrueWhenEqual(Cond)));
4896 else if (NumDifferences == 1) {
4897 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4898 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4899 // Make sure we do a signed comparison here.
4900 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4904 // Only lower this if the icmp is the only user of the GEP or if we expect
4905 // the result to fold to a constant!
4906 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4907 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4908 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4909 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4910 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4911 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4917 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
4919 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
4922 if (!isa<ConstantFP>(RHSC)) return 0;
4923 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4925 // Get the width of the mantissa. We don't want to hack on conversions that
4926 // might lose information from the integer, e.g. "i64 -> float"
4927 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4928 if (MantissaWidth == -1) return 0; // Unknown.
4930 // Check to see that the input is converted from an integer type that is small
4931 // enough that preserves all bits. TODO: check here for "known" sign bits.
4932 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4933 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
4935 // If this is a uitofp instruction, we need an extra bit to hold the sign.
4936 if (isa<UIToFPInst>(LHSI))
4939 // If the conversion would lose info, don't hack on this.
4940 if ((int)InputSize > MantissaWidth)
4943 // Otherwise, we can potentially simplify the comparison. We know that it
4944 // will always come through as an integer value and we know the constant is
4945 // not a NAN (it would have been previously simplified).
4946 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4948 ICmpInst::Predicate Pred;
4949 switch (I.getPredicate()) {
4950 default: assert(0 && "Unexpected predicate!");
4951 case FCmpInst::FCMP_UEQ:
4952 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
4953 case FCmpInst::FCMP_UGT:
4954 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
4955 case FCmpInst::FCMP_UGE:
4956 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
4957 case FCmpInst::FCMP_ULT:
4958 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
4959 case FCmpInst::FCMP_ULE:
4960 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
4961 case FCmpInst::FCMP_UNE:
4962 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
4963 case FCmpInst::FCMP_ORD:
4964 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4965 case FCmpInst::FCMP_UNO:
4966 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4969 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4971 // Now we know that the APFloat is a normal number, zero or inf.
4973 // See if the FP constant is too large for the integer. For example,
4974 // comparing an i8 to 300.0.
4975 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
4977 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4978 // and large values.
4979 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
4980 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4981 APFloat::rmNearestTiesToEven);
4982 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4983 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4984 Pred == ICmpInst::ICMP_SLE)
4985 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4986 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4989 // See if the RHS value is < SignedMin.
4990 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
4991 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4992 APFloat::rmNearestTiesToEven);
4993 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4994 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4995 Pred == ICmpInst::ICMP_SGE)
4996 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4997 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5000 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
5001 // it may still be fractional. See if it is fractional by casting the FP
5002 // value to the integer value and back, checking for equality. Don't do this
5003 // for zero, because -0.0 is not fractional.
5004 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5005 if (!RHS.isZero() &&
5006 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5007 // If we had a comparison against a fractional value, we have to adjust
5008 // the compare predicate and sometimes the value. RHSC is rounded towards
5009 // zero at this point.
5011 default: assert(0 && "Unexpected integer comparison!");
5012 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5013 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5014 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5015 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5016 case ICmpInst::ICMP_SLE:
5017 // (float)int <= 4.4 --> int <= 4
5018 // (float)int <= -4.4 --> int < -4
5019 if (RHS.isNegative())
5020 Pred = ICmpInst::ICMP_SLT;
5022 case ICmpInst::ICMP_SLT:
5023 // (float)int < -4.4 --> int < -4
5024 // (float)int < 4.4 --> int <= 4
5025 if (!RHS.isNegative())
5026 Pred = ICmpInst::ICMP_SLE;
5028 case ICmpInst::ICMP_SGT:
5029 // (float)int > 4.4 --> int > 4
5030 // (float)int > -4.4 --> int >= -4
5031 if (RHS.isNegative())
5032 Pred = ICmpInst::ICMP_SGE;
5034 case ICmpInst::ICMP_SGE:
5035 // (float)int >= -4.4 --> int >= -4
5036 // (float)int >= 4.4 --> int > 4
5037 if (!RHS.isNegative())
5038 Pred = ICmpInst::ICMP_SGT;
5043 // Lower this FP comparison into an appropriate integer version of the
5045 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5048 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5049 bool Changed = SimplifyCompare(I);
5050 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5052 // Fold trivial predicates.
5053 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5054 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5055 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5056 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5058 // Simplify 'fcmp pred X, X'
5060 switch (I.getPredicate()) {
5061 default: assert(0 && "Unknown predicate!");
5062 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5063 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5064 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5065 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5066 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5067 case FCmpInst::FCMP_OLT: // True if ordered and less than
5068 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5069 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5071 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5072 case FCmpInst::FCMP_ULT: // True if unordered or less than
5073 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5074 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5075 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5076 I.setPredicate(FCmpInst::FCMP_UNO);
5077 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5080 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5081 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5082 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5083 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5084 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5085 I.setPredicate(FCmpInst::FCMP_ORD);
5086 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5091 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5092 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5094 // Handle fcmp with constant RHS
5095 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5096 // If the constant is a nan, see if we can fold the comparison based on it.
5097 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5098 if (CFP->getValueAPF().isNaN()) {
5099 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5100 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5101 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5102 "Comparison must be either ordered or unordered!");
5103 // True if unordered.
5104 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5108 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5109 switch (LHSI->getOpcode()) {
5110 case Instruction::PHI:
5111 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5112 // block. If in the same block, we're encouraging jump threading. If
5113 // not, we are just pessimizing the code by making an i1 phi.
5114 if (LHSI->getParent() == I.getParent())
5115 if (Instruction *NV = FoldOpIntoPhi(I))
5118 case Instruction::SIToFP:
5119 case Instruction::UIToFP:
5120 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5123 case Instruction::Select:
5124 // If either operand of the select is a constant, we can fold the
5125 // comparison into the select arms, which will cause one to be
5126 // constant folded and the select turned into a bitwise or.
5127 Value *Op1 = 0, *Op2 = 0;
5128 if (LHSI->hasOneUse()) {
5129 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5130 // Fold the known value into the constant operand.
5131 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5132 // Insert a new FCmp of the other select operand.
5133 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5134 LHSI->getOperand(2), RHSC,
5136 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5137 // Fold the known value into the constant operand.
5138 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5139 // Insert a new FCmp of the other select operand.
5140 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5141 LHSI->getOperand(1), RHSC,
5147 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5152 return Changed ? &I : 0;
5155 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5156 bool Changed = SimplifyCompare(I);
5157 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5158 const Type *Ty = Op0->getType();
5162 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5163 I.isTrueWhenEqual()));
5165 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5166 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5168 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5169 // addresses never equal each other! We already know that Op0 != Op1.
5170 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5171 isa<ConstantPointerNull>(Op0)) &&
5172 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5173 isa<ConstantPointerNull>(Op1)))
5174 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5175 !I.isTrueWhenEqual()));
5177 // icmp's with boolean values can always be turned into bitwise operations
5178 if (Ty == Type::Int1Ty) {
5179 switch (I.getPredicate()) {
5180 default: assert(0 && "Invalid icmp instruction!");
5181 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5182 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5183 InsertNewInstBefore(Xor, I);
5184 return BinaryOperator::CreateNot(Xor);
5186 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5187 return BinaryOperator::CreateXor(Op0, Op1);
5189 case ICmpInst::ICMP_UGT:
5190 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5192 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5193 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5194 InsertNewInstBefore(Not, I);
5195 return BinaryOperator::CreateAnd(Not, Op1);
5197 case ICmpInst::ICMP_SGT:
5198 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5200 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5201 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5202 InsertNewInstBefore(Not, I);
5203 return BinaryOperator::CreateAnd(Not, Op0);
5205 case ICmpInst::ICMP_UGE:
5206 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5208 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5209 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5210 InsertNewInstBefore(Not, I);
5211 return BinaryOperator::CreateOr(Not, Op1);
5213 case ICmpInst::ICMP_SGE:
5214 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5216 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5217 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5218 InsertNewInstBefore(Not, I);
5219 return BinaryOperator::CreateOr(Not, Op0);
5224 // See if we are doing a comparison between a constant and an instruction that
5225 // can be folded into the comparison.
5226 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5229 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5230 if (I.isEquality() && CI->isNullValue() &&
5231 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5232 // (icmp cond A B) if cond is equality
5233 return new ICmpInst(I.getPredicate(), A, B);
5236 // If we have a icmp le or icmp ge instruction, turn it into the appropriate
5237 // icmp lt or icmp gt instruction. This allows us to rely on them being
5238 // folded in the code below.
5239 switch (I.getPredicate()) {
5241 case ICmpInst::ICMP_ULE:
5242 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5243 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5244 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5245 case ICmpInst::ICMP_SLE:
5246 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5247 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5248 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5249 case ICmpInst::ICMP_UGE:
5250 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5251 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5252 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5253 case ICmpInst::ICMP_SGE:
5254 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5255 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5256 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5259 // See if we can fold the comparison based on range information we can get
5260 // by checking whether bits are known to be zero or one in the input.
5261 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5262 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5264 // If this comparison is a normal comparison, it demands all
5265 // bits, if it is a sign bit comparison, it only demands the sign bit.
5267 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5269 if (SimplifyDemandedBits(Op0,
5270 isSignBit ? APInt::getSignBit(BitWidth)
5271 : APInt::getAllOnesValue(BitWidth),
5272 KnownZero, KnownOne, 0))
5275 // Given the known and unknown bits, compute a range that the LHS could be
5276 // in. Compute the Min, Max and RHS values based on the known bits. For the
5277 // EQ and NE we use unsigned values.
5278 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5279 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5280 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5282 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5284 // If Min and Max are known to be the same, then SimplifyDemandedBits
5285 // figured out that the LHS is a constant. Just constant fold this now so
5286 // that code below can assume that Min != Max.
5288 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5289 ConstantInt::get(Min),
5292 // Based on the range information we know about the LHS, see if we can
5293 // simplify this comparison. For example, (x&4) < 8 is always true.
5294 const APInt &RHSVal = CI->getValue();
5295 switch (I.getPredicate()) { // LE/GE have been folded already.
5296 default: assert(0 && "Unknown icmp opcode!");
5297 case ICmpInst::ICMP_EQ:
5298 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5299 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5301 case ICmpInst::ICMP_NE:
5302 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5303 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5305 case ICmpInst::ICMP_ULT:
5306 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5307 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5308 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5309 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5310 if (RHSVal == Max) // A <u MAX -> A != MAX
5311 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5312 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5313 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5315 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5316 if (CI->isMinValue(true))
5317 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5318 ConstantInt::getAllOnesValue(Op0->getType()));
5320 case ICmpInst::ICMP_UGT:
5321 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5322 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5323 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5324 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5326 if (RHSVal == Min) // A >u MIN -> A != MIN
5327 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5328 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5329 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5331 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5332 if (CI->isMaxValue(true))
5333 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5334 ConstantInt::getNullValue(Op0->getType()));
5336 case ICmpInst::ICMP_SLT:
5337 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5338 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5339 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5340 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5341 if (RHSVal == Max) // A <s MAX -> A != MAX
5342 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5343 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5344 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5346 case ICmpInst::ICMP_SGT:
5347 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5348 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5349 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5350 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5352 if (RHSVal == Min) // A >s MIN -> A != MIN
5353 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5354 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5355 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5359 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5360 // instruction, see if that instruction also has constants so that the
5361 // instruction can be folded into the icmp
5362 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5363 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5367 // Handle icmp with constant (but not simple integer constant) RHS
5368 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5369 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5370 switch (LHSI->getOpcode()) {
5371 case Instruction::GetElementPtr:
5372 if (RHSC->isNullValue()) {
5373 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5374 bool isAllZeros = true;
5375 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5376 if (!isa<Constant>(LHSI->getOperand(i)) ||
5377 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5382 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5383 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5387 case Instruction::PHI:
5388 // Only fold icmp into the PHI if the phi and fcmp are in the same
5389 // block. If in the same block, we're encouraging jump threading. If
5390 // not, we are just pessimizing the code by making an i1 phi.
5391 if (LHSI->getParent() == I.getParent())
5392 if (Instruction *NV = FoldOpIntoPhi(I))
5395 case Instruction::Select: {
5396 // If either operand of the select is a constant, we can fold the
5397 // comparison into the select arms, which will cause one to be
5398 // constant folded and the select turned into a bitwise or.
5399 Value *Op1 = 0, *Op2 = 0;
5400 if (LHSI->hasOneUse()) {
5401 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5402 // Fold the known value into the constant operand.
5403 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5404 // Insert a new ICmp of the other select operand.
5405 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5406 LHSI->getOperand(2), RHSC,
5408 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5409 // Fold the known value into the constant operand.
5410 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5411 // Insert a new ICmp of the other select operand.
5412 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5413 LHSI->getOperand(1), RHSC,
5419 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5422 case Instruction::Malloc:
5423 // If we have (malloc != null), and if the malloc has a single use, we
5424 // can assume it is successful and remove the malloc.
5425 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5426 AddToWorkList(LHSI);
5427 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5428 !I.isTrueWhenEqual()));
5434 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5435 if (User *GEP = dyn_castGetElementPtr(Op0))
5436 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5438 if (User *GEP = dyn_castGetElementPtr(Op1))
5439 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5440 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5443 // Test to see if the operands of the icmp are casted versions of other
5444 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5446 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5447 if (isa<PointerType>(Op0->getType()) &&
5448 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5449 // We keep moving the cast from the left operand over to the right
5450 // operand, where it can often be eliminated completely.
5451 Op0 = CI->getOperand(0);
5453 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5454 // so eliminate it as well.
5455 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5456 Op1 = CI2->getOperand(0);
5458 // If Op1 is a constant, we can fold the cast into the constant.
5459 if (Op0->getType() != Op1->getType()) {
5460 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5461 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5463 // Otherwise, cast the RHS right before the icmp
5464 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5467 return new ICmpInst(I.getPredicate(), Op0, Op1);
5471 if (isa<CastInst>(Op0)) {
5472 // Handle the special case of: icmp (cast bool to X), <cst>
5473 // This comes up when you have code like
5476 // For generality, we handle any zero-extension of any operand comparison
5477 // with a constant or another cast from the same type.
5478 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5479 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5483 // See if it's the same type of instruction on the left and right.
5484 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5485 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
5486 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
5487 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
5489 switch (Op0I->getOpcode()) {
5491 case Instruction::Add:
5492 case Instruction::Sub:
5493 case Instruction::Xor:
5494 // a+x icmp eq/ne b+x --> a icmp b
5495 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
5496 Op1I->getOperand(0));
5498 case Instruction::Mul:
5499 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5500 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
5501 if (!CI->isZero() && !CI->isOne()) {
5502 const APInt &AP = CI->getValue();
5503 ConstantInt *Mask = ConstantInt::get(
5504 APInt::getLowBitsSet(AP.getBitWidth(),
5506 AP.countTrailingZeros()));
5507 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
5509 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
5511 InsertNewInstBefore(And1, I);
5512 InsertNewInstBefore(And2, I);
5513 return new ICmpInst(I.getPredicate(), And1, And2);
5522 // ~x < ~y --> y < x
5524 if (match(Op0, m_Not(m_Value(A))) &&
5525 match(Op1, m_Not(m_Value(B))))
5526 return new ICmpInst(I.getPredicate(), B, A);
5529 if (I.isEquality()) {
5530 Value *A, *B, *C, *D;
5532 // -x == -y --> x == y
5533 if (match(Op0, m_Neg(m_Value(A))) &&
5534 match(Op1, m_Neg(m_Value(B))))
5535 return new ICmpInst(I.getPredicate(), A, B);
5537 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5538 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5539 Value *OtherVal = A == Op1 ? B : A;
5540 return new ICmpInst(I.getPredicate(), OtherVal,
5541 Constant::getNullValue(A->getType()));
5544 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5545 // A^c1 == C^c2 --> A == C^(c1^c2)
5546 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5547 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5548 if (Op1->hasOneUse()) {
5549 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5550 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5551 return new ICmpInst(I.getPredicate(), A,
5552 InsertNewInstBefore(Xor, I));
5555 // A^B == A^D -> B == D
5556 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5557 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5558 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5559 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5563 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5564 (A == Op0 || B == Op0)) {
5565 // A == (A^B) -> B == 0
5566 Value *OtherVal = A == Op0 ? B : A;
5567 return new ICmpInst(I.getPredicate(), OtherVal,
5568 Constant::getNullValue(A->getType()));
5570 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5571 // (A-B) == A -> B == 0
5572 return new ICmpInst(I.getPredicate(), B,
5573 Constant::getNullValue(B->getType()));
5575 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5576 // A == (A-B) -> B == 0
5577 return new ICmpInst(I.getPredicate(), B,
5578 Constant::getNullValue(B->getType()));
5581 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5582 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5583 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5584 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5585 Value *X = 0, *Y = 0, *Z = 0;
5588 X = B; Y = D; Z = A;
5589 } else if (A == D) {
5590 X = B; Y = C; Z = A;
5591 } else if (B == C) {
5592 X = A; Y = D; Z = B;
5593 } else if (B == D) {
5594 X = A; Y = C; Z = B;
5597 if (X) { // Build (X^Y) & Z
5598 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
5599 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
5600 I.setOperand(0, Op1);
5601 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5606 return Changed ? &I : 0;
5610 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5611 /// and CmpRHS are both known to be integer constants.
5612 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5613 ConstantInt *DivRHS) {
5614 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5615 const APInt &CmpRHSV = CmpRHS->getValue();
5617 // FIXME: If the operand types don't match the type of the divide
5618 // then don't attempt this transform. The code below doesn't have the
5619 // logic to deal with a signed divide and an unsigned compare (and
5620 // vice versa). This is because (x /s C1) <s C2 produces different
5621 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5622 // (x /u C1) <u C2. Simply casting the operands and result won't
5623 // work. :( The if statement below tests that condition and bails
5625 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5626 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5628 if (DivRHS->isZero())
5629 return 0; // The ProdOV computation fails on divide by zero.
5631 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5632 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5633 // C2 (CI). By solving for X we can turn this into a range check
5634 // instead of computing a divide.
5635 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5637 // Determine if the product overflows by seeing if the product is
5638 // not equal to the divide. Make sure we do the same kind of divide
5639 // as in the LHS instruction that we're folding.
5640 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5641 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5643 // Get the ICmp opcode
5644 ICmpInst::Predicate Pred = ICI.getPredicate();
5646 // Figure out the interval that is being checked. For example, a comparison
5647 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5648 // Compute this interval based on the constants involved and the signedness of
5649 // the compare/divide. This computes a half-open interval, keeping track of
5650 // whether either value in the interval overflows. After analysis each
5651 // overflow variable is set to 0 if it's corresponding bound variable is valid
5652 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5653 int LoOverflow = 0, HiOverflow = 0;
5654 ConstantInt *LoBound = 0, *HiBound = 0;
5657 if (!DivIsSigned) { // udiv
5658 // e.g. X/5 op 3 --> [15, 20)
5660 HiOverflow = LoOverflow = ProdOV;
5662 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5663 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5664 if (CmpRHSV == 0) { // (X / pos) op 0
5665 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5666 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5668 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5669 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5670 HiOverflow = LoOverflow = ProdOV;
5672 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5673 } else { // (X / pos) op neg
5674 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5675 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5676 LoOverflow = AddWithOverflow(LoBound, Prod,
5677 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5678 HiBound = AddOne(Prod);
5679 HiOverflow = ProdOV ? -1 : 0;
5681 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5682 if (CmpRHSV == 0) { // (X / neg) op 0
5683 // e.g. X/-5 op 0 --> [-4, 5)
5684 LoBound = AddOne(DivRHS);
5685 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5686 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5687 HiOverflow = 1; // [INTMIN+1, overflow)
5688 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5690 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5691 // e.g. X/-5 op 3 --> [-19, -14)
5692 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5694 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5695 HiBound = AddOne(Prod);
5696 } else { // (X / neg) op neg
5697 // e.g. X/-5 op -3 --> [15, 20)
5699 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5700 HiBound = Subtract(Prod, DivRHS);
5703 // Dividing by a negative swaps the condition. LT <-> GT
5704 Pred = ICmpInst::getSwappedPredicate(Pred);
5707 Value *X = DivI->getOperand(0);
5709 default: assert(0 && "Unhandled icmp opcode!");
5710 case ICmpInst::ICMP_EQ:
5711 if (LoOverflow && HiOverflow)
5712 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5713 else if (HiOverflow)
5714 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5715 ICmpInst::ICMP_UGE, X, LoBound);
5716 else if (LoOverflow)
5717 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5718 ICmpInst::ICMP_ULT, X, HiBound);
5720 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5721 case ICmpInst::ICMP_NE:
5722 if (LoOverflow && HiOverflow)
5723 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5724 else if (HiOverflow)
5725 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5726 ICmpInst::ICMP_ULT, X, LoBound);
5727 else if (LoOverflow)
5728 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5729 ICmpInst::ICMP_UGE, X, HiBound);
5731 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5732 case ICmpInst::ICMP_ULT:
5733 case ICmpInst::ICMP_SLT:
5734 if (LoOverflow == +1) // Low bound is greater than input range.
