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())))
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 BasicBlock *BB = Root.getParent();
1660 // Now all of the instructions are in the current basic block, go ahead
1661 // and perform the reassociation.
1662 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1664 // First move the selected RHS to the LHS of the root...
1665 Root.setOperand(0, LHSI->getOperand(1));
1667 // Make what used to be the LHS of the root be the user of the root...
1668 Value *ExtraOperand = TmpLHSI->getOperand(1);
1669 if (&Root == TmpLHSI) {
1670 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1673 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1674 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1675 BasicBlock::iterator ARI = &Root; ++ARI;
1676 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1679 // Now propagate the ExtraOperand down the chain of instructions until we
1681 while (TmpLHSI != LHSI) {
1682 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1683 // Move the instruction to immediately before the chain we are
1684 // constructing to avoid breaking dominance properties.
1685 NextLHSI->moveBefore(ARI);
1688 Value *NextOp = NextLHSI->getOperand(1);
1689 NextLHSI->setOperand(1, ExtraOperand);
1691 ExtraOperand = NextOp;
1694 // Now that the instructions are reassociated, have the functor perform
1695 // the transformation...
1696 return F.apply(Root);
1699 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1706 // AddRHS - Implements: X + X --> X << 1
1709 AddRHS(Value *rhs) : RHS(rhs) {}
1710 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1711 Instruction *apply(BinaryOperator &Add) const {
1712 return BinaryOperator::CreateShl(Add.getOperand(0),
1713 ConstantInt::get(Add.getType(), 1));
1717 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1719 struct AddMaskingAnd {
1721 AddMaskingAnd(Constant *c) : C2(c) {}
1722 bool shouldApply(Value *LHS) const {
1724 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1725 ConstantExpr::getAnd(C1, C2)->isNullValue();
1727 Instruction *apply(BinaryOperator &Add) const {
1728 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1734 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1736 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1737 if (Constant *SOC = dyn_cast<Constant>(SO))
1738 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1740 return IC->InsertNewInstBefore(CastInst::Create(
1741 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1744 // Figure out if the constant is the left or the right argument.
1745 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1746 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1748 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1750 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1751 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1754 Value *Op0 = SO, *Op1 = ConstOperand;
1756 std::swap(Op0, Op1);
1758 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1759 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1760 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1761 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1762 SO->getName()+".cmp");
1764 assert(0 && "Unknown binary instruction type!");
1767 return IC->InsertNewInstBefore(New, I);
1770 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1771 // constant as the other operand, try to fold the binary operator into the
1772 // select arguments. This also works for Cast instructions, which obviously do
1773 // not have a second operand.
1774 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1776 // Don't modify shared select instructions
1777 if (!SI->hasOneUse()) return 0;
1778 Value *TV = SI->getOperand(1);
1779 Value *FV = SI->getOperand(2);
1781 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1782 // Bool selects with constant operands can be folded to logical ops.
1783 if (SI->getType() == Type::Int1Ty) return 0;
1785 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1786 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1788 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1795 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1796 /// node as operand #0, see if we can fold the instruction into the PHI (which
1797 /// is only possible if all operands to the PHI are constants).
1798 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1799 PHINode *PN = cast<PHINode>(I.getOperand(0));
1800 unsigned NumPHIValues = PN->getNumIncomingValues();
1801 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1803 // Check to see if all of the operands of the PHI are constants. If there is
1804 // one non-constant value, remember the BB it is. If there is more than one
1805 // or if *it* is a PHI, bail out.
1806 BasicBlock *NonConstBB = 0;
1807 for (unsigned i = 0; i != NumPHIValues; ++i)
1808 if (!isa<Constant>(PN->getIncomingValue(i))) {
1809 if (NonConstBB) return 0; // More than one non-const value.
1810 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1811 NonConstBB = PN->getIncomingBlock(i);
1813 // If the incoming non-constant value is in I's block, we have an infinite
1815 if (NonConstBB == I.getParent())
1819 // If there is exactly one non-constant value, we can insert a copy of the
1820 // operation in that block. However, if this is a critical edge, we would be
1821 // inserting the computation one some other paths (e.g. inside a loop). Only
1822 // do this if the pred block is unconditionally branching into the phi block.
1824 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1825 if (!BI || !BI->isUnconditional()) return 0;
1828 // Okay, we can do the transformation: create the new PHI node.
1829 PHINode *NewPN = PHINode::Create(I.getType(), "");
1830 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1831 InsertNewInstBefore(NewPN, *PN);
1832 NewPN->takeName(PN);
1834 // Next, add all of the operands to the PHI.
1835 if (I.getNumOperands() == 2) {
1836 Constant *C = cast<Constant>(I.getOperand(1));
1837 for (unsigned i = 0; i != NumPHIValues; ++i) {
1839 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1840 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1841 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1843 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1845 assert(PN->getIncomingBlock(i) == NonConstBB);
1846 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1847 InV = BinaryOperator::Create(BO->getOpcode(),
1848 PN->getIncomingValue(i), C, "phitmp",
1849 NonConstBB->getTerminator());
1850 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1851 InV = CmpInst::Create(CI->getOpcode(),
1853 PN->getIncomingValue(i), C, "phitmp",
1854 NonConstBB->getTerminator());
1856 assert(0 && "Unknown binop!");
1858 AddToWorkList(cast<Instruction>(InV));
1860 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1863 CastInst *CI = cast<CastInst>(&I);
1864 const Type *RetTy = CI->getType();
1865 for (unsigned i = 0; i != NumPHIValues; ++i) {
1867 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1868 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1870 assert(PN->getIncomingBlock(i) == NonConstBB);
1871 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1872 I.getType(), "phitmp",
1873 NonConstBB->getTerminator());
1874 AddToWorkList(cast<Instruction>(InV));
1876 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1879 return ReplaceInstUsesWith(I, NewPN);
1883 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1884 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1885 /// This basically requires proving that the add in the original type would not
1886 /// overflow to change the sign bit or have a carry out.
1887 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1888 // There are different heuristics we can use for this. Here are some simple
1891 // Add has the property that adding any two 2's complement numbers can only
1892 // have one carry bit which can change a sign. As such, if LHS and RHS each
1893 // have at least two sign bits, we know that the addition of the two values will
1894 // sign extend fine.
1895 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1899 // If one of the operands only has one non-zero bit, and if the other operand
1900 // has a known-zero bit in a more significant place than it (not including the
1901 // sign bit) the ripple may go up to and fill the zero, but won't change the
1902 // sign. For example, (X & ~4) + 1.
1910 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1911 bool Changed = SimplifyCommutative(I);
1912 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1914 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1915 // X + undef -> undef
1916 if (isa<UndefValue>(RHS))
1917 return ReplaceInstUsesWith(I, RHS);
1920 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1921 if (RHSC->isNullValue())
1922 return ReplaceInstUsesWith(I, LHS);
1923 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1924 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1925 (I.getType())->getValueAPF()))
1926 return ReplaceInstUsesWith(I, LHS);
1929 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1930 // X + (signbit) --> X ^ signbit
1931 const APInt& Val = CI->getValue();
1932 uint32_t BitWidth = Val.getBitWidth();
1933 if (Val == APInt::getSignBit(BitWidth))
1934 return BinaryOperator::CreateXor(LHS, RHS);
1936 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1937 // (X & 254)+1 -> (X&254)|1
1938 if (!isa<VectorType>(I.getType())) {
1939 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1940 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1941 KnownZero, KnownOne))
1946 if (isa<PHINode>(LHS))
1947 if (Instruction *NV = FoldOpIntoPhi(I))
1950 ConstantInt *XorRHS = 0;
1952 if (isa<ConstantInt>(RHSC) &&
1953 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1954 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1955 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1957 uint32_t Size = TySizeBits / 2;
1958 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1959 APInt CFF80Val(-C0080Val);
1961 if (TySizeBits > Size) {
1962 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1963 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1964 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1965 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1966 // This is a sign extend if the top bits are known zero.
1967 if (!MaskedValueIsZero(XorLHS,
1968 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1969 Size = 0; // Not a sign ext, but can't be any others either.
1974 C0080Val = APIntOps::lshr(C0080Val, Size);
1975 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1976 } while (Size >= 1);
1978 // FIXME: This shouldn't be necessary. When the backends can handle types
1979 // with funny bit widths then this switch statement should be removed. It
1980 // is just here to get the size of the "middle" type back up to something
1981 // that the back ends can handle.
1982 const Type *MiddleType = 0;
1985 case 32: MiddleType = Type::Int32Ty; break;
1986 case 16: MiddleType = Type::Int16Ty; break;
1987 case 8: MiddleType = Type::Int8Ty; break;
1990 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1991 InsertNewInstBefore(NewTrunc, I);
1992 return new SExtInst(NewTrunc, I.getType(), I.getName());
1997 if (I.getType() == Type::Int1Ty)
1998 return BinaryOperator::CreateXor(LHS, RHS);
2001 if (I.getType()->isInteger()) {
2002 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2004 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2005 if (RHSI->getOpcode() == Instruction::Sub)
2006 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2007 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2009 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2010 if (LHSI->getOpcode() == Instruction::Sub)
2011 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2012 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2017 // -A + -B --> -(A + B)
2018 if (Value *LHSV = dyn_castNegVal(LHS)) {
2019 if (LHS->getType()->isIntOrIntVector()) {
2020 if (Value *RHSV = dyn_castNegVal(RHS)) {
2021 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2022 InsertNewInstBefore(NewAdd, I);
2023 return BinaryOperator::CreateNeg(NewAdd);
2027 return BinaryOperator::CreateSub(RHS, LHSV);
2031 if (!isa<Constant>(RHS))
2032 if (Value *V = dyn_castNegVal(RHS))
2033 return BinaryOperator::CreateSub(LHS, V);
2037 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2038 if (X == RHS) // X*C + X --> X * (C+1)
2039 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2041 // X*C1 + X*C2 --> X * (C1+C2)
2043 if (X == dyn_castFoldableMul(RHS, C1))
2044 return BinaryOperator::CreateMul(X, Add(C1, C2));
2047 // X + X*C --> X * (C+1)
2048 if (dyn_castFoldableMul(RHS, C2) == LHS)
2049 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2051 // X + ~X --> -1 since ~X = -X-1
2052 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2053 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2056 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2057 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2058 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2061 // A+B --> A|B iff A and B have no bits set in common.
2062 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2063 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2064 APInt LHSKnownOne(IT->getBitWidth(), 0);
2065 APInt LHSKnownZero(IT->getBitWidth(), 0);
2066 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2067 if (LHSKnownZero != 0) {
2068 APInt RHSKnownOne(IT->getBitWidth(), 0);
2069 APInt RHSKnownZero(IT->getBitWidth(), 0);
2070 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2072 // No bits in common -> bitwise or.
2073 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2074 return BinaryOperator::CreateOr(LHS, RHS);
2078 // W*X + Y*Z --> W * (X+Z) iff W == Y
2079 if (I.getType()->isIntOrIntVector()) {
2080 Value *W, *X, *Y, *Z;
2081 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2082 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2086 } else if (Y == X) {
2088 } else if (X == Z) {
2095 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2096 LHS->getName()), I);
2097 return BinaryOperator::CreateMul(W, NewAdd);
2102 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2104 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2105 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2107 // (X & FF00) + xx00 -> (X+xx00) & FF00
2108 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2109 Constant *Anded = And(CRHS, C2);
2110 if (Anded == CRHS) {
2111 // See if all bits from the first bit set in the Add RHS up are included
2112 // in the mask. First, get the rightmost bit.
2113 const APInt& AddRHSV = CRHS->getValue();
2115 // Form a mask of all bits from the lowest bit added through the top.
2116 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2118 // See if the and mask includes all of these bits.
2119 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2121 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2122 // Okay, the xform is safe. Insert the new add pronto.
2123 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2124 LHS->getName()), I);
2125 return BinaryOperator::CreateAnd(NewAdd, C2);
2130 // Try to fold constant add into select arguments.
2131 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2132 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2136 // add (cast *A to intptrtype) B ->
2137 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2139 CastInst *CI = dyn_cast<CastInst>(LHS);
2142 CI = dyn_cast<CastInst>(RHS);
2145 if (CI && CI->getType()->isSized() &&
2146 (CI->getType()->getPrimitiveSizeInBits() ==
2147 TD->getIntPtrType()->getPrimitiveSizeInBits())
2148 && isa<PointerType>(CI->getOperand(0)->getType())) {
2150 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2151 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2152 PointerType::get(Type::Int8Ty, AS), I);
2153 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2154 return new PtrToIntInst(I2, CI->getType());
2158 // add (select X 0 (sub n A)) A --> select X A n
2160 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2163 SI = dyn_cast<SelectInst>(RHS);
2166 if (SI && SI->hasOneUse()) {
2167 Value *TV = SI->getTrueValue();
2168 Value *FV = SI->getFalseValue();
2171 // Can we fold the add into the argument of the select?
2172 // We check both true and false select arguments for a matching subtract.
2173 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2174 A == Other) // Fold the add into the true select value.
2175 return SelectInst::Create(SI->getCondition(), N, A);
2176 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2177 A == Other) // Fold the add into the false select value.
2178 return SelectInst::Create(SI->getCondition(), A, N);
2182 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2183 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2184 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2185 return ReplaceInstUsesWith(I, LHS);
2187 // Check for (add (sext x), y), see if we can merge this into an
2188 // integer add followed by a sext.
2189 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2190 // (add (sext x), cst) --> (sext (add x, cst'))
2191 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2193 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2194 if (LHSConv->hasOneUse() &&
2195 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2196 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2197 // Insert the new, smaller add.
2198 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2200 InsertNewInstBefore(NewAdd, I);
2201 return new SExtInst(NewAdd, I.getType());
2205 // (add (sext x), (sext y)) --> (sext (add int x, y))
2206 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2207 // Only do this if x/y have the same type, if at last one of them has a
2208 // single use (so we don't increase the number of sexts), and if the
2209 // integer add will not overflow.
2210 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2211 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2212 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2213 RHSConv->getOperand(0))) {
2214 // Insert the new integer add.
2215 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2216 RHSConv->getOperand(0),
2218 InsertNewInstBefore(NewAdd, I);
2219 return new SExtInst(NewAdd, I.getType());
2224 // Check for (add double (sitofp x), y), see if we can merge this into an
2225 // integer add followed by a promotion.
2226 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2227 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2228 // ... if the constant fits in the integer value. This is useful for things
2229 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2230 // requires a constant pool load, and generally allows the add to be better
2232 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2234 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2235 if (LHSConv->hasOneUse() &&
2236 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2237 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2238 // Insert the new integer add.
2239 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2241 InsertNewInstBefore(NewAdd, I);
2242 return new SIToFPInst(NewAdd, I.getType());
2246 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2247 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2248 // Only do this if x/y have the same type, if at last one of them has a
2249 // single use (so we don't increase the number of int->fp conversions),
2250 // and if the integer add will not overflow.
2251 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2252 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2253 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2254 RHSConv->getOperand(0))) {
2255 // Insert the new integer add.
2256 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2257 RHSConv->getOperand(0),
2259 InsertNewInstBefore(NewAdd, I);
2260 return new SIToFPInst(NewAdd, I.getType());
2265 return Changed ? &I : 0;
2268 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2269 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2271 if (Op0 == Op1) // sub X, X -> 0
2272 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2274 // If this is a 'B = x-(-A)', change to B = x+A...
2275 if (Value *V = dyn_castNegVal(Op1))
2276 return BinaryOperator::CreateAdd(Op0, V);
2278 if (isa<UndefValue>(Op0))
2279 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2280 if (isa<UndefValue>(Op1))
2281 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2283 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2284 // Replace (-1 - A) with (~A)...
2285 if (C->isAllOnesValue())
2286 return BinaryOperator::CreateNot(Op1);
2288 // C - ~X == X + (1+C)
2290 if (match(Op1, m_Not(m_Value(X))))
2291 return BinaryOperator::CreateAdd(X, AddOne(C));
2293 // -(X >>u 31) -> (X >>s 31)
2294 // -(X >>s 31) -> (X >>u 31)
2296 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2297 if (SI->getOpcode() == Instruction::LShr) {
2298 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2299 // Check to see if we are shifting out everything but the sign bit.
2300 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2301 SI->getType()->getPrimitiveSizeInBits()-1) {
2302 // Ok, the transformation is safe. Insert AShr.
2303 return BinaryOperator::Create(Instruction::AShr,
2304 SI->getOperand(0), CU, SI->getName());
2308 else if (SI->getOpcode() == Instruction::AShr) {
2309 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2310 // Check to see if we are shifting out everything but the sign bit.
2311 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2312 SI->getType()->getPrimitiveSizeInBits()-1) {
2313 // Ok, the transformation is safe. Insert LShr.
2314 return BinaryOperator::CreateLShr(
2315 SI->getOperand(0), CU, SI->getName());
2322 // Try to fold constant sub into select arguments.
2323 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2324 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2327 if (isa<PHINode>(Op0))
2328 if (Instruction *NV = FoldOpIntoPhi(I))
2332 if (I.getType() == Type::Int1Ty)
2333 return BinaryOperator::CreateXor(Op0, Op1);
2335 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2336 if (Op1I->getOpcode() == Instruction::Add &&
2337 !Op0->getType()->isFPOrFPVector()) {
2338 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2339 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2340 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2341 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2342 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2343 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2344 // C1-(X+C2) --> (C1-C2)-X
2345 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2346 Op1I->getOperand(0));
2350 if (Op1I->hasOneUse()) {
2351 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2352 // is not used by anyone else...
2354 if (Op1I->getOpcode() == Instruction::Sub &&
2355 !Op1I->getType()->isFPOrFPVector()) {
2356 // Swap the two operands of the subexpr...
2357 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2358 Op1I->setOperand(0, IIOp1);
2359 Op1I->setOperand(1, IIOp0);
2361 // Create the new top level add instruction...
2362 return BinaryOperator::CreateAdd(Op0, Op1);
2365 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2367 if (Op1I->getOpcode() == Instruction::And &&
2368 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2369 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2372 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2373 return BinaryOperator::CreateAnd(Op0, NewNot);
2376 // 0 - (X sdiv C) -> (X sdiv -C)
2377 if (Op1I->getOpcode() == Instruction::SDiv)
2378 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2380 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2381 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2382 ConstantExpr::getNeg(DivRHS));
2384 // X - X*C --> X * (1-C)
2385 ConstantInt *C2 = 0;
2386 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2387 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2388 return BinaryOperator::CreateMul(Op0, CP1);
2391 // X - ((X / Y) * Y) --> X % Y
2392 if (Op1I->getOpcode() == Instruction::Mul)
2393 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2394 if (Op0 == I->getOperand(0) &&
2395 Op1I->getOperand(1) == I->getOperand(1)) {
2396 if (I->getOpcode() == Instruction::SDiv)
2397 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2398 if (I->getOpcode() == Instruction::UDiv)
2399 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2404 if (!Op0->getType()->isFPOrFPVector())
2405 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2406 if (Op0I->getOpcode() == Instruction::Add) {
2407 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2408 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2409 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2410 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2411 } else if (Op0I->getOpcode() == Instruction::Sub) {
2412 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2413 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2418 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2419 if (X == Op1) // X*C - X --> X * (C-1)
2420 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2422 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2423 if (X == dyn_castFoldableMul(Op1, C2))
2424 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2429 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2430 /// comparison only checks the sign bit. If it only checks the sign bit, set
2431 /// TrueIfSigned if the result of the comparison is true when the input value is
2433 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2434 bool &TrueIfSigned) {
2436 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2437 TrueIfSigned = true;
2438 return RHS->isZero();
2439 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2440 TrueIfSigned = true;
2441 return RHS->isAllOnesValue();
2442 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2443 TrueIfSigned = false;
2444 return RHS->isAllOnesValue();
2445 case ICmpInst::ICMP_UGT:
2446 // True if LHS u> RHS and RHS == high-bit-mask - 1
2447 TrueIfSigned = true;
2448 return RHS->getValue() ==
2449 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2450 case ICmpInst::ICMP_UGE:
2451 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2452 TrueIfSigned = true;
2453 return RHS->getValue().isSignBit();
2459 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2460 bool Changed = SimplifyCommutative(I);
2461 Value *Op0 = I.getOperand(0);
2463 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2464 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2466 // Simplify mul instructions with a constant RHS...
2467 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2468 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2470 // ((X << C1)*C2) == (X * (C2 << C1))
2471 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2472 if (SI->getOpcode() == Instruction::Shl)
2473 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2474 return BinaryOperator::CreateMul(SI->getOperand(0),
2475 ConstantExpr::getShl(CI, ShOp));
2478 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2479 if (CI->equalsInt(1)) // X * 1 == X
2480 return ReplaceInstUsesWith(I, Op0);
2481 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2482 return BinaryOperator::CreateNeg(Op0, I.getName());
2484 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2485 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2486 return BinaryOperator::CreateShl(Op0,
2487 ConstantInt::get(Op0->getType(), Val.logBase2()));
2489 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2490 if (Op1F->isNullValue())
2491 return ReplaceInstUsesWith(I, Op1);
2493 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2494 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2495 // We need a better interface for long double here.
2496 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2497 if (Op1F->isExactlyValue(1.0))
2498 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2501 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2502 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2503 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2504 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2505 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2507 InsertNewInstBefore(Add, I);
2508 Value *C1C2 = ConstantExpr::getMul(Op1,
2509 cast<Constant>(Op0I->getOperand(1)));
2510 return BinaryOperator::CreateAdd(Add, C1C2);
2514 // Try to fold constant mul into select arguments.
2515 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2516 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2519 if (isa<PHINode>(Op0))
2520 if (Instruction *NV = FoldOpIntoPhi(I))
2524 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2525 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2526 return BinaryOperator::CreateMul(Op0v, Op1v);
2528 if (I.getType() == Type::Int1Ty)
2529 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2531 // If one of the operands of the multiply is a cast from a boolean value, then
2532 // we know the bool is either zero or one, so this is a 'masking' multiply.
2533 // See if we can simplify things based on how the boolean was originally
2535 CastInst *BoolCast = 0;
2536 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2537 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2540 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2541 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2544 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2545 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2546 const Type *SCOpTy = SCIOp0->getType();
2549 // If the icmp is true iff the sign bit of X is set, then convert this
2550 // multiply into a shift/and combination.
2551 if (isa<ConstantInt>(SCIOp1) &&
2552 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2554 // Shift the X value right to turn it into "all signbits".
2555 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2556 SCOpTy->getPrimitiveSizeInBits()-1);
2558 InsertNewInstBefore(
2559 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2560 BoolCast->getOperand(0)->getName()+
2563 // If the multiply type is not the same as the source type, sign extend
2564 // or truncate to the multiply type.
2565 if (I.getType() != V->getType()) {
2566 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2567 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2568 Instruction::CastOps opcode =
2569 (SrcBits == DstBits ? Instruction::BitCast :
2570 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2571 V = InsertCastBefore(opcode, V, I.getType(), I);
2574 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2575 return BinaryOperator::CreateAnd(V, OtherOp);
2580 return Changed ? &I : 0;
2583 /// This function implements the transforms on div instructions that work
2584 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2585 /// used by the visitors to those instructions.
2586 /// @brief Transforms common to all three div instructions
2587 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2588 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2590 // undef / X -> 0 for integer.
2591 // undef / X -> undef for FP (the undef could be a snan).
2592 if (isa<UndefValue>(Op0)) {
2593 if (Op0->getType()->isFPOrFPVector())
2594 return ReplaceInstUsesWith(I, Op0);
2595 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2598 // X / undef -> undef
2599 if (isa<UndefValue>(Op1))
2600 return ReplaceInstUsesWith(I, Op1);
2602 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2603 // This does not apply for fdiv.
2604 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2605 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2606 // the same basic block, then we replace the select with Y, and the
2607 // condition of the select with false (if the cond value is in the same BB).
2608 // If the select has uses other than the div, this allows them to be
2609 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2610 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2611 if (ST->isNullValue()) {
2612 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2613 if (CondI && CondI->getParent() == I.getParent())
2614 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2615 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2616 I.setOperand(1, SI->getOperand(2));
2618 UpdateValueUsesWith(SI, SI->getOperand(2));
2622 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2623 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2624 if (ST->isNullValue()) {
2625 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2626 if (CondI && CondI->getParent() == I.getParent())
2627 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2628 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2629 I.setOperand(1, SI->getOperand(1));
2631 UpdateValueUsesWith(SI, SI->getOperand(1));
2639 /// This function implements the transforms common to both integer division
2640 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2641 /// division instructions.
2642 /// @brief Common integer divide transforms
2643 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2644 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2646 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2648 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2649 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2650 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2651 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2654 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2655 return ReplaceInstUsesWith(I, CI);
2658 if (Instruction *Common = commonDivTransforms(I))
2661 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2663 if (RHS->equalsInt(1))
2664 return ReplaceInstUsesWith(I, Op0);
2666 // (X / C1) / C2 -> X / (C1*C2)
2667 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2668 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2669 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2670 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2671 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2673 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2674 Multiply(RHS, LHSRHS));
2677 if (!RHS->isZero()) { // avoid X udiv 0
2678 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2679 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2681 if (isa<PHINode>(Op0))
2682 if (Instruction *NV = FoldOpIntoPhi(I))
2687 // 0 / X == 0, we don't need to preserve faults!
2688 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2689 if (LHS->equalsInt(0))
2690 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2692 // It can't be division by zero, hence it must be division by one.
2693 if (I.getType() == Type::Int1Ty)
2694 return ReplaceInstUsesWith(I, Op0);
2699 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2700 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2702 // Handle the integer div common cases
2703 if (Instruction *Common = commonIDivTransforms(I))
2706 // X udiv C^2 -> X >> C
2707 // Check to see if this is an unsigned division with an exact power of 2,
2708 // if so, convert to a right shift.
2709 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2710 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2711 return BinaryOperator::CreateLShr(Op0,
2712 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2715 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2716 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2717 if (RHSI->getOpcode() == Instruction::Shl &&
2718 isa<ConstantInt>(RHSI->getOperand(0))) {
2719 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2720 if (C1.isPowerOf2()) {
2721 Value *N = RHSI->getOperand(1);
2722 const Type *NTy = N->getType();
2723 if (uint32_t C2 = C1.logBase2()) {
2724 Constant *C2V = ConstantInt::get(NTy, C2);
2725 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2727 return BinaryOperator::CreateLShr(Op0, N);
2732 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2733 // where C1&C2 are powers of two.
2734 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2735 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2736 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2737 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2738 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2739 // Compute the shift amounts
2740 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2741 // Construct the "on true" case of the select
2742 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2743 Instruction *TSI = BinaryOperator::CreateLShr(
2744 Op0, TC, SI->getName()+".t");
2745 TSI = InsertNewInstBefore(TSI, I);
2747 // Construct the "on false" case of the select
2748 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2749 Instruction *FSI = BinaryOperator::CreateLShr(
2750 Op0, FC, SI->getName()+".f");
2751 FSI = InsertNewInstBefore(FSI, I);
2753 // construct the select instruction and return it.
2754 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2760 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2761 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2763 // Handle the integer div common cases
2764 if (Instruction *Common = commonIDivTransforms(I))
2767 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2769 if (RHS->isAllOnesValue())
2770 return BinaryOperator::CreateNeg(Op0);
2773 if (Value *LHSNeg = dyn_castNegVal(Op0))
2774 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2777 // If the sign bits of both operands are zero (i.e. we can prove they are
2778 // unsigned inputs), turn this into a udiv.
2779 if (I.getType()->isInteger()) {
2780 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2781 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2782 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2783 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2790 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2791 return commonDivTransforms(I);
2794 /// This function implements the transforms on rem instructions that work
2795 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2796 /// is used by the visitors to those instructions.
2797 /// @brief Transforms common to all three rem instructions
2798 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2799 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2801 // 0 % X == 0 for integer, we don't need to preserve faults!
2802 if (Constant *LHS = dyn_cast<Constant>(Op0))
2803 if (LHS->isNullValue())
2804 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2806 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2807 if (I.getType()->isFPOrFPVector())
2808 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2809 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2811 if (isa<UndefValue>(Op1))
2812 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2814 // Handle cases involving: rem X, (select Cond, Y, Z)
2815 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2816 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2817 // the same basic block, then we replace the select with Y, and the
2818 // condition of the select with false (if the cond value is in the same
2819 // BB). If the select has uses other than the div, this allows them to be
2821 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2822 if (ST->isNullValue()) {
2823 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2824 if (CondI && CondI->getParent() == I.getParent())
2825 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2826 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2827 I.setOperand(1, SI->getOperand(2));
2829 UpdateValueUsesWith(SI, SI->getOperand(2));
2832 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2833 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2834 if (ST->isNullValue()) {
2835 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2836 if (CondI && CondI->getParent() == I.getParent())
2837 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2838 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2839 I.setOperand(1, SI->getOperand(1));
2841 UpdateValueUsesWith(SI, SI->getOperand(1));
2849 /// This function implements the transforms common to both integer remainder
2850 /// instructions (urem and srem). It is called by the visitors to those integer
2851 /// remainder instructions.
2852 /// @brief Common integer remainder transforms
2853 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2854 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2856 if (Instruction *common = commonRemTransforms(I))
2859 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2860 // X % 0 == undef, we don't need to preserve faults!
2861 if (RHS->equalsInt(0))
2862 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2864 if (RHS->equalsInt(1)) // X % 1 == 0
2865 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2867 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2868 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2869 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2871 } else if (isa<PHINode>(Op0I)) {
2872 if (Instruction *NV = FoldOpIntoPhi(I))
2876 // See if we can fold away this rem instruction.
2877 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2878 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2879 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2880 KnownZero, KnownOne))
2888 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2889 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2891 if (Instruction *common = commonIRemTransforms(I))
2894 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2895 // X urem C^2 -> X and C
2896 // Check to see if this is an unsigned remainder with an exact power of 2,
2897 // if so, convert to a bitwise and.
2898 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2899 if (C->getValue().isPowerOf2())
2900 return BinaryOperator::CreateAnd(Op0, SubOne(C));
2903 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2904 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2905 if (RHSI->getOpcode() == Instruction::Shl &&
2906 isa<ConstantInt>(RHSI->getOperand(0))) {
2907 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2908 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2909 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
2911 return BinaryOperator::CreateAnd(Op0, Add);
2916 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2917 // where C1&C2 are powers of two.
2918 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2919 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2920 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2921 // STO == 0 and SFO == 0 handled above.
2922 if ((STO->getValue().isPowerOf2()) &&
2923 (SFO->getValue().isPowerOf2())) {
2924 Value *TrueAnd = InsertNewInstBefore(
2925 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2926 Value *FalseAnd = InsertNewInstBefore(
2927 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2928 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
2936 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2937 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2939 // Handle the integer rem common cases
2940 if (Instruction *common = commonIRemTransforms(I))
2943 if (Value *RHSNeg = dyn_castNegVal(Op1))
2944 if (!isa<ConstantInt>(RHSNeg) ||
2945 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2947 AddUsesToWorkList(I);
2948 I.setOperand(1, RHSNeg);
2952 // If the sign bits of both operands are zero (i.e. we can prove they are
2953 // unsigned inputs), turn this into a urem.
2954 if (I.getType()->isInteger()) {
2955 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2956 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2957 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2958 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
2965 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2966 return commonRemTransforms(I);
2969 // isMaxValueMinusOne - return true if this is Max-1
2970 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2971 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2973 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2974 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2977 // isMinValuePlusOne - return true if this is Min+1
2978 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2980 return C->getValue() == 1; // unsigned
2982 // Calculate 1111111111000000000000
2983 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2984 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2987 // isOneBitSet - Return true if there is exactly one bit set in the specified
2989 static bool isOneBitSet(const ConstantInt *CI) {
2990 return CI->getValue().isPowerOf2();
2993 // isHighOnes - Return true if the constant is of the form 1+0+.
2994 // This is the same as lowones(~X).
2995 static bool isHighOnes(const ConstantInt *CI) {
2996 return (~CI->getValue() + 1).isPowerOf2();
2999 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3000 /// are carefully arranged to allow folding of expressions such as:
3002 /// (A < B) | (A > B) --> (A != B)
3004 /// Note that this is only valid if the first and second predicates have the
3005 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3007 /// Three bits are used to represent the condition, as follows:
3012 /// <=> Value Definition
3013 /// 000 0 Always false
3020 /// 111 7 Always true
3022 static unsigned getICmpCode(const ICmpInst *ICI) {
3023 switch (ICI->getPredicate()) {
3025 case ICmpInst::ICMP_UGT: return 1; // 001
3026 case ICmpInst::ICMP_SGT: return 1; // 001
3027 case ICmpInst::ICMP_EQ: return 2; // 010
3028 case ICmpInst::ICMP_UGE: return 3; // 011
3029 case ICmpInst::ICMP_SGE: return 3; // 011
3030 case ICmpInst::ICMP_ULT: return 4; // 100
3031 case ICmpInst::ICMP_SLT: return 4; // 100
3032 case ICmpInst::ICMP_NE: return 5; // 101
3033 case ICmpInst::ICMP_ULE: return 6; // 110
3034 case ICmpInst::ICMP_SLE: return 6; // 110
3037 assert(0 && "Invalid ICmp predicate!");
3042 /// getICmpValue - This is the complement of getICmpCode, which turns an
3043 /// opcode and two operands into either a constant true or false, or a brand
3044 /// new ICmp instruction. The sign is passed in to determine which kind
3045 /// of predicate to use in new icmp instructions.
3046 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3048 default: assert(0 && "Illegal ICmp code!");
3049 case 0: return ConstantInt::getFalse();
3052 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3054 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3055 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3058 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3060 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3063 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3065 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3066 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3069 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3071 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3072 case 7: return ConstantInt::getTrue();
3076 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3077 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3078 (ICmpInst::isSignedPredicate(p1) &&
3079 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3080 (ICmpInst::isSignedPredicate(p2) &&
3081 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3085 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3086 struct FoldICmpLogical {
3089 ICmpInst::Predicate pred;
3090 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3091 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3092 pred(ICI->getPredicate()) {}
3093 bool shouldApply(Value *V) const {
3094 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3095 if (PredicatesFoldable(pred, ICI->getPredicate()))
3096 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3097 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3100 Instruction *apply(Instruction &Log) const {
3101 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3102 if (ICI->getOperand(0) != LHS) {
3103 assert(ICI->getOperand(1) == LHS);
3104 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3107 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3108 unsigned LHSCode = getICmpCode(ICI);
3109 unsigned RHSCode = getICmpCode(RHSICI);
3111 switch (Log.getOpcode()) {
3112 case Instruction::And: Code = LHSCode & RHSCode; break;
3113 case Instruction::Or: Code = LHSCode | RHSCode; break;
3114 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3115 default: assert(0 && "Illegal logical opcode!"); return 0;
3118 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3119 ICmpInst::isSignedPredicate(ICI->getPredicate());
3121 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3122 if (Instruction *I = dyn_cast<Instruction>(RV))
3124 // Otherwise, it's a constant boolean value...
3125 return IC.ReplaceInstUsesWith(Log, RV);
3128 } // end anonymous namespace
3130 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3131 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3132 // guaranteed to be a binary operator.
3133 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3135 ConstantInt *AndRHS,
3136 BinaryOperator &TheAnd) {
3137 Value *X = Op->getOperand(0);
3138 Constant *Together = 0;
3140 Together = And(AndRHS, OpRHS);
3142 switch (Op->getOpcode()) {
3143 case Instruction::Xor:
3144 if (Op->hasOneUse()) {
3145 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3146 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3147 InsertNewInstBefore(And, TheAnd);
3149 return BinaryOperator::CreateXor(And, Together);
3152 case Instruction::Or:
3153 if (Together == AndRHS) // (X | C) & C --> C
3154 return ReplaceInstUsesWith(TheAnd, AndRHS);
3156 if (Op->hasOneUse() && Together != OpRHS) {
3157 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3158 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3159 InsertNewInstBefore(Or, TheAnd);
3161 return BinaryOperator::CreateAnd(Or, AndRHS);
3164 case Instruction::Add:
3165 if (Op->hasOneUse()) {
3166 // Adding a one to a single bit bit-field should be turned into an XOR
3167 // of the bit. First thing to check is to see if this AND is with a
3168 // single bit constant.
3169 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3171 // If there is only one bit set...
3172 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3173 // Ok, at this point, we know that we are masking the result of the
3174 // ADD down to exactly one bit. If the constant we are adding has
3175 // no bits set below this bit, then we can eliminate the ADD.
3176 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3178 // Check to see if any bits below the one bit set in AndRHSV are set.
3179 if ((AddRHS & (AndRHSV-1)) == 0) {
3180 // If not, the only thing that can effect the output of the AND is
3181 // the bit specified by AndRHSV. If that bit is set, the effect of
3182 // the XOR is to toggle the bit. If it is clear, then the ADD has
3184 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3185 TheAnd.setOperand(0, X);
3188 // Pull the XOR out of the AND.
3189 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3190 InsertNewInstBefore(NewAnd, TheAnd);
3191 NewAnd->takeName(Op);
3192 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3199 case Instruction::Shl: {
3200 // We know that the AND will not produce any of the bits shifted in, so if
3201 // the anded constant includes them, clear them now!
3203 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3204 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3205 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3206 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3208 if (CI->getValue() == ShlMask) {
3209 // Masking out bits that the shift already masks
3210 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3211 } else if (CI != AndRHS) { // Reducing bits set in and.
3212 TheAnd.setOperand(1, CI);
3217 case Instruction::LShr:
3219 // We know that the AND will not produce any of the bits shifted in, so if
3220 // the anded constant includes them, clear them now! This only applies to
3221 // unsigned shifts, because a signed shr may bring in set bits!
3223 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3224 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3225 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3226 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3228 if (CI->getValue() == ShrMask) {
3229 // Masking out bits that the shift already masks.
3230 return ReplaceInstUsesWith(TheAnd, Op);
3231 } else if (CI != AndRHS) {
3232 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3237 case Instruction::AShr:
3239 // See if this is shifting in some sign extension, then masking it out
3241 if (Op->hasOneUse()) {
3242 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3243 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3244 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3245 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3246 if (C == AndRHS) { // Masking out bits shifted in.
3247 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3248 // Make the argument unsigned.
3249 Value *ShVal = Op->getOperand(0);
3250 ShVal = InsertNewInstBefore(
3251 BinaryOperator::CreateLShr(ShVal, OpRHS,
3252 Op->getName()), TheAnd);
3253 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3262 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3263 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3264 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3265 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3266 /// insert new instructions.
3267 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3268 bool isSigned, bool Inside,
3270 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3271 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3272 "Lo is not <= Hi in range emission code!");
3275 if (Lo == Hi) // Trivially false.
3276 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3278 // V >= Min && V < Hi --> V < Hi
3279 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3280 ICmpInst::Predicate pred = (isSigned ?
3281 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3282 return new ICmpInst(pred, V, Hi);
3285 // Emit V-Lo <u Hi-Lo
3286 Constant *NegLo = ConstantExpr::getNeg(Lo);
3287 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3288 InsertNewInstBefore(Add, IB);
3289 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3290 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3293 if (Lo == Hi) // Trivially true.
3294 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3296 // V < Min || V >= Hi -> V > Hi-1
3297 Hi = SubOne(cast<ConstantInt>(Hi));
3298 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3299 ICmpInst::Predicate pred = (isSigned ?
3300 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3301 return new ICmpInst(pred, V, Hi);
3304 // Emit V-Lo >u Hi-1-Lo
3305 // Note that Hi has already had one subtracted from it, above.
3306 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3307 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3308 InsertNewInstBefore(Add, IB);
3309 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3310 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3313 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3314 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3315 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3316 // not, since all 1s are not contiguous.
3317 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3318 const APInt& V = Val->getValue();
3319 uint32_t BitWidth = Val->getType()->getBitWidth();
3320 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3322 // look for the first zero bit after the run of ones
3323 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3324 // look for the first non-zero bit
3325 ME = V.getActiveBits();
3329 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3330 /// where isSub determines whether the operator is a sub. If we can fold one of
3331 /// the following xforms:
3333 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3334 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3335 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3337 /// return (A +/- B).
3339 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3340 ConstantInt *Mask, bool isSub,
3342 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3343 if (!LHSI || LHSI->getNumOperands() != 2 ||
3344 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3346 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3348 switch (LHSI->getOpcode()) {
3350 case Instruction::And:
3351 if (And(N, Mask) == Mask) {
3352 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3353 if ((Mask->getValue().countLeadingZeros() +
3354 Mask->getValue().countPopulation()) ==
3355 Mask->getValue().getBitWidth())
3358 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3359 // part, we don't need any explicit masks to take them out of A. If that
3360 // is all N is, ignore it.
3361 uint32_t MB = 0, ME = 0;
3362 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3363 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3364 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3365 if (MaskedValueIsZero(RHS, Mask))
3370 case Instruction::Or:
3371 case Instruction::Xor:
3372 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3373 if ((Mask->getValue().countLeadingZeros() +
3374 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3375 && And(N, Mask)->isZero())
3382 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3384 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3385 return InsertNewInstBefore(New, I);
3388 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3389 bool Changed = SimplifyCommutative(I);
3390 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3392 if (isa<UndefValue>(Op1)) // X & undef -> 0
3393 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3397 return ReplaceInstUsesWith(I, Op1);
3399 // See if we can simplify any instructions used by the instruction whose sole
3400 // purpose is to compute bits we don't care about.
3401 if (!isa<VectorType>(I.getType())) {
3402 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3403 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3404 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3405 KnownZero, KnownOne))
3408 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3409 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3410 return ReplaceInstUsesWith(I, I.getOperand(0));
3411 } else if (isa<ConstantAggregateZero>(Op1)) {
3412 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3416 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3417 const APInt& AndRHSMask = AndRHS->getValue();
3418 APInt NotAndRHS(~AndRHSMask);
3420 // Optimize a variety of ((val OP C1) & C2) combinations...
3421 if (isa<BinaryOperator>(Op0)) {
3422 Instruction *Op0I = cast<Instruction>(Op0);
3423 Value *Op0LHS = Op0I->getOperand(0);
3424 Value *Op0RHS = Op0I->getOperand(1);
3425 switch (Op0I->getOpcode()) {
3426 case Instruction::Xor:
3427 case Instruction::Or:
3428 // If the mask is only needed on one incoming arm, push it up.
3429 if (Op0I->hasOneUse()) {
3430 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3431 // Not masking anything out for the LHS, move to RHS.
3432 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3433 Op0RHS->getName()+".masked");
3434 InsertNewInstBefore(NewRHS, I);
3435 return BinaryOperator::Create(
3436 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3438 if (!isa<Constant>(Op0RHS) &&
3439 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3440 // Not masking anything out for the RHS, move to LHS.
3441 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3442 Op0LHS->getName()+".masked");
3443 InsertNewInstBefore(NewLHS, I);
3444 return BinaryOperator::Create(
3445 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3450 case Instruction::Add:
3451 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3452 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3453 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3454 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3455 return BinaryOperator::CreateAnd(V, AndRHS);
3456 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3457 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3460 case Instruction::Sub:
3461 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3462 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3463 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3464 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3465 return BinaryOperator::CreateAnd(V, AndRHS);
3469 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3470 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3472 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3473 // If this is an integer truncation or change from signed-to-unsigned, and
3474 // if the source is an and/or with immediate, transform it. This
3475 // frequently occurs for bitfield accesses.
3476 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3477 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3478 CastOp->getNumOperands() == 2)
3479 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3480 if (CastOp->getOpcode() == Instruction::And) {
3481 // Change: and (cast (and X, C1) to T), C2
3482 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3483 // This will fold the two constants together, which may allow
3484 // other simplifications.
3485 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3486 CastOp->getOperand(0), I.getType(),
3487 CastOp->getName()+".shrunk");
3488 NewCast = InsertNewInstBefore(NewCast, I);
3489 // trunc_or_bitcast(C1)&C2
3490 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3491 C3 = ConstantExpr::getAnd(C3, AndRHS);
3492 return BinaryOperator::CreateAnd(NewCast, C3);
3493 } else if (CastOp->getOpcode() == Instruction::Or) {
3494 // Change: and (cast (or X, C1) to T), C2
3495 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3496 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3497 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3498 return ReplaceInstUsesWith(I, AndRHS);
3504 // Try to fold constant and into select arguments.
3505 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3506 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3508 if (isa<PHINode>(Op0))
3509 if (Instruction *NV = FoldOpIntoPhi(I))
3513 Value *Op0NotVal = dyn_castNotVal(Op0);
3514 Value *Op1NotVal = dyn_castNotVal(Op1);
3516 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3517 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3519 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3520 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3521 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3522 I.getName()+".demorgan");
3523 InsertNewInstBefore(Or, I);
3524 return BinaryOperator::CreateNot(Or);
3528 Value *A = 0, *B = 0, *C = 0, *D = 0;
3529 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3530 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3531 return ReplaceInstUsesWith(I, Op1);
3533 // (A|B) & ~(A&B) -> A^B
3534 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3535 if ((A == C && B == D) || (A == D && B == C))
3536 return BinaryOperator::CreateXor(A, B);
3540 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3541 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3542 return ReplaceInstUsesWith(I, Op0);
3544 // ~(A&B) & (A|B) -> A^B
3545 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3546 if ((A == C && B == D) || (A == D && B == C))
3547 return BinaryOperator::CreateXor(A, B);
3551 if (Op0->hasOneUse() &&
3552 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3553 if (A == Op1) { // (A^B)&A -> A&(A^B)
3554 I.swapOperands(); // Simplify below
3555 std::swap(Op0, Op1);
3556 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3557 cast<BinaryOperator>(Op0)->swapOperands();
3558 I.swapOperands(); // Simplify below
3559 std::swap(Op0, Op1);
3562 if (Op1->hasOneUse() &&
3563 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3564 if (B == Op0) { // B&(A^B) -> B&(B^A)
3565 cast<BinaryOperator>(Op1)->swapOperands();
3568 if (A == Op0) { // A&(A^B) -> A & ~B
3569 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3570 InsertNewInstBefore(NotB, I);
3571 return BinaryOperator::CreateAnd(A, NotB);
3576 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3577 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3578 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3581 Value *LHSVal, *RHSVal;
3582 ConstantInt *LHSCst, *RHSCst;
3583 ICmpInst::Predicate LHSCC, RHSCC;
3584 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3585 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3586 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3587 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3588 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3589 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3590 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3591 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3593 // Don't try to fold ICMP_SLT + ICMP_ULT.
3594 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3595 ICmpInst::isSignedPredicate(LHSCC) ==
3596 ICmpInst::isSignedPredicate(RHSCC))) {
3597 // Ensure that the larger constant is on the RHS.