5735 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5736 if (LoOverflow == -1) // Low bound is less than input range.
5737 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5738 return new ICmpInst(Pred, X, LoBound);
5739 case ICmpInst::ICMP_UGT:
5740 case ICmpInst::ICMP_SGT:
5741 if (HiOverflow == +1) // High bound greater than input range.
5742 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5743 else if (HiOverflow == -1) // High bound less than input range.
5744 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5745 if (Pred == ICmpInst::ICMP_UGT)
5746 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5748 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5753 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5755 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5758 const APInt &RHSV = RHS->getValue();
5760 switch (LHSI->getOpcode()) {
5761 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5762 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5763 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5765 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5766 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5767 Value *CompareVal = LHSI->getOperand(0);
5769 // If the sign bit of the XorCST is not set, there is no change to
5770 // the operation, just stop using the Xor.
5771 if (!XorCST->getValue().isNegative()) {
5772 ICI.setOperand(0, CompareVal);
5773 AddToWorkList(LHSI);
5777 // Was the old condition true if the operand is positive?
5778 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5780 // If so, the new one isn't.
5781 isTrueIfPositive ^= true;
5783 if (isTrueIfPositive)
5784 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5786 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5790 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5791 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5792 LHSI->getOperand(0)->hasOneUse()) {
5793 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5795 // If the LHS is an AND of a truncating cast, we can widen the
5796 // and/compare to be the input width without changing the value
5797 // produced, eliminating a cast.
5798 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5799 // We can do this transformation if either the AND constant does not
5800 // have its sign bit set or if it is an equality comparison.
5801 // Extending a relational comparison when we're checking the sign
5802 // bit would not work.
5803 if (Cast->hasOneUse() &&
5804 (ICI.isEquality() ||
5805 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5807 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5808 APInt NewCST = AndCST->getValue();
5809 NewCST.zext(BitWidth);
5811 NewCI.zext(BitWidth);
5812 Instruction *NewAnd =
5813 BinaryOperator::CreateAnd(Cast->getOperand(0),
5814 ConstantInt::get(NewCST),LHSI->getName());
5815 InsertNewInstBefore(NewAnd, ICI);
5816 return new ICmpInst(ICI.getPredicate(), NewAnd,
5817 ConstantInt::get(NewCI));
5821 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5822 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5823 // happens a LOT in code produced by the C front-end, for bitfield
5825 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5826 if (Shift && !Shift->isShift())
5830 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5831 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5832 const Type *AndTy = AndCST->getType(); // Type of the and.
5834 // We can fold this as long as we can't shift unknown bits
5835 // into the mask. This can only happen with signed shift
5836 // rights, as they sign-extend.
5838 bool CanFold = Shift->isLogicalShift();
5840 // To test for the bad case of the signed shr, see if any
5841 // of the bits shifted in could be tested after the mask.
5842 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5843 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5845 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5846 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5847 AndCST->getValue()) == 0)
5853 if (Shift->getOpcode() == Instruction::Shl)
5854 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5856 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5858 // Check to see if we are shifting out any of the bits being
5860 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5861 // If we shifted bits out, the fold is not going to work out.
5862 // As a special case, check to see if this means that the
5863 // result is always true or false now.
5864 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5865 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5866 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5867 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5869 ICI.setOperand(1, NewCst);
5870 Constant *NewAndCST;
5871 if (Shift->getOpcode() == Instruction::Shl)
5872 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5874 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5875 LHSI->setOperand(1, NewAndCST);
5876 LHSI->setOperand(0, Shift->getOperand(0));
5877 AddToWorkList(Shift); // Shift is dead.
5878 AddUsesToWorkList(ICI);
5884 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5885 // preferable because it allows the C<<Y expression to be hoisted out
5886 // of a loop if Y is invariant and X is not.
5887 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5888 ICI.isEquality() && !Shift->isArithmeticShift() &&
5889 isa<Instruction>(Shift->getOperand(0))) {
5892 if (Shift->getOpcode() == Instruction::LShr) {
5893 NS = BinaryOperator::CreateShl(AndCST,
5894 Shift->getOperand(1), "tmp");
5896 // Insert a logical shift.
5897 NS = BinaryOperator::CreateLShr(AndCST,
5898 Shift->getOperand(1), "tmp");
5900 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5902 // Compute X & (C << Y).
5903 Instruction *NewAnd =
5904 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
5905 InsertNewInstBefore(NewAnd, ICI);
5907 ICI.setOperand(0, NewAnd);
5913 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5914 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5917 uint32_t TypeBits = RHSV.getBitWidth();
5919 // Check that the shift amount is in range. If not, don't perform
5920 // undefined shifts. When the shift is visited it will be
5922 if (ShAmt->uge(TypeBits))
5925 if (ICI.isEquality()) {
5926 // If we are comparing against bits always shifted out, the
5927 // comparison cannot succeed.
5929 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5930 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5931 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5932 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5933 return ReplaceInstUsesWith(ICI, Cst);
5936 if (LHSI->hasOneUse()) {
5937 // Otherwise strength reduce the shift into an and.
5938 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5940 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5943 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5944 Mask, LHSI->getName()+".mask");
5945 Value *And = InsertNewInstBefore(AndI, ICI);
5946 return new ICmpInst(ICI.getPredicate(), And,
5947 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5951 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5952 bool TrueIfSigned = false;
5953 if (LHSI->hasOneUse() &&
5954 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5955 // (X << 31) <s 0 --> (X&1) != 0
5956 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5957 (TypeBits-ShAmt->getZExtValue()-1));
5959 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5960 Mask, LHSI->getName()+".mask");
5961 Value *And = InsertNewInstBefore(AndI, ICI);
5963 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5964 And, Constant::getNullValue(And->getType()));
5969 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5970 case Instruction::AShr: {
5971 // Only handle equality comparisons of shift-by-constant.
5972 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5973 if (!ShAmt || !ICI.isEquality()) break;
5975 // Check that the shift amount is in range. If not, don't perform
5976 // undefined shifts. When the shift is visited it will be
5978 uint32_t TypeBits = RHSV.getBitWidth();
5979 if (ShAmt->uge(TypeBits))
5982 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5984 // If we are comparing against bits always shifted out, the
5985 // comparison cannot succeed.
5986 APInt Comp = RHSV << ShAmtVal;
5987 if (LHSI->getOpcode() == Instruction::LShr)
5988 Comp = Comp.lshr(ShAmtVal);
5990 Comp = Comp.ashr(ShAmtVal);
5992 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5993 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5994 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5995 return ReplaceInstUsesWith(ICI, Cst);
5998 // Otherwise, check to see if the bits shifted out are known to be zero.
5999 // If so, we can compare against the unshifted value:
6000 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6001 if (LHSI->hasOneUse() &&
6002 MaskedValueIsZero(LHSI->getOperand(0),
6003 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6004 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6005 ConstantExpr::getShl(RHS, ShAmt));
6008 if (LHSI->hasOneUse()) {
6009 // Otherwise strength reduce the shift into an and.
6010 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6011 Constant *Mask = ConstantInt::get(Val);
6014 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6015 Mask, LHSI->getName()+".mask");
6016 Value *And = InsertNewInstBefore(AndI, ICI);
6017 return new ICmpInst(ICI.getPredicate(), And,
6018 ConstantExpr::getShl(RHS, ShAmt));
6023 case Instruction::SDiv:
6024 case Instruction::UDiv:
6025 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6026 // Fold this div into the comparison, producing a range check.
6027 // Determine, based on the divide type, what the range is being
6028 // checked. If there is an overflow on the low or high side, remember
6029 // it, otherwise compute the range [low, hi) bounding the new value.
6030 // See: InsertRangeTest above for the kinds of replacements possible.
6031 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6032 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6037 case Instruction::Add:
6038 // Fold: icmp pred (add, X, C1), C2
6040 if (!ICI.isEquality()) {
6041 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6043 const APInt &LHSV = LHSC->getValue();
6045 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6048 if (ICI.isSignedPredicate()) {
6049 if (CR.getLower().isSignBit()) {
6050 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6051 ConstantInt::get(CR.getUpper()));
6052 } else if (CR.getUpper().isSignBit()) {
6053 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6054 ConstantInt::get(CR.getLower()));
6057 if (CR.getLower().isMinValue()) {
6058 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6059 ConstantInt::get(CR.getUpper()));
6060 } else if (CR.getUpper().isMinValue()) {
6061 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6062 ConstantInt::get(CR.getLower()));
6069 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6070 if (ICI.isEquality()) {
6071 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6073 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6074 // the second operand is a constant, simplify a bit.
6075 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6076 switch (BO->getOpcode()) {
6077 case Instruction::SRem:
6078 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6079 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6080 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6081 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6082 Instruction *NewRem =
6083 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6085 InsertNewInstBefore(NewRem, ICI);
6086 return new ICmpInst(ICI.getPredicate(), NewRem,
6087 Constant::getNullValue(BO->getType()));
6091 case Instruction::Add:
6092 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6093 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6094 if (BO->hasOneUse())
6095 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6096 Subtract(RHS, BOp1C));
6097 } else if (RHSV == 0) {
6098 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6099 // efficiently invertible, or if the add has just this one use.
6100 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6102 if (Value *NegVal = dyn_castNegVal(BOp1))
6103 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6104 else if (Value *NegVal = dyn_castNegVal(BOp0))
6105 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6106 else if (BO->hasOneUse()) {
6107 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6108 InsertNewInstBefore(Neg, ICI);
6110 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6114 case Instruction::Xor:
6115 // For the xor case, we can xor two constants together, eliminating
6116 // the explicit xor.
6117 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6118 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6119 ConstantExpr::getXor(RHS, BOC));
6122 case Instruction::Sub:
6123 // Replace (([sub|xor] A, B) != 0) with (A != B)
6125 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6129 case Instruction::Or:
6130 // If bits are being or'd in that are not present in the constant we
6131 // are comparing against, then the comparison could never succeed!
6132 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6133 Constant *NotCI = ConstantExpr::getNot(RHS);
6134 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6135 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6140 case Instruction::And:
6141 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6142 // If bits are being compared against that are and'd out, then the
6143 // comparison can never succeed!
6144 if ((RHSV & ~BOC->getValue()) != 0)
6145 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6148 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6149 if (RHS == BOC && RHSV.isPowerOf2())
6150 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6151 ICmpInst::ICMP_NE, LHSI,
6152 Constant::getNullValue(RHS->getType()));
6154 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6155 if (BOC->getValue().isSignBit()) {
6156 Value *X = BO->getOperand(0);
6157 Constant *Zero = Constant::getNullValue(X->getType());
6158 ICmpInst::Predicate pred = isICMP_NE ?
6159 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6160 return new ICmpInst(pred, X, Zero);
6163 // ((X & ~7) == 0) --> X < 8
6164 if (RHSV == 0 && isHighOnes(BOC)) {
6165 Value *X = BO->getOperand(0);
6166 Constant *NegX = ConstantExpr::getNeg(BOC);
6167 ICmpInst::Predicate pred = isICMP_NE ?
6168 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6169 return new ICmpInst(pred, X, NegX);
6174 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6175 // Handle icmp {eq|ne} <intrinsic>, intcst.
6176 if (II->getIntrinsicID() == Intrinsic::bswap) {
6178 ICI.setOperand(0, II->getOperand(1));
6179 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6183 } else { // Not a ICMP_EQ/ICMP_NE
6184 // If the LHS is a cast from an integral value of the same size,
6185 // then since we know the RHS is a constant, try to simlify.
6186 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6187 Value *CastOp = Cast->getOperand(0);
6188 const Type *SrcTy = CastOp->getType();
6189 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6190 if (SrcTy->isInteger() &&
6191 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6192 // If this is an unsigned comparison, try to make the comparison use
6193 // smaller constant values.
6194 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6195 // X u< 128 => X s> -1
6196 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6197 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6198 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6199 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6200 // X u> 127 => X s< 0
6201 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6202 Constant::getNullValue(SrcTy));
6210 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6211 /// We only handle extending casts so far.
6213 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6214 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6215 Value *LHSCIOp = LHSCI->getOperand(0);
6216 const Type *SrcTy = LHSCIOp->getType();
6217 const Type *DestTy = LHSCI->getType();
6220 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6221 // integer type is the same size as the pointer type.
6222 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6223 getTargetData().getPointerSizeInBits() ==
6224 cast<IntegerType>(DestTy)->getBitWidth()) {
6226 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6227 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6228 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6229 RHSOp = RHSC->getOperand(0);
6230 // If the pointer types don't match, insert a bitcast.
6231 if (LHSCIOp->getType() != RHSOp->getType())
6232 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6236 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6239 // The code below only handles extension cast instructions, so far.
6241 if (LHSCI->getOpcode() != Instruction::ZExt &&
6242 LHSCI->getOpcode() != Instruction::SExt)
6245 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6246 bool isSignedCmp = ICI.isSignedPredicate();
6248 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6249 // Not an extension from the same type?
6250 RHSCIOp = CI->getOperand(0);
6251 if (RHSCIOp->getType() != LHSCIOp->getType())
6254 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6255 // and the other is a zext), then we can't handle this.
6256 if (CI->getOpcode() != LHSCI->getOpcode())
6259 // Deal with equality cases early.
6260 if (ICI.isEquality())
6261 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6263 // A signed comparison of sign extended values simplifies into a
6264 // signed comparison.
6265 if (isSignedCmp && isSignedExt)
6266 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6268 // The other three cases all fold into an unsigned comparison.
6269 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6272 // If we aren't dealing with a constant on the RHS, exit early
6273 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6277 // Compute the constant that would happen if we truncated to SrcTy then
6278 // reextended to DestTy.
6279 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6280 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6282 // If the re-extended constant didn't change...
6284 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6285 // For example, we might have:
6286 // %A = sext short %X to uint
6287 // %B = icmp ugt uint %A, 1330
6288 // It is incorrect to transform this into
6289 // %B = icmp ugt short %X, 1330
6290 // because %A may have negative value.
6292 // However, we allow this when the compare is EQ/NE, because they are
6294 if (isSignedExt == isSignedCmp || ICI.isEquality())
6295 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6299 // The re-extended constant changed so the constant cannot be represented
6300 // in the shorter type. Consequently, we cannot emit a simple comparison.
6302 // First, handle some easy cases. We know the result cannot be equal at this
6303 // point so handle the ICI.isEquality() cases
6304 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6305 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6306 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6307 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6309 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6310 // should have been folded away previously and not enter in here.
6313 // We're performing a signed comparison.
6314 if (cast<ConstantInt>(CI)->getValue().isNegative())
6315 Result = ConstantInt::getFalse(); // X < (small) --> false
6317 Result = ConstantInt::getTrue(); // X < (large) --> true
6319 // We're performing an unsigned comparison.
6321 // We're performing an unsigned comp with a sign extended value.
6322 // This is true if the input is >= 0. [aka >s -1]
6323 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6324 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6325 NegOne, ICI.getName()), ICI);
6327 // Unsigned extend & unsigned compare -> always true.
6328 Result = ConstantInt::getTrue();
6332 // Finally, return the value computed.
6333 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6334 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6335 return ReplaceInstUsesWith(ICI, Result);
6337 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6338 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6339 "ICmp should be folded!");
6340 if (Constant *CI = dyn_cast<Constant>(Result))
6341 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6342 return BinaryOperator::CreateNot(Result);
6345 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6346 return commonShiftTransforms(I);
6349 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6350 return commonShiftTransforms(I);
6353 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6354 if (Instruction *R = commonShiftTransforms(I))
6357 Value *Op0 = I.getOperand(0);
6359 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6360 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6361 if (CSI->isAllOnesValue())
6362 return ReplaceInstUsesWith(I, CSI);
6364 // See if we can turn a signed shr into an unsigned shr.
6365 if (MaskedValueIsZero(Op0,
6366 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6367 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6372 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6373 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6374 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6376 // shl X, 0 == X and shr X, 0 == X
6377 // shl 0, X == 0 and shr 0, X == 0
6378 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6379 Op0 == Constant::getNullValue(Op0->getType()))
6380 return ReplaceInstUsesWith(I, Op0);
6382 if (isa<UndefValue>(Op0)) {
6383 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6384 return ReplaceInstUsesWith(I, Op0);
6385 else // undef << X -> 0, undef >>u X -> 0
6386 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6388 if (isa<UndefValue>(Op1)) {
6389 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6390 return ReplaceInstUsesWith(I, Op0);
6391 else // X << undef, X >>u undef -> 0
6392 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6395 // Try to fold constant and into select arguments.
6396 if (isa<Constant>(Op0))
6397 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6398 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6401 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6402 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6407 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6408 BinaryOperator &I) {
6409 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6411 // See if we can simplify any instructions used by the instruction whose sole
6412 // purpose is to compute bits we don't care about.
6413 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6414 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6415 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6416 KnownZero, KnownOne))
6419 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6420 // of a signed value.
6422 if (Op1->uge(TypeBits)) {
6423 if (I.getOpcode() != Instruction::AShr)
6424 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6426 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6431 // ((X*C1) << C2) == (X * (C1 << C2))
6432 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6433 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6434 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6435 return BinaryOperator::CreateMul(BO->getOperand(0),
6436 ConstantExpr::getShl(BOOp, Op1));
6438 // Try to fold constant and into select arguments.
6439 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6440 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6442 if (isa<PHINode>(Op0))
6443 if (Instruction *NV = FoldOpIntoPhi(I))
6446 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6447 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6448 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6449 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6450 // place. Don't try to do this transformation in this case. Also, we
6451 // require that the input operand is a shift-by-constant so that we have
6452 // confidence that the shifts will get folded together. We could do this
6453 // xform in more cases, but it is unlikely to be profitable.
6454 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6455 isa<ConstantInt>(TrOp->getOperand(1))) {
6456 // Okay, we'll do this xform. Make the shift of shift.
6457 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6458 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6460 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6462 // For logical shifts, the truncation has the effect of making the high
6463 // part of the register be zeros. Emulate this by inserting an AND to
6464 // clear the top bits as needed. This 'and' will usually be zapped by
6465 // other xforms later if dead.
6466 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6467 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6468 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6470 // The mask we constructed says what the trunc would do if occurring
6471 // between the shifts. We want to know the effect *after* the second
6472 // shift. We know that it is a logical shift by a constant, so adjust the
6473 // mask as appropriate.
6474 if (I.getOpcode() == Instruction::Shl)
6475 MaskV <<= Op1->getZExtValue();
6477 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6478 MaskV = MaskV.lshr(Op1->getZExtValue());
6481 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6483 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6485 // Return the value truncated to the interesting size.
6486 return new TruncInst(And, I.getType());
6490 if (Op0->hasOneUse()) {
6491 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6492 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6495 switch (Op0BO->getOpcode()) {
6497 case Instruction::Add:
6498 case Instruction::And:
6499 case Instruction::Or:
6500 case Instruction::Xor: {
6501 // These operators commute.
6502 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6503 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6504 match(Op0BO->getOperand(1),
6505 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6506 Instruction *YS = BinaryOperator::CreateShl(
6507 Op0BO->getOperand(0), Op1,
6509 InsertNewInstBefore(YS, I); // (Y << C)
6511 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6512 Op0BO->getOperand(1)->getName());
6513 InsertNewInstBefore(X, I); // (X + (Y << C))
6514 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6515 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6516 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6519 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6520 Value *Op0BOOp1 = Op0BO->getOperand(1);
6521 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6523 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6524 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6526 Instruction *YS = BinaryOperator::CreateShl(
6527 Op0BO->getOperand(0), Op1,
6529 InsertNewInstBefore(YS, I); // (Y << C)
6531 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6532 V1->getName()+".mask");
6533 InsertNewInstBefore(XM, I); // X & (CC << C)
6535 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6540 case Instruction::Sub: {
6541 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6542 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6543 match(Op0BO->getOperand(0),
6544 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6545 Instruction *YS = BinaryOperator::CreateShl(
6546 Op0BO->getOperand(1), Op1,
6548 InsertNewInstBefore(YS, I); // (Y << C)
6550 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6551 Op0BO->getOperand(0)->getName());
6552 InsertNewInstBefore(X, I); // (X + (Y << C))
6553 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6554 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6555 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6558 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6559 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6560 match(Op0BO->getOperand(0),
6561 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6562 m_ConstantInt(CC))) && V2 == Op1 &&
6563 cast<BinaryOperator>(Op0BO->getOperand(0))
6564 ->getOperand(0)->hasOneUse()) {
6565 Instruction *YS = BinaryOperator::CreateShl(
6566 Op0BO->getOperand(1), Op1,
6568 InsertNewInstBefore(YS, I); // (Y << C)
6570 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6571 V1->getName()+".mask");
6572 InsertNewInstBefore(XM, I); // X & (CC << C)
6574 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
6582 // If the operand is an bitwise operator with a constant RHS, and the
6583 // shift is the only use, we can pull it out of the shift.
6584 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6585 bool isValid = true; // Valid only for And, Or, Xor
6586 bool highBitSet = false; // Transform if high bit of constant set?
6588 switch (Op0BO->getOpcode()) {
6589 default: isValid = false; break; // Do not perform transform!
6590 case Instruction::Add:
6591 isValid = isLeftShift;
6593 case Instruction::Or:
6594 case Instruction::Xor:
6597 case Instruction::And:
6602 // If this is a signed shift right, and the high bit is modified
6603 // by the logical operation, do not perform the transformation.
6604 // The highBitSet boolean indicates the value of the high bit of
6605 // the constant which would cause it to be modified for this
6608 if (isValid && I.getOpcode() == Instruction::AShr)
6609 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6612 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6614 Instruction *NewShift =
6615 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6616 InsertNewInstBefore(NewShift, I);
6617 NewShift->takeName(Op0BO);
6619 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
6626 // Find out if this is a shift of a shift by a constant.
6627 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6628 if (ShiftOp && !ShiftOp->isShift())
6631 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6632 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6633 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6634 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6635 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6636 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6637 Value *X = ShiftOp->getOperand(0);
6639 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6640 if (AmtSum > TypeBits)
6643 const IntegerType *Ty = cast<IntegerType>(I.getType());
6645 // Check for (X << c1) << c2 and (X >> c1) >> c2
6646 if (I.getOpcode() == ShiftOp->getOpcode()) {
6647 return BinaryOperator::Create(I.getOpcode(), X,
6648 ConstantInt::get(Ty, AmtSum));
6649 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6650 I.getOpcode() == Instruction::AShr) {
6651 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6652 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
6653 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6654 I.getOpcode() == Instruction::LShr) {
6655 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6656 Instruction *Shift =
6657 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
6658 InsertNewInstBefore(Shift, I);
6660 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6661 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6664 // Okay, if we get here, one shift must be left, and the other shift must be
6665 // right. See if the amounts are equal.
6666 if (ShiftAmt1 == ShiftAmt2) {
6667 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6668 if (I.getOpcode() == Instruction::Shl) {
6669 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6670 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6672 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6673 if (I.getOpcode() == Instruction::LShr) {
6674 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6675 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6677 // We can simplify ((X << C) >>s C) into a trunc + sext.
6678 // NOTE: we could do this for any C, but that would make 'unusual' integer
6679 // types. For now, just stick to ones well-supported by the code
6681 const Type *SExtType = 0;
6682 switch (Ty->getBitWidth() - ShiftAmt1) {
6689 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6694 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6695 InsertNewInstBefore(NewTrunc, I);
6696 return new SExtInst(NewTrunc, Ty);
6698 // Otherwise, we can't handle it yet.
6699 } else if (ShiftAmt1 < ShiftAmt2) {
6700 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6702 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6703 if (I.getOpcode() == Instruction::Shl) {
6704 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6705 ShiftOp->getOpcode() == Instruction::AShr);
6706 Instruction *Shift =
6707 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6708 InsertNewInstBefore(Shift, I);
6710 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6711 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6714 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6715 if (I.getOpcode() == Instruction::LShr) {
6716 assert(ShiftOp->getOpcode() == Instruction::Shl);
6717 Instruction *Shift =
6718 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
6719 InsertNewInstBefore(Shift, I);
6721 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6722 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6725 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6727 assert(ShiftAmt2 < ShiftAmt1);
6728 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6730 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6731 if (I.getOpcode() == Instruction::Shl) {
6732 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6733 ShiftOp->getOpcode() == Instruction::AShr);
6734 Instruction *Shift =
6735 BinaryOperator::Create(ShiftOp->getOpcode(), X,
6736 ConstantInt::get(Ty, ShiftDiff));
6737 InsertNewInstBefore(Shift, I);
6739 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6740 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6743 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6744 if (I.getOpcode() == Instruction::LShr) {
6745 assert(ShiftOp->getOpcode() == Instruction::Shl);
6746 Instruction *Shift =
6747 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6748 InsertNewInstBefore(Shift, I);
6750 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6751 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6754 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6761 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6762 /// expression. If so, decompose it, returning some value X, such that Val is
6765 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6767 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6768 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6769 Offset = CI->getZExtValue();
6771 return ConstantInt::get(Type::Int32Ty, 0);
6772 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6773 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6774 if (I->getOpcode() == Instruction::Shl) {
6775 // This is a value scaled by '1 << the shift amt'.
6776 Scale = 1U << RHS->getZExtValue();
6778 return I->getOperand(0);
6779 } else if (I->getOpcode() == Instruction::Mul) {
6780 // This value is scaled by 'RHS'.
6781 Scale = RHS->getZExtValue();
6783 return I->getOperand(0);
6784 } else if (I->getOpcode() == Instruction::Add) {
6785 // We have X+C. Check to see if we really have (X*C2)+C1,
6786 // where C1 is divisible by C2.
6789 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6790 Offset += RHS->getZExtValue();
6797 // Otherwise, we can't look past this.
6804 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6805 /// try to eliminate the cast by moving the type information into the alloc.
6806 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6807 AllocationInst &AI) {
6808 const PointerType *PTy = cast<PointerType>(CI.getType());
6810 // Remove any uses of AI that are dead.
6811 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6813 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6814 Instruction *User = cast<Instruction>(*UI++);
6815 if (isInstructionTriviallyDead(User)) {
6816 while (UI != E && *UI == User)
6817 ++UI; // If this instruction uses AI more than once, don't break UI.
6820 DOUT << "IC: DCE: " << *User;
6821 EraseInstFromFunction(*User);
6825 // Get the type really allocated and the type casted to.
6826 const Type *AllocElTy = AI.getAllocatedType();
6827 const Type *CastElTy = PTy->getElementType();
6828 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6830 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6831 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6832 if (CastElTyAlign < AllocElTyAlign) return 0;
6834 // If the allocation has multiple uses, only promote it if we are strictly
6835 // increasing the alignment of the resultant allocation. If we keep it the
6836 // same, we open the door to infinite loops of various kinds.
6837 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6839 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6840 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6841 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6843 // See if we can satisfy the modulus by pulling a scale out of the array
6845 unsigned ArraySizeScale;
6847 Value *NumElements = // See if the array size is a decomposable linear expr.
6848 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6850 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6852 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6853 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6855 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6860 // If the allocation size is constant, form a constant mul expression
6861 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6862 if (isa<ConstantInt>(NumElements))
6863 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6864 // otherwise multiply the amount and the number of elements
6865 else if (Scale != 1) {
6866 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
6867 Amt = InsertNewInstBefore(Tmp, AI);
6871 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6872 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6873 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
6874 Amt = InsertNewInstBefore(Tmp, AI);
6877 AllocationInst *New;
6878 if (isa<MallocInst>(AI))
6879 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6881 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6882 InsertNewInstBefore(New, AI);
6885 // If the allocation has multiple uses, insert a cast and change all things
6886 // that used it to use the new cast. This will also hack on CI, but it will
6888 if (!AI.hasOneUse()) {
6889 AddUsesToWorkList(AI);
6890 // New is the allocation instruction, pointer typed. AI is the original
6891 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6892 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6893 InsertNewInstBefore(NewCast, AI);
6894 AI.replaceAllUsesWith(NewCast);
6896 return ReplaceInstUsesWith(CI, New);
6899 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6900 /// and return it as type Ty without inserting any new casts and without
6901 /// changing the computed value. This is used by code that tries to decide
6902 /// whether promoting or shrinking integer operations to wider or smaller types
6903 /// will allow us to eliminate a truncate or extend.
6905 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6906 /// extension operation if Ty is larger.
6908 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
6909 /// should return true if trunc(V) can be computed by computing V in the smaller
6910 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
6911 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
6912 /// efficiently truncated.
6914 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
6915 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
6916 /// the final result.
6917 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6919 int &NumCastsRemoved) {
6920 // We can always evaluate constants in another type.
6921 if (isa<ConstantInt>(V))
6924 Instruction *I = dyn_cast<Instruction>(V);
6925 if (!I) return false;
6927 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6929 // If this is an extension or truncate, we can often eliminate it.
6930 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6931 // If this is a cast from the destination type, we can trivially eliminate
6932 // it, and this will remove a cast overall.
6933 if (I->getOperand(0)->getType() == Ty) {
6934 // If the first operand is itself a cast, and is eliminable, do not count
6935 // this as an eliminable cast. We would prefer to eliminate those two
6937 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
6943 // We can't extend or shrink something that has multiple uses: doing so would
6944 // require duplicating the instruction in general, which isn't profitable.
6945 if (!I->hasOneUse()) return false;
6947 switch (I->getOpcode()) {
6948 case Instruction::Add:
6949 case Instruction::Sub:
6950 case Instruction::Mul:
6951 case Instruction::And:
6952 case Instruction::Or:
6953 case Instruction::Xor:
6954 // These operators can all arbitrarily be extended or truncated.
6955 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6957 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6960 case Instruction::Shl:
6961 // If we are truncating the result of this SHL, and if it's a shift of a
6962 // constant amount, we can always perform a SHL in a smaller type.
6963 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6964 uint32_t BitWidth = Ty->getBitWidth();
6965 if (BitWidth < OrigTy->getBitWidth() &&
6966 CI->getLimitedValue(BitWidth) < BitWidth)
6967 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6971 case Instruction::LShr:
6972 // If this is a truncate of a logical shr, we can truncate it to a smaller
6973 // lshr iff we know that the bits we would otherwise be shifting in are
6975 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6976 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6977 uint32_t BitWidth = Ty->getBitWidth();
6978 if (BitWidth < OrigBitWidth &&
6979 MaskedValueIsZero(I->getOperand(0),
6980 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6981 CI->getLimitedValue(BitWidth) < BitWidth) {
6982 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6987 case Instruction::ZExt:
6988 case Instruction::SExt:
6989 case Instruction::Trunc:
6990 // If this is the same kind of case as our original (e.g. zext+zext), we
6991 // can safely replace it. Note that replacing it does not reduce the number
6992 // of casts in the input.
6993 if (I->getOpcode() == CastOpc)
6996 case Instruction::Select: {
6997 SelectInst *SI = cast<SelectInst>(I);
6998 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7000 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7003 case Instruction::PHI: {
7004 // We can change a phi if we can change all operands.
7005 PHINode *PN = cast<PHINode>(I);
7006 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7007 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7013 // TODO: Can handle more cases here.
7020 /// EvaluateInDifferentType - Given an expression that
7021 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7022 /// evaluate the expression.
7023 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7025 if (Constant *C = dyn_cast<Constant>(V))
7026 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7028 // Otherwise, it must be an instruction.
7029 Instruction *I = cast<Instruction>(V);
7030 Instruction *Res = 0;
7031 switch (I->getOpcode()) {
7032 case Instruction::Add:
7033 case Instruction::Sub:
7034 case Instruction::Mul:
7035 case Instruction::And:
7036 case Instruction::Or:
7037 case Instruction::Xor:
7038 case Instruction::AShr:
7039 case Instruction::LShr:
7040 case Instruction::Shl: {
7041 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7042 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7043 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7047 case Instruction::Trunc:
7048 case Instruction::ZExt:
7049 case Instruction::SExt:
7050 // If the source type of the cast is the type we're trying for then we can
7051 // just return the source. There's no need to insert it because it is not
7053 if (I->getOperand(0)->getType() == Ty)
7054 return I->getOperand(0);
7056 // Otherwise, must be the same type of cast, so just reinsert a new one.
7057 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7060 case Instruction::Select: {
7061 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7062 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7063 Res = SelectInst::Create(I->getOperand(0), True, False);
7066 case Instruction::PHI: {
7067 PHINode *OPN = cast<PHINode>(I);
7068 PHINode *NPN = PHINode::Create(Ty);
7069 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7070 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7071 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7077 // TODO: Can handle more cases here.
7078 assert(0 && "Unreachable!");
7083 return InsertNewInstBefore(Res, *I);
7086 /// @brief Implement the transforms common to all CastInst visitors.
7087 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7088 Value *Src = CI.getOperand(0);
7090 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7091 // eliminate it now.
7092 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7093 if (Instruction::CastOps opc =
7094 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7095 // The first cast (CSrc) is eliminable so we need to fix up or replace
7096 // the second cast (CI). CSrc will then have a good chance of being dead.
7097 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7101 // If we are casting a select then fold the cast into the select
7102 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7103 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7106 // If we are casting a PHI then fold the cast into the PHI
7107 if (isa<PHINode>(Src))
7108 if (Instruction *NV = FoldOpIntoPhi(CI))
7114 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7115 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7116 Value *Src = CI.getOperand(0);
7118 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7119 // If casting the result of a getelementptr instruction with no offset, turn
7120 // this into a cast of the original pointer!
7121 if (GEP->hasAllZeroIndices()) {
7122 // Changing the cast operand is usually not a good idea but it is safe
7123 // here because the pointer operand is being replaced with another
7124 // pointer operand so the opcode doesn't need to change.
7126 CI.setOperand(0, GEP->getOperand(0));
7130 // If the GEP has a single use, and the base pointer is a bitcast, and the
7131 // GEP computes a constant offset, see if we can convert these three
7132 // instructions into fewer. This typically happens with unions and other
7133 // non-type-safe code.
7134 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7135 if (GEP->hasAllConstantIndices()) {
7136 // We are guaranteed to get a constant from EmitGEPOffset.
7137 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7138 int64_t Offset = OffsetV->getSExtValue();
7140 // Get the base pointer input of the bitcast, and the type it points to.
7141 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7142 const Type *GEPIdxTy =
7143 cast<PointerType>(OrigBase->getType())->getElementType();
7144 if (GEPIdxTy->isSized()) {
7145 SmallVector<Value*, 8> NewIndices;
7147 // Start with the index over the outer type. Note that the type size
7148 // might be zero (even if the offset isn't zero) if the indexed type
7149 // is something like [0 x {int, int}]
7150 const Type *IntPtrTy = TD->getIntPtrType();
7151 int64_t FirstIdx = 0;
7152 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7153 FirstIdx = Offset/TySize;
7156 // Handle silly modulus not returning values values [0..TySize).
7160 assert(Offset >= 0);
7162 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7165 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7167 // Index into the types. If we fail, set OrigBase to null.
7169 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7170 const StructLayout *SL = TD->getStructLayout(STy);
7171 if (Offset < (int64_t)SL->getSizeInBytes()) {
7172 unsigned Elt = SL->getElementContainingOffset(Offset);
7173 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7175 Offset -= SL->getElementOffset(Elt);
7176 GEPIdxTy = STy->getElementType(Elt);
7178 // Otherwise, we can't index into this, bail out.
7182 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7183 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7184 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7185 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7188 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7190 GEPIdxTy = STy->getElementType();
7192 // Otherwise, we can't index into this, bail out.
7198 // If we were able to index down into an element, create the GEP
7199 // and bitcast the result. This eliminates one bitcast, potentially
7201 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7203 NewIndices.end(), "");
7204 InsertNewInstBefore(NGEP, CI);
7205 NGEP->takeName(GEP);
7207 if (isa<BitCastInst>(CI))
7208 return new BitCastInst(NGEP, CI.getType());
7209 assert(isa<PtrToIntInst>(CI));
7210 return new PtrToIntInst(NGEP, CI.getType());
7217 return commonCastTransforms(CI);
7222 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7223 /// integer types. This function implements the common transforms for all those
7225 /// @brief Implement the transforms common to CastInst with integer operands
7226 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7227 if (Instruction *Result = commonCastTransforms(CI))
7230 Value *Src = CI.getOperand(0);
7231 const Type *SrcTy = Src->getType();
7232 const Type *DestTy = CI.getType();
7233 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7234 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7236 // See if we can simplify any instructions used by the LHS whose sole
7237 // purpose is to compute bits we don't care about.
7238 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7239 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7240 KnownZero, KnownOne))
7243 // If the source isn't an instruction or has more than one use then we
7244 // can't do anything more.
7245 Instruction *SrcI = dyn_cast<Instruction>(Src);
7246 if (!SrcI || !Src->hasOneUse())
7249 // Attempt to propagate the cast into the instruction for int->int casts.
7250 int NumCastsRemoved = 0;
7251 if (!isa<BitCastInst>(CI) &&
7252 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7253 CI.getOpcode(), NumCastsRemoved)) {
7254 // If this cast is a truncate, evaluting in a different type always
7255 // eliminates the cast, so it is always a win. If this is a zero-extension,
7256 // we need to do an AND to maintain the clear top-part of the computation,
7257 // so we require that the input have eliminated at least one cast. If this
7258 // is a sign extension, we insert two new casts (to do the extension) so we
7259 // require that two casts have been eliminated.
7261 switch (CI.getOpcode()) {
7263 // All the others use floating point so we shouldn't actually
7264 // get here because of the check above.
7265 assert(0 && "Unknown cast type");
7266 case Instruction::Trunc:
7269 case Instruction::ZExt:
7270 DoXForm = NumCastsRemoved >= 1;
7272 case Instruction::SExt:
7273 DoXForm = NumCastsRemoved >= 2;
7278 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7279 CI.getOpcode() == Instruction::SExt);
7280 assert(Res->getType() == DestTy);
7281 switch (CI.getOpcode()) {
7282 default: assert(0 && "Unknown cast type!");
7283 case Instruction::Trunc:
7284 case Instruction::BitCast:
7285 // Just replace this cast with the result.
7286 return ReplaceInstUsesWith(CI, Res);
7287 case Instruction::ZExt: {
7288 // We need to emit an AND to clear the high bits.
7289 assert(SrcBitSize < DestBitSize && "Not a zext?");
7290 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7292 return BinaryOperator::CreateAnd(Res, C);
7294 case Instruction::SExt:
7295 // We need to emit a cast to truncate, then a cast to sext.
7296 return CastInst::Create(Instruction::SExt,
7297 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7303 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7304 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7306 switch (SrcI->getOpcode()) {
7307 case Instruction::Add:
7308 case Instruction::Mul:
7309 case Instruction::And:
7310 case Instruction::Or:
7311 case Instruction::Xor:
7312 // If we are discarding information, rewrite.
7313 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7314 // Don't insert two casts if they cannot be eliminated. We allow
7315 // two casts to be inserted if the sizes are the same. This could
7316 // only be converting signedness, which is a noop.
7317 if (DestBitSize == SrcBitSize ||
7318 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7319 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7320 Instruction::CastOps opcode = CI.getOpcode();
7321 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7322 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7323 return BinaryOperator::Create(
7324 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7328 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7329 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7330 SrcI->getOpcode() == Instruction::Xor &&
7331 Op1 == ConstantInt::getTrue() &&
7332 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7333 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7334 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7337 case Instruction::SDiv:
7338 case Instruction::UDiv:
7339 case Instruction::SRem:
7340 case Instruction::URem:
7341 // If we are just changing the sign, rewrite.
7342 if (DestBitSize == SrcBitSize) {
7343 // Don't insert two casts if they cannot be eliminated. We allow
7344 // two casts to be inserted if the sizes are the same. This could
7345 // only be converting signedness, which is a noop.
7346 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7347 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7348 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7350 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7352 return BinaryOperator::Create(
7353 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7358 case Instruction::Shl:
7359 // Allow changing the sign of the source operand. Do not allow
7360 // changing the size of the shift, UNLESS the shift amount is a
7361 // constant. We must not change variable sized shifts to a smaller
7362 // size, because it is undefined to shift more bits out than exist
7364 if (DestBitSize == SrcBitSize ||
7365 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7366 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7367 Instruction::BitCast : Instruction::Trunc);
7368 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7369 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7370 return BinaryOperator::CreateShl(Op0c, Op1c);
7373 case Instruction::AShr:
7374 // If this is a signed shr, and if all bits shifted in are about to be
7375 // truncated off, turn it into an unsigned shr to allow greater
7377 if (DestBitSize < SrcBitSize &&
7378 isa<ConstantInt>(Op1)) {
7379 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7380 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7381 // Insert the new logical shift right.
7382 return BinaryOperator::CreateLShr(Op0, Op1);
7390 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7391 if (Instruction *Result = commonIntCastTransforms(CI))
7394 Value *Src = CI.getOperand(0);
7395 const Type *Ty = CI.getType();
7396 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7397 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7399 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7400 switch (SrcI->getOpcode()) {
7402 case Instruction::LShr:
7403 // We can shrink lshr to something smaller if we know the bits shifted in
7404 // are already zeros.
7405 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7406 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7408 // Get a mask for the bits shifting in.
7409 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7410 Value* SrcIOp0 = SrcI->getOperand(0);
7411 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7412 if (ShAmt >= DestBitWidth) // All zeros.
7413 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7415 // Okay, we can shrink this. Truncate the input, then return a new
7417 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7418 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7420 return BinaryOperator::CreateLShr(V1, V2);
7422 } else { // This is a variable shr.
7424 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7425 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7426 // loop-invariant and CSE'd.
7427 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7428 Value *One = ConstantInt::get(SrcI->getType(), 1);
7430 Value *V = InsertNewInstBefore(
7431 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7433 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7434 SrcI->getOperand(0),
7436 Value *Zero = Constant::getNullValue(V->getType());
7437 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7447 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7448 /// in order to eliminate the icmp.