3598 ICmpInst::Predicate GT;
3599 if (ICmpInst::isSignedPredicate(LHSCC) ||
3600 (ICmpInst::isEquality(LHSCC) &&
3601 ICmpInst::isSignedPredicate(RHSCC)))
3602 GT = ICmpInst::ICMP_SGT;
3604 GT = ICmpInst::ICMP_UGT;
3606 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3607 ICmpInst *LHS = cast<ICmpInst>(Op0);
3608 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3609 std::swap(LHS, RHS);
3610 std::swap(LHSCst, RHSCst);
3611 std::swap(LHSCC, RHSCC);
3614 // At this point, we know we have have two icmp instructions
3615 // comparing a value against two constants and and'ing the result
3616 // together. Because of the above check, we know that we only have
3617 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3618 // (from the FoldICmpLogical check above), that the two constants
3619 // are not equal and that the larger constant is on the RHS
3620 assert(LHSCst != RHSCst && "Compares not folded above?");
3623 default: assert(0 && "Unknown integer condition code!");
3624 case ICmpInst::ICMP_EQ:
3626 default: assert(0 && "Unknown integer condition code!");
3627 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3628 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3629 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3630 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3631 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3632 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3633 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3634 return ReplaceInstUsesWith(I, LHS);
3636 case ICmpInst::ICMP_NE:
3638 default: assert(0 && "Unknown integer condition code!");
3639 case ICmpInst::ICMP_ULT:
3640 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3641 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3642 break; // (X != 13 & X u< 15) -> no change
3643 case ICmpInst::ICMP_SLT:
3644 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3645 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3646 break; // (X != 13 & X s< 15) -> no change
3647 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3648 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3649 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3650 return ReplaceInstUsesWith(I, RHS);
3651 case ICmpInst::ICMP_NE:
3652 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3653 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3654 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
3655 LHSVal->getName()+".off");
3656 InsertNewInstBefore(Add, I);
3657 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3658 ConstantInt::get(Add->getType(), 1));
3660 break; // (X != 13 & X != 15) -> no change
3663 case ICmpInst::ICMP_ULT:
3665 default: assert(0 && "Unknown integer condition code!");
3666 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3667 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3668 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3669 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3671 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3672 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3673 return ReplaceInstUsesWith(I, LHS);
3674 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3678 case ICmpInst::ICMP_SLT:
3680 default: assert(0 && "Unknown integer condition code!");
3681 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3682 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3683 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3684 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3686 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3687 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3688 return ReplaceInstUsesWith(I, LHS);
3689 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3693 case ICmpInst::ICMP_UGT:
3695 default: assert(0 && "Unknown integer condition code!");
3696 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3697 return ReplaceInstUsesWith(I, LHS);
3698 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3699 return ReplaceInstUsesWith(I, RHS);
3700 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3702 case ICmpInst::ICMP_NE:
3703 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3704 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3705 break; // (X u> 13 & X != 15) -> no change
3706 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3707 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3709 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3713 case ICmpInst::ICMP_SGT:
3715 default: assert(0 && "Unknown integer condition code!");
3716 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3717 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3718 return ReplaceInstUsesWith(I, RHS);
3719 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3721 case ICmpInst::ICMP_NE:
3722 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3723 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3724 break; // (X s> 13 & X != 15) -> no change
3725 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3726 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3728 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3736 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3737 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3738 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3739 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3740 const Type *SrcTy = Op0C->getOperand(0)->getType();
3741 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3742 // Only do this if the casts both really cause code to be generated.
3743 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3745 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3747 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3748 Op1C->getOperand(0),
3750 InsertNewInstBefore(NewOp, I);
3751 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3755 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3756 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3757 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3758 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3759 SI0->getOperand(1) == SI1->getOperand(1) &&
3760 (SI0->hasOneUse() || SI1->hasOneUse())) {
3761 Instruction *NewOp =
3762 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3764 SI0->getName()), I);
3765 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3766 SI1->getOperand(1));
3770 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3771 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3772 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3773 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3774 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3775 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3776 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3777 // If either of the constants are nans, then the whole thing returns
3779 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3780 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3781 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3782 RHS->getOperand(0));
3787 return Changed ? &I : 0;
3790 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3791 /// in the result. If it does, and if the specified byte hasn't been filled in
3792 /// yet, fill it in and return false.
3793 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3794 Instruction *I = dyn_cast<Instruction>(V);
3795 if (I == 0) return true;
3797 // If this is an or instruction, it is an inner node of the bswap.
3798 if (I->getOpcode() == Instruction::Or)
3799 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3800 CollectBSwapParts(I->getOperand(1), ByteValues);
3802 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3803 // If this is a shift by a constant int, and it is "24", then its operand
3804 // defines a byte. We only handle unsigned types here.
3805 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3806 // Not shifting the entire input by N-1 bytes?
3807 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3808 8*(ByteValues.size()-1))
3812 if (I->getOpcode() == Instruction::Shl) {
3813 // X << 24 defines the top byte with the lowest of the input bytes.
3814 DestNo = ByteValues.size()-1;
3816 // X >>u 24 defines the low byte with the highest of the input bytes.
3820 // If the destination byte value is already defined, the values are or'd
3821 // together, which isn't a bswap (unless it's an or of the same bits).
3822 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3824 ByteValues[DestNo] = I->getOperand(0);
3828 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3830 Value *Shift = 0, *ShiftLHS = 0;
3831 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3832 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3833 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3835 Instruction *SI = cast<Instruction>(Shift);
3837 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3838 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3839 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3842 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3844 if (AndAmt->getValue().getActiveBits() > 64)
3846 uint64_t AndAmtVal = AndAmt->getZExtValue();
3847 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3848 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3850 // Unknown mask for bswap.
3851 if (DestByte == ByteValues.size()) return true;
3853 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3855 if (SI->getOpcode() == Instruction::Shl)
3856 SrcByte = DestByte - ShiftBytes;
3858 SrcByte = DestByte + ShiftBytes;
3860 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3861 if (SrcByte != ByteValues.size()-DestByte-1)
3864 // If the destination byte value is already defined, the values are or'd
3865 // together, which isn't a bswap (unless it's an or of the same bits).
3866 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3868 ByteValues[DestByte] = SI->getOperand(0);
3872 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3873 /// If so, insert the new bswap intrinsic and return it.
3874 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3875 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3876 if (!ITy || ITy->getBitWidth() % 16)
3877 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3879 /// ByteValues - For each byte of the result, we keep track of which value
3880 /// defines each byte.
3881 SmallVector<Value*, 8> ByteValues;
3882 ByteValues.resize(ITy->getBitWidth()/8);
3884 // Try to find all the pieces corresponding to the bswap.
3885 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3886 CollectBSwapParts(I.getOperand(1), ByteValues))
3889 // Check to see if all of the bytes come from the same value.
3890 Value *V = ByteValues[0];
3891 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3893 // Check to make sure that all of the bytes come from the same value.
3894 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3895 if (ByteValues[i] != V)
3897 const Type *Tys[] = { ITy };
3898 Module *M = I.getParent()->getParent()->getParent();
3899 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3900 return CallInst::Create(F, V);
3904 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3905 bool Changed = SimplifyCommutative(I);
3906 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3908 if (isa<UndefValue>(Op1)) // X | undef -> -1
3909 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3913 return ReplaceInstUsesWith(I, Op0);
3915 // See if we can simplify any instructions used by the instruction whose sole
3916 // purpose is to compute bits we don't care about.
3917 if (!isa<VectorType>(I.getType())) {
3918 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3919 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3920 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3921 KnownZero, KnownOne))
3923 } else if (isa<ConstantAggregateZero>(Op1)) {
3924 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3925 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3926 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3927 return ReplaceInstUsesWith(I, I.getOperand(1));
3933 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3934 ConstantInt *C1 = 0; Value *X = 0;
3935 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3936 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3937 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3938 InsertNewInstBefore(Or, I);
3940 return BinaryOperator::CreateAnd(Or,
3941 ConstantInt::get(RHS->getValue() | C1->getValue()));
3944 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3945 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3946 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3947 InsertNewInstBefore(Or, I);
3949 return BinaryOperator::CreateXor(Or,
3950 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3953 // Try to fold constant and into select arguments.
3954 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3955 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3957 if (isa<PHINode>(Op0))
3958 if (Instruction *NV = FoldOpIntoPhi(I))
3962 Value *A = 0, *B = 0;
3963 ConstantInt *C1 = 0, *C2 = 0;
3965 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3966 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3967 return ReplaceInstUsesWith(I, Op1);
3968 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3969 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3970 return ReplaceInstUsesWith(I, Op0);
3972 // (A | B) | C and A | (B | C) -> bswap if possible.
3973 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3974 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3975 match(Op1, m_Or(m_Value(), m_Value())) ||
3976 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3977 match(Op1, m_Shift(m_Value(), m_Value())))) {
3978 if (Instruction *BSwap = MatchBSwap(I))
3982 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3983 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3984 MaskedValueIsZero(Op1, C1->getValue())) {
3985 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
3986 InsertNewInstBefore(NOr, I);
3988 return BinaryOperator::CreateXor(NOr, C1);
3991 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3992 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3993 MaskedValueIsZero(Op0, C1->getValue())) {
3994 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
3995 InsertNewInstBefore(NOr, I);
3997 return BinaryOperator::CreateXor(NOr, C1);
4001 Value *C = 0, *D = 0;
4002 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4003 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4004 Value *V1 = 0, *V2 = 0, *V3 = 0;
4005 C1 = dyn_cast<ConstantInt>(C);
4006 C2 = dyn_cast<ConstantInt>(D);
4007 if (C1 && C2) { // (A & C1)|(B & C2)
4008 // If we have: ((V + N) & C1) | (V & C2)
4009 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4010 // replace with V+N.
4011 if (C1->getValue() == ~C2->getValue()) {
4012 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4013 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4014 // Add commutes, try both ways.
4015 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4016 return ReplaceInstUsesWith(I, A);
4017 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4018 return ReplaceInstUsesWith(I, A);
4020 // Or commutes, try both ways.
4021 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4022 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4023 // Add commutes, try both ways.
4024 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4025 return ReplaceInstUsesWith(I, B);
4026 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4027 return ReplaceInstUsesWith(I, B);
4030 V1 = 0; V2 = 0; V3 = 0;
4033 // Check to see if we have any common things being and'ed. If so, find the
4034 // terms for V1 & (V2|V3).
4035 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4036 if (A == B) // (A & C)|(A & D) == A & (C|D)
4037 V1 = A, V2 = C, V3 = D;
4038 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4039 V1 = A, V2 = B, V3 = C;
4040 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4041 V1 = C, V2 = A, V3 = D;
4042 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4043 V1 = C, V2 = A, V3 = B;
4047 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4048 return BinaryOperator::CreateAnd(V1, Or);
4053 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4054 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4055 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4056 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4057 SI0->getOperand(1) == SI1->getOperand(1) &&
4058 (SI0->hasOneUse() || SI1->hasOneUse())) {
4059 Instruction *NewOp =
4060 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4062 SI0->getName()), I);
4063 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4064 SI1->getOperand(1));
4068 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4069 if (A == Op1) // ~A | A == -1
4070 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4074 // Note, A is still live here!
4075 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4077 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4079 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4080 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4081 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4082 I.getName()+".demorgan"), I);
4083 return BinaryOperator::CreateNot(And);
4087 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4088 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4089 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4092 Value *LHSVal, *RHSVal;
4093 ConstantInt *LHSCst, *RHSCst;
4094 ICmpInst::Predicate LHSCC, RHSCC;
4095 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4096 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4097 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4098 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4099 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4100 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4101 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4102 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4103 // We can't fold (ugt x, C) | (sgt x, C2).
4104 PredicatesFoldable(LHSCC, RHSCC)) {
4105 // Ensure that the larger constant is on the RHS.
4106 ICmpInst *LHS = cast<ICmpInst>(Op0);
4108 if (ICmpInst::isSignedPredicate(LHSCC))
4109 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4111 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4114 std::swap(LHS, RHS);
4115 std::swap(LHSCst, RHSCst);
4116 std::swap(LHSCC, RHSCC);
4119 // At this point, we know we have have two icmp instructions
4120 // comparing a value against two constants and or'ing the result
4121 // together. Because of the above check, we know that we only have
4122 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4123 // FoldICmpLogical check above), that the two constants are not
4125 assert(LHSCst != RHSCst && "Compares not folded above?");
4128 default: assert(0 && "Unknown integer condition code!");
4129 case ICmpInst::ICMP_EQ:
4131 default: assert(0 && "Unknown integer condition code!");
4132 case ICmpInst::ICMP_EQ:
4133 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4134 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4135 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4136 LHSVal->getName()+".off");
4137 InsertNewInstBefore(Add, I);
4138 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4139 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4141 break; // (X == 13 | X == 15) -> no change
4142 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4143 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4145 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4146 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4147 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4148 return ReplaceInstUsesWith(I, RHS);
4151 case ICmpInst::ICMP_NE:
4153 default: assert(0 && "Unknown integer condition code!");
4154 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4155 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4156 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4157 return ReplaceInstUsesWith(I, LHS);
4158 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4159 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4160 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4161 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4164 case ICmpInst::ICMP_ULT:
4166 default: assert(0 && "Unknown integer condition code!");
4167 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4169 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4170 // If RHSCst is [us]MAXINT, it is always false. Not handling
4171 // this can cause overflow.
4172 if (RHSCst->isMaxValue(false))
4173 return ReplaceInstUsesWith(I, LHS);
4174 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4176 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4178 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4179 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4180 return ReplaceInstUsesWith(I, RHS);
4181 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4185 case ICmpInst::ICMP_SLT:
4187 default: assert(0 && "Unknown integer condition code!");
4188 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4190 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4191 // If RHSCst is [us]MAXINT, it is always false. Not handling
4192 // this can cause overflow.
4193 if (RHSCst->isMaxValue(true))
4194 return ReplaceInstUsesWith(I, LHS);
4195 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4197 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4199 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4200 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4201 return ReplaceInstUsesWith(I, RHS);
4202 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4206 case ICmpInst::ICMP_UGT:
4208 default: assert(0 && "Unknown integer condition code!");
4209 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4210 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4211 return ReplaceInstUsesWith(I, LHS);
4212 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4214 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4215 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4216 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4217 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4221 case ICmpInst::ICMP_SGT:
4223 default: assert(0 && "Unknown integer condition code!");
4224 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4225 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4226 return ReplaceInstUsesWith(I, LHS);
4227 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4229 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4230 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4231 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4232 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4240 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4241 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4242 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4243 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4244 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4245 !isa<ICmpInst>(Op1C->getOperand(0))) {
4246 const Type *SrcTy = Op0C->getOperand(0)->getType();
4247 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4248 // Only do this if the casts both really cause code to be
4250 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4252 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4254 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4255 Op1C->getOperand(0),
4257 InsertNewInstBefore(NewOp, I);
4258 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4265 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4266 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4267 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4268 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4269 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4270 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4271 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4272 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4273 // If either of the constants are nans, then the whole thing returns
4275 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4276 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4278 // Otherwise, no need to compare the two constants, compare the
4280 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4281 RHS->getOperand(0));
4286 return Changed ? &I : 0;
4291 // XorSelf - Implements: X ^ X --> 0
4294 XorSelf(Value *rhs) : RHS(rhs) {}
4295 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4296 Instruction *apply(BinaryOperator &Xor) const {
4303 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4304 bool Changed = SimplifyCommutative(I);
4305 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4307 if (isa<UndefValue>(Op1)) {
4308 if (isa<UndefValue>(Op0))
4309 // Handle undef ^ undef -> 0 special case. This is a common
4311 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4312 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4315 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4316 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4317 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4318 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4321 // See if we can simplify any instructions used by the instruction whose sole
4322 // purpose is to compute bits we don't care about.
4323 if (!isa<VectorType>(I.getType())) {
4324 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4325 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4326 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4327 KnownZero, KnownOne))
4329 } else if (isa<ConstantAggregateZero>(Op1)) {
4330 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4333 // Is this a ~ operation?
4334 if (Value *NotOp = dyn_castNotVal(&I)) {
4335 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4336 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4337 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4338 if (Op0I->getOpcode() == Instruction::And ||
4339 Op0I->getOpcode() == Instruction::Or) {
4340 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4341 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4343 BinaryOperator::CreateNot(Op0I->getOperand(1),
4344 Op0I->getOperand(1)->getName()+".not");
4345 InsertNewInstBefore(NotY, I);
4346 if (Op0I->getOpcode() == Instruction::And)
4347 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4349 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4356 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4357 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4358 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4359 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4360 return new ICmpInst(ICI->getInversePredicate(),
4361 ICI->getOperand(0), ICI->getOperand(1));
4363 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4364 return new FCmpInst(FCI->getInversePredicate(),
4365 FCI->getOperand(0), FCI->getOperand(1));
4368 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4369 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4370 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4371 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4372 Instruction::CastOps Opcode = Op0C->getOpcode();
4373 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4374 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4375 Op0C->getDestTy())) {
4376 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4377 CI->getOpcode(), CI->getInversePredicate(),
4378 CI->getOperand(0), CI->getOperand(1)), I);
4379 NewCI->takeName(CI);
4380 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4387 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4388 // ~(c-X) == X-c-1 == X+(-c-1)
4389 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4390 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4391 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4392 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4393 ConstantInt::get(I.getType(), 1));
4394 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4397 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4398 if (Op0I->getOpcode() == Instruction::Add) {
4399 // ~(X-c) --> (-c-1)-X
4400 if (RHS->isAllOnesValue()) {
4401 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4402 return BinaryOperator::CreateSub(
4403 ConstantExpr::getSub(NegOp0CI,
4404 ConstantInt::get(I.getType(), 1)),
4405 Op0I->getOperand(0));
4406 } else if (RHS->getValue().isSignBit()) {
4407 // (X + C) ^ signbit -> (X + C + signbit)
4408 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4409 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4412 } else if (Op0I->getOpcode() == Instruction::Or) {
4413 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4414 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4415 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4416 // Anything in both C1 and C2 is known to be zero, remove it from
4418 Constant *CommonBits = And(Op0CI, RHS);
4419 NewRHS = ConstantExpr::getAnd(NewRHS,
4420 ConstantExpr::getNot(CommonBits));
4421 AddToWorkList(Op0I);
4422 I.setOperand(0, Op0I->getOperand(0));
4423 I.setOperand(1, NewRHS);
4430 // Try to fold constant and into select arguments.
4431 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4432 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4434 if (isa<PHINode>(Op0))
4435 if (Instruction *NV = FoldOpIntoPhi(I))
4439 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4441 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4443 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4445 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4448 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4451 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4452 if (A == Op0) { // B^(B|A) == (A|B)^B
4453 Op1I->swapOperands();
4455 std::swap(Op0, Op1);
4456 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4457 I.swapOperands(); // Simplified below.
4458 std::swap(Op0, Op1);
4460 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4461 if (Op0 == A) // A^(A^B) == B
4462 return ReplaceInstUsesWith(I, B);
4463 else if (Op0 == B) // A^(B^A) == B
4464 return ReplaceInstUsesWith(I, A);
4465 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4466 if (A == Op0) { // A^(A&B) -> A^(B&A)
4467 Op1I->swapOperands();
4470 if (B == Op0) { // A^(B&A) -> (B&A)^A
4471 I.swapOperands(); // Simplified below.
4472 std::swap(Op0, Op1);
4477 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4480 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4481 if (A == Op1) // (B|A)^B == (A|B)^B
4483 if (B == Op1) { // (A|B)^B == A & ~B
4485 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4486 return BinaryOperator::CreateAnd(A, NotB);
4488 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4489 if (Op1 == A) // (A^B)^A == B
4490 return ReplaceInstUsesWith(I, B);
4491 else if (Op1 == B) // (B^A)^A == B
4492 return ReplaceInstUsesWith(I, A);
4493 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4494 if (A == Op1) // (A&B)^A -> (B&A)^A
4496 if (B == Op1 && // (B&A)^A == ~B & A
4497 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4499 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4500 return BinaryOperator::CreateAnd(N, Op1);
4505 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4506 if (Op0I && Op1I && Op0I->isShift() &&
4507 Op0I->getOpcode() == Op1I->getOpcode() &&
4508 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4509 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4510 Instruction *NewOp =
4511 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4512 Op1I->getOperand(0),
4513 Op0I->getName()), I);
4514 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4515 Op1I->getOperand(1));
4519 Value *A, *B, *C, *D;
4520 // (A & B)^(A | B) -> A ^ B
4521 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4522 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4523 if ((A == C && B == D) || (A == D && B == C))
4524 return BinaryOperator::CreateXor(A, B);
4526 // (A | B)^(A & B) -> A ^ B
4527 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4528 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4529 if ((A == C && B == D) || (A == D && B == C))
4530 return BinaryOperator::CreateXor(A, B);
4534 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4535 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4536 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4537 // (X & Y)^(X & Y) -> (Y^Z) & X
4538 Value *X = 0, *Y = 0, *Z = 0;
4540 X = A, Y = B, Z = D;
4542 X = A, Y = B, Z = C;
4544 X = B, Y = A, Z = D;
4546 X = B, Y = A, Z = C;
4549 Instruction *NewOp =
4550 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4551 return BinaryOperator::CreateAnd(NewOp, X);
4556 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4557 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4558 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4561 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4562 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4563 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4564 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4565 const Type *SrcTy = Op0C->getOperand(0)->getType();
4566 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4567 // Only do this if the casts both really cause code to be generated.
4568 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4570 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4572 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4573 Op1C->getOperand(0),
4575 InsertNewInstBefore(NewOp, I);
4576 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4581 return Changed ? &I : 0;
4584 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4585 /// overflowed for this type.
4586 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4587 ConstantInt *In2, bool IsSigned = false) {
4588 Result = cast<ConstantInt>(Add(In1, In2));
4591 if (In2->getValue().isNegative())
4592 return Result->getValue().sgt(In1->getValue());
4594 return Result->getValue().slt(In1->getValue());
4596 return Result->getValue().ult(In1->getValue());
4599 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4600 /// code necessary to compute the offset from the base pointer (without adding
4601 /// in the base pointer). Return the result as a signed integer of intptr size.
4602 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4603 TargetData &TD = IC.getTargetData();
4604 gep_type_iterator GTI = gep_type_begin(GEP);
4605 const Type *IntPtrTy = TD.getIntPtrType();
4606 Value *Result = Constant::getNullValue(IntPtrTy);
4608 // Build a mask for high order bits.
4609 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4610 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4612 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
4615 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4616 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4617 if (OpC->isZero()) continue;
4619 // Handle a struct index, which adds its field offset to the pointer.
4620 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4621 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4623 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4624 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4626 Result = IC.InsertNewInstBefore(
4627 BinaryOperator::CreateAdd(Result,
4628 ConstantInt::get(IntPtrTy, Size),
4629 GEP->getName()+".offs"), I);
4633 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4634 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4635 Scale = ConstantExpr::getMul(OC, Scale);
4636 if (Constant *RC = dyn_cast<Constant>(Result))
4637 Result = ConstantExpr::getAdd(RC, Scale);
4639 // Emit an add instruction.
4640 Result = IC.InsertNewInstBefore(
4641 BinaryOperator::CreateAdd(Result, Scale,
4642 GEP->getName()+".offs"), I);
4646 // Convert to correct type.
4647 if (Op->getType() != IntPtrTy) {
4648 if (Constant *OpC = dyn_cast<Constant>(Op))
4649 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4651 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4652 Op->getName()+".c"), I);
4655 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4656 if (Constant *OpC = dyn_cast<Constant>(Op))
4657 Op = ConstantExpr::getMul(OpC, Scale);
4658 else // We'll let instcombine(mul) convert this to a shl if possible.
4659 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
4660 GEP->getName()+".idx"), I);
4663 // Emit an add instruction.
4664 if (isa<Constant>(Op) && isa<Constant>(Result))
4665 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4666 cast<Constant>(Result));
4668 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
4669 GEP->getName()+".offs"), I);
4675 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4676 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4677 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4678 /// complex, and scales are involved. The above expression would also be legal
4679 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4680 /// later form is less amenable to optimization though, and we are allowed to
4681 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4683 /// If we can't emit an optimized form for this expression, this returns null.
4685 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4687 TargetData &TD = IC.getTargetData();
4688 gep_type_iterator GTI = gep_type_begin(GEP);
4690 // Check to see if this gep only has a single variable index. If so, and if
4691 // any constant indices are a multiple of its scale, then we can compute this
4692 // in terms of the scale of the variable index. For example, if the GEP
4693 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4694 // because the expression will cross zero at the same point.
4695 unsigned i, e = GEP->getNumOperands();
4697 for (i = 1; i != e; ++i, ++GTI) {
4698 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4699 // Compute the aggregate offset of constant indices.
4700 if (CI->isZero()) continue;
4702 // Handle a struct index, which adds its field offset to the pointer.
4703 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4704 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4706 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4707 Offset += Size*CI->getSExtValue();
4710 // Found our variable index.
4715 // If there are no variable indices, we must have a constant offset, just
4716 // evaluate it the general way.
4717 if (i == e) return 0;
4719 Value *VariableIdx = GEP->getOperand(i);
4720 // Determine the scale factor of the variable element. For example, this is
4721 // 4 if the variable index is into an array of i32.
4722 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4724 // Verify that there are no other variable indices. If so, emit the hard way.
4725 for (++i, ++GTI; i != e; ++i, ++GTI) {
4726 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4729 // Compute the aggregate offset of constant indices.
4730 if (CI->isZero()) continue;
4732 // Handle a struct index, which adds its field offset to the pointer.