7449 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7451 // If we are just checking for a icmp eq of a single bit and zext'ing it
7452 // to an integer, then shift the bit to the appropriate place and then
7453 // cast to integer to avoid the comparison.
7454 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7455 const APInt &Op1CV = Op1C->getValue();
7457 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7458 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7459 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7460 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7461 if (!DoXform) return ICI;
7463 Value *In = ICI->getOperand(0);
7464 Value *Sh = ConstantInt::get(In->getType(),
7465 In->getType()->getPrimitiveSizeInBits()-1);
7466 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7467 In->getName()+".lobit"),
7469 if (In->getType() != CI.getType())
7470 In = CastInst::CreateIntegerCast(In, CI.getType(),
7471 false/*ZExt*/, "tmp", &CI);
7473 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7474 Constant *One = ConstantInt::get(In->getType(), 1);
7475 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7476 In->getName()+".not"),
7480 return ReplaceInstUsesWith(CI, In);
7485 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7486 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7487 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7488 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7489 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7490 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7491 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7492 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7493 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7494 // This only works for EQ and NE
7495 ICI->isEquality()) {
7496 // If Op1C some other power of two, convert:
7497 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7498 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7499 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7500 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7502 APInt KnownZeroMask(~KnownZero);
7503 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7504 if (!DoXform) return ICI;
7506 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7507 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7508 // (X&4) == 2 --> false
7509 // (X&4) != 2 --> true
7510 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7511 Res = ConstantExpr::getZExt(Res, CI.getType());
7512 return ReplaceInstUsesWith(CI, Res);
7515 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7516 Value *In = ICI->getOperand(0);
7518 // Perform a logical shr by shiftamt.
7519 // Insert the shift to put the result in the low bit.
7520 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7521 ConstantInt::get(In->getType(), ShiftAmt),
7522 In->getName()+".lobit"), CI);
7525 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7526 Constant *One = ConstantInt::get(In->getType(), 1);
7527 In = BinaryOperator::CreateXor(In, One, "tmp");
7528 InsertNewInstBefore(cast<Instruction>(In), CI);
7531 if (CI.getType() == In->getType())
7532 return ReplaceInstUsesWith(CI, In);
7534 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7542 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7543 // If one of the common conversion will work ..
7544 if (Instruction *Result = commonIntCastTransforms(CI))
7547 Value *Src = CI.getOperand(0);
7549 // If this is a cast of a cast
7550 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7551 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7552 // types and if the sizes are just right we can convert this into a logical
7553 // 'and' which will be much cheaper than the pair of casts.
7554 if (isa<TruncInst>(CSrc)) {
7555 // Get the sizes of the types involved
7556 Value *A = CSrc->getOperand(0);
7557 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7558 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7559 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7560 // If we're actually extending zero bits and the trunc is a no-op
7561 if (MidSize < DstSize && SrcSize == DstSize) {
7562 // Replace both of the casts with an And of the type mask.
7563 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7564 Constant *AndConst = ConstantInt::get(AndValue);
7566 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
7567 // Unfortunately, if the type changed, we need to cast it back.
7568 if (And->getType() != CI.getType()) {
7569 And->setName(CSrc->getName()+".mask");
7570 InsertNewInstBefore(And, CI);
7571 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
7578 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7579 return transformZExtICmp(ICI, CI);
7581 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7582 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7583 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7584 // of the (zext icmp) will be transformed.
7585 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7586 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7587 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7588 (transformZExtICmp(LHS, CI, false) ||
7589 transformZExtICmp(RHS, CI, false))) {
7590 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7591 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7592 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
7599 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7600 if (Instruction *I = commonIntCastTransforms(CI))
7603 Value *Src = CI.getOperand(0);
7605 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7606 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7607 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7608 // If we are just checking for a icmp eq of a single bit and zext'ing it
7609 // to an integer, then shift the bit to the appropriate place and then
7610 // cast to integer to avoid the comparison.
7611 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7612 const APInt &Op1CV = Op1C->getValue();
7614 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7615 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7616 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7617 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7618 Value *In = ICI->getOperand(0);
7619 Value *Sh = ConstantInt::get(In->getType(),
7620 In->getType()->getPrimitiveSizeInBits()-1);
7621 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
7622 In->getName()+".lobit"),
7624 if (In->getType() != CI.getType())
7625 In = CastInst::CreateIntegerCast(In, CI.getType(),
7626 true/*SExt*/, "tmp", &CI);
7628 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7629 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
7630 In->getName()+".not"), CI);
7632 return ReplaceInstUsesWith(CI, In);
7637 // See if the value being truncated is already sign extended. If so, just
7638 // eliminate the trunc/sext pair.
7639 if (getOpcode(Src) == Instruction::Trunc) {
7640 Value *Op = cast<User>(Src)->getOperand(0);
7641 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
7642 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
7643 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
7644 unsigned NumSignBits = ComputeNumSignBits(Op);
7646 if (OpBits == DestBits) {
7647 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7648 // bits, it is already ready.
7649 if (NumSignBits > DestBits-MidBits)
7650 return ReplaceInstUsesWith(CI, Op);
7651 } else if (OpBits < DestBits) {
7652 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7653 // bits, just sext from i32.
7654 if (NumSignBits > OpBits-MidBits)
7655 return new SExtInst(Op, CI.getType(), "tmp");
7657 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7658 // bits, just truncate to i32.
7659 if (NumSignBits > OpBits-MidBits)
7660 return new TruncInst(Op, CI.getType(), "tmp");
7667 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7668 /// in the specified FP type without changing its value.
7669 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7670 APFloat F = CFP->getValueAPF();
7671 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7672 return ConstantFP::get(F);
7676 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7677 /// through it until we get the source value.
7678 static Value *LookThroughFPExtensions(Value *V) {
7679 if (Instruction *I = dyn_cast<Instruction>(V))
7680 if (I->getOpcode() == Instruction::FPExt)
7681 return LookThroughFPExtensions(I->getOperand(0));
7683 // If this value is a constant, return the constant in the smallest FP type
7684 // that can accurately represent it. This allows us to turn
7685 // (float)((double)X+2.0) into x+2.0f.
7686 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7687 if (CFP->getType() == Type::PPC_FP128Ty)
7688 return V; // No constant folding of this.
7689 // See if the value can be truncated to float and then reextended.
7690 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7692 if (CFP->getType() == Type::DoubleTy)
7693 return V; // Won't shrink.
7694 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7696 // Don't try to shrink to various long double types.
7702 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7703 if (Instruction *I = commonCastTransforms(CI))
7706 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7707 // smaller than the destination type, we can eliminate the truncate by doing
7708 // the add as the smaller type. This applies to add/sub/mul/div as well as
7709 // many builtins (sqrt, etc).
7710 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7711 if (OpI && OpI->hasOneUse()) {
7712 switch (OpI->getOpcode()) {
7714 case Instruction::Add:
7715 case Instruction::Sub:
7716 case Instruction::Mul:
7717 case Instruction::FDiv:
7718 case Instruction::FRem:
7719 const Type *SrcTy = OpI->getType();
7720 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7721 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7722 if (LHSTrunc->getType() != SrcTy &&
7723 RHSTrunc->getType() != SrcTy) {
7724 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7725 // If the source types were both smaller than the destination type of
7726 // the cast, do this xform.
7727 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7728 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7729 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7731 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7733 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7742 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7743 return commonCastTransforms(CI);
7746 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
7747 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
7748 // mantissa to accurately represent all values of X. For example, do not
7749 // do this with i64->float->i64.
7750 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
7751 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7752 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
7753 SrcI->getType()->getFPMantissaWidth())
7754 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7756 return commonCastTransforms(FI);
7759 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
7760 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
7761 // mantissa to accurately represent all values of X. For example, do not
7762 // do this with i64->float->i64.
7763 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
7764 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7765 (int)FI.getType()->getPrimitiveSizeInBits() <=
7766 SrcI->getType()->getFPMantissaWidth())
7767 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7769 return commonCastTransforms(FI);
7772 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7773 return commonCastTransforms(CI);
7776 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7777 return commonCastTransforms(CI);
7780 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7781 return commonPointerCastTransforms(CI);
7784 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7785 if (Instruction *I = commonCastTransforms(CI))
7788 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7789 if (!DestPointee->isSized()) return 0;
7791 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7794 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7795 m_ConstantInt(Cst)))) {
7796 // If the source and destination operands have the same type, see if this
7797 // is a single-index GEP.
7798 if (X->getType() == CI.getType()) {
7799 // Get the size of the pointee type.
7800 uint64_t Size = TD->getABITypeSize(DestPointee);
7802 // Convert the constant to intptr type.
7803 APInt Offset = Cst->getValue();
7804 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7806 // If Offset is evenly divisible by Size, we can do this xform.
7807 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7808 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7809 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7812 // TODO: Could handle other cases, e.g. where add is indexing into field of
7814 } else if (CI.getOperand(0)->hasOneUse() &&
7815 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7816 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7817 // "inttoptr+GEP" instead of "add+intptr".
7819 // Get the size of the pointee type.
7820 uint64_t Size = TD->getABITypeSize(DestPointee);
7822 // Convert the constant to intptr type.
7823 APInt Offset = Cst->getValue();
7824 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7826 // If Offset is evenly divisible by Size, we can do this xform.
7827 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7828 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7830 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7832 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7838 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7839 // If the operands are integer typed then apply the integer transforms,
7840 // otherwise just apply the common ones.
7841 Value *Src = CI.getOperand(0);
7842 const Type *SrcTy = Src->getType();
7843 const Type *DestTy = CI.getType();
7845 if (SrcTy->isInteger() && DestTy->isInteger()) {
7846 if (Instruction *Result = commonIntCastTransforms(CI))
7848 } else if (isa<PointerType>(SrcTy)) {
7849 if (Instruction *I = commonPointerCastTransforms(CI))
7852 if (Instruction *Result = commonCastTransforms(CI))
7857 // Get rid of casts from one type to the same type. These are useless and can
7858 // be replaced by the operand.
7859 if (DestTy == Src->getType())
7860 return ReplaceInstUsesWith(CI, Src);
7862 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7863 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7864 const Type *DstElTy = DstPTy->getElementType();
7865 const Type *SrcElTy = SrcPTy->getElementType();
7867 // If the address spaces don't match, don't eliminate the bitcast, which is
7868 // required for changing types.
7869 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7872 // If we are casting a malloc or alloca to a pointer to a type of the same
7873 // size, rewrite the allocation instruction to allocate the "right" type.
7874 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7875 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7878 // If the source and destination are pointers, and this cast is equivalent
7879 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7880 // This can enhance SROA and other transforms that want type-safe pointers.
7881 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7882 unsigned NumZeros = 0;
7883 while (SrcElTy != DstElTy &&
7884 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7885 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7886 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7890 // If we found a path from the src to dest, create the getelementptr now.
7891 if (SrcElTy == DstElTy) {
7892 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7893 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7894 ((Instruction*) NULL));
7898 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7899 if (SVI->hasOneUse()) {
7900 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7901 // a bitconvert to a vector with the same # elts.
7902 if (isa<VectorType>(DestTy) &&
7903 cast<VectorType>(DestTy)->getNumElements() ==
7904 SVI->getType()->getNumElements()) {
7906 // If either of the operands is a cast from CI.getType(), then
7907 // evaluating the shuffle in the casted destination's type will allow
7908 // us to eliminate at least one cast.
7909 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7910 Tmp->getOperand(0)->getType() == DestTy) ||
7911 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7912 Tmp->getOperand(0)->getType() == DestTy)) {
7913 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7914 SVI->getOperand(0), DestTy, &CI);
7915 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7916 SVI->getOperand(1), DestTy, &CI);
7917 // Return a new shuffle vector. Use the same element ID's, as we
7918 // know the vector types match #elts.
7919 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7927 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7929 /// %D = select %cond, %C, %A
7931 /// %C = select %cond, %B, 0
7934 /// Assuming that the specified instruction is an operand to the select, return
7935 /// a bitmask indicating which operands of this instruction are foldable if they
7936 /// equal the other incoming value of the select.
7938 static unsigned GetSelectFoldableOperands(Instruction *I) {
7939 switch (I->getOpcode()) {
7940 case Instruction::Add:
7941 case Instruction::Mul:
7942 case Instruction::And:
7943 case Instruction::Or:
7944 case Instruction::Xor:
7945 return 3; // Can fold through either operand.
7946 case Instruction::Sub: // Can only fold on the amount subtracted.
7947 case Instruction::Shl: // Can only fold on the shift amount.
7948 case Instruction::LShr:
7949 case Instruction::AShr:
7952 return 0; // Cannot fold
7956 /// GetSelectFoldableConstant - For the same transformation as the previous
7957 /// function, return the identity constant that goes into the select.
7958 static Constant *GetSelectFoldableConstant(Instruction *I) {
7959 switch (I->getOpcode()) {
7960 default: assert(0 && "This cannot happen!"); abort();
7961 case Instruction::Add:
7962 case Instruction::Sub:
7963 case Instruction::Or:
7964 case Instruction::Xor:
7965 case Instruction::Shl:
7966 case Instruction::LShr:
7967 case Instruction::AShr:
7968 return Constant::getNullValue(I->getType());
7969 case Instruction::And:
7970 return Constant::getAllOnesValue(I->getType());
7971 case Instruction::Mul:
7972 return ConstantInt::get(I->getType(), 1);
7976 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7977 /// have the same opcode and only one use each. Try to simplify this.
7978 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7980 if (TI->getNumOperands() == 1) {
7981 // If this is a non-volatile load or a cast from the same type,
7984 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7987 return 0; // unknown unary op.
7990 // Fold this by inserting a select from the input values.
7991 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
7992 FI->getOperand(0), SI.getName()+".v");
7993 InsertNewInstBefore(NewSI, SI);
7994 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
7998 // Only handle binary operators here.
7999 if (!isa<BinaryOperator>(TI))
8002 // Figure out if the operations have any operands in common.
8003 Value *MatchOp, *OtherOpT, *OtherOpF;
8005 if (TI->getOperand(0) == FI->getOperand(0)) {
8006 MatchOp = TI->getOperand(0);
8007 OtherOpT = TI->getOperand(1);
8008 OtherOpF = FI->getOperand(1);
8009 MatchIsOpZero = true;
8010 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8011 MatchOp = TI->getOperand(1);
8012 OtherOpT = TI->getOperand(0);
8013 OtherOpF = FI->getOperand(0);
8014 MatchIsOpZero = false;
8015 } else if (!TI->isCommutative()) {
8017 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8018 MatchOp = TI->getOperand(0);
8019 OtherOpT = TI->getOperand(1);
8020 OtherOpF = FI->getOperand(0);
8021 MatchIsOpZero = true;
8022 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8023 MatchOp = TI->getOperand(1);
8024 OtherOpT = TI->getOperand(0);
8025 OtherOpF = FI->getOperand(1);
8026 MatchIsOpZero = true;
8031 // If we reach here, they do have operations in common.
8032 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8033 OtherOpF, SI.getName()+".v");
8034 InsertNewInstBefore(NewSI, SI);
8036 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8038 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8040 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8042 assert(0 && "Shouldn't get here");
8046 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8047 Value *CondVal = SI.getCondition();
8048 Value *TrueVal = SI.getTrueValue();
8049 Value *FalseVal = SI.getFalseValue();
8051 // select true, X, Y -> X
8052 // select false, X, Y -> Y
8053 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8054 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8056 // select C, X, X -> X
8057 if (TrueVal == FalseVal)
8058 return ReplaceInstUsesWith(SI, TrueVal);
8060 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8061 return ReplaceInstUsesWith(SI, FalseVal);
8062 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8063 return ReplaceInstUsesWith(SI, TrueVal);
8064 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8065 if (isa<Constant>(TrueVal))
8066 return ReplaceInstUsesWith(SI, TrueVal);
8068 return ReplaceInstUsesWith(SI, FalseVal);
8071 if (SI.getType() == Type::Int1Ty) {
8072 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8073 if (C->getZExtValue()) {
8074 // Change: A = select B, true, C --> A = or B, C
8075 return BinaryOperator::CreateOr(CondVal, FalseVal);
8077 // Change: A = select B, false, C --> A = and !B, C
8079 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8080 "not."+CondVal->getName()), SI);
8081 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8083 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8084 if (C->getZExtValue() == false) {
8085 // Change: A = select B, C, false --> A = and B, C
8086 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8088 // Change: A = select B, C, true --> A = or !B, C
8090 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8091 "not."+CondVal->getName()), SI);
8092 return BinaryOperator::CreateOr(NotCond, TrueVal);
8096 // select a, b, a -> a&b
8097 // select a, a, b -> a|b
8098 if (CondVal == TrueVal)
8099 return BinaryOperator::CreateOr(CondVal, FalseVal);
8100 else if (CondVal == FalseVal)
8101 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8104 // Selecting between two integer constants?
8105 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8106 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8107 // select C, 1, 0 -> zext C to int
8108 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8109 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8110 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8111 // select C, 0, 1 -> zext !C to int
8113 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8114 "not."+CondVal->getName()), SI);
8115 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8118 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8120 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8122 // (x <s 0) ? -1 : 0 -> ashr x, 31
8123 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8124 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8125 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8126 // The comparison constant and the result are not neccessarily the
8127 // same width. Make an all-ones value by inserting a AShr.
8128 Value *X = IC->getOperand(0);
8129 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8130 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8131 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8133 InsertNewInstBefore(SRA, SI);
8135 // Finally, convert to the type of the select RHS. We figure out
8136 // if this requires a SExt, Trunc or BitCast based on the sizes.
8137 Instruction::CastOps opc = Instruction::BitCast;
8138 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8139 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8140 if (SRASize < SISize)
8141 opc = Instruction::SExt;
8142 else if (SRASize > SISize)
8143 opc = Instruction::Trunc;
8144 return CastInst::Create(opc, SRA, SI.getType());
8149 // If one of the constants is zero (we know they can't both be) and we
8150 // have an icmp instruction with zero, and we have an 'and' with the
8151 // non-constant value, eliminate this whole mess. This corresponds to
8152 // cases like this: ((X & 27) ? 27 : 0)
8153 if (TrueValC->isZero() || FalseValC->isZero())
8154 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8155 cast<Constant>(IC->getOperand(1))->isNullValue())
8156 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8157 if (ICA->getOpcode() == Instruction::And &&
8158 isa<ConstantInt>(ICA->getOperand(1)) &&
8159 (ICA->getOperand(1) == TrueValC ||
8160 ICA->getOperand(1) == FalseValC) &&
8161 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8162 // Okay, now we know that everything is set up, we just don't
8163 // know whether we have a icmp_ne or icmp_eq and whether the
8164 // true or false val is the zero.
8165 bool ShouldNotVal = !TrueValC->isZero();
8166 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8169 V = InsertNewInstBefore(BinaryOperator::Create(
8170 Instruction::Xor, V, ICA->getOperand(1)), SI);
8171 return ReplaceInstUsesWith(SI, V);
8176 // See if we are selecting two values based on a comparison of the two values.
8177 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8178 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8179 // Transform (X == Y) ? X : Y -> Y
8180 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8181 // This is not safe in general for floating point:
8182 // consider X== -0, Y== +0.
8183 // It becomes safe if either operand is a nonzero constant.
8184 ConstantFP *CFPt, *CFPf;
8185 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8186 !CFPt->getValueAPF().isZero()) ||
8187 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8188 !CFPf->getValueAPF().isZero()))
8189 return ReplaceInstUsesWith(SI, FalseVal);
8191 // Transform (X != Y) ? X : Y -> X
8192 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8193 return ReplaceInstUsesWith(SI, TrueVal);
8194 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8196 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8197 // Transform (X == Y) ? Y : X -> X
8198 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8199 // This is not safe in general for floating point:
8200 // consider X== -0, Y== +0.
8201 // It becomes safe if either operand is a nonzero constant.
8202 ConstantFP *CFPt, *CFPf;
8203 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8204 !CFPt->getValueAPF().isZero()) ||
8205 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8206 !CFPf->getValueAPF().isZero()))
8207 return ReplaceInstUsesWith(SI, FalseVal);
8209 // Transform (X != Y) ? Y : X -> Y
8210 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8211 return ReplaceInstUsesWith(SI, TrueVal);
8212 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8216 // See if we are selecting two values based on a comparison of the two values.
8217 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8218 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8219 // Transform (X == Y) ? X : Y -> Y
8220 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8221 return ReplaceInstUsesWith(SI, FalseVal);
8222 // Transform (X != Y) ? X : Y -> X
8223 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8224 return ReplaceInstUsesWith(SI, TrueVal);
8225 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8227 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8228 // Transform (X == Y) ? Y : X -> X
8229 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8230 return ReplaceInstUsesWith(SI, FalseVal);
8231 // Transform (X != Y) ? Y : X -> Y
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.
8238 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8239 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8240 if (TI->hasOneUse() && FI->hasOneUse()) {
8241 Instruction *AddOp = 0, *SubOp = 0;
8243 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8244 if (TI->getOpcode() == FI->getOpcode())
8245 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8248 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8249 // even legal for FP.
8250 if (TI->getOpcode() == Instruction::Sub &&
8251 FI->getOpcode() == Instruction::Add) {
8252 AddOp = FI; SubOp = TI;
8253 } else if (FI->getOpcode() == Instruction::Sub &&
8254 TI->getOpcode() == Instruction::Add) {
8255 AddOp = TI; SubOp = FI;
8259 Value *OtherAddOp = 0;
8260 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8261 OtherAddOp = AddOp->getOperand(1);
8262 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8263 OtherAddOp = AddOp->getOperand(0);
8267 // So at this point we know we have (Y -> OtherAddOp):
8268 // select C, (add X, Y), (sub X, Z)
8269 Value *NegVal; // Compute -Z
8270 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8271 NegVal = ConstantExpr::getNeg(C);
8273 NegVal = InsertNewInstBefore(
8274 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8277 Value *NewTrueOp = OtherAddOp;
8278 Value *NewFalseOp = NegVal;
8280 std::swap(NewTrueOp, NewFalseOp);
8281 Instruction *NewSel =
8282 SelectInst::Create(CondVal, NewTrueOp,
8283 NewFalseOp, SI.getName() + ".p");
8285 NewSel = InsertNewInstBefore(NewSel, SI);
8286 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8291 // See if we can fold the select into one of our operands.
8292 if (SI.getType()->isInteger()) {
8293 // See the comment above GetSelectFoldableOperands for a description of the
8294 // transformation we are doing here.
8295 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8296 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8297 !isa<Constant>(FalseVal))
8298 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8299 unsigned OpToFold = 0;
8300 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8302 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8307 Constant *C = GetSelectFoldableConstant(TVI);
8308 Instruction *NewSel =
8309 SelectInst::Create(SI.getCondition(),
8310 TVI->getOperand(2-OpToFold), C);
8311 InsertNewInstBefore(NewSel, SI);
8312 NewSel->takeName(TVI);
8313 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8314 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8316 assert(0 && "Unknown instruction!!");
8321 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8322 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8323 !isa<Constant>(TrueVal))
8324 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8325 unsigned OpToFold = 0;
8326 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8328 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8333 Constant *C = GetSelectFoldableConstant(FVI);
8334 Instruction *NewSel =
8335 SelectInst::Create(SI.getCondition(), C,
8336 FVI->getOperand(2-OpToFold));
8337 InsertNewInstBefore(NewSel, SI);
8338 NewSel->takeName(FVI);
8339 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8340 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8342 assert(0 && "Unknown instruction!!");
8347 if (BinaryOperator::isNot(CondVal)) {
8348 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8349 SI.setOperand(1, FalseVal);
8350 SI.setOperand(2, TrueVal);
8357 /// EnforceKnownAlignment - If the specified pointer points to an object that
8358 /// we control, modify the object's alignment to PrefAlign. This isn't
8359 /// often possible though. If alignment is important, a more reliable approach
8360 /// is to simply align all global variables and allocation instructions to
8361 /// their preferred alignment from the beginning.
8363 static unsigned EnforceKnownAlignment(Value *V,
8364 unsigned Align, unsigned PrefAlign) {
8366 User *U = dyn_cast<User>(V);
8367 if (!U) return Align;
8369 switch (getOpcode(U)) {
8371 case Instruction::BitCast:
8372 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8373 case Instruction::GetElementPtr: {
8374 // If all indexes are zero, it is just the alignment of the base pointer.
8375 bool AllZeroOperands = true;
8376 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8377 if (!isa<Constant>(*i) ||
8378 !cast<Constant>(*i)->isNullValue()) {
8379 AllZeroOperands = false;
8383 if (AllZeroOperands) {
8384 // Treat this like a bitcast.
8385 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8391 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8392 // If there is a large requested alignment and we can, bump up the alignment
8394 if (!GV->isDeclaration()) {
8395 GV->setAlignment(PrefAlign);
8398 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8399 // If there is a requested alignment and if this is an alloca, round up. We
8400 // don't do this for malloc, because some systems can't respect the request.
8401 if (isa<AllocaInst>(AI)) {
8402 AI->setAlignment(PrefAlign);
8410 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8411 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8412 /// and it is more than the alignment of the ultimate object, see if we can
8413 /// increase the alignment of the ultimate object, making this check succeed.
8414 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8415 unsigned PrefAlign) {
8416 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8417 sizeof(PrefAlign) * CHAR_BIT;
8418 APInt Mask = APInt::getAllOnesValue(BitWidth);
8419 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8420 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8421 unsigned TrailZ = KnownZero.countTrailingOnes();
8422 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8424 if (PrefAlign > Align)
8425 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8427 // We don't need to make any adjustment.
8431 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8432 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8433 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8434 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8435 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8437 if (CopyAlign < MinAlign) {
8438 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8442 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8444 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8445 if (MemOpLength == 0) return 0;
8447 // Source and destination pointer types are always "i8*" for intrinsic. See
8448 // if the size is something we can handle with a single primitive load/store.
8449 // A single load+store correctly handles overlapping memory in the memmove
8451 unsigned Size = MemOpLength->getZExtValue();
8452 if (Size == 0) return MI; // Delete this mem transfer.
8454 if (Size > 8 || (Size&(Size-1)))
8455 return 0; // If not 1/2/4/8 bytes, exit.
8457 // Use an integer load+store unless we can find something better.
8458 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8460 // Memcpy forces the use of i8* for the source and destination. That means
8461 // that if you're using memcpy to move one double around, you'll get a cast
8462 // from double* to i8*. We'd much rather use a double load+store rather than
8463 // an i64 load+store, here because this improves the odds that the source or
8464 // dest address will be promotable. See if we can find a better type than the
8465 // integer datatype.
8466 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8467 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8468 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8469 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8470 // down through these levels if so.
8471 while (!SrcETy->isSingleValueType()) {
8472 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8473 if (STy->getNumElements() == 1)
8474 SrcETy = STy->getElementType(0);
8477 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8478 if (ATy->getNumElements() == 1)
8479 SrcETy = ATy->getElementType();
8486 if (SrcETy->isSingleValueType())
8487 NewPtrTy = PointerType::getUnqual(SrcETy);
8492 // If the memcpy/memmove provides better alignment info than we can
8494 SrcAlign = std::max(SrcAlign, CopyAlign);
8495 DstAlign = std::max(DstAlign, CopyAlign);
8497 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8498 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8499 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8500 InsertNewInstBefore(L, *MI);
8501 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8503 // Set the size of the copy to 0, it will be deleted on the next iteration.
8504 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8508 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8509 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8510 if (MI->getAlignment()->getZExtValue() < Alignment) {
8511 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8515 // Extract the length and alignment and fill if they are constant.
8516 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8517 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8518 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8520 uint64_t Len = LenC->getZExtValue();
8521 Alignment = MI->getAlignment()->getZExtValue();
8523 // If the length is zero, this is a no-op
8524 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8526 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8527 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8528 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8530 Value *Dest = MI->getDest();
8531 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8533 // Alignment 0 is identity for alignment 1 for memset, but not store.
8534 if (Alignment == 0) Alignment = 1;
8536 // Extract the fill value and store.
8537 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8538 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8541 // Set the size of the copy to 0, it will be deleted on the next iteration.
8542 MI->setLength(Constant::getNullValue(LenC->getType()));
8550 /// visitCallInst - CallInst simplification. This mostly only handles folding
8551 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8552 /// the heavy lifting.
8554 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8555 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8556 if (!II) return visitCallSite(&CI);
8558 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8560 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8561 bool Changed = false;
8563 // memmove/cpy/set of zero bytes is a noop.
8564 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8565 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8567 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8568 if (CI->getZExtValue() == 1) {
8569 // Replace the instruction with just byte operations. We would
8570 // transform other cases to loads/stores, but we don't know if
8571 // alignment is sufficient.
8575 // If we have a memmove and the source operation is a constant global,
8576 // then the source and dest pointers can't alias, so we can change this
8577 // into a call to memcpy.
8578 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8579 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8580 if (GVSrc->isConstant()) {
8581 Module *M = CI.getParent()->getParent()->getParent();
8582 Intrinsic::ID MemCpyID;
8583 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8584 MemCpyID = Intrinsic::memcpy_i32;
8586 MemCpyID = Intrinsic::memcpy_i64;
8587 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8591 // memmove(x,x,size) -> noop.
8592 if (MMI->getSource() == MMI->getDest())
8593 return EraseInstFromFunction(CI);
8596 // If we can determine a pointer alignment that is bigger than currently
8597 // set, update the alignment.
8598 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8599 if (Instruction *I = SimplifyMemTransfer(MI))
8601 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8602 if (Instruction *I = SimplifyMemSet(MSI))
8606 if (Changed) return II;
8609 switch (II->getIntrinsicID()) {
8611 case Intrinsic::bswap:
8612 // bswap(bswap(x)) -> x
8613 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
8614 if (Operand->getIntrinsicID() == Intrinsic::bswap)
8615 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
8617 case Intrinsic::ppc_altivec_lvx:
8618 case Intrinsic::ppc_altivec_lvxl:
8619 case Intrinsic::x86_sse_loadu_ps:
8620 case Intrinsic::x86_sse2_loadu_pd:
8621 case Intrinsic::x86_sse2_loadu_dq:
8622 // Turn PPC lvx -> load if the pointer is known aligned.
8623 // Turn X86 loadups -> load if the pointer is known aligned.
8624 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8625 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8626 PointerType::getUnqual(II->getType()),
8628 return new LoadInst(Ptr);
8631 case Intrinsic::ppc_altivec_stvx:
8632 case Intrinsic::ppc_altivec_stvxl:
8633 // Turn stvx -> store if the pointer is known aligned.
8634 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8635 const Type *OpPtrTy =
8636 PointerType::getUnqual(II->getOperand(1)->getType());
8637 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8638 return new StoreInst(II->getOperand(1), Ptr);
8641 case Intrinsic::x86_sse_storeu_ps:
8642 case Intrinsic::x86_sse2_storeu_pd:
8643 case Intrinsic::x86_sse2_storeu_dq:
8644 case Intrinsic::x86_sse2_storel_dq:
8645 // Turn X86 storeu -> store if the pointer is known aligned.
8646 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8647 const Type *OpPtrTy =
8648 PointerType::getUnqual(II->getOperand(2)->getType());
8649 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8650 return new StoreInst(II->getOperand(2), Ptr);
8654 case Intrinsic::x86_sse_cvttss2si: {
8655 // These intrinsics only demands the 0th element of its input vector. If
8656 // we can simplify the input based on that, do so now.
8658 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8660 II->setOperand(1, V);
8666 case Intrinsic::ppc_altivec_vperm:
8667 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8668 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8669 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8671 // Check that all of the elements are integer constants or undefs.
8672 bool AllEltsOk = true;
8673 for (unsigned i = 0; i != 16; ++i) {
8674 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8675 !isa<UndefValue>(Mask->getOperand(i))) {
8682 // Cast the input vectors to byte vectors.
8683 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8684 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8685 Value *Result = UndefValue::get(Op0->getType());
8687 // Only extract each element once.
8688 Value *ExtractedElts[32];
8689 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8691 for (unsigned i = 0; i != 16; ++i) {
8692 if (isa<UndefValue>(Mask->getOperand(i)))
8694 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8695 Idx &= 31; // Match the hardware behavior.
8697 if (ExtractedElts[Idx] == 0) {
8699 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8700 InsertNewInstBefore(Elt, CI);
8701 ExtractedElts[Idx] = Elt;
8704 // Insert this value into the result vector.
8705 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
8707 InsertNewInstBefore(cast<Instruction>(Result), CI);
8709 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
8714 case Intrinsic::stackrestore: {
8715 // If the save is right next to the restore, remove the restore. This can
8716 // happen when variable allocas are DCE'd.
8717 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8718 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8719 BasicBlock::iterator BI = SS;
8721 return EraseInstFromFunction(CI);
8725 // Scan down this block to see if there is another stack restore in the
8726 // same block without an intervening call/alloca.
8727 BasicBlock::iterator BI = II;
8728 TerminatorInst *TI = II->getParent()->getTerminator();
8729 bool CannotRemove = false;
8730 for (++BI; &*BI != TI; ++BI) {
8731 if (isa<AllocaInst>(BI)) {
8732 CannotRemove = true;
8735 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
8736 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
8737 // If there is a stackrestore below this one, remove this one.
8738 if (II->getIntrinsicID() == Intrinsic::stackrestore)
8739 return EraseInstFromFunction(CI);
8740 // Otherwise, ignore the intrinsic.
8742 // If we found a non-intrinsic call, we can't remove the stack
8744 CannotRemove = true;
8750 // If the stack restore is in a return/unwind block and if there are no
8751 // allocas or calls between the restore and the return, nuke the restore.
8752 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8753 return EraseInstFromFunction(CI);
8758 return visitCallSite(II);
8761 // InvokeInst simplification
8763 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8764 return visitCallSite(&II);
8767 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8768 /// passed through the varargs area, we can eliminate the use of the cast.
8769 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8770 const CastInst * const CI,
8771 const TargetData * const TD,
8773 if (!CI->isLosslessCast())
8776 // The size of ByVal arguments is derived from the type, so we
8777 // can't change to a type with a different size. If the size were
8778 // passed explicitly we could avoid this check.
8779 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
8783 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
8784 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
8785 if (!SrcTy->isSized() || !DstTy->isSized())
8787 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
8792 // visitCallSite - Improvements for call and invoke instructions.
8794 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8795 bool Changed = false;
8797 // If the callee is a constexpr cast of a function, attempt to move the cast
8798 // to the arguments of the call/invoke.
8799 if (transformConstExprCastCall(CS)) return 0;
8801 Value *Callee = CS.getCalledValue();
8803 if (Function *CalleeF = dyn_cast<Function>(Callee))
8804 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8805 Instruction *OldCall = CS.getInstruction();
8806 // If the call and callee calling conventions don't match, this call must
8807 // be unreachable, as the call is undefined.
8808 new StoreInst(ConstantInt::getTrue(),
8809 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8811 if (!OldCall->use_empty())
8812 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8813 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8814 return EraseInstFromFunction(*OldCall);
8818 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8819 // This instruction is not reachable, just remove it. We insert a store to
8820 // undef so that we know that this code is not reachable, despite the fact
8821 // that we can't modify the CFG here.
8822 new StoreInst(ConstantInt::getTrue(),
8823 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8824 CS.getInstruction());
8826 if (!CS.getInstruction()->use_empty())
8827 CS.getInstruction()->
8828 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8830 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8831 // Don't break the CFG, insert a dummy cond branch.
8832 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8833 ConstantInt::getTrue(), II);
8835 return EraseInstFromFunction(*CS.getInstruction());
8838 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8839 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8840 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8841 return transformCallThroughTrampoline(CS);
8843 const PointerType *PTy = cast<PointerType>(Callee->getType());
8844 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8845 if (FTy->isVarArg()) {
8846 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
8847 // See if we can optimize any arguments passed through the varargs area of
8849 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8850 E = CS.arg_end(); I != E; ++I, ++ix) {
8851 CastInst *CI = dyn_cast<CastInst>(*I);
8852 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
8853 *I = CI->getOperand(0);
8859 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8860 // Inline asm calls cannot throw - mark them 'nounwind'.
8861 CS.setDoesNotThrow();
8865 return Changed ? CS.getInstruction() : 0;
8868 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8869 // attempt to move the cast to the arguments of the call/invoke.
8871 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8872 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8873 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8874 if (CE->getOpcode() != Instruction::BitCast ||
8875 !isa<Function>(CE->getOperand(0)))
8877 Function *Callee = cast<Function>(CE->getOperand(0));
8878 Instruction *Caller = CS.getInstruction();
8879 const PAListPtr &CallerPAL = CS.getParamAttrs();
8881 // Okay, this is a cast from a function to a different type. Unless doing so
8882 // would cause a type conversion of one of our arguments, change this call to
8883 // be a direct call with arguments casted to the appropriate types.
8885 const FunctionType *FT = Callee->getFunctionType();
8886 const Type *OldRetTy = Caller->getType();
8887 const Type *NewRetTy = FT->getReturnType();
8889 if (isa<StructType>(NewRetTy))
8890 return false; // TODO: Handle multiple return values.
8892 // Check to see if we are changing the return type...
8893 if (OldRetTy != NewRetTy) {
8894 if (Callee->isDeclaration() &&
8895 // Conversion is ok if changing from one pointer type to another or from
8896 // a pointer to an integer of the same size.
8897 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
8898 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
8899 return false; // Cannot transform this return value.
8901 if (!Caller->use_empty() &&
8902 // void -> non-void is handled specially
8903 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
8904 return false; // Cannot transform this return value.
8906 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8907 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8908 if (RAttrs & ParamAttr::typeIncompatible(NewRetTy))
8909 return false; // Attribute not compatible with transformed value.
8912 // If the callsite is an invoke instruction, and the return value is used by
8913 // a PHI node in a successor, we cannot change the return type of the call
8914 // because there is no place to put the cast instruction (without breaking
8915 // the critical edge). Bail out in this case.
8916 if (!Caller->use_empty())
8917 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8918 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8920 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8921 if (PN->getParent() == II->getNormalDest() ||
8922 PN->getParent() == II->getUnwindDest())
8926 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8927 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8929 CallSite::arg_iterator AI = CS.arg_begin();
8930 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8931 const Type *ParamTy = FT->getParamType(i);
8932 const Type *ActTy = (*AI)->getType();
8934 if (!CastInst::isCastable(ActTy, ParamTy))
8935 return false; // Cannot transform this parameter value.
8937 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8938 return false; // Attribute not compatible with transformed value.
8940 // Converting from one pointer type to another or between a pointer and an
8941 // integer of the same size is safe even if we do not have a body.
8942 bool isConvertible = ActTy == ParamTy ||
8943 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
8944 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
8945 if (Callee->isDeclaration() && !isConvertible) return false;
8948 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8949 Callee->isDeclaration())
8950 return false; // Do not delete arguments unless we have a function body.
8952 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
8953 !CallerPAL.isEmpty())
8954 // In this case we have more arguments than the new function type, but we
8955 // won't be dropping them. Check that these extra arguments have attributes
8956 // that are compatible with being a vararg call argument.
8957 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
8958 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
8960 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
8961 if (PAttrs & ParamAttr::VarArgsIncompatible)
8965 // Okay, we decided that this is a safe thing to do: go ahead and start
8966 // inserting cast instructions as necessary...
8967 std::vector<Value*> Args;
8968 Args.reserve(NumActualArgs);
8969 SmallVector<ParamAttrsWithIndex, 8> attrVec;
8970 attrVec.reserve(NumCommonArgs);
8972 // Get any return attributes.
8973 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8975 // If the return value is not being used, the type may not be compatible
8976 // with the existing attributes. Wipe out any problematic attributes.
8977 RAttrs &= ~ParamAttr::typeIncompatible(NewRetTy);
8979 // Add the new return attributes.
8981 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8983 AI = CS.arg_begin();
8984 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8985 const Type *ParamTy = FT->getParamType(i);
8986 if ((*AI)->getType() == ParamTy) {
8987 Args.push_back(*AI);
8989 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8990 false, ParamTy, false);
8991 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
8992 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8995 // Add any parameter attributes.
8996 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8997 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9000 // If the function takes more arguments than the call was taking, add them
9002 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9003 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9005 // If we are removing arguments to the function, emit an obnoxious warning...
9006 if (FT->getNumParams() < NumActualArgs) {
9007 if (!FT->isVarArg()) {
9008 cerr << "WARNING: While resolving call to function '"
9009 << Callee->getName() << "' arguments were dropped!\n";
9011 // Add all of the arguments in their promoted form to the arg list...
9012 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9013 const Type *PTy = getPromotedType((*AI)->getType());
9014 if (PTy != (*AI)->getType()) {
9015 // Must promote to pass through va_arg area!
9016 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9018 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9019 InsertNewInstBefore(Cast, *Caller);
9020 Args.push_back(Cast);
9022 Args.push_back(*AI);
9025 // Add any parameter attributes.
9026 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9027 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9032 if (NewRetTy == Type::VoidTy)
9033 Caller->setName(""); // Void type should not have a name.
9035 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9038 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9039 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9040 Args.begin(), Args.end(),
9041 Caller->getName(), Caller);
9042 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9043 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9045 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9046 Caller->getName(), Caller);
9047 CallInst *CI = cast<CallInst>(Caller);
9048 if (CI->isTailCall())
9049 cast<CallInst>(NC)->setTailCall();
9050 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9051 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9054 // Insert a cast of the return type as necessary.
9056 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9057 if (NV->getType() != Type::VoidTy) {
9058 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9060 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9062 // If this is an invoke instruction, we should insert it after the first
9063 // non-phi, instruction in the normal successor block.