4733 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4734 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4736 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4737 Offset += Size*CI->getSExtValue();
4741 // Okay, we know we have a single variable index, which must be a
4742 // pointer/array/vector index. If there is no offset, life is simple, return
4744 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4746 // Cast to intptrty in case a truncation occurs. If an extension is needed,
4747 // we don't need to bother extending: the extension won't affect where the
4748 // computation crosses zero.
4749 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
4750 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
4751 VariableIdx->getNameStart(), &I);
4755 // Otherwise, there is an index. The computation we will do will be modulo
4756 // the pointer size, so get it.
4757 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4759 Offset &= PtrSizeMask;
4760 VariableScale &= PtrSizeMask;
4762 // To do this transformation, any constant index must be a multiple of the
4763 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
4764 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
4765 // multiple of the variable scale.
4766 int64_t NewOffs = Offset / (int64_t)VariableScale;
4767 if (Offset != NewOffs*(int64_t)VariableScale)
4770 // Okay, we can do this evaluation. Start by converting the index to intptr.
4771 const Type *IntPtrTy = TD.getIntPtrType();
4772 if (VariableIdx->getType() != IntPtrTy)
4773 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
4775 VariableIdx->getNameStart(), &I);
4776 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
4777 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
4781 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4782 /// else. At this point we know that the GEP is on the LHS of the comparison.
4783 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4784 ICmpInst::Predicate Cond,
4786 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4788 // Look through bitcasts.
4789 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
4790 RHS = BCI->getOperand(0);
4792 Value *PtrBase = GEPLHS->getOperand(0);
4793 if (PtrBase == RHS) {
4794 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4795 // This transformation (ignoring the base and scales) is valid because we
4796 // know pointers can't overflow. See if we can output an optimized form.
4797 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
4799 // If not, synthesize the offset the hard way.
4801 Offset = EmitGEPOffset(GEPLHS, I, *this);
4802 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4803 Constant::getNullValue(Offset->getType()));
4804 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4805 // If the base pointers are different, but the indices are the same, just
4806 // compare the base pointer.
4807 if (PtrBase != GEPRHS->getOperand(0)) {
4808 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4809 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4810 GEPRHS->getOperand(0)->getType();
4812 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4813 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4814 IndicesTheSame = false;
4818 // If all indices are the same, just compare the base pointers.
4820 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4821 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4823 // Otherwise, the base pointers are different and the indices are
4824 // different, bail out.
4828 // If one of the GEPs has all zero indices, recurse.
4829 bool AllZeros = true;
4830 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4831 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4832 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4837 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4838 ICmpInst::getSwappedPredicate(Cond), I);
4840 // If the other GEP has all zero indices, recurse.
4842 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4843 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4844 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4849 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4851 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4852 // If the GEPs only differ by one index, compare it.
4853 unsigned NumDifferences = 0; // Keep track of # differences.
4854 unsigned DiffOperand = 0; // The operand that differs.
4855 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4856 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4857 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4858 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4859 // Irreconcilable differences.
4863 if (NumDifferences++) break;
4868 if (NumDifferences == 0) // SAME GEP?
4869 return ReplaceInstUsesWith(I, // No comparison is needed here.
4870 ConstantInt::get(Type::Int1Ty,
4871 ICmpInst::isTrueWhenEqual(Cond)));
4873 else if (NumDifferences == 1) {
4874 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4875 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4876 // Make sure we do a signed comparison here.
4877 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4881 // Only lower this if the icmp is the only user of the GEP or if we expect
4882 // the result to fold to a constant!
4883 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4884 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4885 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4886 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4887 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4888 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4894 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
4896 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
4899 if (!isa<ConstantFP>(RHSC)) return 0;
4900 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4902 // Get the width of the mantissa. We don't want to hack on conversions that
4903 // might lose information from the integer, e.g. "i64 -> float"
4904 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4905 if (MantissaWidth == -1) return 0; // Unknown.
4907 // Check to see that the input is converted from an integer type that is small
4908 // enough that preserves all bits. TODO: check here for "known" sign bits.
4909 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4910 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
4912 // If this is a uitofp instruction, we need an extra bit to hold the sign.
4913 if (isa<UIToFPInst>(LHSI))
4916 // If the conversion would lose info, don't hack on this.
4917 if ((int)InputSize > MantissaWidth)
4920 // Otherwise, we can potentially simplify the comparison. We know that it
4921 // will always come through as an integer value and we know the constant is
4922 // not a NAN (it would have been previously simplified).
4923 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4925 ICmpInst::Predicate Pred;
4926 switch (I.getPredicate()) {
4927 default: assert(0 && "Unexpected predicate!");
4928 case FCmpInst::FCMP_UEQ:
4929 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
4930 case FCmpInst::FCMP_UGT:
4931 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
4932 case FCmpInst::FCMP_UGE:
4933 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
4934 case FCmpInst::FCMP_ULT:
4935 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
4936 case FCmpInst::FCMP_ULE:
4937 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
4938 case FCmpInst::FCMP_UNE:
4939 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
4940 case FCmpInst::FCMP_ORD:
4941 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4942 case FCmpInst::FCMP_UNO:
4943 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4946 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4948 // Now we know that the APFloat is a normal number, zero or inf.
4950 // See if the FP constant is too large for the integer. For example,
4951 // comparing an i8 to 300.0.
4952 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
4954 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4955 // and large values.
4956 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
4957 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4958 APFloat::rmNearestTiesToEven);
4959 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4960 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4961 Pred == ICmpInst::ICMP_SLE)
4962 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4963 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4966 // See if the RHS value is < SignedMin.
4967 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
4968 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4969 APFloat::rmNearestTiesToEven);
4970 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4971 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4972 Pred == ICmpInst::ICMP_SGE)
4973 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4974 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4977 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
4978 // it may still be fractional. See if it is fractional by casting the FP
4979 // value to the integer value and back, checking for equality. Don't do this
4980 // for zero, because -0.0 is not fractional.
4981 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
4982 if (!RHS.isZero() &&
4983 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
4984 // If we had a comparison against a fractional value, we have to adjust
4985 // the compare predicate and sometimes the value. RHSC is rounded towards
4986 // zero at this point.
4988 default: assert(0 && "Unexpected integer comparison!");
4989 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
4990 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4991 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
4992 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4993 case ICmpInst::ICMP_SLE:
4994 // (float)int <= 4.4 --> int <= 4
4995 // (float)int <= -4.4 --> int < -4
4996 if (RHS.isNegative())
4997 Pred = ICmpInst::ICMP_SLT;
4999 case ICmpInst::ICMP_SLT:
5000 // (float)int < -4.4 --> int < -4
5001 // (float)int < 4.4 --> int <= 4
5002 if (!RHS.isNegative())
5003 Pred = ICmpInst::ICMP_SLE;
5005 case ICmpInst::ICMP_SGT:
5006 // (float)int > 4.4 --> int > 4
5007 // (float)int > -4.4 --> int >= -4
5008 if (RHS.isNegative())
5009 Pred = ICmpInst::ICMP_SGE;
5011 case ICmpInst::ICMP_SGE:
5012 // (float)int >= -4.4 --> int >= -4
5013 // (float)int >= 4.4 --> int > 4
5014 if (!RHS.isNegative())
5015 Pred = ICmpInst::ICMP_SGT;
5020 // Lower this FP comparison into an appropriate integer version of the
5022 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5025 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5026 bool Changed = SimplifyCompare(I);
5027 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5029 // Fold trivial predicates.
5030 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5031 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5032 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5033 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5035 // Simplify 'fcmp pred X, X'
5037 switch (I.getPredicate()) {
5038 default: assert(0 && "Unknown predicate!");
5039 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5040 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5041 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5042 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5043 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5044 case FCmpInst::FCMP_OLT: // True if ordered and less than
5045 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5046 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5048 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5049 case FCmpInst::FCMP_ULT: // True if unordered or less than
5050 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5051 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5052 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5053 I.setPredicate(FCmpInst::FCMP_UNO);
5054 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5057 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5058 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5059 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5060 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5061 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5062 I.setPredicate(FCmpInst::FCMP_ORD);
5063 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5068 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5069 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5071 // Handle fcmp with constant RHS
5072 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5073 // If the constant is a nan, see if we can fold the comparison based on it.
5074 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5075 if (CFP->getValueAPF().isNaN()) {
5076 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5077 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5078 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5079 "Comparison must be either ordered or unordered!");
5080 // True if unordered.
5081 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5085 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5086 switch (LHSI->getOpcode()) {
5087 case Instruction::PHI:
5088 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5089 // block. If in the same block, we're encouraging jump threading. If
5090 // not, we are just pessimizing the code by making an i1 phi.
5091 if (LHSI->getParent() == I.getParent())
5092 if (Instruction *NV = FoldOpIntoPhi(I))
5095 case Instruction::SIToFP:
5096 case Instruction::UIToFP:
5097 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5100 case Instruction::Select:
5101 // If either operand of the select is a constant, we can fold the
5102 // comparison into the select arms, which will cause one to be
5103 // constant folded and the select turned into a bitwise or.
5104 Value *Op1 = 0, *Op2 = 0;
5105 if (LHSI->hasOneUse()) {
5106 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5107 // Fold the known value into the constant operand.
5108 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5109 // Insert a new FCmp of the other select operand.
5110 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5111 LHSI->getOperand(2), RHSC,
5113 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5114 // Fold the known value into the constant operand.
5115 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5116 // Insert a new FCmp of the other select operand.
5117 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5118 LHSI->getOperand(1), RHSC,
5124 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5129 return Changed ? &I : 0;
5132 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5133 bool Changed = SimplifyCompare(I);
5134 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5135 const Type *Ty = Op0->getType();
5139 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5140 I.isTrueWhenEqual()));
5142 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5143 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5145 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5146 // addresses never equal each other! We already know that Op0 != Op1.
5147 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5148 isa<ConstantPointerNull>(Op0)) &&
5149 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5150 isa<ConstantPointerNull>(Op1)))
5151 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5152 !I.isTrueWhenEqual()));
5154 // icmp's with boolean values can always be turned into bitwise operations
5155 if (Ty == Type::Int1Ty) {
5156 switch (I.getPredicate()) {
5157 default: assert(0 && "Invalid icmp instruction!");
5158 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5159 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5160 InsertNewInstBefore(Xor, I);
5161 return BinaryOperator::CreateNot(Xor);
5163 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5164 return BinaryOperator::CreateXor(Op0, Op1);
5166 case ICmpInst::ICMP_UGT:
5167 case ICmpInst::ICMP_SGT:
5168 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5170 case ICmpInst::ICMP_ULT:
5171 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5172 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5173 InsertNewInstBefore(Not, I);
5174 return BinaryOperator::CreateAnd(Not, Op1);
5176 case ICmpInst::ICMP_UGE:
5177 case ICmpInst::ICMP_SGE:
5178 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5180 case ICmpInst::ICMP_ULE:
5181 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5182 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5183 InsertNewInstBefore(Not, I);
5184 return BinaryOperator::CreateOr(Not, Op1);
5189 // See if we are doing a comparison between a constant and an instruction that
5190 // can be folded into the comparison.
5191 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5194 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5195 if (I.isEquality() && CI->isNullValue() &&
5196 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5197 // (icmp cond A B) if cond is equality
5198 return new ICmpInst(I.getPredicate(), A, B);
5201 switch (I.getPredicate()) {
5203 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5204 if (CI->isMinValue(false))
5205 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5206 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5207 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5208 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5209 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5210 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5211 if (CI->isMinValue(true))
5212 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5213 ConstantInt::getAllOnesValue(Op0->getType()));
5217 case ICmpInst::ICMP_SLT:
5218 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5219 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5220 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5221 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5222 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5223 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5226 case ICmpInst::ICMP_UGT:
5227 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5228 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5229 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5230 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5231 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5232 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5234 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5235 if (CI->isMaxValue(true))
5236 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5237 ConstantInt::getNullValue(Op0->getType()));
5240 case ICmpInst::ICMP_SGT:
5241 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5242 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5243 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5244 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5245 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5246 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5249 case ICmpInst::ICMP_ULE:
5250 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5251 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5252 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5253 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5254 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5255 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5258 case ICmpInst::ICMP_SLE:
5259 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5260 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5261 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5262 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5263 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5264 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5267 case ICmpInst::ICMP_UGE:
5268 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5269 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5270 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5271 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5272 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5273 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5276 case ICmpInst::ICMP_SGE:
5277 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5278 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5279 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5280 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5281 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5282 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5286 // If we still have a icmp le or icmp ge instruction, turn it into the
5287 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5288 // already been handled above, this requires little checking.
5290 switch (I.getPredicate()) {
5292 case ICmpInst::ICMP_ULE:
5293 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5294 case ICmpInst::ICMP_SLE:
5295 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5296 case ICmpInst::ICMP_UGE:
5297 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5298 case ICmpInst::ICMP_SGE:
5299 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5302 // See if we can fold the comparison based on bits known to be zero or one
5303 // in the input. If this comparison is a normal comparison, it demands all
5304 // bits, if it is a sign bit comparison, it only demands the sign bit.
5307 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5309 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5310 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5311 if (SimplifyDemandedBits(Op0,
5312 isSignBit ? APInt::getSignBit(BitWidth)
5313 : APInt::getAllOnesValue(BitWidth),
5314 KnownZero, KnownOne, 0))
5317 // Given the known and unknown bits, compute a range that the LHS could be
5319 if ((KnownOne | KnownZero) != 0) {
5320 // Compute the Min, Max and RHS values based on the known bits. For the
5321 // EQ and NE we use unsigned values.
5322 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5323 const APInt& RHSVal = CI->getValue();
5324 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5325 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5328 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5331 switch (I.getPredicate()) { // LE/GE have been folded already.
5332 default: assert(0 && "Unknown icmp opcode!");
5333 case ICmpInst::ICMP_EQ:
5334 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5335 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5337 case ICmpInst::ICMP_NE:
5338 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5339 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5341 case ICmpInst::ICMP_ULT:
5342 if (Max.ult(RHSVal))
5343 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5344 if (Min.uge(RHSVal))
5345 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5347 case ICmpInst::ICMP_UGT:
5348 if (Min.ugt(RHSVal))
5349 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5350 if (Max.ule(RHSVal))
5351 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5353 case ICmpInst::ICMP_SLT:
5354 if (Max.slt(RHSVal))
5355 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5356 if (Min.sgt(RHSVal))
5357 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5359 case ICmpInst::ICMP_SGT:
5360 if (Min.sgt(RHSVal))
5361 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5362 if (Max.sle(RHSVal))
5363 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5368 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5369 // instruction, see if that instruction also has constants so that the
5370 // instruction can be folded into the icmp
5371 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5372 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5376 // Handle icmp with constant (but not simple integer constant) RHS
5377 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5378 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5379 switch (LHSI->getOpcode()) {
5380 case Instruction::GetElementPtr:
5381 if (RHSC->isNullValue()) {
5382 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5383 bool isAllZeros = true;
5384 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5385 if (!isa<Constant>(LHSI->getOperand(i)) ||
5386 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5391 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5392 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5396 case Instruction::PHI:
5397 // Only fold icmp into the PHI if the phi and fcmp are in the same
5398 // block. If in the same block, we're encouraging jump threading. If
5399 // not, we are just pessimizing the code by making an i1 phi.
5400 if (LHSI->getParent() == I.getParent())
5401 if (Instruction *NV = FoldOpIntoPhi(I))
5404 case Instruction::Select: {
5405 // If either operand of the select is a constant, we can fold the
5406 // comparison into the select arms, which will cause one to be
5407 // constant folded and the select turned into a bitwise or.
5408 Value *Op1 = 0, *Op2 = 0;
5409 if (LHSI->hasOneUse()) {
5410 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5411 // Fold the known value into the constant operand.
5412 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5413 // Insert a new ICmp of the other select operand.
5414 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5415 LHSI->getOperand(2), RHSC,
5417 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5418 // Fold the known value into the constant operand.
5419 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5420 // Insert a new ICmp of the other select operand.
5421 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5422 LHSI->getOperand(1), RHSC,
5428 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5431 case Instruction::Malloc:
5432 // If we have (malloc != null), and if the malloc has a single use, we
5433 // can assume it is successful and remove the malloc.
5434 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5435 AddToWorkList(LHSI);
5436 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5437 !I.isTrueWhenEqual()));
5443 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5444 if (User *GEP = dyn_castGetElementPtr(Op0))
5445 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5447 if (User *GEP = dyn_castGetElementPtr(Op1))
5448 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5449 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5452 // Test to see if the operands of the icmp are casted versions of other
5453 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5455 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5456 if (isa<PointerType>(Op0->getType()) &&
5457 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5458 // We keep moving the cast from the left operand over to the right
5459 // operand, where it can often be eliminated completely.
5460 Op0 = CI->getOperand(0);
5462 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5463 // so eliminate it as well.
5464 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5465 Op1 = CI2->getOperand(0);
5467 // If Op1 is a constant, we can fold the cast into the constant.
5468 if (Op0->getType() != Op1->getType()) {
5469 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5470 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5472 // Otherwise, cast the RHS right before the icmp
5473 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5476 return new ICmpInst(I.getPredicate(), Op0, Op1);
5480 if (isa<CastInst>(Op0)) {
5481 // Handle the special case of: icmp (cast bool to X), <cst>
5482 // This comes up when you have code like
5485 // For generality, we handle any zero-extension of any operand comparison
5486 // with a constant or another cast from the same type.
5487 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5488 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5492 // ~x < ~y --> y < x
5494 if (match(Op0, m_Not(m_Value(A))) &&
5495 match(Op1, m_Not(m_Value(B))))
5496 return new ICmpInst(I.getPredicate(), B, A);
5499 if (I.isEquality()) {
5500 Value *A, *B, *C, *D;
5502 // -x == -y --> x == y
5503 if (match(Op0, m_Neg(m_Value(A))) &&
5504 match(Op1, m_Neg(m_Value(B))))
5505 return new ICmpInst(I.getPredicate(), A, B);
5507 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5508 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5509 Value *OtherVal = A == Op1 ? B : A;
5510 return new ICmpInst(I.getPredicate(), OtherVal,
5511 Constant::getNullValue(A->getType()));
5514 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5515 // A^c1 == C^c2 --> A == C^(c1^c2)
5516 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5517 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5518 if (Op1->hasOneUse()) {
5519 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5520 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5521 return new ICmpInst(I.getPredicate(), A,
5522 InsertNewInstBefore(Xor, I));
5525 // A^B == A^D -> B == D
5526 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5527 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5528 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5529 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5533 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5534 (A == Op0 || B == Op0)) {
5535 // A == (A^B) -> B == 0
5536 Value *OtherVal = A == Op0 ? B : A;
5537 return new ICmpInst(I.getPredicate(), OtherVal,
5538 Constant::getNullValue(A->getType()));
5540 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5541 // (A-B) == A -> B == 0
5542 return new ICmpInst(I.getPredicate(), B,
5543 Constant::getNullValue(B->getType()));
5545 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5546 // A == (A-B) -> B == 0
5547 return new ICmpInst(I.getPredicate(), B,
5548 Constant::getNullValue(B->getType()));
5551 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5552 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5553 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5554 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5555 Value *X = 0, *Y = 0, *Z = 0;
5558 X = B; Y = D; Z = A;
5559 } else if (A == D) {
5560 X = B; Y = C; Z = A;
5561 } else if (B == C) {
5562 X = A; Y = D; Z = B;
5563 } else if (B == D) {
5564 X = A; Y = C; Z = B;
5567 if (X) { // Build (X^Y) & Z
5568 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
5569 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
5570 I.setOperand(0, Op1);
5571 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5576 return Changed ? &I : 0;
5580 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5581 /// and CmpRHS are both known to be integer constants.
5582 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5583 ConstantInt *DivRHS) {
5584 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5585 const APInt &CmpRHSV = CmpRHS->getValue();
5587 // FIXME: If the operand types don't match the type of the divide
5588 // then don't attempt this transform. The code below doesn't have the
5589 // logic to deal with a signed divide and an unsigned compare (and
5590 // vice versa). This is because (x /s C1) <s C2 produces different
5591 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5592 // (x /u C1) <u C2. Simply casting the operands and result won't
5593 // work. :( The if statement below tests that condition and bails
5595 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5596 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5598 if (DivRHS->isZero())
5599 return 0; // The ProdOV computation fails on divide by zero.
5601 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5602 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5603 // C2 (CI). By solving for X we can turn this into a range check
5604 // instead of computing a divide.
5605 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5607 // Determine if the product overflows by seeing if the product is
5608 // not equal to the divide. Make sure we do the same kind of divide
5609 // as in the LHS instruction that we're folding.
5610 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5611 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5613 // Get the ICmp opcode
5614 ICmpInst::Predicate Pred = ICI.getPredicate();
5616 // Figure out the interval that is being checked. For example, a comparison
5617 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5618 // Compute this interval based on the constants involved and the signedness of
5619 // the compare/divide. This computes a half-open interval, keeping track of
5620 // whether either value in the interval overflows. After analysis each
5621 // overflow variable is set to 0 if it's corresponding bound variable is valid
5622 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5623 int LoOverflow = 0, HiOverflow = 0;
5624 ConstantInt *LoBound = 0, *HiBound = 0;
5627 if (!DivIsSigned) { // udiv
5628 // e.g. X/5 op 3 --> [15, 20)
5630 HiOverflow = LoOverflow = ProdOV;
5632 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5633 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5634 if (CmpRHSV == 0) { // (X / pos) op 0
5635 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5636 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5638 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5639 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5640 HiOverflow = LoOverflow = ProdOV;
5642 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5643 } else { // (X / pos) op neg
5644 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5645 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5646 LoOverflow = AddWithOverflow(LoBound, Prod,
5647 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5648 HiBound = AddOne(Prod);
5649 HiOverflow = ProdOV ? -1 : 0;
5651 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5652 if (CmpRHSV == 0) { // (X / neg) op 0
5653 // e.g. X/-5 op 0 --> [-4, 5)
5654 LoBound = AddOne(DivRHS);
5655 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5656 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5657 HiOverflow = 1; // [INTMIN+1, overflow)
5658 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5660 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5661 // e.g. X/-5 op 3 --> [-19, -14)
5662 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5664 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5665 HiBound = AddOne(Prod);
5666 } else { // (X / neg) op neg
5667 // e.g. X/-5 op -3 --> [15, 20)
5669 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5670 HiBound = Subtract(Prod, DivRHS);
5673 // Dividing by a negative swaps the condition. LT <-> GT
5674 Pred = ICmpInst::getSwappedPredicate(Pred);
5677 Value *X = DivI->getOperand(0);
5679 default: assert(0 && "Unhandled icmp opcode!");
5680 case ICmpInst::ICMP_EQ:
5681 if (LoOverflow && HiOverflow)
5682 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5683 else if (HiOverflow)
5684 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5685 ICmpInst::ICMP_UGE, X, LoBound);
5686 else if (LoOverflow)
5687 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5688 ICmpInst::ICMP_ULT, X, HiBound);
5690 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5691 case ICmpInst::ICMP_NE:
5692 if (LoOverflow && HiOverflow)
5693 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5694 else if (HiOverflow)
5695 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5696 ICmpInst::ICMP_ULT, X, LoBound);
5697 else if (LoOverflow)
5698 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5699 ICmpInst::ICMP_UGE, X, HiBound);
5701 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5702 case ICmpInst::ICMP_ULT:
5703 case ICmpInst::ICMP_SLT:
5704 if (LoOverflow == +1) // Low bound is greater than input range.
5705 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5706 if (LoOverflow == -1) // Low bound is less than input range.
5707 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5708 return new ICmpInst(Pred, X, LoBound);
5709 case ICmpInst::ICMP_UGT:
5710 case ICmpInst::ICMP_SGT:
5711 if (HiOverflow == +1) // High bound greater than input range.
5712 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5713 else if (HiOverflow == -1) // High bound less than input range.
5714 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5715 if (Pred == ICmpInst::ICMP_UGT)
5716 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5718 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5723 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5725 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5728 const APInt &RHSV = RHS->getValue();
5730 switch (LHSI->getOpcode()) {
5731 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5732 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5733 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5735 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5736 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5737 Value *CompareVal = LHSI->getOperand(0);
5739 // If the sign bit of the XorCST is not set, there is no change to
5740 // the operation, just stop using the Xor.