9064 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9065 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9066 InsertNewInstBefore(NC, *I);
9068 // Otherwise, it's a call, just insert cast right after the call instr
9069 InsertNewInstBefore(NC, *Caller);
9071 AddUsersToWorkList(*Caller);
9073 NV = UndefValue::get(Caller->getType());
9077 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9078 Caller->replaceAllUsesWith(NV);
9079 Caller->eraseFromParent();
9080 RemoveFromWorkList(Caller);
9084 // transformCallThroughTrampoline - Turn a call to a function created by the
9085 // init_trampoline intrinsic into a direct call to the underlying function.
9087 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9088 Value *Callee = CS.getCalledValue();
9089 const PointerType *PTy = cast<PointerType>(Callee->getType());
9090 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9091 const PAListPtr &Attrs = CS.getParamAttrs();
9093 // If the call already has the 'nest' attribute somewhere then give up -
9094 // otherwise 'nest' would occur twice after splicing in the chain.
9095 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9098 IntrinsicInst *Tramp =
9099 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9101 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9102 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9103 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9105 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9106 if (!NestAttrs.isEmpty()) {
9107 unsigned NestIdx = 1;
9108 const Type *NestTy = 0;
9109 ParameterAttributes NestAttr = ParamAttr::None;
9111 // Look for a parameter marked with the 'nest' attribute.
9112 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9113 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9114 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9115 // Record the parameter type and any other attributes.
9117 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9122 Instruction *Caller = CS.getInstruction();
9123 std::vector<Value*> NewArgs;
9124 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9126 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9127 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9129 // Insert the nest argument into the call argument list, which may
9130 // mean appending it. Likewise for attributes.
9132 // Add any function result attributes.
9133 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9134 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9138 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9140 if (Idx == NestIdx) {
9141 // Add the chain argument and attributes.
9142 Value *NestVal = Tramp->getOperand(3);
9143 if (NestVal->getType() != NestTy)
9144 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9145 NewArgs.push_back(NestVal);
9146 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9152 // Add the original argument and attributes.
9153 NewArgs.push_back(*I);
9154 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9156 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9162 // The trampoline may have been bitcast to a bogus type (FTy).
9163 // Handle this by synthesizing a new function type, equal to FTy
9164 // with the chain parameter inserted.
9166 std::vector<const Type*> NewTypes;
9167 NewTypes.reserve(FTy->getNumParams()+1);
9169 // Insert the chain's type into the list of parameter types, which may
9170 // mean appending it.
9173 FunctionType::param_iterator I = FTy->param_begin(),
9174 E = FTy->param_end();
9178 // Add the chain's type.
9179 NewTypes.push_back(NestTy);
9184 // Add the original type.
9185 NewTypes.push_back(*I);
9191 // Replace the trampoline call with a direct call. Let the generic
9192 // code sort out any function type mismatches.
9193 FunctionType *NewFTy =
9194 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9195 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9196 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9197 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9199 Instruction *NewCaller;
9200 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9201 NewCaller = InvokeInst::Create(NewCallee,
9202 II->getNormalDest(), II->getUnwindDest(),
9203 NewArgs.begin(), NewArgs.end(),
9204 Caller->getName(), Caller);
9205 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9206 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9208 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9209 Caller->getName(), Caller);
9210 if (cast<CallInst>(Caller)->isTailCall())
9211 cast<CallInst>(NewCaller)->setTailCall();
9212 cast<CallInst>(NewCaller)->
9213 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9214 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9216 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9217 Caller->replaceAllUsesWith(NewCaller);
9218 Caller->eraseFromParent();
9219 RemoveFromWorkList(Caller);
9224 // Replace the trampoline call with a direct call. Since there is no 'nest'
9225 // parameter, there is no need to adjust the argument list. Let the generic
9226 // code sort out any function type mismatches.
9227 Constant *NewCallee =
9228 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9229 CS.setCalledFunction(NewCallee);
9230 return CS.getInstruction();
9233 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9234 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9235 /// and a single binop.
9236 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9237 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9238 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9239 isa<CmpInst>(FirstInst));
9240 unsigned Opc = FirstInst->getOpcode();
9241 Value *LHSVal = FirstInst->getOperand(0);
9242 Value *RHSVal = FirstInst->getOperand(1);
9244 const Type *LHSType = LHSVal->getType();
9245 const Type *RHSType = RHSVal->getType();
9247 // Scan to see if all operands are the same opcode, all have one use, and all
9248 // kill their operands (i.e. the operands have one use).
9249 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9250 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9251 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9252 // Verify type of the LHS matches so we don't fold cmp's of different
9253 // types or GEP's with different index types.
9254 I->getOperand(0)->getType() != LHSType ||
9255 I->getOperand(1)->getType() != RHSType)
9258 // If they are CmpInst instructions, check their predicates
9259 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9260 if (cast<CmpInst>(I)->getPredicate() !=
9261 cast<CmpInst>(FirstInst)->getPredicate())
9264 // Keep track of which operand needs a phi node.
9265 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9266 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9269 // Otherwise, this is safe to transform, determine if it is profitable.
9271 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9272 // Indexes are often folded into load/store instructions, so we don't want to
9273 // hide them behind a phi.
9274 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9277 Value *InLHS = FirstInst->getOperand(0);
9278 Value *InRHS = FirstInst->getOperand(1);
9279 PHINode *NewLHS = 0, *NewRHS = 0;
9281 NewLHS = PHINode::Create(LHSType,
9282 FirstInst->getOperand(0)->getName() + ".pn");
9283 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9284 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9285 InsertNewInstBefore(NewLHS, PN);
9290 NewRHS = PHINode::Create(RHSType,
9291 FirstInst->getOperand(1)->getName() + ".pn");
9292 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9293 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9294 InsertNewInstBefore(NewRHS, PN);
9298 // Add all operands to the new PHIs.
9299 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9301 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9302 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9305 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9306 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9310 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9311 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9312 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9313 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9316 assert(isa<GetElementPtrInst>(FirstInst));
9317 return GetElementPtrInst::Create(LHSVal, RHSVal);
9321 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9322 /// of the block that defines it. This means that it must be obvious the value
9323 /// of the load is not changed from the point of the load to the end of the
9326 /// Finally, it is safe, but not profitable, to sink a load targetting a
9327 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9329 static bool isSafeToSinkLoad(LoadInst *L) {
9330 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9332 for (++BBI; BBI != E; ++BBI)
9333 if (BBI->mayWriteToMemory())
9336 // Check for non-address taken alloca. If not address-taken already, it isn't
9337 // profitable to do this xform.
9338 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9339 bool isAddressTaken = false;
9340 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9342 if (isa<LoadInst>(UI)) continue;
9343 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9344 // If storing TO the alloca, then the address isn't taken.
9345 if (SI->getOperand(1) == AI) continue;
9347 isAddressTaken = true;
9351 if (!isAddressTaken)
9359 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9360 // operator and they all are only used by the PHI, PHI together their
9361 // inputs, and do the operation once, to the result of the PHI.
9362 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9363 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9365 // Scan the instruction, looking for input operations that can be folded away.
9366 // If all input operands to the phi are the same instruction (e.g. a cast from
9367 // the same type or "+42") we can pull the operation through the PHI, reducing
9368 // code size and simplifying code.
9369 Constant *ConstantOp = 0;
9370 const Type *CastSrcTy = 0;
9371 bool isVolatile = false;
9372 if (isa<CastInst>(FirstInst)) {
9373 CastSrcTy = FirstInst->getOperand(0)->getType();
9374 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9375 // Can fold binop, compare or shift here if the RHS is a constant,
9376 // otherwise call FoldPHIArgBinOpIntoPHI.
9377 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9378 if (ConstantOp == 0)
9379 return FoldPHIArgBinOpIntoPHI(PN);
9380 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9381 isVolatile = LI->isVolatile();
9382 // We can't sink the load if the loaded value could be modified between the
9383 // load and the PHI.
9384 if (LI->getParent() != PN.getIncomingBlock(0) ||
9385 !isSafeToSinkLoad(LI))
9388 // If the PHI is of volatile loads and the load block has multiple
9389 // successors, sinking it would remove a load of the volatile value from
9390 // the path through the other successor.
9392 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9395 } else if (isa<GetElementPtrInst>(FirstInst)) {
9396 if (FirstInst->getNumOperands() == 2)
9397 return FoldPHIArgBinOpIntoPHI(PN);
9398 // Can't handle general GEPs yet.
9401 return 0; // Cannot fold this operation.
9404 // Check to see if all arguments are the same operation.
9405 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9406 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9407 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9408 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9411 if (I->getOperand(0)->getType() != CastSrcTy)
9412 return 0; // Cast operation must match.
9413 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9414 // We can't sink the load if the loaded value could be modified between
9415 // the load and the PHI.
9416 if (LI->isVolatile() != isVolatile ||
9417 LI->getParent() != PN.getIncomingBlock(i) ||
9418 !isSafeToSinkLoad(LI))
9421 // If the PHI is of volatile loads and the load block has multiple
9422 // successors, sinking it would remove a load of the volatile value from
9423 // the path through the other successor.
9425 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9429 } else if (I->getOperand(1) != ConstantOp) {
9434 // Okay, they are all the same operation. Create a new PHI node of the
9435 // correct type, and PHI together all of the LHS's of the instructions.
9436 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9437 PN.getName()+".in");
9438 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9440 Value *InVal = FirstInst->getOperand(0);
9441 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9443 // Add all operands to the new PHI.
9444 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9445 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9446 if (NewInVal != InVal)
9448 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9453 // The new PHI unions all of the same values together. This is really
9454 // common, so we handle it intelligently here for compile-time speed.
9458 InsertNewInstBefore(NewPN, PN);
9462 // Insert and return the new operation.
9463 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9464 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9465 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9466 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9467 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9468 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9469 PhiVal, ConstantOp);
9470 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9472 // If this was a volatile load that we are merging, make sure to loop through
9473 // and mark all the input loads as non-volatile. If we don't do this, we will
9474 // insert a new volatile load and the old ones will not be deletable.
9476 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9477 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9479 return new LoadInst(PhiVal, "", isVolatile);
9482 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9484 static bool DeadPHICycle(PHINode *PN,
9485 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9486 if (PN->use_empty()) return true;
9487 if (!PN->hasOneUse()) return false;
9489 // Remember this node, and if we find the cycle, return.
9490 if (!PotentiallyDeadPHIs.insert(PN))
9493 // Don't scan crazily complex things.
9494 if (PotentiallyDeadPHIs.size() == 16)
9497 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9498 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9503 /// PHIsEqualValue - Return true if this phi node is always equal to
9504 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9505 /// z = some value; x = phi (y, z); y = phi (x, z)
9506 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9507 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9508 // See if we already saw this PHI node.
9509 if (!ValueEqualPHIs.insert(PN))
9512 // Don't scan crazily complex things.
9513 if (ValueEqualPHIs.size() == 16)
9516 // Scan the operands to see if they are either phi nodes or are equal to
9518 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9519 Value *Op = PN->getIncomingValue(i);
9520 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9521 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9523 } else if (Op != NonPhiInVal)
9531 // PHINode simplification
9533 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9534 // If LCSSA is around, don't mess with Phi nodes
9535 if (MustPreserveLCSSA) return 0;
9537 if (Value *V = PN.hasConstantValue())
9538 return ReplaceInstUsesWith(PN, V);
9540 // If all PHI operands are the same operation, pull them through the PHI,
9541 // reducing code size.
9542 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9543 PN.getIncomingValue(0)->hasOneUse())
9544 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9547 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9548 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9549 // PHI)... break the cycle.
9550 if (PN.hasOneUse()) {
9551 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9552 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9553 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9554 PotentiallyDeadPHIs.insert(&PN);
9555 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9556 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9559 // If this phi has a single use, and if that use just computes a value for
9560 // the next iteration of a loop, delete the phi. This occurs with unused
9561 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9562 // common case here is good because the only other things that catch this
9563 // are induction variable analysis (sometimes) and ADCE, which is only run
9565 if (PHIUser->hasOneUse() &&
9566 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9567 PHIUser->use_back() == &PN) {
9568 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9572 // We sometimes end up with phi cycles that non-obviously end up being the
9573 // same value, for example:
9574 // z = some value; x = phi (y, z); y = phi (x, z)
9575 // where the phi nodes don't necessarily need to be in the same block. Do a
9576 // quick check to see if the PHI node only contains a single non-phi value, if
9577 // so, scan to see if the phi cycle is actually equal to that value.
9579 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9580 // Scan for the first non-phi operand.
9581 while (InValNo != NumOperandVals &&
9582 isa<PHINode>(PN.getIncomingValue(InValNo)))
9585 if (InValNo != NumOperandVals) {
9586 Value *NonPhiInVal = PN.getOperand(InValNo);
9588 // Scan the rest of the operands to see if there are any conflicts, if so
9589 // there is no need to recursively scan other phis.
9590 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9591 Value *OpVal = PN.getIncomingValue(InValNo);
9592 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9596 // If we scanned over all operands, then we have one unique value plus
9597 // phi values. Scan PHI nodes to see if they all merge in each other or
9599 if (InValNo == NumOperandVals) {
9600 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9601 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9602 return ReplaceInstUsesWith(PN, NonPhiInVal);
9609 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9610 Instruction *InsertPoint,
9612 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9613 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9614 // We must cast correctly to the pointer type. Ensure that we
9615 // sign extend the integer value if it is smaller as this is
9616 // used for address computation.
9617 Instruction::CastOps opcode =
9618 (VTySize < PtrSize ? Instruction::SExt :
9619 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9620 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9624 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9625 Value *PtrOp = GEP.getOperand(0);
9626 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9627 // If so, eliminate the noop.
9628 if (GEP.getNumOperands() == 1)
9629 return ReplaceInstUsesWith(GEP, PtrOp);
9631 if (isa<UndefValue>(GEP.getOperand(0)))
9632 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9634 bool HasZeroPointerIndex = false;
9635 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9636 HasZeroPointerIndex = C->isNullValue();
9638 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9639 return ReplaceInstUsesWith(GEP, PtrOp);
9641 // Eliminate unneeded casts for indices.
9642 bool MadeChange = false;
9644 gep_type_iterator GTI = gep_type_begin(GEP);
9645 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
9646 i != e; ++i, ++GTI) {
9647 if (isa<SequentialType>(*GTI)) {
9648 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
9649 if (CI->getOpcode() == Instruction::ZExt ||
9650 CI->getOpcode() == Instruction::SExt) {
9651 const Type *SrcTy = CI->getOperand(0)->getType();
9652 // We can eliminate a cast from i32 to i64 iff the target
9653 // is a 32-bit pointer target.
9654 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9656 *i = CI->getOperand(0);
9660 // If we are using a wider index than needed for this platform, shrink it
9661 // to what we need. If the incoming value needs a cast instruction,
9662 // insert it. This explicit cast can make subsequent optimizations more
9665 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9666 if (Constant *C = dyn_cast<Constant>(Op)) {
9667 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
9670 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9678 if (MadeChange) return &GEP;
9680 // If this GEP instruction doesn't move the pointer, and if the input operand
9681 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9682 // real input to the dest type.
9683 if (GEP.hasAllZeroIndices()) {
9684 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9685 // If the bitcast is of an allocation, and the allocation will be
9686 // converted to match the type of the cast, don't touch this.
9687 if (isa<AllocationInst>(BCI->getOperand(0))) {
9688 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9689 if (Instruction *I = visitBitCast(*BCI)) {
9692 BCI->getParent()->getInstList().insert(BCI, I);
9693 ReplaceInstUsesWith(*BCI, I);
9698 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9702 // Combine Indices - If the source pointer to this getelementptr instruction
9703 // is a getelementptr instruction, combine the indices of the two
9704 // getelementptr instructions into a single instruction.
9706 SmallVector<Value*, 8> SrcGEPOperands;
9707 if (User *Src = dyn_castGetElementPtr(PtrOp))
9708 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9710 if (!SrcGEPOperands.empty()) {
9711 // Note that if our source is a gep chain itself that we wait for that
9712 // chain to be resolved before we perform this transformation. This
9713 // avoids us creating a TON of code in some cases.
9715 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9716 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9717 return 0; // Wait until our source is folded to completion.
9719 SmallVector<Value*, 8> Indices;
9721 // Find out whether the last index in the source GEP is a sequential idx.
9722 bool EndsWithSequential = false;
9723 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9724 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9725 EndsWithSequential = !isa<StructType>(*I);
9727 // Can we combine the two pointer arithmetics offsets?
9728 if (EndsWithSequential) {
9729 // Replace: gep (gep %P, long B), long A, ...
9730 // With: T = long A+B; gep %P, T, ...
9732 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9733 if (SO1 == Constant::getNullValue(SO1->getType())) {
9735 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9738 // If they aren't the same type, convert both to an integer of the
9739 // target's pointer size.
9740 if (SO1->getType() != GO1->getType()) {
9741 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9742 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9743 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9744 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9746 unsigned PS = TD->getPointerSizeInBits();
9747 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9748 // Convert GO1 to SO1's type.
9749 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9751 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9752 // Convert SO1 to GO1's type.
9753 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9755 const Type *PT = TD->getIntPtrType();
9756 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9757 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9761 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9762 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9764 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
9765 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9769 // Recycle the GEP we already have if possible.
9770 if (SrcGEPOperands.size() == 2) {
9771 GEP.setOperand(0, SrcGEPOperands[0]);
9772 GEP.setOperand(1, Sum);
9775 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9776 SrcGEPOperands.end()-1);
9777 Indices.push_back(Sum);
9778 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9780 } else if (isa<Constant>(*GEP.idx_begin()) &&
9781 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9782 SrcGEPOperands.size() != 1) {
9783 // Otherwise we can do the fold if the first index of the GEP is a zero
9784 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9785 SrcGEPOperands.end());
9786 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9789 if (!Indices.empty())
9790 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9791 Indices.end(), GEP.getName());
9793 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9794 // GEP of global variable. If all of the indices for this GEP are
9795 // constants, we can promote this to a constexpr instead of an instruction.
9797 // Scan for nonconstants...
9798 SmallVector<Constant*, 8> Indices;
9799 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9800 for (; I != E && isa<Constant>(*I); ++I)
9801 Indices.push_back(cast<Constant>(*I));
9803 if (I == E) { // If they are all constants...
9804 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9805 &Indices[0],Indices.size());
9807 // Replace all uses of the GEP with the new constexpr...
9808 return ReplaceInstUsesWith(GEP, CE);
9810 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9811 if (!isa<PointerType>(X->getType())) {
9812 // Not interesting. Source pointer must be a cast from pointer.
9813 } else if (HasZeroPointerIndex) {
9814 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9815 // into : GEP [10 x i8]* X, i32 0, ...
9817 // This occurs when the program declares an array extern like "int X[];"
9819 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9820 const PointerType *XTy = cast<PointerType>(X->getType());
9821 if (const ArrayType *XATy =
9822 dyn_cast<ArrayType>(XTy->getElementType()))
9823 if (const ArrayType *CATy =
9824 dyn_cast<ArrayType>(CPTy->getElementType()))
9825 if (CATy->getElementType() == XATy->getElementType()) {
9826 // At this point, we know that the cast source type is a pointer
9827 // to an array of the same type as the destination pointer
9828 // array. Because the array type is never stepped over (there
9829 // is a leading zero) we can fold the cast into this GEP.
9830 GEP.setOperand(0, X);
9833 } else if (GEP.getNumOperands() == 2) {
9834 // Transform things like:
9835 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9836 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9837 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9838 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9839 if (isa<ArrayType>(SrcElTy) &&
9840 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9841 TD->getABITypeSize(ResElTy)) {
9843 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9844 Idx[1] = GEP.getOperand(1);
9845 Value *V = InsertNewInstBefore(
9846 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9847 // V and GEP are both pointer types --> BitCast
9848 return new BitCastInst(V, GEP.getType());
9851 // Transform things like:
9852 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9853 // (where tmp = 8*tmp2) into:
9854 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9856 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9857 uint64_t ArrayEltSize =
9858 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9860 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9861 // allow either a mul, shift, or constant here.
9863 ConstantInt *Scale = 0;
9864 if (ArrayEltSize == 1) {
9865 NewIdx = GEP.getOperand(1);
9866 Scale = ConstantInt::get(NewIdx->getType(), 1);
9867 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9868 NewIdx = ConstantInt::get(CI->getType(), 1);
9870 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9871 if (Inst->getOpcode() == Instruction::Shl &&
9872 isa<ConstantInt>(Inst->getOperand(1))) {
9873 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9874 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9875 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9876 NewIdx = Inst->getOperand(0);
9877 } else if (Inst->getOpcode() == Instruction::Mul &&
9878 isa<ConstantInt>(Inst->getOperand(1))) {
9879 Scale = cast<ConstantInt>(Inst->getOperand(1));
9880 NewIdx = Inst->getOperand(0);
9884 // If the index will be to exactly the right offset with the scale taken
9885 // out, perform the transformation. Note, we don't know whether Scale is
9886 // signed or not. We'll use unsigned version of division/modulo
9887 // operation after making sure Scale doesn't have the sign bit set.
9888 if (Scale && Scale->getSExtValue() >= 0LL &&
9889 Scale->getZExtValue() % ArrayEltSize == 0) {
9890 Scale = ConstantInt::get(Scale->getType(),
9891 Scale->getZExtValue() / ArrayEltSize);
9892 if (Scale->getZExtValue() != 1) {
9893 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9895 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
9896 NewIdx = InsertNewInstBefore(Sc, GEP);
9899 // Insert the new GEP instruction.
9901 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9903 Instruction *NewGEP =
9904 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9905 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9906 // The NewGEP must be pointer typed, so must the old one -> BitCast
9907 return new BitCastInst(NewGEP, GEP.getType());
9916 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9917 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9918 if (AI.isArrayAllocation()) { // Check C != 1
9919 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9921 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9922 AllocationInst *New = 0;
9924 // Create and insert the replacement instruction...
9925 if (isa<MallocInst>(AI))
9926 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9928 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9929 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9932 InsertNewInstBefore(New, AI);
9934 // Scan to the end of the allocation instructions, to skip over a block of
9935 // allocas if possible...
9937 BasicBlock::iterator It = New;
9938 while (isa<AllocationInst>(*It)) ++It;
9940 // Now that I is pointing to the first non-allocation-inst in the block,
9941 // insert our getelementptr instruction...
9943 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9947 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9948 New->getName()+".sub", It);
9950 // Now make everything use the getelementptr instead of the original
9952 return ReplaceInstUsesWith(AI, V);
9953 } else if (isa<UndefValue>(AI.getArraySize())) {
9954 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9958 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9959 // Note that we only do this for alloca's, because malloc should allocate and
9960 // return a unique pointer, even for a zero byte allocation.
9961 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9962 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9963 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9968 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9969 Value *Op = FI.getOperand(0);
9971 // free undef -> unreachable.
9972 if (isa<UndefValue>(Op)) {
9973 // Insert a new store to null because we cannot modify the CFG here.
9974 new StoreInst(ConstantInt::getTrue(),
9975 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9976 return EraseInstFromFunction(FI);
9979 // If we have 'free null' delete the instruction. This can happen in stl code
9980 // when lots of inlining happens.
9981 if (isa<ConstantPointerNull>(Op))
9982 return EraseInstFromFunction(FI);
9984 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9985 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9986 FI.setOperand(0, CI->getOperand(0));
9990 // Change free (gep X, 0,0,0,0) into free(X)
9991 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9992 if (GEPI->hasAllZeroIndices()) {
9993 AddToWorkList(GEPI);
9994 FI.setOperand(0, GEPI->getOperand(0));
9999 // Change free(malloc) into nothing, if the malloc has a single use.
10000 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10001 if (MI->hasOneUse()) {
10002 EraseInstFromFunction(FI);
10003 return EraseInstFromFunction(*MI);
10010 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10011 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10012 const TargetData *TD) {
10013 User *CI = cast<User>(LI.getOperand(0));
10014 Value *CastOp = CI->getOperand(0);
10016 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10017 // Instead of loading constant c string, use corresponding integer value
10018 // directly if string length is small enough.
10020 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10021 unsigned len = Str.length();
10022 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10023 unsigned numBits = Ty->getPrimitiveSizeInBits();
10024 // Replace LI with immediate integer store.
10025 if ((numBits >> 3) == len + 1) {
10026 APInt StrVal(numBits, 0);
10027 APInt SingleChar(numBits, 0);
10028 if (TD->isLittleEndian()) {
10029 for (signed i = len-1; i >= 0; i--) {
10030 SingleChar = (uint64_t) Str[i];
10031 StrVal = (StrVal << 8) | SingleChar;
10034 for (unsigned i = 0; i < len; i++) {
10035 SingleChar = (uint64_t) Str[i];
10036 StrVal = (StrVal << 8) | SingleChar;
10038 // Append NULL at the end.
10040 StrVal = (StrVal << 8) | SingleChar;
10042 Value *NL = ConstantInt::get(StrVal);
10043 return IC.ReplaceInstUsesWith(LI, NL);
10048 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10049 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10050 const Type *SrcPTy = SrcTy->getElementType();
10052 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10053 isa<VectorType>(DestPTy)) {
10054 // If the source is an array, the code below will not succeed. Check to
10055 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10057 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10058 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10059 if (ASrcTy->getNumElements() != 0) {
10061 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10062 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10063 SrcTy = cast<PointerType>(CastOp->getType());
10064 SrcPTy = SrcTy->getElementType();
10067 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10068 isa<VectorType>(SrcPTy)) &&
10069 // Do not allow turning this into a load of an integer, which is then
10070 // casted to a pointer, this pessimizes pointer analysis a lot.
10071 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10072 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10073 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10075 // Okay, we are casting from one integer or pointer type to another of
10076 // the same size. Instead of casting the pointer before the load, cast
10077 // the result of the loaded value.
10078 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10080 LI.isVolatile()),LI);
10081 // Now cast the result of the load.
10082 return new BitCastInst(NewLoad, LI.getType());
10089 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10090 /// from this value cannot trap. If it is not obviously safe to load from the
10091 /// specified pointer, we do a quick local scan of the basic block containing
10092 /// ScanFrom, to determine if the address is already accessed.
10093 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10094 // If it is an alloca it is always safe to load from.
10095 if (isa<AllocaInst>(V)) return true;
10097 // If it is a global variable it is mostly safe to load from.
10098 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10099 // Don't try to evaluate aliases. External weak GV can be null.
10100 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10102 // Otherwise, be a little bit agressive by scanning the local block where we
10103 // want to check to see if the pointer is already being loaded or stored
10104 // from/to. If so, the previous load or store would have already trapped,
10105 // so there is no harm doing an extra load (also, CSE will later eliminate
10106 // the load entirely).
10107 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10112 // If we see a free or a call (which might do a free) the pointer could be
10114 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10117 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10118 if (LI->getOperand(0) == V) return true;
10119 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10120 if (SI->getOperand(1) == V) return true;
10127 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10128 /// until we find the underlying object a pointer is referring to or something
10129 /// we don't understand. Note that the returned pointer may be offset from the
10130 /// input, because we ignore GEP indices.
10131 static Value *GetUnderlyingObject(Value *Ptr) {
10133 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10134 if (CE->getOpcode() == Instruction::BitCast ||
10135 CE->getOpcode() == Instruction::GetElementPtr)
10136 Ptr = CE->getOperand(0);
10139 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10140 Ptr = BCI->getOperand(0);
10141 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10142 Ptr = GEP->getOperand(0);
10149 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10150 Value *Op = LI.getOperand(0);
10152 // Attempt to improve the alignment.
10153 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10155 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10156 LI.getAlignment()))
10157 LI.setAlignment(KnownAlign);
10159 // load (cast X) --> cast (load X) iff safe
10160 if (isa<CastInst>(Op))
10161 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10164 // None of the following transforms are legal for volatile loads.
10165 if (LI.isVolatile()) return 0;
10167 if (&LI.getParent()->front() != &LI) {
10168 BasicBlock::iterator BBI = &LI; --BBI;
10169 // If the instruction immediately before this is a store to the same
10170 // address, do a simple form of store->load forwarding.
10171 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10172 if (SI->getOperand(1) == LI.getOperand(0))
10173 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10174 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10175 if (LIB->getOperand(0) == LI.getOperand(0))
10176 return ReplaceInstUsesWith(LI, LIB);
10179 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10180 const Value *GEPI0 = GEPI->getOperand(0);
10181 // TODO: Consider a target hook for valid address spaces for this xform.
10182 if (isa<ConstantPointerNull>(GEPI0) &&
10183 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10184 // Insert a new store to null instruction before the load to indicate
10185 // that this code is not reachable. We do this instead of inserting
10186 // an unreachable instruction directly because we cannot modify the
10188 new StoreInst(UndefValue::get(LI.getType()),
10189 Constant::getNullValue(Op->getType()), &LI);
10190 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10194 if (Constant *C = dyn_cast<Constant>(Op)) {
10195 // load null/undef -> undef
10196 // TODO: Consider a target hook for valid address spaces for this xform.
10197 if (isa<UndefValue>(C) || (C->isNullValue() &&
10198 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10199 // Insert a new store to null instruction before the load to indicate that
10200 // this code is not reachable. We do this instead of inserting an
10201 // unreachable instruction directly because we cannot modify the CFG.
10202 new StoreInst(UndefValue::get(LI.getType()),
10203 Constant::getNullValue(Op->getType()), &LI);
10204 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10207 // Instcombine load (constant global) into the value loaded.
10208 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10209 if (GV->isConstant() && !GV->isDeclaration())
10210 return ReplaceInstUsesWith(LI, GV->getInitializer());
10212 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10213 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10214 if (CE->getOpcode() == Instruction::GetElementPtr) {
10215 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10216 if (GV->isConstant() && !GV->isDeclaration())
10218 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10219 return ReplaceInstUsesWith(LI, V);
10220 if (CE->getOperand(0)->isNullValue()) {
10221 // Insert a new store to null instruction before the load to indicate
10222 // that this code is not reachable. We do this instead of inserting
10223 // an unreachable instruction directly because we cannot modify the
10225 new StoreInst(UndefValue::get(LI.getType()),
10226 Constant::getNullValue(Op->getType()), &LI);
10227 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10230 } else if (CE->isCast()) {
10231 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10237 // If this load comes from anywhere in a constant global, and if the global
10238 // is all undef or zero, we know what it loads.
10239 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10240 if (GV->isConstant() && GV->hasInitializer()) {
10241 if (GV->getInitializer()->isNullValue())
10242 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10243 else if (isa<UndefValue>(GV->getInitializer()))
10244 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10248 if (Op->hasOneUse()) {
10249 // Change select and PHI nodes to select values instead of addresses: this
10250 // helps alias analysis out a lot, allows many others simplifications, and
10251 // exposes redundancy in the code.
10253 // Note that we cannot do the transformation unless we know that the
10254 // introduced loads cannot trap! Something like this is valid as long as
10255 // the condition is always false: load (select bool %C, int* null, int* %G),
10256 // but it would not be valid if we transformed it to load from null
10257 // unconditionally.
10259 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10260 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10261 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10262 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10263 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10264 SI->getOperand(1)->getName()+".val"), LI);
10265 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10266 SI->getOperand(2)->getName()+".val"), LI);
10267 return SelectInst::Create(SI->getCondition(), V1, V2);
10270 // load (select (cond, null, P)) -> load P
10271 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10272 if (C->isNullValue()) {
10273 LI.setOperand(0, SI->getOperand(2));
10277 // load (select (cond, P, null)) -> load P
10278 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10279 if (C->isNullValue()) {
10280 LI.setOperand(0, SI->getOperand(1));
10288 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10290 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10291 User *CI = cast<User>(SI.getOperand(1));
10292 Value *CastOp = CI->getOperand(0);
10294 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10295 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10296 const Type *SrcPTy = SrcTy->getElementType();
10298 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10299 // If the source is an array, the code below will not succeed. Check to
10300 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10302 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10303 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10304 if (ASrcTy->getNumElements() != 0) {
10306 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10307 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10308 SrcTy = cast<PointerType>(CastOp->getType());
10309 SrcPTy = SrcTy->getElementType();
10312 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10313 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10314 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10316 // Okay, we are casting from one integer or pointer type to another of
10317 // the same size. Instead of casting the pointer before
10318 // the store, cast the value to be stored.
10320 Value *SIOp0 = SI.getOperand(0);
10321 Instruction::CastOps opcode = Instruction::BitCast;
10322 const Type* CastSrcTy = SIOp0->getType();
10323 const Type* CastDstTy = SrcPTy;
10324 if (isa<PointerType>(CastDstTy)) {
10325 if (CastSrcTy->isInteger())
10326 opcode = Instruction::IntToPtr;
10327 } else if (isa<IntegerType>(CastDstTy)) {
10328 if (isa<PointerType>(SIOp0->getType()))
10329 opcode = Instruction::PtrToInt;
10331 if (Constant *C = dyn_cast<Constant>(SIOp0))
10332 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10334 NewCast = IC.InsertNewInstBefore(
10335 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10337 return new StoreInst(NewCast, CastOp);
10344 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10345 Value *Val = SI.getOperand(0);
10346 Value *Ptr = SI.getOperand(1);
10348 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10349 EraseInstFromFunction(SI);
10354 // If the RHS is an alloca with a single use, zapify the store, making the
10356 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10357 if (isa<AllocaInst>(Ptr)) {
10358 EraseInstFromFunction(SI);
10363 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10364 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10365 GEP->getOperand(0)->hasOneUse()) {
10366 EraseInstFromFunction(SI);
10372 // Attempt to improve the alignment.
10373 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10375 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10376 SI.getAlignment()))
10377 SI.setAlignment(KnownAlign);
10379 // Do really simple DSE, to catch cases where there are several consequtive
10380 // stores to the same location, separated by a few arithmetic operations. This
10381 // situation often occurs with bitfield accesses.
10382 BasicBlock::iterator BBI = &SI;
10383 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10387 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10388 // Prev store isn't volatile, and stores to the same location?
10389 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10392 EraseInstFromFunction(*PrevSI);
10398 // If this is a load, we have to stop. However, if the loaded value is from
10399 // the pointer we're loading and is producing the pointer we're storing,
10400 // then *this* store is dead (X = load P; store X -> P).
10401 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10402 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10403 EraseInstFromFunction(SI);
10407 // Otherwise, this is a load from some other location. Stores before it
10408 // may not be dead.
10412 // Don't skip over loads or things that can modify memory.
10413 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10418 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10420 // store X, null -> turns into 'unreachable' in SimplifyCFG
10421 if (isa<ConstantPointerNull>(Ptr)) {
10422 if (!isa<UndefValue>(Val)) {
10423 SI.setOperand(0, UndefValue::get(Val->getType()));
10424 if (Instruction *U = dyn_cast<Instruction>(Val))
10425 AddToWorkList(U); // Dropped a use.
10428 return 0; // Do not modify these!
10431 // store undef, Ptr -> noop
10432 if (isa<UndefValue>(Val)) {
10433 EraseInstFromFunction(SI);
10438 // If the pointer destination is a cast, see if we can fold the cast into the
10440 if (isa<CastInst>(Ptr))
10441 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10443 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10445 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10449 // If this store is the last instruction in the basic block, and if the block
10450 // ends with an unconditional branch, try to move it to the successor block.
10452 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10453 if (BI->isUnconditional())
10454 if (SimplifyStoreAtEndOfBlock(SI))
10455 return 0; // xform done!
10460 /// SimplifyStoreAtEndOfBlock - Turn things like:
10461 /// if () { *P = v1; } else { *P = v2 }
10462 /// into a phi node with a store in the successor.
10464 /// Simplify things like:
10465 /// *P = v1; if () { *P = v2; }
10466 /// into a phi node with a store in the successor.
10468 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10469 BasicBlock *StoreBB = SI.getParent();
10471 // Check to see if the successor block has exactly two incoming edges. If
10472 // so, see if the other predecessor contains a store to the same location.
10473 // if so, insert a PHI node (if needed) and move the stores down.
10474 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10476 // Determine whether Dest has exactly two predecessors and, if so, compute
10477 // the other predecessor.
10478 pred_iterator PI = pred_begin(DestBB);
10479 BasicBlock *OtherBB = 0;
10480 if (*PI != StoreBB)
10483 if (PI == pred_end(DestBB))
10486 if (*PI != StoreBB) {
10491 if (++PI != pred_end(DestBB))
10494 // Bail out if all the relevant blocks aren't distinct (this can happen,
10495 // for example, if SI is in an infinite loop)
10496 if (StoreBB == DestBB || OtherBB == DestBB)
10499 // Verify that the other block ends in a branch and is not otherwise empty.
10500 BasicBlock::iterator BBI = OtherBB->getTerminator();
10501 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10502 if (!OtherBr || BBI == OtherBB->begin())
10505 // If the other block ends in an unconditional branch, check for the 'if then
10506 // else' case. there is an instruction before the branch.
10507 StoreInst *OtherStore = 0;
10508 if (OtherBr->isUnconditional()) {
10509 // If this isn't a store, or isn't a store to the same location, bail out.
10511 OtherStore = dyn_cast<StoreInst>(BBI);
10512 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10515 // Otherwise, the other block ended with a conditional branch. If one of the
10516 // destinations is StoreBB, then we have the if/then case.
10517 if (OtherBr->getSuccessor(0) != StoreBB &&
10518 OtherBr->getSuccessor(1) != StoreBB)
10521 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10522 // if/then triangle. See if there is a store to the same ptr as SI that
10523 // lives in OtherBB.
10525 // Check to see if we find the matching store.
10526 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10527 if (OtherStore->getOperand(1) != SI.getOperand(1))
10531 // If we find something that may be using or overwriting the stored
10532 // value, or if we run out of instructions, we can't do the xform.
10533 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
10534 BBI == OtherBB->begin())
10538 // In order to eliminate the store in OtherBr, we have to
10539 // make sure nothing reads or overwrites the stored value in
10541 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10542 // FIXME: This should really be AA driven.
10543 if (I->mayReadFromMemory() || I->mayWriteToMemory())
10548 // Insert a PHI node now if we need it.
10549 Value *MergedVal = OtherStore->getOperand(0);
10550 if (MergedVal != SI.getOperand(0)) {
10551 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10552 PN->reserveOperandSpace(2);
10553 PN->addIncoming(SI.getOperand(0), SI.getParent());
10554 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10555 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10558 // Advance to a place where it is safe to insert the new store and
10560 BBI = DestBB->getFirstNonPHI();
10561 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10562 OtherStore->isVolatile()), *BBI);
10564 // Nuke the old stores.
10565 EraseInstFromFunction(SI);
10566 EraseInstFromFunction(*OtherStore);
10572 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10573 // Change br (not X), label True, label False to: br X, label False, True
10575 BasicBlock *TrueDest;
10576 BasicBlock *FalseDest;
10577 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10578 !isa<Constant>(X)) {
10579 // Swap Destinations and condition...
10580 BI.setCondition(X);
10581 BI.setSuccessor(0, FalseDest);
10582 BI.setSuccessor(1, TrueDest);
10586 // Cannonicalize fcmp_one -> fcmp_oeq
10587 FCmpInst::Predicate FPred; Value *Y;
10588 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10589 TrueDest, FalseDest)))
10590 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10591 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10592 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10593 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10594 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10595 NewSCC->takeName(I);
10596 // Swap Destinations and condition...
10597 BI.setCondition(NewSCC);
10598 BI.setSuccessor(0, FalseDest);
10599 BI.setSuccessor(1, TrueDest);
10600 RemoveFromWorkList(I);
10601 I->eraseFromParent();
10602 AddToWorkList(NewSCC);
10606 // Cannonicalize icmp_ne -> icmp_eq
10607 ICmpInst::Predicate IPred;
10608 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10609 TrueDest, FalseDest)))
10610 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10611 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10612 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10613 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10614 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10615 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10616 NewSCC->takeName(I);
10617 // Swap Destinations and condition...
10618 BI.setCondition(NewSCC);
10619 BI.setSuccessor(0, FalseDest);
10620 BI.setSuccessor(1, TrueDest);
10621 RemoveFromWorkList(I);
10622 I->eraseFromParent();;
10623 AddToWorkList(NewSCC);
10630 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10631 Value *Cond = SI.getCondition();
10632 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10633 if (I->getOpcode() == Instruction::Add)
10634 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10635 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10636 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10637 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10639 SI.setOperand(0, I->getOperand(0));
10647 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
10648 // See if we are trying to extract a known value. If so, use that instead.
10649 if (Value *Elt = FindInsertedValue(EV.getOperand(0), EV.idx_begin(),
10650 EV.idx_end(), &EV))
10651 return ReplaceInstUsesWith(EV, Elt);
10657 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10658 /// is to leave as a vector operation.
10659 static bool CheapToScalarize(Value *V, bool isConstant) {
10660 if (isa<ConstantAggregateZero>(V))
10662 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10663 if (isConstant) return true;
10664 // If all elts are the same, we can extract.
10665 Constant *Op0 = C->getOperand(0);
10666 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10667 if (C->getOperand(i) != Op0)
10671 Instruction *I = dyn_cast<Instruction>(V);
10672 if (!I) return false;
10674 // Insert element gets simplified to the inserted element or is deleted if
10675 // this is constant idx extract element and its a constant idx insertelt.
10676 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10677 isa<ConstantInt>(I->getOperand(2)))
10679 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10681 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10682 if (BO->hasOneUse() &&
10683 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10684 CheapToScalarize(BO->getOperand(1), isConstant)))
10686 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10687 if (CI->hasOneUse() &&
10688 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10689 CheapToScalarize(CI->getOperand(1), isConstant)))
10695 /// Read and decode a shufflevector mask.
10697 /// It turns undef elements into values that are larger than the number of
10698 /// elements in the input.
10699 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10700 unsigned NElts = SVI->getType()->getNumElements();
10701 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10702 return std::vector<unsigned>(NElts, 0);
10703 if (isa<UndefValue>(SVI->getOperand(2)))
10704 return std::vector<unsigned>(NElts, 2*NElts);
10706 std::vector<unsigned> Result;
10707 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10708 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
10709 if (isa<UndefValue>(*i))
10710 Result.push_back(NElts*2); // undef -> 8
10712 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
10716 /// FindScalarElement - Given a vector and an element number, see if the scalar
10717 /// value is already around as a register, for example if it were inserted then
10718 /// extracted from the vector.