5741 if (!XorCST->getValue().isNegative()) {
5742 ICI.setOperand(0, CompareVal);
5743 AddToWorkList(LHSI);
5747 // Was the old condition true if the operand is positive?
5748 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5750 // If so, the new one isn't.
5751 isTrueIfPositive ^= true;
5753 if (isTrueIfPositive)
5754 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5756 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5760 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5761 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5762 LHSI->getOperand(0)->hasOneUse()) {
5763 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5765 // If the LHS is an AND of a truncating cast, we can widen the
5766 // and/compare to be the input width without changing the value
5767 // produced, eliminating a cast.
5768 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5769 // We can do this transformation if either the AND constant does not
5770 // have its sign bit set or if it is an equality comparison.
5771 // Extending a relational comparison when we're checking the sign
5772 // bit would not work.
5773 if (Cast->hasOneUse() &&
5774 (ICI.isEquality() ||
5775 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5777 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5778 APInt NewCST = AndCST->getValue();
5779 NewCST.zext(BitWidth);
5781 NewCI.zext(BitWidth);
5782 Instruction *NewAnd =
5783 BinaryOperator::CreateAnd(Cast->getOperand(0),
5784 ConstantInt::get(NewCST),LHSI->getName());
5785 InsertNewInstBefore(NewAnd, ICI);
5786 return new ICmpInst(ICI.getPredicate(), NewAnd,
5787 ConstantInt::get(NewCI));
5791 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5792 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5793 // happens a LOT in code produced by the C front-end, for bitfield
5795 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5796 if (Shift && !Shift->isShift())
5800 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5801 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5802 const Type *AndTy = AndCST->getType(); // Type of the and.
5804 // We can fold this as long as we can't shift unknown bits
5805 // into the mask. This can only happen with signed shift
5806 // rights, as they sign-extend.
5808 bool CanFold = Shift->isLogicalShift();
5810 // To test for the bad case of the signed shr, see if any
5811 // of the bits shifted in could be tested after the mask.
5812 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5813 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5815 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5816 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5817 AndCST->getValue()) == 0)
5823 if (Shift->getOpcode() == Instruction::Shl)
5824 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5826 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5828 // Check to see if we are shifting out any of the bits being
5830 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5831 // If we shifted bits out, the fold is not going to work out.
5832 // As a special case, check to see if this means that the
5833 // result is always true or false now.
5834 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5835 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5836 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5837 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5839 ICI.setOperand(1, NewCst);
5840 Constant *NewAndCST;
5841 if (Shift->getOpcode() == Instruction::Shl)
5842 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5844 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5845 LHSI->setOperand(1, NewAndCST);
5846 LHSI->setOperand(0, Shift->getOperand(0));
5847 AddToWorkList(Shift); // Shift is dead.
5848 AddUsesToWorkList(ICI);
5854 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5855 // preferable because it allows the C<<Y expression to be hoisted out
5856 // of a loop if Y is invariant and X is not.
5857 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5858 ICI.isEquality() && !Shift->isArithmeticShift() &&
5859 isa<Instruction>(Shift->getOperand(0))) {
5862 if (Shift->getOpcode() == Instruction::LShr) {
5863 NS = BinaryOperator::CreateShl(AndCST,
5864 Shift->getOperand(1), "tmp");
5866 // Insert a logical shift.
5867 NS = BinaryOperator::CreateLShr(AndCST,
5868 Shift->getOperand(1), "tmp");
5870 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5872 // Compute X & (C << Y).
5873 Instruction *NewAnd =
5874 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
5875 InsertNewInstBefore(NewAnd, ICI);
5877 ICI.setOperand(0, NewAnd);
5883 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5884 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5887 uint32_t TypeBits = RHSV.getBitWidth();
5889 // Check that the shift amount is in range. If not, don't perform
5890 // undefined shifts. When the shift is visited it will be
5892 if (ShAmt->uge(TypeBits))
5895 if (ICI.isEquality()) {
5896 // If we are comparing against bits always shifted out, the
5897 // comparison cannot succeed.
5899 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5900 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5901 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5902 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5903 return ReplaceInstUsesWith(ICI, Cst);
5906 if (LHSI->hasOneUse()) {
5907 // Otherwise strength reduce the shift into an and.
5908 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5910 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5913 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5914 Mask, LHSI->getName()+".mask");
5915 Value *And = InsertNewInstBefore(AndI, ICI);
5916 return new ICmpInst(ICI.getPredicate(), And,
5917 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5921 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5922 bool TrueIfSigned = false;
5923 if (LHSI->hasOneUse() &&
5924 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5925 // (X << 31) <s 0 --> (X&1) != 0
5926 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5927 (TypeBits-ShAmt->getZExtValue()-1));
5929 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5930 Mask, LHSI->getName()+".mask");
5931 Value *And = InsertNewInstBefore(AndI, ICI);
5933 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5934 And, Constant::getNullValue(And->getType()));
5939 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5940 case Instruction::AShr: {
5941 // Only handle equality comparisons of shift-by-constant.
5942 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5943 if (!ShAmt || !ICI.isEquality()) break;
5945 // Check that the shift amount is in range. If not, don't perform
5946 // undefined shifts. When the shift is visited it will be
5948 uint32_t TypeBits = RHSV.getBitWidth();
5949 if (ShAmt->uge(TypeBits))
5952 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5954 // If we are comparing against bits always shifted out, the
5955 // comparison cannot succeed.
5956 APInt Comp = RHSV << ShAmtVal;
5957 if (LHSI->getOpcode() == Instruction::LShr)
5958 Comp = Comp.lshr(ShAmtVal);
5960 Comp = Comp.ashr(ShAmtVal);
5962 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5963 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5964 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5965 return ReplaceInstUsesWith(ICI, Cst);
5968 // Otherwise, check to see if the bits shifted out are known to be zero.
5969 // If so, we can compare against the unshifted value:
5970 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
5971 if (LHSI->hasOneUse() &&
5972 MaskedValueIsZero(LHSI->getOperand(0),
5973 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
5974 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
5975 ConstantExpr::getShl(RHS, ShAmt));
5978 if (LHSI->hasOneUse()) {
5979 // Otherwise strength reduce the shift into an and.
5980 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5981 Constant *Mask = ConstantInt::get(Val);
5984 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5985 Mask, LHSI->getName()+".mask");
5986 Value *And = InsertNewInstBefore(AndI, ICI);
5987 return new ICmpInst(ICI.getPredicate(), And,
5988 ConstantExpr::getShl(RHS, ShAmt));
5993 case Instruction::SDiv:
5994 case Instruction::UDiv:
5995 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5996 // Fold this div into the comparison, producing a range check.
5997 // Determine, based on the divide type, what the range is being
5998 // checked. If there is an overflow on the low or high side, remember
5999 // it, otherwise compute the range [low, hi) bounding the new value.
6000 // See: InsertRangeTest above for the kinds of replacements possible.
6001 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6002 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6007 case Instruction::Add:
6008 // Fold: icmp pred (add, X, C1), C2
6010 if (!ICI.isEquality()) {
6011 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6013 const APInt &LHSV = LHSC->getValue();
6015 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6018 if (ICI.isSignedPredicate()) {
6019 if (CR.getLower().isSignBit()) {
6020 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6021 ConstantInt::get(CR.getUpper()));
6022 } else if (CR.getUpper().isSignBit()) {
6023 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6024 ConstantInt::get(CR.getLower()));
6027 if (CR.getLower().isMinValue()) {
6028 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6029 ConstantInt::get(CR.getUpper()));
6030 } else if (CR.getUpper().isMinValue()) {
6031 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6032 ConstantInt::get(CR.getLower()));
6039 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6040 if (ICI.isEquality()) {
6041 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6043 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6044 // the second operand is a constant, simplify a bit.
6045 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6046 switch (BO->getOpcode()) {
6047 case Instruction::SRem:
6048 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6049 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6050 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6051 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6052 Instruction *NewRem =
6053 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6055 InsertNewInstBefore(NewRem, ICI);
6056 return new ICmpInst(ICI.getPredicate(), NewRem,
6057 Constant::getNullValue(BO->getType()));
6061 case Instruction::Add:
6062 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6063 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6064 if (BO->hasOneUse())
6065 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6066 Subtract(RHS, BOp1C));
6067 } else if (RHSV == 0) {
6068 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6069 // efficiently invertible, or if the add has just this one use.
6070 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6072 if (Value *NegVal = dyn_castNegVal(BOp1))
6073 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6074 else if (Value *NegVal = dyn_castNegVal(BOp0))
6075 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6076 else if (BO->hasOneUse()) {
6077 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6078 InsertNewInstBefore(Neg, ICI);
6080 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6084 case Instruction::Xor:
6085 // For the xor case, we can xor two constants together, eliminating
6086 // the explicit xor.
6087 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6088 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6089 ConstantExpr::getXor(RHS, BOC));
6092 case Instruction::Sub:
6093 // Replace (([sub|xor] A, B) != 0) with (A != B)
6095 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6099 case Instruction::Or:
6100 // If bits are being or'd in that are not present in the constant we
6101 // are comparing against, then the comparison could never succeed!
6102 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6103 Constant *NotCI = ConstantExpr::getNot(RHS);
6104 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6105 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6110 case Instruction::And:
6111 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6112 // If bits are being compared against that are and'd out, then the
6113 // comparison can never succeed!
6114 if ((RHSV & ~BOC->getValue()) != 0)
6115 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6118 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6119 if (RHS == BOC && RHSV.isPowerOf2())
6120 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6121 ICmpInst::ICMP_NE, LHSI,
6122 Constant::getNullValue(RHS->getType()));
6124 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6125 if (BOC->getValue().isSignBit()) {
6126 Value *X = BO->getOperand(0);
6127 Constant *Zero = Constant::getNullValue(X->getType());
6128 ICmpInst::Predicate pred = isICMP_NE ?
6129 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6130 return new ICmpInst(pred, X, Zero);
6133 // ((X & ~7) == 0) --> X < 8
6134 if (RHSV == 0 && isHighOnes(BOC)) {
6135 Value *X = BO->getOperand(0);
6136 Constant *NegX = ConstantExpr::getNeg(BOC);
6137 ICmpInst::Predicate pred = isICMP_NE ?
6138 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6139 return new ICmpInst(pred, X, NegX);
6144 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6145 // Handle icmp {eq|ne} <intrinsic>, intcst.
6146 if (II->getIntrinsicID() == Intrinsic::bswap) {
6148 ICI.setOperand(0, II->getOperand(1));
6149 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6153 } else { // Not a ICMP_EQ/ICMP_NE
6154 // If the LHS is a cast from an integral value of the same size,
6155 // then since we know the RHS is a constant, try to simlify.
6156 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6157 Value *CastOp = Cast->getOperand(0);
6158 const Type *SrcTy = CastOp->getType();
6159 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6160 if (SrcTy->isInteger() &&
6161 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6162 // If this is an unsigned comparison, try to make the comparison use
6163 // smaller constant values.
6164 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6165 // X u< 128 => X s> -1
6166 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6167 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6168 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6169 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6170 // X u> 127 => X s< 0
6171 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6172 Constant::getNullValue(SrcTy));
6180 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6181 /// We only handle extending casts so far.
6183 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6184 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6185 Value *LHSCIOp = LHSCI->getOperand(0);
6186 const Type *SrcTy = LHSCIOp->getType();
6187 const Type *DestTy = LHSCI->getType();
6190 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6191 // integer type is the same size as the pointer type.
6192 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6193 getTargetData().getPointerSizeInBits() ==
6194 cast<IntegerType>(DestTy)->getBitWidth()) {
6196 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6197 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6198 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6199 RHSOp = RHSC->getOperand(0);
6200 // If the pointer types don't match, insert a bitcast.
6201 if (LHSCIOp->getType() != RHSOp->getType())
6202 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6206 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6209 // The code below only handles extension cast instructions, so far.
6211 if (LHSCI->getOpcode() != Instruction::ZExt &&
6212 LHSCI->getOpcode() != Instruction::SExt)
6215 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6216 bool isSignedCmp = ICI.isSignedPredicate();
6218 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6219 // Not an extension from the same type?
6220 RHSCIOp = CI->getOperand(0);
6221 if (RHSCIOp->getType() != LHSCIOp->getType())
6224 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6225 // and the other is a zext), then we can't handle this.
6226 if (CI->getOpcode() != LHSCI->getOpcode())
6229 // Deal with equality cases early.
6230 if (ICI.isEquality())
6231 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6233 // A signed comparison of sign extended values simplifies into a
6234 // signed comparison.
6235 if (isSignedCmp && isSignedExt)
6236 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6238 // The other three cases all fold into an unsigned comparison.
6239 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6242 // If we aren't dealing with a constant on the RHS, exit early
6243 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6247 // Compute the constant that would happen if we truncated to SrcTy then
6248 // reextended to DestTy.
6249 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6250 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6252 // If the re-extended constant didn't change...
6254 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6255 // For example, we might have:
6256 // %A = sext short %X to uint
6257 // %B = icmp ugt uint %A, 1330
6258 // It is incorrect to transform this into
6259 // %B = icmp ugt short %X, 1330
6260 // because %A may have negative value.
6262 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6263 // OR operation is EQ/NE.
6264 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6265 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6270 // The re-extended constant changed so the constant cannot be represented
6271 // in the shorter type. Consequently, we cannot emit a simple comparison.
6273 // First, handle some easy cases. We know the result cannot be equal at this
6274 // point so handle the ICI.isEquality() cases
6275 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6276 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6277 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6278 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6280 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6281 // should have been folded away previously and not enter in here.
6284 // We're performing a signed comparison.
6285 if (cast<ConstantInt>(CI)->getValue().isNegative())
6286 Result = ConstantInt::getFalse(); // X < (small) --> false
6288 Result = ConstantInt::getTrue(); // X < (large) --> true
6290 // We're performing an unsigned comparison.
6292 // We're performing an unsigned comp with a sign extended value.
6293 // This is true if the input is >= 0. [aka >s -1]
6294 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6295 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6296 NegOne, ICI.getName()), ICI);
6298 // Unsigned extend & unsigned compare -> always true.
6299 Result = ConstantInt::getTrue();
6303 // Finally, return the value computed.
6304 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6305 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6306 return ReplaceInstUsesWith(ICI, Result);
6308 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6309 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6310 "ICmp should be folded!");
6311 if (Constant *CI = dyn_cast<Constant>(Result))
6312 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6314 return BinaryOperator::CreateNot(Result);
6318 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6319 return commonShiftTransforms(I);
6322 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6323 return commonShiftTransforms(I);
6326 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6327 if (Instruction *R = commonShiftTransforms(I))
6330 Value *Op0 = I.getOperand(0);
6332 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6333 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6334 if (CSI->isAllOnesValue())
6335 return ReplaceInstUsesWith(I, CSI);
6337 // See if we can turn a signed shr into an unsigned shr.
6338 if (MaskedValueIsZero(Op0,
6339 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6340 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6345 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6346 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6347 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6349 // shl X, 0 == X and shr X, 0 == X
6350 // shl 0, X == 0 and shr 0, X == 0
6351 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6352 Op0 == Constant::getNullValue(Op0->getType()))
6353 return ReplaceInstUsesWith(I, Op0);
6355 if (isa<UndefValue>(Op0)) {
6356 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6357 return ReplaceInstUsesWith(I, Op0);
6358 else // undef << X -> 0, undef >>u X -> 0
6359 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6361 if (isa<UndefValue>(Op1)) {
6362 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6363 return ReplaceInstUsesWith(I, Op0);
6364 else // X << undef, X >>u undef -> 0
6365 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6368 // Try to fold constant and into select arguments.
6369 if (isa<Constant>(Op0))
6370 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6371 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6374 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6375 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6380 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6381 BinaryOperator &I) {
6382 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6384 // See if we can simplify any instructions used by the instruction whose sole
6385 // purpose is to compute bits we don't care about.
6386 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6387 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6388 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6389 KnownZero, KnownOne))
6392 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6393 // of a signed value.
6395 if (Op1->uge(TypeBits)) {
6396 if (I.getOpcode() != Instruction::AShr)
6397 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6399 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6404 // ((X*C1) << C2) == (X * (C1 << C2))
6405 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6406 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6407 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6408 return BinaryOperator::CreateMul(BO->getOperand(0),
6409 ConstantExpr::getShl(BOOp, Op1));
6411 // Try to fold constant and into select arguments.
6412 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6413 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6415 if (isa<PHINode>(Op0))
6416 if (Instruction *NV = FoldOpIntoPhi(I))
6419 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6420 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6421 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6422 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6423 // place. Don't try to do this transformation in this case. Also, we
6424 // require that the input operand is a shift-by-constant so that we have
6425 // confidence that the shifts will get folded together. We could do this
6426 // xform in more cases, but it is unlikely to be profitable.
6427 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6428 isa<ConstantInt>(TrOp->getOperand(1))) {
6429 // Okay, we'll do this xform. Make the shift of shift.
6430 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6431 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6433 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6435 // For logical shifts, the truncation has the effect of making the high
6436 // part of the register be zeros. Emulate this by inserting an AND to
6437 // clear the top bits as needed. This 'and' will usually be zapped by
6438 // other xforms later if dead.
6439 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6440 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6441 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6443 // The mask we constructed says what the trunc would do if occurring
6444 // between the shifts. We want to know the effect *after* the second
6445 // shift. We know that it is a logical shift by a constant, so adjust the
6446 // mask as appropriate.
6447 if (I.getOpcode() == Instruction::Shl)
6448 MaskV <<= Op1->getZExtValue();
6450 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6451 MaskV = MaskV.lshr(Op1->getZExtValue());
6454 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6456 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6458 // Return the value truncated to the interesting size.
6459 return new TruncInst(And, I.getType());
6463 if (Op0->hasOneUse()) {
6464 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6465 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6468 switch (Op0BO->getOpcode()) {
6470 case Instruction::Add:
6471 case Instruction::And:
6472 case Instruction::Or:
6473 case Instruction::Xor: {
6474 // These operators commute.
6475 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6476 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6477 match(Op0BO->getOperand(1),
6478 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6479 Instruction *YS = BinaryOperator::CreateShl(
6480 Op0BO->getOperand(0), Op1,
6482 InsertNewInstBefore(YS, I); // (Y << C)
6484 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6485 Op0BO->getOperand(1)->getName());
6486 InsertNewInstBefore(X, I); // (X + (Y << C))
6487 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6488 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6489 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6492 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6493 Value *Op0BOOp1 = Op0BO->getOperand(1);
6494 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6496 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6497 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6499 Instruction *YS = BinaryOperator::CreateShl(
6500 Op0BO->getOperand(0), Op1,
6502 InsertNewInstBefore(YS, I); // (Y << C)
6504 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6505 V1->getName()+".mask");
6506 InsertNewInstBefore(XM, I); // X & (CC << C)
6508 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6513 case Instruction::Sub: {
6514 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6515 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6516 match(Op0BO->getOperand(0),
6517 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6518 Instruction *YS = BinaryOperator::CreateShl(
6519 Op0BO->getOperand(1), Op1,
6521 InsertNewInstBefore(YS, I); // (Y << C)
6523 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6524 Op0BO->getOperand(0)->getName());
6525 InsertNewInstBefore(X, I); // (X + (Y << C))
6526 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6527 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6528 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6531 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6532 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6533 match(Op0BO->getOperand(0),
6534 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6535 m_ConstantInt(CC))) && V2 == Op1 &&
6536 cast<BinaryOperator>(Op0BO->getOperand(0))
6537 ->getOperand(0)->hasOneUse()) {
6538 Instruction *YS = BinaryOperator::CreateShl(
6539 Op0BO->getOperand(1), Op1,
6541 InsertNewInstBefore(YS, I); // (Y << C)
6543 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6544 V1->getName()+".mask");
6545 InsertNewInstBefore(XM, I); // X & (CC << C)
6547 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
6555 // If the operand is an bitwise operator with a constant RHS, and the
6556 // shift is the only use, we can pull it out of the shift.
6557 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6558 bool isValid = true; // Valid only for And, Or, Xor
6559 bool highBitSet = false; // Transform if high bit of constant set?
6561 switch (Op0BO->getOpcode()) {
6562 default: isValid = false; break; // Do not perform transform!
6563 case Instruction::Add:
6564 isValid = isLeftShift;
6566 case Instruction::Or:
6567 case Instruction::Xor:
6570 case Instruction::And:
6575 // If this is a signed shift right, and the high bit is modified
6576 // by the logical operation, do not perform the transformation.
6577 // The highBitSet boolean indicates the value of the high bit of
6578 // the constant which would cause it to be modified for this
6581 if (isValid && I.getOpcode() == Instruction::AShr)
6582 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6585 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6587 Instruction *NewShift =
6588 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6589 InsertNewInstBefore(NewShift, I);
6590 NewShift->takeName(Op0BO);
6592 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
6599 // Find out if this is a shift of a shift by a constant.
6600 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6601 if (ShiftOp && !ShiftOp->isShift())
6604 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6605 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6606 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6607 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6608 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6609 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6610 Value *X = ShiftOp->getOperand(0);
6612 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6613 if (AmtSum > TypeBits)
6616 const IntegerType *Ty = cast<IntegerType>(I.getType());
6618 // Check for (X << c1) << c2 and (X >> c1) >> c2
6619 if (I.getOpcode() == ShiftOp->getOpcode()) {
6620 return BinaryOperator::Create(I.getOpcode(), X,
6621 ConstantInt::get(Ty, AmtSum));
6622 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6623 I.getOpcode() == Instruction::AShr) {
6624 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6625 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
6626 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6627 I.getOpcode() == Instruction::LShr) {
6628 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6629 Instruction *Shift =
6630 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
6631 InsertNewInstBefore(Shift, I);
6633 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6634 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6637 // Okay, if we get here, one shift must be left, and the other shift must be
6638 // right. See if the amounts are equal.
6639 if (ShiftAmt1 == ShiftAmt2) {
6640 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6641 if (I.getOpcode() == Instruction::Shl) {
6642 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6643 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6645 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6646 if (I.getOpcode() == Instruction::LShr) {
6647 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6648 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6650 // We can simplify ((X << C) >>s C) into a trunc + sext.
6651 // NOTE: we could do this for any C, but that would make 'unusual' integer
6652 // types. For now, just stick to ones well-supported by the code
6654 const Type *SExtType = 0;
6655 switch (Ty->getBitWidth() - ShiftAmt1) {
6662 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6667 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6668 InsertNewInstBefore(NewTrunc, I);
6669 return new SExtInst(NewTrunc, Ty);
6671 // Otherwise, we can't handle it yet.
6672 } else if (ShiftAmt1 < ShiftAmt2) {
6673 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6675 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6676 if (I.getOpcode() == Instruction::Shl) {
6677 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6678 ShiftOp->getOpcode() == Instruction::AShr);
6679 Instruction *Shift =
6680 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6681 InsertNewInstBefore(Shift, I);
6683 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6684 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6687 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6688 if (I.getOpcode() == Instruction::LShr) {
6689 assert(ShiftOp->getOpcode() == Instruction::Shl);
6690 Instruction *Shift =
6691 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
6692 InsertNewInstBefore(Shift, I);
6694 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6695 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6698 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6700 assert(ShiftAmt2 < ShiftAmt1);
6701 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6703 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6704 if (I.getOpcode() == Instruction::Shl) {
6705 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6706 ShiftOp->getOpcode() == Instruction::AShr);
6707 Instruction *Shift =
6708 BinaryOperator::Create(ShiftOp->getOpcode(), X,
6709 ConstantInt::get(Ty, ShiftDiff));
6710 InsertNewInstBefore(Shift, I);
6712 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6713 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6716 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6717 if (I.getOpcode() == Instruction::LShr) {
6718 assert(ShiftOp->getOpcode() == Instruction::Shl);
6719 Instruction *Shift =
6720 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6721 InsertNewInstBefore(Shift, I);
6723 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6724 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6727 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6734 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6735 /// expression. If so, decompose it, returning some value X, such that Val is
6738 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6740 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6741 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6742 Offset = CI->getZExtValue();
6744 return ConstantInt::get(Type::Int32Ty, 0);
6745 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6746 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6747 if (I->getOpcode() == Instruction::Shl) {
6748 // This is a value scaled by '1 << the shift amt'.
6749 Scale = 1U << RHS->getZExtValue();
6751 return I->getOperand(0);
6752 } else if (I->getOpcode() == Instruction::Mul) {
6753 // This value is scaled by 'RHS'.
6754 Scale = RHS->getZExtValue();
6756 return I->getOperand(0);
6757 } else if (I->getOpcode() == Instruction::Add) {
6758 // We have X+C. Check to see if we really have (X*C2)+C1,
6759 // where C1 is divisible by C2.
6762 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6763 Offset += RHS->getZExtValue();
6770 // Otherwise, we can't look past this.
6777 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6778 /// try to eliminate the cast by moving the type information into the alloc.
6779 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6780 AllocationInst &AI) {
6781 const PointerType *PTy = cast<PointerType>(CI.getType());
6783 // Remove any uses of AI that are dead.
6784 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6786 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6787 Instruction *User = cast<Instruction>(*UI++);
6788 if (isInstructionTriviallyDead(User)) {
6789 while (UI != E && *UI == User)
6790 ++UI; // If this instruction uses AI more than once, don't break UI.
6793 DOUT << "IC: DCE: " << *User;
6794 EraseInstFromFunction(*User);
6798 // Get the type really allocated and the type casted to.
6799 const Type *AllocElTy = AI.getAllocatedType();
6800 const Type *CastElTy = PTy->getElementType();
6801 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6803 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6804 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6805 if (CastElTyAlign < AllocElTyAlign) return 0;
6807 // If the allocation has multiple uses, only promote it if we are strictly
6808 // increasing the alignment of the resultant allocation. If we keep it the
6809 // same, we open the door to infinite loops of various kinds.
6810 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6812 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6813 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6814 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6816 // See if we can satisfy the modulus by pulling a scale out of the array
6818 unsigned ArraySizeScale;
6820 Value *NumElements = // See if the array size is a decomposable linear expr.
6821 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6823 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6825 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6826 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6828 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6833 // If the allocation size is constant, form a constant mul expression
6834 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6835 if (isa<ConstantInt>(NumElements))
6836 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6837 // otherwise multiply the amount and the number of elements
6838 else if (Scale != 1) {
6839 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
6840 Amt = InsertNewInstBefore(Tmp, AI);
6844 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6845 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6846 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
6847 Amt = InsertNewInstBefore(Tmp, AI);
6850 AllocationInst *New;
6851 if (isa<MallocInst>(AI))
6852 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6854 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6855 InsertNewInstBefore(New, AI);
6858 // If the allocation has multiple uses, insert a cast and change all things
6859 // that used it to use the new cast. This will also hack on CI, but it will
6861 if (!AI.hasOneUse()) {
6862 AddUsesToWorkList(AI);
6863 // New is the allocation instruction, pointer typed. AI is the original
6864 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6865 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6866 InsertNewInstBefore(NewCast, AI);
6867 AI.replaceAllUsesWith(NewCast);
6869 return ReplaceInstUsesWith(CI, New);
6872 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6873 /// and return it as type Ty without inserting any new casts and without
6874 /// changing the computed value. This is used by code that tries to decide
6875 /// whether promoting or shrinking integer operations to wider or smaller types
6876 /// will allow us to eliminate a truncate or extend.
6878 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6879 /// extension operation if Ty is larger.
6881 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
6882 /// should return true if trunc(V) can be computed by computing V in the smaller
6883 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
6884 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
6885 /// efficiently truncated.
6887 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
6888 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
6889 /// the final result.
6890 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6892 int &NumCastsRemoved) {
6893 // We can always evaluate constants in another type.
6894 if (isa<ConstantInt>(V))
6897 Instruction *I = dyn_cast<Instruction>(V);
6898 if (!I) return false;
6900 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6902 // If this is an extension or truncate, we can often eliminate it.
6903 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6904 // If this is a cast from the destination type, we can trivially eliminate
6905 // it, and this will remove a cast overall.
6906 if (I->getOperand(0)->getType() == Ty) {
6907 // If the first operand is itself a cast, and is eliminable, do not count
6908 // this as an eliminable cast. We would prefer to eliminate those two
6910 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
6916 // We can't extend or shrink something that has multiple uses: doing so would
6917 // require duplicating the instruction in general, which isn't profitable.
6918 if (!I->hasOneUse()) return false;
6920 switch (I->getOpcode()) {
6921 case Instruction::Add:
6922 case Instruction::Sub:
6923 case Instruction::And:
6924 case Instruction::Or:
6925 case Instruction::Xor:
6926 // These operators can all arbitrarily be extended or truncated.
6927 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6929 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6932 case Instruction::Mul:
6933 // A multiply can be truncated by truncating its operands.
6934 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6935 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6937 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6940 case Instruction::Shl:
6941 // If we are truncating the result of this SHL, and if it's a shift of a
6942 // constant amount, we can always perform a SHL in a smaller type.
6943 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6944 uint32_t BitWidth = Ty->getBitWidth();
6945 if (BitWidth < OrigTy->getBitWidth() &&
6946 CI->getLimitedValue(BitWidth) < BitWidth)
6947 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6951 case Instruction::LShr:
6952 // If this is a truncate of a logical shr, we can truncate it to a smaller
6953 // lshr iff we know that the bits we would otherwise be shifting in are
6955 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6956 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6957 uint32_t BitWidth = Ty->getBitWidth();
6958 if (BitWidth < OrigBitWidth &&
6959 MaskedValueIsZero(I->getOperand(0),
6960 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6961 CI->getLimitedValue(BitWidth) < BitWidth) {
6962 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6967 case Instruction::ZExt:
6968 case Instruction::SExt:
6969 case Instruction::Trunc:
6970 // If this is the same kind of case as our original (e.g. zext+zext), we
6971 // can safely replace it. Note that replacing it does not reduce the number
6972 // of casts in the input.
6973 if (I->getOpcode() == CastOpc)
6977 case Instruction::PHI: {
6978 // We can change a phi if we can change all operands.
6979 PHINode *PN = cast<PHINode>(I);
6980 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
6981 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
6987 // TODO: Can handle more cases here.
6994 /// EvaluateInDifferentType - Given an expression that
6995 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6996 /// evaluate the expression.
6997 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6999 if (Constant *C = dyn_cast<Constant>(V))
7000 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7002 // Otherwise, it must be an instruction.
7003 Instruction *I = cast<Instruction>(V);
7004 Instruction *Res = 0;
7005 switch (I->getOpcode()) {
7006 case Instruction::Add:
7007 case Instruction::Sub:
7008 case Instruction::Mul:
7009 case Instruction::And:
7010 case Instruction::Or:
7011 case Instruction::Xor:
7012 case Instruction::AShr:
7013 case Instruction::LShr:
7014 case Instruction::Shl: {
7015 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7016 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7017 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7021 case Instruction::Trunc:
7022 case Instruction::ZExt:
7023 case Instruction::SExt:
7024 // If the source type of the cast is the type we're trying for then we can
7025 // just return the source. There's no need to insert it because it is not
7027 if (I->getOperand(0)->getType() == Ty)
7028 return I->getOperand(0);
7030 // Otherwise, must be the same type of cast, so just reinsert a new one.
7031 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7034 case Instruction::PHI: {
7035 PHINode *OPN = cast<PHINode>(I);
7036 PHINode *NPN = PHINode::Create(Ty);
7037 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7038 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7039 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7045 // TODO: Can handle more cases here.
7046 assert(0 && "Unreachable!");
7051 return InsertNewInstBefore(Res, *I);
7054 /// @brief Implement the transforms common to all CastInst visitors.
7055 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7056 Value *Src = CI.getOperand(0);
7058 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7059 // eliminate it now.
7060 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7061 if (Instruction::CastOps opc =
7062 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7063 // The first cast (CSrc) is eliminable so we need to fix up or replace
7064 // the second cast (CI). CSrc will then have a good chance of being dead.
7065 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7069 // If we are casting a select then fold the cast into the select
7070 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7071 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7074 // If we are casting a PHI then fold the cast into the PHI
7075 if (isa<PHINode>(Src))
7076 if (Instruction *NV = FoldOpIntoPhi(CI))
7082 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7083 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7084 Value *Src = CI.getOperand(0);
7086 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7087 // If casting the result of a getelementptr instruction with no offset, turn
7088 // this into a cast of the original pointer!
7089 if (GEP->hasAllZeroIndices()) {
7090 // Changing the cast operand is usually not a good idea but it is safe
7091 // here because the pointer operand is being replaced with another
7092 // pointer operand so the opcode doesn't need to change.
7094 CI.setOperand(0, GEP->getOperand(0));
7098 // If the GEP has a single use, and the base pointer is a bitcast, and the
7099 // GEP computes a constant offset, see if we can convert these three
7100 // instructions into fewer. This typically happens with unions and other
7101 // non-type-safe code.
7102 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7103 if (GEP->hasAllConstantIndices()) {
7104 // We are guaranteed to get a constant from EmitGEPOffset.
7105 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7106 int64_t Offset = OffsetV->getSExtValue();
7108 // Get the base pointer input of the bitcast, and the type it points to.
7109 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7110 const Type *GEPIdxTy =
7111 cast<PointerType>(OrigBase->getType())->getElementType();
7112 if (GEPIdxTy->isSized()) {
7113 SmallVector<Value*, 8> NewIndices;
7115 // Start with the index over the outer type. Note that the type size
7116 // might be zero (even if the offset isn't zero) if the indexed type
7117 // is something like [0 x {int, int}]
7118 const Type *IntPtrTy = TD->getIntPtrType();
7119 int64_t FirstIdx = 0;
7120 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7121 FirstIdx = Offset/TySize;
7124 // Handle silly modulus not returning values values [0..TySize).
7128 assert(Offset >= 0);
7130 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7133 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7135 // Index into the types. If we fail, set OrigBase to null.
7137 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7138 const StructLayout *SL = TD->getStructLayout(STy);
7139 if (Offset < (int64_t)SL->getSizeInBytes()) {
7140 unsigned Elt = SL->getElementContainingOffset(Offset);
7141 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7143 Offset -= SL->getElementOffset(Elt);
7144 GEPIdxTy = STy->getElementType(Elt);
7146 // Otherwise, we can't index into this, bail out.
7150 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7151 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7152 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7153 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7156 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7158 GEPIdxTy = STy->getElementType();
7160 // Otherwise, we can't index into this, bail out.
7166 // If we were able to index down into an element, create the GEP
7167 // and bitcast the result. This eliminates one bitcast, potentially
7169 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7171 NewIndices.end(), "");
7172 InsertNewInstBefore(NGEP, CI);
7173 NGEP->takeName(GEP);
7175 if (isa<BitCastInst>(CI))
7176 return new BitCastInst(NGEP, CI.getType());
7177 assert(isa<PtrToIntInst>(CI));
7178 return new PtrToIntInst(NGEP, CI.getType());
7185 return commonCastTransforms(CI);
7190 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7191 /// integer types. This function implements the common transforms for all those
7193 /// @brief Implement the transforms common to CastInst with integer operands
7194 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7195 if (Instruction *Result = commonCastTransforms(CI))
7198 Value *Src = CI.getOperand(0);
7199 const Type *SrcTy = Src->getType();
7200 const Type *DestTy = CI.getType();
7201 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7202 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7204 // See if we can simplify any instructions used by the LHS whose sole
7205 // purpose is to compute bits we don't care about.
7206 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7207 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7208 KnownZero, KnownOne))
7211 // If the source isn't an instruction or has more than one use then we
7212 // can't do anything more.
7213 Instruction *SrcI = dyn_cast<Instruction>(Src);
7214 if (!SrcI || !Src->hasOneUse())
7217 // Attempt to propagate the cast into the instruction for int->int casts.
7218 int NumCastsRemoved = 0;
7219 if (!isa<BitCastInst>(CI) &&
7220 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7221 CI.getOpcode(), NumCastsRemoved)) {
7222 // If this cast is a truncate, evaluting in a different type always
7223 // eliminates the cast, so it is always a win. If this is a zero-extension,
7224 // we need to do an AND to maintain the clear top-part of the computation,
7225 // so we require that the input have eliminated at least one cast. If this
7226 // is a sign extension, we insert two new casts (to do the extension) so we
7227 // require that two casts have been eliminated.
7229 switch (CI.getOpcode()) {
7231 // All the others use floating point so we shouldn't actually
7232 // get here because of the check above.
7233 assert(0 && "Unknown cast type");
7234 case Instruction::Trunc:
7237 case Instruction::ZExt:
7238 DoXForm = NumCastsRemoved >= 1;
7240 case Instruction::SExt:
7241 DoXForm = NumCastsRemoved >= 2;
7246 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7247 CI.getOpcode() == Instruction::SExt);
7248 assert(Res->getType() == DestTy);
7249 switch (CI.getOpcode()) {
7250 default: assert(0 && "Unknown cast type!");
7251 case Instruction::Trunc:
7252 case Instruction::BitCast:
7253 // Just replace this cast with the result.
7254 return ReplaceInstUsesWith(CI, Res);
7255 case Instruction::ZExt: {
7256 // We need to emit an AND to clear the high bits.
7257 assert(SrcBitSize < DestBitSize && "Not a zext?");
7258 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7260 return BinaryOperator::CreateAnd(Res, C);
7262 case Instruction::SExt:
7263 // We need to emit a cast to truncate, then a cast to sext.
7264 return CastInst::Create(Instruction::SExt,
7265 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7271 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7272 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7274 switch (SrcI->getOpcode()) {
7275 case Instruction::Add:
7276 case Instruction::Mul:
7277 case Instruction::And:
7278 case Instruction::Or:
7279 case Instruction::Xor:
7280 // If we are discarding information, rewrite.
7281 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7282 // Don't insert two casts if they cannot be eliminated. We allow
7283 // two casts to be inserted if the sizes are the same. This could
7284 // only be converting signedness, which is a noop.
7285 if (DestBitSize == SrcBitSize ||
7286 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7287 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7288 Instruction::CastOps opcode = CI.getOpcode();
7289 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7290 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7291 return BinaryOperator::Create(
7292 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7296 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7297 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7298 SrcI->getOpcode() == Instruction::Xor &&
7299 Op1 == ConstantInt::getTrue() &&
7300 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7301 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7302 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7305 case Instruction::SDiv:
7306 case Instruction::UDiv:
7307 case Instruction::SRem:
7308 case Instruction::URem:
7309 // If we are just changing the sign, rewrite.
7310 if (DestBitSize == SrcBitSize) {
7311 // Don't insert two casts if they cannot be eliminated. We allow
7312 // two casts to be inserted if the sizes are the same. This could
7313 // only be converting signedness, which is a noop.
7314 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7315 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7316 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7318 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7320 return BinaryOperator::Create(
7321 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7326 case Instruction::Shl:
7327 // Allow changing the sign of the source operand. Do not allow
7328 // changing the size of the shift, UNLESS the shift amount is a
7329 // constant. We must not change variable sized shifts to a smaller
7330 // size, because it is undefined to shift more bits out than exist
7332 if (DestBitSize == SrcBitSize ||
7333 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7334 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7335 Instruction::BitCast : Instruction::Trunc);
7336 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7337 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7338 return BinaryOperator::CreateShl(Op0c, Op1c);
7341 case Instruction::AShr:
7342 // If this is a signed shr, and if all bits shifted in are about to be
7343 // truncated off, turn it into an unsigned shr to allow greater
7345 if (DestBitSize < SrcBitSize &&
7346 isa<ConstantInt>(Op1)) {
7347 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7348 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7349 // Insert the new logical shift right.
7350 return BinaryOperator::CreateLShr(Op0, Op1);
7358 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7359 if (Instruction *Result = commonIntCastTransforms(CI))
7362 Value *Src = CI.getOperand(0);
7363 const Type *Ty = CI.getType();
7364 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7365 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7367 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7368 switch (SrcI->getOpcode()) {
7370 case Instruction::LShr:
7371 // We can shrink lshr to something smaller if we know the bits shifted in
7372 // are already zeros.
7373 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7374 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7376 // Get a mask for the bits shifting in.
7377 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7378 Value* SrcIOp0 = SrcI->getOperand(0);
7379 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7380 if (ShAmt >= DestBitWidth) // All zeros.
7381 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7383 // Okay, we can shrink this. Truncate the input, then return a new
7385 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7386 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7388 return BinaryOperator::CreateLShr(V1, V2);
7390 } else { // This is a variable shr.
7392 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7393 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7394 // loop-invariant and CSE'd.
7395 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7396 Value *One = ConstantInt::get(SrcI->getType(), 1);
7398 Value *V = InsertNewInstBefore(
7399 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7401 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7402 SrcI->getOperand(0),
7404 Value *Zero = Constant::getNullValue(V->getType());
7405 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7415 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7416 /// in order to eliminate the icmp.
7417 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7419 // If we are just checking for a icmp eq of a single bit and zext'ing it
7420 // to an integer, then shift the bit to the appropriate place and then
7421 // cast to integer to avoid the comparison.
7422 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7423 const APInt &Op1CV = Op1C->getValue();
7425 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7426 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7427 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7428 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7429 if (!DoXform) return ICI;
7431 Value *In = ICI->getOperand(0);
7432 Value *Sh = ConstantInt::get(In->getType(),
7433 In->getType()->getPrimitiveSizeInBits()-1);
7434 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7435 In->getName()+".lobit"),
7437 if (In->getType() != CI.getType())
7438 In = CastInst::CreateIntegerCast(In, CI.getType(),
7439 false/*ZExt*/, "tmp", &CI);
7441 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7442 Constant *One = ConstantInt::get(In->getType(), 1);
7443 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7444 In->getName()+".not"),
7448 return ReplaceInstUsesWith(CI, In);
7453 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7454 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7455 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7456 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7457 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7458 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7459 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7460 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7461 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7462 // This only works for EQ and NE
7463 ICI->isEquality()) {
7464 // If Op1C some other power of two, convert:
7465 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7466 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7467 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7468 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7470 APInt KnownZeroMask(~KnownZero);
7471 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7472 if (!DoXform) return ICI;
7474 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7475 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7476 // (X&4) == 2 --> false
7477 // (X&4) != 2 --> true
7478 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7479 Res = ConstantExpr::getZExt(Res, CI.getType());
7480 return ReplaceInstUsesWith(CI, Res);
7483 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7484 Value *In = ICI->getOperand(0);
7486 // Perform a logical shr by shiftamt.
7487 // Insert the shift to put the result in the low bit.
7488 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7489 ConstantInt::get(In->getType(), ShiftAmt),
7490 In->getName()+".lobit"), CI);
7493 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7494 Constant *One = ConstantInt::get(In->getType(), 1);
7495 In = BinaryOperator::CreateXor(In, One, "tmp");
7496 InsertNewInstBefore(cast<Instruction>(In), CI);
7499 if (CI.getType() == In->getType())
7500 return ReplaceInstUsesWith(CI, In);
7502 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7510 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7511 // If one of the common conversion will work ..
7512 if (Instruction *Result = commonIntCastTransforms(CI))
7515 Value *Src = CI.getOperand(0);
7517 // If this is a cast of a cast
7518 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7519 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7520 // types and if the sizes are just right we can convert this into a logical
7521 // 'and' which will be much cheaper than the pair of casts.
7522 if (isa<TruncInst>(CSrc)) {
7523 // Get the sizes of the types involved
7524 Value *A = CSrc->getOperand(0);
7525 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7526 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7527 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7528 // If we're actually extending zero bits and the trunc is a no-op
7529 if (MidSize < DstSize && SrcSize == DstSize) {
7530 // Replace both of the casts with an And of the type mask.
7531 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7532 Constant *AndConst = ConstantInt::get(AndValue);
7534 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
7535 // Unfortunately, if the type changed, we need to cast it back.
7536 if (And->getType() != CI.getType()) {
7537 And->setName(CSrc->getName()+".mask");
7538 InsertNewInstBefore(And, CI);
7539 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
7546 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7547 return transformZExtICmp(ICI, CI);
7549 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7550 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7551 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7552 // of the (zext icmp) will be transformed.
7553 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7554 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7555 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7556 (transformZExtICmp(LHS, CI, false) ||
7557 transformZExtICmp(RHS, CI, false))) {
7558 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7559 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7560 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
7567 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7568 if (Instruction *I = commonIntCastTransforms(CI))
7571 Value *Src = CI.getOperand(0);
7573 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7574 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7575 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7576 // If we are just checking for a icmp eq of a single bit and zext'ing it
7577 // to an integer, then shift the bit to the appropriate place and then
7578 // cast to integer to avoid the comparison.
7579 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7580 const APInt &Op1CV = Op1C->getValue();
7582 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7583 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7584 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7585 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7586 Value *In = ICI->getOperand(0);
7587 Value *Sh = ConstantInt::get(In->getType(),
7588 In->getType()->getPrimitiveSizeInBits()-1);
7589 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
7590 In->getName()+".lobit"),
7592 if (In->getType() != CI.getType())
7593 In = CastInst::CreateIntegerCast(In, CI.getType(),
7594 true/*SExt*/, "tmp", &CI);
7596 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7597 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
7598 In->getName()+".not"), CI);
7600 return ReplaceInstUsesWith(CI, In);
7605 // See if the value being truncated is already sign extended. If so, just
7606 // eliminate the trunc/sext pair.
7607 if (getOpcode(Src) == Instruction::Trunc) {
7608 Value *Op = cast<User>(Src)->getOperand(0);
7609 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
7610 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
7611 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
7612 unsigned NumSignBits = ComputeNumSignBits(Op);
7614 if (OpBits == DestBits) {
7615 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7616 // bits, it is already ready.
7617 if (NumSignBits > DestBits-MidBits)
7618 return ReplaceInstUsesWith(CI, Op);
7619 } else if (OpBits < DestBits) {
7620 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7621 // bits, just sext from i32.
7622 if (NumSignBits > OpBits-MidBits)
7623 return new SExtInst(Op, CI.getType(), "tmp");
7625 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7626 // bits, just truncate to i32.
7627 if (NumSignBits > OpBits-MidBits)
7628 return new TruncInst(Op, CI.getType(), "tmp");
7635 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7636 /// in the specified FP type without changing its value.
7637 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7638 APFloat F = CFP->getValueAPF();
7639 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7640 return ConstantFP::get(F);
7644 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7645 /// through it until we get the source value.
7646 static Value *LookThroughFPExtensions(Value *V) {
7647 if (Instruction *I = dyn_cast<Instruction>(V))
7648 if (I->getOpcode() == Instruction::FPExt)
7649 return LookThroughFPExtensions(I->getOperand(0));
7651 // If this value is a constant, return the constant in the smallest FP type
7652 // that can accurately represent it. This allows us to turn
7653 // (float)((double)X+2.0) into x+2.0f.
7654 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7655 if (CFP->getType() == Type::PPC_FP128Ty)
7656 return V; // No constant folding of this.
7657 // See if the value can be truncated to float and then reextended.
7658 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7660 if (CFP->getType() == Type::DoubleTy)
7661 return V; // Won't shrink.
7662 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7664 // Don't try to shrink to various long double types.
7670 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7671 if (Instruction *I = commonCastTransforms(CI))
7674 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7675 // smaller than the destination type, we can eliminate the truncate by doing
7676 // the add as the smaller type. This applies to add/sub/mul/div as well as
7677 // many builtins (sqrt, etc).
7678 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7679 if (OpI && OpI->hasOneUse()) {
7680 switch (OpI->getOpcode()) {
7682 case Instruction::Add:
7683 case Instruction::Sub:
7684 case Instruction::Mul:
7685 case Instruction::FDiv:
7686 case Instruction::FRem:
7687 const Type *SrcTy = OpI->getType();
7688 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7689 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7690 if (LHSTrunc->getType() != SrcTy &&
7691 RHSTrunc->getType() != SrcTy) {
7692 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7693 // If the source types were both smaller than the destination type of
7694 // the cast, do this xform.
7695 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7696 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7697 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7699 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7701 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7710 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7711 return commonCastTransforms(CI);
7714 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
7715 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
7716 // mantissa to accurately represent all values of X. For example, do not
7717 // do this with i64->float->i64.