10719 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10720 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10721 const VectorType *PTy = cast<VectorType>(V->getType());
10722 unsigned Width = PTy->getNumElements();
10723 if (EltNo >= Width) // Out of range access.
10724 return UndefValue::get(PTy->getElementType());
10726 if (isa<UndefValue>(V))
10727 return UndefValue::get(PTy->getElementType());
10728 else if (isa<ConstantAggregateZero>(V))
10729 return Constant::getNullValue(PTy->getElementType());
10730 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10731 return CP->getOperand(EltNo);
10732 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10733 // If this is an insert to a variable element, we don't know what it is.
10734 if (!isa<ConstantInt>(III->getOperand(2)))
10736 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10738 // If this is an insert to the element we are looking for, return the
10740 if (EltNo == IIElt)
10741 return III->getOperand(1);
10743 // Otherwise, the insertelement doesn't modify the value, recurse on its
10745 return FindScalarElement(III->getOperand(0), EltNo);
10746 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10747 unsigned InEl = getShuffleMask(SVI)[EltNo];
10749 return FindScalarElement(SVI->getOperand(0), InEl);
10750 else if (InEl < Width*2)
10751 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10753 return UndefValue::get(PTy->getElementType());
10756 // Otherwise, we don't know.
10760 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10761 // If vector val is undef, replace extract with scalar undef.
10762 if (isa<UndefValue>(EI.getOperand(0)))
10763 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10765 // If vector val is constant 0, replace extract with scalar 0.
10766 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10767 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10769 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10770 // If vector val is constant with all elements the same, replace EI with
10771 // that element. When the elements are not identical, we cannot replace yet
10772 // (we do that below, but only when the index is constant).
10773 Constant *op0 = C->getOperand(0);
10774 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10775 if (C->getOperand(i) != op0) {
10780 return ReplaceInstUsesWith(EI, op0);
10783 // If extracting a specified index from the vector, see if we can recursively
10784 // find a previously computed scalar that was inserted into the vector.
10785 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10786 unsigned IndexVal = IdxC->getZExtValue();
10787 unsigned VectorWidth =
10788 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10790 // If this is extracting an invalid index, turn this into undef, to avoid
10791 // crashing the code below.
10792 if (IndexVal >= VectorWidth)
10793 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10795 // This instruction only demands the single element from the input vector.
10796 // If the input vector has a single use, simplify it based on this use
10798 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10799 uint64_t UndefElts;
10800 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10803 EI.setOperand(0, V);
10808 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10809 return ReplaceInstUsesWith(EI, Elt);
10811 // If the this extractelement is directly using a bitcast from a vector of
10812 // the same number of elements, see if we can find the source element from
10813 // it. In this case, we will end up needing to bitcast the scalars.
10814 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10815 if (const VectorType *VT =
10816 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10817 if (VT->getNumElements() == VectorWidth)
10818 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10819 return new BitCastInst(Elt, EI.getType());
10823 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10824 if (I->hasOneUse()) {
10825 // Push extractelement into predecessor operation if legal and
10826 // profitable to do so
10827 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10828 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10829 if (CheapToScalarize(BO, isConstantElt)) {
10830 ExtractElementInst *newEI0 =
10831 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10832 EI.getName()+".lhs");
10833 ExtractElementInst *newEI1 =
10834 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10835 EI.getName()+".rhs");
10836 InsertNewInstBefore(newEI0, EI);
10837 InsertNewInstBefore(newEI1, EI);
10838 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
10840 } else if (isa<LoadInst>(I)) {
10842 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10843 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10844 PointerType::get(EI.getType(), AS),EI);
10845 GetElementPtrInst *GEP =
10846 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
10847 InsertNewInstBefore(GEP, EI);
10848 return new LoadInst(GEP);
10851 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10852 // Extracting the inserted element?
10853 if (IE->getOperand(2) == EI.getOperand(1))
10854 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10855 // If the inserted and extracted elements are constants, they must not
10856 // be the same value, extract from the pre-inserted value instead.
10857 if (isa<Constant>(IE->getOperand(2)) &&
10858 isa<Constant>(EI.getOperand(1))) {
10859 AddUsesToWorkList(EI);
10860 EI.setOperand(0, IE->getOperand(0));
10863 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10864 // If this is extracting an element from a shufflevector, figure out where
10865 // it came from and extract from the appropriate input element instead.
10866 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10867 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10869 if (SrcIdx < SVI->getType()->getNumElements())
10870 Src = SVI->getOperand(0);
10871 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10872 SrcIdx -= SVI->getType()->getNumElements();
10873 Src = SVI->getOperand(1);
10875 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10877 return new ExtractElementInst(Src, SrcIdx);
10884 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10885 /// elements from either LHS or RHS, return the shuffle mask and true.
10886 /// Otherwise, return false.
10887 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10888 std::vector<Constant*> &Mask) {
10889 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10890 "Invalid CollectSingleShuffleElements");
10891 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10893 if (isa<UndefValue>(V)) {
10894 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10896 } else if (V == LHS) {
10897 for (unsigned i = 0; i != NumElts; ++i)
10898 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10900 } else if (V == RHS) {
10901 for (unsigned i = 0; i != NumElts; ++i)
10902 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10904 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10905 // If this is an insert of an extract from some other vector, include it.
10906 Value *VecOp = IEI->getOperand(0);
10907 Value *ScalarOp = IEI->getOperand(1);
10908 Value *IdxOp = IEI->getOperand(2);
10910 if (!isa<ConstantInt>(IdxOp))
10912 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10914 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10915 // Okay, we can handle this if the vector we are insertinting into is
10916 // transitively ok.
10917 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10918 // If so, update the mask to reflect the inserted undef.
10919 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10922 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10923 if (isa<ConstantInt>(EI->getOperand(1)) &&
10924 EI->getOperand(0)->getType() == V->getType()) {
10925 unsigned ExtractedIdx =
10926 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10928 // This must be extracting from either LHS or RHS.
10929 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10930 // Okay, we can handle this if the vector we are insertinting into is
10931 // transitively ok.
10932 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10933 // If so, update the mask to reflect the inserted value.
10934 if (EI->getOperand(0) == LHS) {
10935 Mask[InsertedIdx & (NumElts-1)] =
10936 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10938 assert(EI->getOperand(0) == RHS);
10939 Mask[InsertedIdx & (NumElts-1)] =
10940 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10949 // TODO: Handle shufflevector here!
10954 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10955 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10956 /// that computes V and the LHS value of the shuffle.
10957 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10959 assert(isa<VectorType>(V->getType()) &&
10960 (RHS == 0 || V->getType() == RHS->getType()) &&
10961 "Invalid shuffle!");
10962 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10964 if (isa<UndefValue>(V)) {
10965 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10967 } else if (isa<ConstantAggregateZero>(V)) {
10968 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10970 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10971 // If this is an insert of an extract from some other vector, include it.
10972 Value *VecOp = IEI->getOperand(0);
10973 Value *ScalarOp = IEI->getOperand(1);
10974 Value *IdxOp = IEI->getOperand(2);
10976 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10977 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10978 EI->getOperand(0)->getType() == V->getType()) {
10979 unsigned ExtractedIdx =
10980 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10981 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10983 // Either the extracted from or inserted into vector must be RHSVec,
10984 // otherwise we'd end up with a shuffle of three inputs.
10985 if (EI->getOperand(0) == RHS || RHS == 0) {
10986 RHS = EI->getOperand(0);
10987 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10988 Mask[InsertedIdx & (NumElts-1)] =
10989 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10993 if (VecOp == RHS) {
10994 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10995 // Everything but the extracted element is replaced with the RHS.
10996 for (unsigned i = 0; i != NumElts; ++i) {
10997 if (i != InsertedIdx)
10998 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11003 // If this insertelement is a chain that comes from exactly these two
11004 // vectors, return the vector and the effective shuffle.
11005 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11006 return EI->getOperand(0);
11011 // TODO: Handle shufflevector here!
11013 // Otherwise, can't do anything fancy. Return an identity vector.
11014 for (unsigned i = 0; i != NumElts; ++i)
11015 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11019 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11020 Value *VecOp = IE.getOperand(0);
11021 Value *ScalarOp = IE.getOperand(1);
11022 Value *IdxOp = IE.getOperand(2);
11024 // Inserting an undef or into an undefined place, remove this.
11025 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11026 ReplaceInstUsesWith(IE, VecOp);
11028 // If the inserted element was extracted from some other vector, and if the
11029 // indexes are constant, try to turn this into a shufflevector operation.
11030 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11031 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11032 EI->getOperand(0)->getType() == IE.getType()) {
11033 unsigned NumVectorElts = IE.getType()->getNumElements();
11034 unsigned ExtractedIdx =
11035 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11036 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11038 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11039 return ReplaceInstUsesWith(IE, VecOp);
11041 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11042 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11044 // If we are extracting a value from a vector, then inserting it right
11045 // back into the same place, just use the input vector.
11046 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11047 return ReplaceInstUsesWith(IE, VecOp);
11049 // We could theoretically do this for ANY input. However, doing so could
11050 // turn chains of insertelement instructions into a chain of shufflevector
11051 // instructions, and right now we do not merge shufflevectors. As such,
11052 // only do this in a situation where it is clear that there is benefit.
11053 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11054 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11055 // the values of VecOp, except then one read from EIOp0.
11056 // Build a new shuffle mask.
11057 std::vector<Constant*> Mask;
11058 if (isa<UndefValue>(VecOp))
11059 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11061 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11062 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11065 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11066 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11067 ConstantVector::get(Mask));
11070 // If this insertelement isn't used by some other insertelement, turn it
11071 // (and any insertelements it points to), into one big shuffle.
11072 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11073 std::vector<Constant*> Mask;
11075 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11076 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11077 // We now have a shuffle of LHS, RHS, Mask.
11078 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11087 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11088 Value *LHS = SVI.getOperand(0);
11089 Value *RHS = SVI.getOperand(1);
11090 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11092 bool MadeChange = false;
11094 // Undefined shuffle mask -> undefined value.
11095 if (isa<UndefValue>(SVI.getOperand(2)))
11096 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11098 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11099 // the undef, change them to undefs.
11100 if (isa<UndefValue>(SVI.getOperand(1))) {
11101 // Scan to see if there are any references to the RHS. If so, replace them
11102 // with undef element refs and set MadeChange to true.
11103 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11104 if (Mask[i] >= e && Mask[i] != 2*e) {
11111 // Remap any references to RHS to use LHS.
11112 std::vector<Constant*> Elts;
11113 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11114 if (Mask[i] == 2*e)
11115 Elts.push_back(UndefValue::get(Type::Int32Ty));
11117 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11119 SVI.setOperand(2, ConstantVector::get(Elts));
11123 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11124 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11125 if (LHS == RHS || isa<UndefValue>(LHS)) {
11126 if (isa<UndefValue>(LHS) && LHS == RHS) {
11127 // shuffle(undef,undef,mask) -> undef.
11128 return ReplaceInstUsesWith(SVI, LHS);
11131 // Remap any references to RHS to use LHS.
11132 std::vector<Constant*> Elts;
11133 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11134 if (Mask[i] >= 2*e)
11135 Elts.push_back(UndefValue::get(Type::Int32Ty));
11137 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11138 (Mask[i] < e && isa<UndefValue>(LHS)))
11139 Mask[i] = 2*e; // Turn into undef.
11141 Mask[i] &= (e-1); // Force to LHS.
11142 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11145 SVI.setOperand(0, SVI.getOperand(1));
11146 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11147 SVI.setOperand(2, ConstantVector::get(Elts));
11148 LHS = SVI.getOperand(0);
11149 RHS = SVI.getOperand(1);
11153 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11154 bool isLHSID = true, isRHSID = true;
11156 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11157 if (Mask[i] >= e*2) continue; // Ignore undef values.
11158 // Is this an identity shuffle of the LHS value?
11159 isLHSID &= (Mask[i] == i);
11161 // Is this an identity shuffle of the RHS value?
11162 isRHSID &= (Mask[i]-e == i);
11165 // Eliminate identity shuffles.
11166 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11167 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11169 // If the LHS is a shufflevector itself, see if we can combine it with this
11170 // one without producing an unusual shuffle. Here we are really conservative:
11171 // we are absolutely afraid of producing a shuffle mask not in the input
11172 // program, because the code gen may not be smart enough to turn a merged
11173 // shuffle into two specific shuffles: it may produce worse code. As such,
11174 // we only merge two shuffles if the result is one of the two input shuffle
11175 // masks. In this case, merging the shuffles just removes one instruction,
11176 // which we know is safe. This is good for things like turning:
11177 // (splat(splat)) -> splat.
11178 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11179 if (isa<UndefValue>(RHS)) {
11180 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11182 std::vector<unsigned> NewMask;
11183 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11184 if (Mask[i] >= 2*e)
11185 NewMask.push_back(2*e);
11187 NewMask.push_back(LHSMask[Mask[i]]);
11189 // If the result mask is equal to the src shuffle or this shuffle mask, do
11190 // the replacement.
11191 if (NewMask == LHSMask || NewMask == Mask) {
11192 std::vector<Constant*> Elts;
11193 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11194 if (NewMask[i] >= e*2) {
11195 Elts.push_back(UndefValue::get(Type::Int32Ty));
11197 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11200 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11201 LHSSVI->getOperand(1),
11202 ConstantVector::get(Elts));
11207 return MadeChange ? &SVI : 0;
11213 /// TryToSinkInstruction - Try to move the specified instruction from its
11214 /// current block into the beginning of DestBlock, which can only happen if it's
11215 /// safe to move the instruction past all of the instructions between it and the
11216 /// end of its block.
11217 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11218 assert(I->hasOneUse() && "Invariants didn't hold!");
11220 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11221 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11224 // Do not sink alloca instructions out of the entry block.
11225 if (isa<AllocaInst>(I) && I->getParent() ==
11226 &DestBlock->getParent()->getEntryBlock())
11229 // We can only sink load instructions if there is nothing between the load and
11230 // the end of block that could change the value.
11231 if (I->mayReadFromMemory()) {
11232 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11234 if (Scan->mayWriteToMemory())
11238 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11240 I->moveBefore(InsertPos);
11246 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11247 /// all reachable code to the worklist.
11249 /// This has a couple of tricks to make the code faster and more powerful. In
11250 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11251 /// them to the worklist (this significantly speeds up instcombine on code where
11252 /// many instructions are dead or constant). Additionally, if we find a branch
11253 /// whose condition is a known constant, we only visit the reachable successors.
11255 static void AddReachableCodeToWorklist(BasicBlock *BB,
11256 SmallPtrSet<BasicBlock*, 64> &Visited,
11258 const TargetData *TD) {
11259 std::vector<BasicBlock*> Worklist;
11260 Worklist.push_back(BB);
11262 while (!Worklist.empty()) {
11263 BB = Worklist.back();
11264 Worklist.pop_back();
11266 // We have now visited this block! If we've already been here, ignore it.
11267 if (!Visited.insert(BB)) continue;
11269 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11270 Instruction *Inst = BBI++;
11272 // DCE instruction if trivially dead.
11273 if (isInstructionTriviallyDead(Inst)) {
11275 DOUT << "IC: DCE: " << *Inst;
11276 Inst->eraseFromParent();
11280 // ConstantProp instruction if trivially constant.
11281 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11282 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11283 Inst->replaceAllUsesWith(C);
11285 Inst->eraseFromParent();
11289 IC.AddToWorkList(Inst);
11292 // Recursively visit successors. If this is a branch or switch on a
11293 // constant, only visit the reachable successor.
11294 TerminatorInst *TI = BB->getTerminator();
11295 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11296 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11297 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11298 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11299 Worklist.push_back(ReachableBB);
11302 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11303 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11304 // See if this is an explicit destination.
11305 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11306 if (SI->getCaseValue(i) == Cond) {
11307 BasicBlock *ReachableBB = SI->getSuccessor(i);
11308 Worklist.push_back(ReachableBB);
11312 // Otherwise it is the default destination.
11313 Worklist.push_back(SI->getSuccessor(0));
11318 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11319 Worklist.push_back(TI->getSuccessor(i));
11323 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11324 bool Changed = false;
11325 TD = &getAnalysis<TargetData>();
11327 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11328 << F.getNameStr() << "\n");
11331 // Do a depth-first traversal of the function, populate the worklist with
11332 // the reachable instructions. Ignore blocks that are not reachable. Keep
11333 // track of which blocks we visit.
11334 SmallPtrSet<BasicBlock*, 64> Visited;
11335 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11337 // Do a quick scan over the function. If we find any blocks that are
11338 // unreachable, remove any instructions inside of them. This prevents
11339 // the instcombine code from having to deal with some bad special cases.
11340 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11341 if (!Visited.count(BB)) {
11342 Instruction *Term = BB->getTerminator();
11343 while (Term != BB->begin()) { // Remove instrs bottom-up
11344 BasicBlock::iterator I = Term; --I;
11346 DOUT << "IC: DCE: " << *I;
11349 if (!I->use_empty())
11350 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11351 I->eraseFromParent();
11356 while (!Worklist.empty()) {
11357 Instruction *I = RemoveOneFromWorkList();
11358 if (I == 0) continue; // skip null values.
11360 // Check to see if we can DCE the instruction.
11361 if (isInstructionTriviallyDead(I)) {
11362 // Add operands to the worklist.
11363 if (I->getNumOperands() < 4)
11364 AddUsesToWorkList(*I);
11367 DOUT << "IC: DCE: " << *I;
11369 I->eraseFromParent();
11370 RemoveFromWorkList(I);
11374 // Instruction isn't dead, see if we can constant propagate it.
11375 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11376 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11378 // Add operands to the worklist.
11379 AddUsesToWorkList(*I);
11380 ReplaceInstUsesWith(*I, C);
11383 I->eraseFromParent();
11384 RemoveFromWorkList(I);
11388 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
11389 // See if we can constant fold its operands.
11390 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
11391 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
11392 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
11398 // See if we can trivially sink this instruction to a successor basic block.
11399 // FIXME: Remove GetResultInst test when first class support for aggregates
11401 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11402 BasicBlock *BB = I->getParent();
11403 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11404 if (UserParent != BB) {
11405 bool UserIsSuccessor = false;
11406 // See if the user is one of our successors.
11407 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11408 if (*SI == UserParent) {
11409 UserIsSuccessor = true;
11413 // If the user is one of our immediate successors, and if that successor
11414 // only has us as a predecessors (we'd have to split the critical edge
11415 // otherwise), we can keep going.
11416 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11417 next(pred_begin(UserParent)) == pred_end(UserParent))
11418 // Okay, the CFG is simple enough, try to sink this instruction.
11419 Changed |= TryToSinkInstruction(I, UserParent);
11423 // Now that we have an instruction, try combining it to simplify it...
11427 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11428 if (Instruction *Result = visit(*I)) {
11430 // Should we replace the old instruction with a new one?
11432 DOUT << "IC: Old = " << *I
11433 << " New = " << *Result;
11435 // Everything uses the new instruction now.
11436 I->replaceAllUsesWith(Result);
11438 // Push the new instruction and any users onto the worklist.
11439 AddToWorkList(Result);
11440 AddUsersToWorkList(*Result);
11442 // Move the name to the new instruction first.
11443 Result->takeName(I);
11445 // Insert the new instruction into the basic block...
11446 BasicBlock *InstParent = I->getParent();
11447 BasicBlock::iterator InsertPos = I;
11449 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11450 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11453 InstParent->getInstList().insert(InsertPos, Result);
11455 // Make sure that we reprocess all operands now that we reduced their
11457 AddUsesToWorkList(*I);
11459 // Instructions can end up on the worklist more than once. Make sure
11460 // we do not process an instruction that has been deleted.
11461 RemoveFromWorkList(I);
11463 // Erase the old instruction.
11464 InstParent->getInstList().erase(I);
11467 DOUT << "IC: Mod = " << OrigI
11468 << " New = " << *I;
11471 // If the instruction was modified, it's possible that it is now dead.
11472 // if so, remove it.
11473 if (isInstructionTriviallyDead(I)) {
11474 // Make sure we process all operands now that we are reducing their
11476 AddUsesToWorkList(*I);
11478 // Instructions may end up in the worklist more than once. Erase all
11479 // occurrences of this instruction.
11480 RemoveFromWorkList(I);
11481 I->eraseFromParent();
11484 AddUsersToWorkList(*I);
11491 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11493 // Do an explicit clear, this shrinks the map if needed.
11494 WorklistMap.clear();
11499 bool InstCombiner::runOnFunction(Function &F) {
11500 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11502 bool EverMadeChange = false;
11504 // Iterate while there is work to do.
11505 unsigned Iteration = 0;
11506 while (DoOneIteration(F, Iteration++))
11507 EverMadeChange = true;
11508 return EverMadeChange;
11511 FunctionPass *llvm::createInstructionCombiningPass() {
11512 return new InstCombiner();