7718 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
7719 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7720 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
7721 SrcI->getType()->getFPMantissaWidth())
7722 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7724 return commonCastTransforms(FI);
7727 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
7728 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
7729 // mantissa to accurately represent all values of X. For example, do not
7730 // do this with i64->float->i64.
7731 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
7732 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7733 (int)FI.getType()->getPrimitiveSizeInBits() <=
7734 SrcI->getType()->getFPMantissaWidth())
7735 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7737 return commonCastTransforms(FI);
7740 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7741 return commonCastTransforms(CI);
7744 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7745 return commonCastTransforms(CI);
7748 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7749 return commonPointerCastTransforms(CI);
7752 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7753 if (Instruction *I = commonCastTransforms(CI))
7756 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7757 if (!DestPointee->isSized()) return 0;
7759 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7762 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7763 m_ConstantInt(Cst)))) {
7764 // If the source and destination operands have the same type, see if this
7765 // is a single-index GEP.
7766 if (X->getType() == CI.getType()) {
7767 // Get the size of the pointee type.
7768 uint64_t Size = TD->getABITypeSize(DestPointee);
7770 // Convert the constant to intptr type.
7771 APInt Offset = Cst->getValue();
7772 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7774 // If Offset is evenly divisible by Size, we can do this xform.
7775 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7776 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7777 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7780 // TODO: Could handle other cases, e.g. where add is indexing into field of
7782 } else if (CI.getOperand(0)->hasOneUse() &&
7783 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7784 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7785 // "inttoptr+GEP" instead of "add+intptr".
7787 // Get the size of the pointee type.
7788 uint64_t Size = TD->getABITypeSize(DestPointee);
7790 // Convert the constant to intptr type.
7791 APInt Offset = Cst->getValue();
7792 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7794 // If Offset is evenly divisible by Size, we can do this xform.
7795 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7796 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7798 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7800 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7806 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7807 // If the operands are integer typed then apply the integer transforms,
7808 // otherwise just apply the common ones.
7809 Value *Src = CI.getOperand(0);
7810 const Type *SrcTy = Src->getType();
7811 const Type *DestTy = CI.getType();
7813 if (SrcTy->isInteger() && DestTy->isInteger()) {
7814 if (Instruction *Result = commonIntCastTransforms(CI))
7816 } else if (isa<PointerType>(SrcTy)) {
7817 if (Instruction *I = commonPointerCastTransforms(CI))
7820 if (Instruction *Result = commonCastTransforms(CI))
7825 // Get rid of casts from one type to the same type. These are useless and can
7826 // be replaced by the operand.
7827 if (DestTy == Src->getType())
7828 return ReplaceInstUsesWith(CI, Src);
7830 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7831 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7832 const Type *DstElTy = DstPTy->getElementType();
7833 const Type *SrcElTy = SrcPTy->getElementType();
7835 // If the address spaces don't match, don't eliminate the bitcast, which is
7836 // required for changing types.
7837 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7840 // If we are casting a malloc or alloca to a pointer to a type of the same
7841 // size, rewrite the allocation instruction to allocate the "right" type.
7842 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7843 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7846 // If the source and destination are pointers, and this cast is equivalent
7847 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7848 // This can enhance SROA and other transforms that want type-safe pointers.
7849 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7850 unsigned NumZeros = 0;
7851 while (SrcElTy != DstElTy &&
7852 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7853 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7854 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7858 // If we found a path from the src to dest, create the getelementptr now.
7859 if (SrcElTy == DstElTy) {
7860 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7861 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7862 ((Instruction*) NULL));
7866 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7867 if (SVI->hasOneUse()) {
7868 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7869 // a bitconvert to a vector with the same # elts.
7870 if (isa<VectorType>(DestTy) &&
7871 cast<VectorType>(DestTy)->getNumElements() ==
7872 SVI->getType()->getNumElements()) {
7874 // If either of the operands is a cast from CI.getType(), then
7875 // evaluating the shuffle in the casted destination's type will allow
7876 // us to eliminate at least one cast.
7877 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7878 Tmp->getOperand(0)->getType() == DestTy) ||
7879 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7880 Tmp->getOperand(0)->getType() == DestTy)) {
7881 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7882 SVI->getOperand(0), DestTy, &CI);
7883 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7884 SVI->getOperand(1), DestTy, &CI);
7885 // Return a new shuffle vector. Use the same element ID's, as we
7886 // know the vector types match #elts.
7887 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7895 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7897 /// %D = select %cond, %C, %A
7899 /// %C = select %cond, %B, 0
7902 /// Assuming that the specified instruction is an operand to the select, return
7903 /// a bitmask indicating which operands of this instruction are foldable if they
7904 /// equal the other incoming value of the select.
7906 static unsigned GetSelectFoldableOperands(Instruction *I) {
7907 switch (I->getOpcode()) {
7908 case Instruction::Add:
7909 case Instruction::Mul:
7910 case Instruction::And:
7911 case Instruction::Or:
7912 case Instruction::Xor:
7913 return 3; // Can fold through either operand.
7914 case Instruction::Sub: // Can only fold on the amount subtracted.
7915 case Instruction::Shl: // Can only fold on the shift amount.
7916 case Instruction::LShr:
7917 case Instruction::AShr:
7920 return 0; // Cannot fold
7924 /// GetSelectFoldableConstant - For the same transformation as the previous
7925 /// function, return the identity constant that goes into the select.
7926 static Constant *GetSelectFoldableConstant(Instruction *I) {
7927 switch (I->getOpcode()) {
7928 default: assert(0 && "This cannot happen!"); abort();
7929 case Instruction::Add:
7930 case Instruction::Sub:
7931 case Instruction::Or:
7932 case Instruction::Xor:
7933 case Instruction::Shl:
7934 case Instruction::LShr:
7935 case Instruction::AShr:
7936 return Constant::getNullValue(I->getType());
7937 case Instruction::And:
7938 return Constant::getAllOnesValue(I->getType());
7939 case Instruction::Mul:
7940 return ConstantInt::get(I->getType(), 1);
7944 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7945 /// have the same opcode and only one use each. Try to simplify this.
7946 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7948 if (TI->getNumOperands() == 1) {
7949 // If this is a non-volatile load or a cast from the same type,
7952 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7955 return 0; // unknown unary op.
7958 // Fold this by inserting a select from the input values.
7959 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
7960 FI->getOperand(0), SI.getName()+".v");
7961 InsertNewInstBefore(NewSI, SI);
7962 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
7966 // Only handle binary operators here.
7967 if (!isa<BinaryOperator>(TI))
7970 // Figure out if the operations have any operands in common.
7971 Value *MatchOp, *OtherOpT, *OtherOpF;
7973 if (TI->getOperand(0) == FI->getOperand(0)) {
7974 MatchOp = TI->getOperand(0);
7975 OtherOpT = TI->getOperand(1);
7976 OtherOpF = FI->getOperand(1);
7977 MatchIsOpZero = true;
7978 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7979 MatchOp = TI->getOperand(1);
7980 OtherOpT = TI->getOperand(0);
7981 OtherOpF = FI->getOperand(0);
7982 MatchIsOpZero = false;
7983 } else if (!TI->isCommutative()) {
7985 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7986 MatchOp = TI->getOperand(0);
7987 OtherOpT = TI->getOperand(1);
7988 OtherOpF = FI->getOperand(0);
7989 MatchIsOpZero = true;
7990 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7991 MatchOp = TI->getOperand(1);
7992 OtherOpT = TI->getOperand(0);
7993 OtherOpF = FI->getOperand(1);
7994 MatchIsOpZero = true;
7999 // If we reach here, they do have operations in common.
8000 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8001 OtherOpF, SI.getName()+".v");
8002 InsertNewInstBefore(NewSI, SI);
8004 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8006 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8008 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8010 assert(0 && "Shouldn't get here");
8014 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8015 Value *CondVal = SI.getCondition();
8016 Value *TrueVal = SI.getTrueValue();
8017 Value *FalseVal = SI.getFalseValue();
8019 // select true, X, Y -> X
8020 // select false, X, Y -> Y
8021 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8022 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8024 // select C, X, X -> X
8025 if (TrueVal == FalseVal)
8026 return ReplaceInstUsesWith(SI, TrueVal);
8028 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8029 return ReplaceInstUsesWith(SI, FalseVal);
8030 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8031 return ReplaceInstUsesWith(SI, TrueVal);
8032 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8033 if (isa<Constant>(TrueVal))
8034 return ReplaceInstUsesWith(SI, TrueVal);
8036 return ReplaceInstUsesWith(SI, FalseVal);
8039 if (SI.getType() == Type::Int1Ty) {
8040 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8041 if (C->getZExtValue()) {
8042 // Change: A = select B, true, C --> A = or B, C
8043 return BinaryOperator::CreateOr(CondVal, FalseVal);
8045 // Change: A = select B, false, C --> A = and !B, C
8047 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8048 "not."+CondVal->getName()), SI);
8049 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8051 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8052 if (C->getZExtValue() == false) {
8053 // Change: A = select B, C, false --> A = and B, C
8054 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8056 // Change: A = select B, C, true --> A = or !B, C
8058 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8059 "not."+CondVal->getName()), SI);
8060 return BinaryOperator::CreateOr(NotCond, TrueVal);
8064 // select a, b, a -> a&b
8065 // select a, a, b -> a|b
8066 if (CondVal == TrueVal)
8067 return BinaryOperator::CreateOr(CondVal, FalseVal);
8068 else if (CondVal == FalseVal)
8069 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8072 // Selecting between two integer constants?
8073 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8074 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8075 // select C, 1, 0 -> zext C to int
8076 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8077 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8078 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8079 // select C, 0, 1 -> zext !C to int
8081 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8082 "not."+CondVal->getName()), SI);
8083 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8086 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8088 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8090 // (x <s 0) ? -1 : 0 -> ashr x, 31
8091 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8092 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8093 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8094 // The comparison constant and the result are not neccessarily the
8095 // same width. Make an all-ones value by inserting a AShr.
8096 Value *X = IC->getOperand(0);
8097 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8098 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8099 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8101 InsertNewInstBefore(SRA, SI);
8103 // Finally, convert to the type of the select RHS. We figure out
8104 // if this requires a SExt, Trunc or BitCast based on the sizes.
8105 Instruction::CastOps opc = Instruction::BitCast;
8106 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8107 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8108 if (SRASize < SISize)
8109 opc = Instruction::SExt;
8110 else if (SRASize > SISize)
8111 opc = Instruction::Trunc;
8112 return CastInst::Create(opc, SRA, SI.getType());
8117 // If one of the constants is zero (we know they can't both be) and we
8118 // have an icmp instruction with zero, and we have an 'and' with the
8119 // non-constant value, eliminate this whole mess. This corresponds to
8120 // cases like this: ((X & 27) ? 27 : 0)
8121 if (TrueValC->isZero() || FalseValC->isZero())
8122 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8123 cast<Constant>(IC->getOperand(1))->isNullValue())
8124 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8125 if (ICA->getOpcode() == Instruction::And &&
8126 isa<ConstantInt>(ICA->getOperand(1)) &&
8127 (ICA->getOperand(1) == TrueValC ||
8128 ICA->getOperand(1) == FalseValC) &&
8129 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8130 // Okay, now we know that everything is set up, we just don't
8131 // know whether we have a icmp_ne or icmp_eq and whether the
8132 // true or false val is the zero.
8133 bool ShouldNotVal = !TrueValC->isZero();
8134 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8137 V = InsertNewInstBefore(BinaryOperator::Create(
8138 Instruction::Xor, V, ICA->getOperand(1)), SI);
8139 return ReplaceInstUsesWith(SI, V);
8144 // See if we are selecting two values based on a comparison of the two values.
8145 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8146 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8147 // Transform (X == Y) ? X : Y -> Y
8148 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8149 // This is not safe in general for floating point:
8150 // consider X== -0, Y== +0.
8151 // It becomes safe if either operand is a nonzero constant.
8152 ConstantFP *CFPt, *CFPf;
8153 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8154 !CFPt->getValueAPF().isZero()) ||
8155 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8156 !CFPf->getValueAPF().isZero()))
8157 return ReplaceInstUsesWith(SI, FalseVal);
8159 // Transform (X != Y) ? X : Y -> X
8160 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8161 return ReplaceInstUsesWith(SI, TrueVal);
8162 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8164 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8165 // Transform (X == Y) ? Y : X -> X
8166 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8167 // This is not safe in general for floating point:
8168 // consider X== -0, Y== +0.
8169 // It becomes safe if either operand is a nonzero constant.
8170 ConstantFP *CFPt, *CFPf;
8171 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8172 !CFPt->getValueAPF().isZero()) ||
8173 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8174 !CFPf->getValueAPF().isZero()))
8175 return ReplaceInstUsesWith(SI, FalseVal);
8177 // Transform (X != Y) ? Y : X -> Y
8178 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8179 return ReplaceInstUsesWith(SI, TrueVal);
8180 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8184 // See if we are selecting two values based on a comparison of the two values.
8185 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8186 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8187 // Transform (X == Y) ? X : Y -> Y
8188 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8189 return ReplaceInstUsesWith(SI, FalseVal);
8190 // Transform (X != Y) ? X : Y -> X
8191 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8192 return ReplaceInstUsesWith(SI, TrueVal);
8193 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8195 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8196 // Transform (X == Y) ? Y : X -> X
8197 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8198 return ReplaceInstUsesWith(SI, FalseVal);
8199 // Transform (X != Y) ? Y : X -> Y
8200 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8201 return ReplaceInstUsesWith(SI, TrueVal);
8202 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8206 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8207 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8208 if (TI->hasOneUse() && FI->hasOneUse()) {
8209 Instruction *AddOp = 0, *SubOp = 0;
8211 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8212 if (TI->getOpcode() == FI->getOpcode())
8213 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8216 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8217 // even legal for FP.
8218 if (TI->getOpcode() == Instruction::Sub &&
8219 FI->getOpcode() == Instruction::Add) {
8220 AddOp = FI; SubOp = TI;
8221 } else if (FI->getOpcode() == Instruction::Sub &&
8222 TI->getOpcode() == Instruction::Add) {
8223 AddOp = TI; SubOp = FI;
8227 Value *OtherAddOp = 0;
8228 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8229 OtherAddOp = AddOp->getOperand(1);
8230 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8231 OtherAddOp = AddOp->getOperand(0);
8235 // So at this point we know we have (Y -> OtherAddOp):
8236 // select C, (add X, Y), (sub X, Z)
8237 Value *NegVal; // Compute -Z
8238 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8239 NegVal = ConstantExpr::getNeg(C);
8241 NegVal = InsertNewInstBefore(
8242 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8245 Value *NewTrueOp = OtherAddOp;
8246 Value *NewFalseOp = NegVal;
8248 std::swap(NewTrueOp, NewFalseOp);
8249 Instruction *NewSel =
8250 SelectInst::Create(CondVal, NewTrueOp,
8251 NewFalseOp, SI.getName() + ".p");
8253 NewSel = InsertNewInstBefore(NewSel, SI);
8254 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8259 // See if we can fold the select into one of our operands.
8260 if (SI.getType()->isInteger()) {
8261 // See the comment above GetSelectFoldableOperands for a description of the
8262 // transformation we are doing here.
8263 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8264 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8265 !isa<Constant>(FalseVal))
8266 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8267 unsigned OpToFold = 0;
8268 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8270 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8275 Constant *C = GetSelectFoldableConstant(TVI);
8276 Instruction *NewSel =
8277 SelectInst::Create(SI.getCondition(),
8278 TVI->getOperand(2-OpToFold), C);
8279 InsertNewInstBefore(NewSel, SI);
8280 NewSel->takeName(TVI);
8281 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8282 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8284 assert(0 && "Unknown instruction!!");
8289 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8290 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8291 !isa<Constant>(TrueVal))
8292 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8293 unsigned OpToFold = 0;
8294 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8296 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8301 Constant *C = GetSelectFoldableConstant(FVI);
8302 Instruction *NewSel =
8303 SelectInst::Create(SI.getCondition(), C,
8304 FVI->getOperand(2-OpToFold));
8305 InsertNewInstBefore(NewSel, SI);
8306 NewSel->takeName(FVI);
8307 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8308 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8310 assert(0 && "Unknown instruction!!");
8315 if (BinaryOperator::isNot(CondVal)) {
8316 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8317 SI.setOperand(1, FalseVal);
8318 SI.setOperand(2, TrueVal);
8325 /// EnforceKnownAlignment - If the specified pointer points to an object that
8326 /// we control, modify the object's alignment to PrefAlign. This isn't
8327 /// often possible though. If alignment is important, a more reliable approach
8328 /// is to simply align all global variables and allocation instructions to
8329 /// their preferred alignment from the beginning.
8331 static unsigned EnforceKnownAlignment(Value *V,
8332 unsigned Align, unsigned PrefAlign) {
8334 User *U = dyn_cast<User>(V);
8335 if (!U) return Align;
8337 switch (getOpcode(U)) {
8339 case Instruction::BitCast:
8340 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8341 case Instruction::GetElementPtr: {
8342 // If all indexes are zero, it is just the alignment of the base pointer.
8343 bool AllZeroOperands = true;
8344 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8345 if (!isa<Constant>(*i) ||
8346 !cast<Constant>(*i)->isNullValue()) {
8347 AllZeroOperands = false;
8351 if (AllZeroOperands) {
8352 // Treat this like a bitcast.
8353 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8359 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8360 // If there is a large requested alignment and we can, bump up the alignment
8362 if (!GV->isDeclaration()) {
8363 GV->setAlignment(PrefAlign);
8366 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8367 // If there is a requested alignment and if this is an alloca, round up. We
8368 // don't do this for malloc, because some systems can't respect the request.
8369 if (isa<AllocaInst>(AI)) {
8370 AI->setAlignment(PrefAlign);
8378 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8379 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8380 /// and it is more than the alignment of the ultimate object, see if we can
8381 /// increase the alignment of the ultimate object, making this check succeed.
8382 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8383 unsigned PrefAlign) {
8384 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8385 sizeof(PrefAlign) * CHAR_BIT;
8386 APInt Mask = APInt::getAllOnesValue(BitWidth);
8387 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8388 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8389 unsigned TrailZ = KnownZero.countTrailingOnes();
8390 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8392 if (PrefAlign > Align)
8393 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8395 // We don't need to make any adjustment.
8399 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8400 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8401 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8402 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8403 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8405 if (CopyAlign < MinAlign) {
8406 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8410 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8412 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8413 if (MemOpLength == 0) return 0;
8415 // Source and destination pointer types are always "i8*" for intrinsic. See
8416 // if the size is something we can handle with a single primitive load/store.
8417 // A single load+store correctly handles overlapping memory in the memmove
8419 unsigned Size = MemOpLength->getZExtValue();
8420 if (Size == 0) return MI; // Delete this mem transfer.
8422 if (Size > 8 || (Size&(Size-1)))
8423 return 0; // If not 1/2/4/8 bytes, exit.
8425 // Use an integer load+store unless we can find something better.
8426 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8428 // Memcpy forces the use of i8* for the source and destination. That means
8429 // that if you're using memcpy to move one double around, you'll get a cast
8430 // from double* to i8*. We'd much rather use a double load+store rather than
8431 // an i64 load+store, here because this improves the odds that the source or
8432 // dest address will be promotable. See if we can find a better type than the
8433 // integer datatype.
8434 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8435 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8436 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8437 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8438 // down through these levels if so.
8439 while (!SrcETy->isSingleValueType()) {
8440 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8441 if (STy->getNumElements() == 1)
8442 SrcETy = STy->getElementType(0);
8445 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8446 if (ATy->getNumElements() == 1)
8447 SrcETy = ATy->getElementType();
8454 if (SrcETy->isSingleValueType())
8455 NewPtrTy = PointerType::getUnqual(SrcETy);
8460 // If the memcpy/memmove provides better alignment info than we can
8462 SrcAlign = std::max(SrcAlign, CopyAlign);
8463 DstAlign = std::max(DstAlign, CopyAlign);
8465 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8466 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8467 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8468 InsertNewInstBefore(L, *MI);
8469 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8471 // Set the size of the copy to 0, it will be deleted on the next iteration.
8472 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8476 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8477 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8478 if (MI->getAlignment()->getZExtValue() < Alignment) {
8479 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8483 // Extract the length and alignment and fill if they are constant.
8484 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8485 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8486 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8488 uint64_t Len = LenC->getZExtValue();
8489 Alignment = MI->getAlignment()->getZExtValue();
8491 // If the length is zero, this is a no-op
8492 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8494 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8495 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8496 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8498 Value *Dest = MI->getDest();
8499 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8501 // Alignment 0 is identity for alignment 1 for memset, but not store.
8502 if (Alignment == 0) Alignment = 1;
8504 // Extract the fill value and store.
8505 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8506 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8509 // Set the size of the copy to 0, it will be deleted on the next iteration.
8510 MI->setLength(Constant::getNullValue(LenC->getType()));
8518 /// visitCallInst - CallInst simplification. This mostly only handles folding
8519 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8520 /// the heavy lifting.
8522 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8523 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8524 if (!II) return visitCallSite(&CI);
8526 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8528 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8529 bool Changed = false;
8531 // memmove/cpy/set of zero bytes is a noop.
8532 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8533 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8535 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8536 if (CI->getZExtValue() == 1) {
8537 // Replace the instruction with just byte operations. We would
8538 // transform other cases to loads/stores, but we don't know if
8539 // alignment is sufficient.
8543 // If we have a memmove and the source operation is a constant global,
8544 // then the source and dest pointers can't alias, so we can change this
8545 // into a call to memcpy.
8546 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8547 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8548 if (GVSrc->isConstant()) {
8549 Module *M = CI.getParent()->getParent()->getParent();
8550 Intrinsic::ID MemCpyID;
8551 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8552 MemCpyID = Intrinsic::memcpy_i32;
8554 MemCpyID = Intrinsic::memcpy_i64;
8555 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8559 // memmove(x,x,size) -> noop.
8560 if (MMI->getSource() == MMI->getDest())
8561 return EraseInstFromFunction(CI);
8564 // If we can determine a pointer alignment that is bigger than currently
8565 // set, update the alignment.
8566 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8567 if (Instruction *I = SimplifyMemTransfer(MI))
8569 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8570 if (Instruction *I = SimplifyMemSet(MSI))
8574 if (Changed) return II;
8577 switch (II->getIntrinsicID()) {
8579 case Intrinsic::bswap:
8580 // bswap(bswap(x)) -> x
8581 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
8582 if (Operand->getIntrinsicID() == Intrinsic::bswap)
8583 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
8585 case Intrinsic::ppc_altivec_lvx:
8586 case Intrinsic::ppc_altivec_lvxl:
8587 case Intrinsic::x86_sse_loadu_ps:
8588 case Intrinsic::x86_sse2_loadu_pd:
8589 case Intrinsic::x86_sse2_loadu_dq:
8590 // Turn PPC lvx -> load if the pointer is known aligned.
8591 // Turn X86 loadups -> load if the pointer is known aligned.
8592 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8593 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8594 PointerType::getUnqual(II->getType()),
8596 return new LoadInst(Ptr);
8599 case Intrinsic::ppc_altivec_stvx:
8600 case Intrinsic::ppc_altivec_stvxl:
8601 // Turn stvx -> store if the pointer is known aligned.
8602 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8603 const Type *OpPtrTy =
8604 PointerType::getUnqual(II->getOperand(1)->getType());
8605 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8606 return new StoreInst(II->getOperand(1), Ptr);
8609 case Intrinsic::x86_sse_storeu_ps:
8610 case Intrinsic::x86_sse2_storeu_pd:
8611 case Intrinsic::x86_sse2_storeu_dq:
8612 case Intrinsic::x86_sse2_storel_dq:
8613 // Turn X86 storeu -> store if the pointer is known aligned.
8614 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8615 const Type *OpPtrTy =
8616 PointerType::getUnqual(II->getOperand(2)->getType());
8617 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8618 return new StoreInst(II->getOperand(2), Ptr);
8622 case Intrinsic::x86_sse_cvttss2si: {
8623 // These intrinsics only demands the 0th element of its input vector. If
8624 // we can simplify the input based on that, do so now.
8626 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8628 II->setOperand(1, V);
8634 case Intrinsic::ppc_altivec_vperm:
8635 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8636 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8637 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8639 // Check that all of the elements are integer constants or undefs.
8640 bool AllEltsOk = true;
8641 for (unsigned i = 0; i != 16; ++i) {
8642 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8643 !isa<UndefValue>(Mask->getOperand(i))) {
8650 // Cast the input vectors to byte vectors.
8651 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8652 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8653 Value *Result = UndefValue::get(Op0->getType());
8655 // Only extract each element once.
8656 Value *ExtractedElts[32];
8657 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8659 for (unsigned i = 0; i != 16; ++i) {
8660 if (isa<UndefValue>(Mask->getOperand(i)))
8662 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8663 Idx &= 31; // Match the hardware behavior.
8665 if (ExtractedElts[Idx] == 0) {
8667 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8668 InsertNewInstBefore(Elt, CI);
8669 ExtractedElts[Idx] = Elt;
8672 // Insert this value into the result vector.
8673 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
8675 InsertNewInstBefore(cast<Instruction>(Result), CI);
8677 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
8682 case Intrinsic::stackrestore: {
8683 // If the save is right next to the restore, remove the restore. This can
8684 // happen when variable allocas are DCE'd.
8685 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8686 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8687 BasicBlock::iterator BI = SS;
8689 return EraseInstFromFunction(CI);
8693 // Scan down this block to see if there is another stack restore in the
8694 // same block without an intervening call/alloca.
8695 BasicBlock::iterator BI = II;
8696 TerminatorInst *TI = II->getParent()->getTerminator();
8697 bool CannotRemove = false;
8698 for (++BI; &*BI != TI; ++BI) {
8699 if (isa<AllocaInst>(BI)) {
8700 CannotRemove = true;
8703 if (isa<CallInst>(BI)) {
8704 if (!isa<IntrinsicInst>(BI)) {
8705 CannotRemove = true;
8708 // If there is a stackrestore below this one, remove this one.
8709 return EraseInstFromFunction(CI);
8713 // If the stack restore is in a return/unwind block and if there are no
8714 // allocas or calls between the restore and the return, nuke the restore.
8715 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8716 return EraseInstFromFunction(CI);
8721 return visitCallSite(II);
8724 // InvokeInst simplification
8726 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8727 return visitCallSite(&II);
8730 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8731 /// passed through the varargs area, we can eliminate the use of the cast.
8732 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8733 const CastInst * const CI,
8734 const TargetData * const TD,
8736 if (!CI->isLosslessCast())
8739 // The size of ByVal arguments is derived from the type, so we
8740 // can't change to a type with a different size. If the size were
8741 // passed explicitly we could avoid this check.
8742 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
8746 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
8747 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
8748 if (!SrcTy->isSized() || !DstTy->isSized())
8750 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
8755 // visitCallSite - Improvements for call and invoke instructions.
8757 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8758 bool Changed = false;
8760 // If the callee is a constexpr cast of a function, attempt to move the cast
8761 // to the arguments of the call/invoke.
8762 if (transformConstExprCastCall(CS)) return 0;
8764 Value *Callee = CS.getCalledValue();
8766 if (Function *CalleeF = dyn_cast<Function>(Callee))
8767 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8768 Instruction *OldCall = CS.getInstruction();
8769 // If the call and callee calling conventions don't match, this call must
8770 // be unreachable, as the call is undefined.
8771 new StoreInst(ConstantInt::getTrue(),
8772 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8774 if (!OldCall->use_empty())
8775 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8776 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8777 return EraseInstFromFunction(*OldCall);
8781 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8782 // This instruction is not reachable, just remove it. We insert a store to
8783 // undef so that we know that this code is not reachable, despite the fact
8784 // that we can't modify the CFG here.
8785 new StoreInst(ConstantInt::getTrue(),
8786 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8787 CS.getInstruction());
8789 if (!CS.getInstruction()->use_empty())
8790 CS.getInstruction()->
8791 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8793 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8794 // Don't break the CFG, insert a dummy cond branch.
8795 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8796 ConstantInt::getTrue(), II);
8798 return EraseInstFromFunction(*CS.getInstruction());
8801 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8802 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8803 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8804 return transformCallThroughTrampoline(CS);
8806 const PointerType *PTy = cast<PointerType>(Callee->getType());
8807 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8808 if (FTy->isVarArg()) {
8809 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
8810 // See if we can optimize any arguments passed through the varargs area of
8812 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8813 E = CS.arg_end(); I != E; ++I, ++ix) {
8814 CastInst *CI = dyn_cast<CastInst>(*I);
8815 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
8816 *I = CI->getOperand(0);
8822 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8823 // Inline asm calls cannot throw - mark them 'nounwind'.
8824 CS.setDoesNotThrow();
8828 return Changed ? CS.getInstruction() : 0;
8831 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8832 // attempt to move the cast to the arguments of the call/invoke.
8834 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8835 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8836 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8837 if (CE->getOpcode() != Instruction::BitCast ||
8838 !isa<Function>(CE->getOperand(0)))
8840 Function *Callee = cast<Function>(CE->getOperand(0));
8841 Instruction *Caller = CS.getInstruction();
8842 const PAListPtr &CallerPAL = CS.getParamAttrs();
8844 // Okay, this is a cast from a function to a different type. Unless doing so
8845 // would cause a type conversion of one of our arguments, change this call to
8846 // be a direct call with arguments casted to the appropriate types.
8848 const FunctionType *FT = Callee->getFunctionType();
8849 const Type *OldRetTy = Caller->getType();
8850 const Type *NewRetTy = FT->getReturnType();
8852 if (isa<StructType>(NewRetTy))
8853 return false; // TODO: Handle multiple return values.
8855 // Check to see if we are changing the return type...
8856 if (OldRetTy != NewRetTy) {
8857 if (Callee->isDeclaration() &&
8858 // Conversion is ok if changing from one pointer type to another or from
8859 // a pointer to an integer of the same size.
8860 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
8861 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
8862 return false; // Cannot transform this return value.
8864 if (!Caller->use_empty() &&
8865 // void -> non-void is handled specially
8866 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
8867 return false; // Cannot transform this return value.
8869 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8870 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8871 if (RAttrs & ParamAttr::typeIncompatible(NewRetTy))
8872 return false; // Attribute not compatible with transformed value.
8875 // If the callsite is an invoke instruction, and the return value is used by
8876 // a PHI node in a successor, we cannot change the return type of the call
8877 // because there is no place to put the cast instruction (without breaking
8878 // the critical edge). Bail out in this case.
8879 if (!Caller->use_empty())
8880 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8881 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8883 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8884 if (PN->getParent() == II->getNormalDest() ||
8885 PN->getParent() == II->getUnwindDest())
8889 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8890 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8892 CallSite::arg_iterator AI = CS.arg_begin();
8893 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8894 const Type *ParamTy = FT->getParamType(i);
8895 const Type *ActTy = (*AI)->getType();
8897 if (!CastInst::isCastable(ActTy, ParamTy))
8898 return false; // Cannot transform this parameter value.
8900 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8901 return false; // Attribute not compatible with transformed value.
8903 // Converting from one pointer type to another or between a pointer and an
8904 // integer of the same size is safe even if we do not have a body.
8905 bool isConvertible = ActTy == ParamTy ||
8906 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
8907 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
8908 if (Callee->isDeclaration() && !isConvertible) return false;
8911 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8912 Callee->isDeclaration())
8913 return false; // Do not delete arguments unless we have a function body.
8915 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
8916 !CallerPAL.isEmpty())
8917 // In this case we have more arguments than the new function type, but we
8918 // won't be dropping them. Check that these extra arguments have attributes
8919 // that are compatible with being a vararg call argument.
8920 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
8921 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
8923 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
8924 if (PAttrs & ParamAttr::VarArgsIncompatible)
8928 // Okay, we decided that this is a safe thing to do: go ahead and start
8929 // inserting cast instructions as necessary...
8930 std::vector<Value*> Args;
8931 Args.reserve(NumActualArgs);
8932 SmallVector<ParamAttrsWithIndex, 8> attrVec;
8933 attrVec.reserve(NumCommonArgs);
8935 // Get any return attributes.
8936 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8938 // If the return value is not being used, the type may not be compatible
8939 // with the existing attributes. Wipe out any problematic attributes.
8940 RAttrs &= ~ParamAttr::typeIncompatible(NewRetTy);
8942 // Add the new return attributes.
8944 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8946 AI = CS.arg_begin();
8947 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8948 const Type *ParamTy = FT->getParamType(i);
8949 if ((*AI)->getType() == ParamTy) {
8950 Args.push_back(*AI);
8952 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8953 false, ParamTy, false);
8954 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
8955 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8958 // Add any parameter attributes.
8959 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8960 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8963 // If the function takes more arguments than the call was taking, add them
8965 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8966 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8968 // If we are removing arguments to the function, emit an obnoxious warning...
8969 if (FT->getNumParams() < NumActualArgs) {
8970 if (!FT->isVarArg()) {
8971 cerr << "WARNING: While resolving call to function '"
8972 << Callee->getName() << "' arguments were dropped!\n";
8974 // Add all of the arguments in their promoted form to the arg list...
8975 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8976 const Type *PTy = getPromotedType((*AI)->getType());
8977 if (PTy != (*AI)->getType()) {
8978 // Must promote to pass through va_arg area!
8979 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8981 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
8982 InsertNewInstBefore(Cast, *Caller);
8983 Args.push_back(Cast);
8985 Args.push_back(*AI);
8988 // Add any parameter attributes.
8989 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8990 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8995 if (NewRetTy == Type::VoidTy)
8996 Caller->setName(""); // Void type should not have a name.
8998 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9001 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9002 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9003 Args.begin(), Args.end(),
9004 Caller->getName(), Caller);
9005 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9006 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9008 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9009 Caller->getName(), Caller);
9010 CallInst *CI = cast<CallInst>(Caller);
9011 if (CI->isTailCall())
9012 cast<CallInst>(NC)->setTailCall();
9013 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9014 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9017 // Insert a cast of the return type as necessary.
9019 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9020 if (NV->getType() != Type::VoidTy) {
9021 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9023 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9025 // If this is an invoke instruction, we should insert it after the first
9026 // non-phi, instruction in the normal successor block.
9027 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9028 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9029 InsertNewInstBefore(NC, *I);
9031 // Otherwise, it's a call, just insert cast right after the call instr
9032 InsertNewInstBefore(NC, *Caller);
9034 AddUsersToWorkList(*Caller);
9036 NV = UndefValue::get(Caller->getType());
9040 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9041 Caller->replaceAllUsesWith(NV);
9042 Caller->eraseFromParent();
9043 RemoveFromWorkList(Caller);
9047 // transformCallThroughTrampoline - Turn a call to a function created by the
9048 // init_trampoline intrinsic into a direct call to the underlying function.
9050 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9051 Value *Callee = CS.getCalledValue();
9052 const PointerType *PTy = cast<PointerType>(Callee->getType());
9053 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9054 const PAListPtr &Attrs = CS.getParamAttrs();
9056 // If the call already has the 'nest' attribute somewhere then give up -
9057 // otherwise 'nest' would occur twice after splicing in the chain.
9058 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9061 IntrinsicInst *Tramp =
9062 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9064 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9065 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9066 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9068 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9069 if (!NestAttrs.isEmpty()) {
9070 unsigned NestIdx = 1;
9071 const Type *NestTy = 0;
9072 ParameterAttributes NestAttr = ParamAttr::None;
9074 // Look for a parameter marked with the 'nest' attribute.
9075 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9076 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9077 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9078 // Record the parameter type and any other attributes.
9080 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9085 Instruction *Caller = CS.getInstruction();
9086 std::vector<Value*> NewArgs;
9087 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9089 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9090 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9092 // Insert the nest argument into the call argument list, which may
9093 // mean appending it. Likewise for attributes.
9095 // Add any function result attributes.
9096 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9097 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9101 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9103 if (Idx == NestIdx) {
9104 // Add the chain argument and attributes.
9105 Value *NestVal = Tramp->getOperand(3);
9106 if (NestVal->getType() != NestTy)
9107 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9108 NewArgs.push_back(NestVal);
9109 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9115 // Add the original argument and attributes.
9116 NewArgs.push_back(*I);
9117 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9119 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9125 // The trampoline may have been bitcast to a bogus type (FTy).
9126 // Handle this by synthesizing a new function type, equal to FTy
9127 // with the chain parameter inserted.
9129 std::vector<const Type*> NewTypes;
9130 NewTypes.reserve(FTy->getNumParams()+1);
9132 // Insert the chain's type into the list of parameter types, which may
9133 // mean appending it.
9136 FunctionType::param_iterator I = FTy->param_begin(),
9137 E = FTy->param_end();
9141 // Add the chain's type.
9142 NewTypes.push_back(NestTy);
9147 // Add the original type.
9148 NewTypes.push_back(*I);
9154 // Replace the trampoline call with a direct call. Let the generic
9155 // code sort out any function type mismatches.
9156 FunctionType *NewFTy =
9157 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9158 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9159 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9160 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9162 Instruction *NewCaller;
9163 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9164 NewCaller = InvokeInst::Create(NewCallee,
9165 II->getNormalDest(), II->getUnwindDest(),
9166 NewArgs.begin(), NewArgs.end(),
9167 Caller->getName(), Caller);
9168 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9169 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9171 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9172 Caller->getName(), Caller);
9173 if (cast<CallInst>(Caller)->isTailCall())
9174 cast<CallInst>(NewCaller)->setTailCall();
9175 cast<CallInst>(NewCaller)->
9176 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9177 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9179 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9180 Caller->replaceAllUsesWith(NewCaller);
9181 Caller->eraseFromParent();
9182 RemoveFromWorkList(Caller);
9187 // Replace the trampoline call with a direct call. Since there is no 'nest'
9188 // parameter, there is no need to adjust the argument list. Let the generic
9189 // code sort out any function type mismatches.
9190 Constant *NewCallee =
9191 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9192 CS.setCalledFunction(NewCallee);
9193 return CS.getInstruction();
9196 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9197 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9198 /// and a single binop.
9199 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9200 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9201 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9202 isa<CmpInst>(FirstInst));
9203 unsigned Opc = FirstInst->getOpcode();
9204 Value *LHSVal = FirstInst->getOperand(0);
9205 Value *RHSVal = FirstInst->getOperand(1);
9207 const Type *LHSType = LHSVal->getType();
9208 const Type *RHSType = RHSVal->getType();
9210 // Scan to see if all operands are the same opcode, all have one use, and all
9211 // kill their operands (i.e. the operands have one use).
9212 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9213 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9214 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9215 // Verify type of the LHS matches so we don't fold cmp's of different
9216 // types or GEP's with different index types.
9217 I->getOperand(0)->getType() != LHSType ||
9218 I->getOperand(1)->getType() != RHSType)
9221 // If they are CmpInst instructions, check their predicates
9222 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9223 if (cast<CmpInst>(I)->getPredicate() !=
9224 cast<CmpInst>(FirstInst)->getPredicate())
9227 // Keep track of which operand needs a phi node.
9228 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9229 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9232 // Otherwise, this is safe to transform, determine if it is profitable.
9234 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9235 // Indexes are often folded into load/store instructions, so we don't want to
9236 // hide them behind a phi.
9237 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9240 Value *InLHS = FirstInst->getOperand(0);
9241 Value *InRHS = FirstInst->getOperand(1);
9242 PHINode *NewLHS = 0, *NewRHS = 0;
9244 NewLHS = PHINode::Create(LHSType,
9245 FirstInst->getOperand(0)->getName() + ".pn");
9246 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9247 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9248 InsertNewInstBefore(NewLHS, PN);
9253 NewRHS = PHINode::Create(RHSType,
9254 FirstInst->getOperand(1)->getName() + ".pn");
9255 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9256 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9257 InsertNewInstBefore(NewRHS, PN);
9261 // Add all operands to the new PHIs.
9262 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9264 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9265 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9268 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9269 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9273 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9274 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9275 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9276 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9279 assert(isa<GetElementPtrInst>(FirstInst));
9280 return GetElementPtrInst::Create(LHSVal, RHSVal);
9284 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9285 /// of the block that defines it. This means that it must be obvious the value
9286 /// of the load is not changed from the point of the load to the end of the
9289 /// Finally, it is safe, but not profitable, to sink a load targetting a
9290 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9292 static bool isSafeToSinkLoad(LoadInst *L) {
9293 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9295 for (++BBI; BBI != E; ++BBI)
9296 if (BBI->mayWriteToMemory())
9299 // Check for non-address taken alloca. If not address-taken already, it isn't
9300 // profitable to do this xform.
9301 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9302 bool isAddressTaken = false;
9303 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9305 if (isa<LoadInst>(UI)) continue;
9306 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9307 // If storing TO the alloca, then the address isn't taken.
9308 if (SI->getOperand(1) == AI) continue;
9310 isAddressTaken = true;
9314 if (!isAddressTaken)
9322 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9323 // operator and they all are only used by the PHI, PHI together their
9324 // inputs, and do the operation once, to the result of the PHI.
9325 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9326 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9328 // Scan the instruction, looking for input operations that can be folded away.
9329 // If all input operands to the phi are the same instruction (e.g. a cast from
9330 // the same type or "+42") we can pull the operation through the PHI, reducing
9331 // code size and simplifying code.
9332 Constant *ConstantOp = 0;
9333 const Type *CastSrcTy = 0;
9334 bool isVolatile = false;
9335 if (isa<CastInst>(FirstInst)) {
9336 CastSrcTy = FirstInst->getOperand(0)->getType();
9337 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9338 // Can fold binop, compare or shift here if the RHS is a constant,
9339 // otherwise call FoldPHIArgBinOpIntoPHI.
9340 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9341 if (ConstantOp == 0)
9342 return FoldPHIArgBinOpIntoPHI(PN);
9343 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9344 isVolatile = LI->isVolatile();
9345 // We can't sink the load if the loaded value could be modified between the
9346 // load and the PHI.
9347 if (LI->getParent() != PN.getIncomingBlock(0) ||
9348 !isSafeToSinkLoad(LI))
9350 } else if (isa<GetElementPtrInst>(FirstInst)) {
9351 if (FirstInst->getNumOperands() == 2)
9352 return FoldPHIArgBinOpIntoPHI(PN);
9353 // Can't handle general GEPs yet.
9356 return 0; // Cannot fold this operation.
9359 // Check to see if all arguments are the same operation.
9360 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9361 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9362 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9363 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9366 if (I->getOperand(0)->getType() != CastSrcTy)
9367 return 0; // Cast operation must match.
9368 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9369 // We can't sink the load if the loaded value could be modified between
9370 // the load and the PHI.
9371 if (LI->isVolatile() != isVolatile ||
9372 LI->getParent() != PN.getIncomingBlock(i) ||
9373 !isSafeToSinkLoad(LI))
9376 // If the PHI is volatile and its block has multiple successors, sinking
9377 // it would remove a load of the volatile value from the path through the
9380 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9384 } else if (I->getOperand(1) != ConstantOp) {
9389 // Okay, they are all the same operation. Create a new PHI node of the
9390 // correct type, and PHI together all of the LHS's of the instructions.
9391 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9392 PN.getName()+".in");
9393 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9395 Value *InVal = FirstInst->getOperand(0);
9396 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9398 // Add all operands to the new PHI.
9399 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9400 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9401 if (NewInVal != InVal)
9403 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9408 // The new PHI unions all of the same values together. This is really
9409 // common, so we handle it intelligently here for compile-time speed.
9413 InsertNewInstBefore(NewPN, PN);
9417 // Insert and return the new operation.
9418 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9419 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9420 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9421 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9422 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9423 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9424 PhiVal, ConstantOp);
9425 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9427 // If this was a volatile load that we are merging, make sure to loop through
9428 // and mark all the input loads as non-volatile. If we don't do this, we will
9429 // insert a new volatile load and the old ones will not be deletable.
9431 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9432 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9434 return new LoadInst(PhiVal, "", isVolatile);
9437 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9439 static bool DeadPHICycle(PHINode *PN,
9440 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9441 if (PN->use_empty()) return true;
9442 if (!PN->hasOneUse()) return false;
9444 // Remember this node, and if we find the cycle, return.
9445 if (!PotentiallyDeadPHIs.insert(PN))
9448 // Don't scan crazily complex things.
9449 if (PotentiallyDeadPHIs.size() == 16)
9452 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9453 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9458 /// PHIsEqualValue - Return true if this phi node is always equal to
9459 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9460 /// z = some value; x = phi (y, z); y = phi (x, z)
9461 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9462 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9463 // See if we already saw this PHI node.
9464 if (!ValueEqualPHIs.insert(PN))
9467 // Don't scan crazily complex things.
9468 if (ValueEqualPHIs.size() == 16)
9471 // Scan the operands to see if they are either phi nodes or are equal to
9473 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9474 Value *Op = PN->getIncomingValue(i);
9475 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9476 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9478 } else if (Op != NonPhiInVal)
9486 // PHINode simplification
9488 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9489 // If LCSSA is around, don't mess with Phi nodes
9490 if (MustPreserveLCSSA) return 0;
9492 if (Value *V = PN.hasConstantValue())
9493 return ReplaceInstUsesWith(PN, V);
9495 // If all PHI operands are the same operation, pull them through the PHI,
9496 // reducing code size.
9497 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9498 PN.getIncomingValue(0)->hasOneUse())
9499 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9502 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9503 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9504 // PHI)... break the cycle.
9505 if (PN.hasOneUse()) {
9506 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9507 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9508 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9509 PotentiallyDeadPHIs.insert(&PN);
9510 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9511 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9514 // If this phi has a single use, and if that use just computes a value for
9515 // the next iteration of a loop, delete the phi. This occurs with unused
9516 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9517 // common case here is good because the only other things that catch this
9518 // are induction variable analysis (sometimes) and ADCE, which is only run
9520 if (PHIUser->hasOneUse() &&
9521 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9522 PHIUser->use_back() == &PN) {
9523 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9527 // We sometimes end up with phi cycles that non-obviously end up being the
9528 // same value, for example:
9529 // z = some value; x = phi (y, z); y = phi (x, z)
9530 // where the phi nodes don't necessarily need to be in the same block. Do a
9531 // quick check to see if the PHI node only contains a single non-phi value, if
9532 // so, scan to see if the phi cycle is actually equal to that value.
9534 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9535 // Scan for the first non-phi operand.
9536 while (InValNo != NumOperandVals &&
9537 isa<PHINode>(PN.getIncomingValue(InValNo)))
9540 if (InValNo != NumOperandVals) {
9541 Value *NonPhiInVal = PN.getOperand(InValNo);
9543 // Scan the rest of the operands to see if there are any conflicts, if so
9544 // there is no need to recursively scan other phis.
9545 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9546 Value *OpVal = PN.getIncomingValue(InValNo);
9547 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9551 // If we scanned over all operands, then we have one unique value plus
9552 // phi values. Scan PHI nodes to see if they all merge in each other or
9554 if (InValNo == NumOperandVals) {
9555 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9556 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9557 return ReplaceInstUsesWith(PN, NonPhiInVal);
9564 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9565 Instruction *InsertPoint,
9567 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9568 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9569 // We must cast correctly to the pointer type. Ensure that we
9570 // sign extend the integer value if it is smaller as this is
9571 // used for address computation.
9572 Instruction::CastOps opcode =
9573 (VTySize < PtrSize ? Instruction::SExt :
9574 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9575 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9579 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9580 Value *PtrOp = GEP.getOperand(0);
9581 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9582 // If so, eliminate the noop.
9583 if (GEP.getNumOperands() == 1)
9584 return ReplaceInstUsesWith(GEP, PtrOp);
9586 if (isa<UndefValue>(GEP.getOperand(0)))
9587 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9589 bool HasZeroPointerIndex = false;
9590 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9591 HasZeroPointerIndex = C->isNullValue();
9593 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9594 return ReplaceInstUsesWith(GEP, PtrOp);
9596 // Eliminate unneeded casts for indices.
9597 bool MadeChange = false;
9599 gep_type_iterator GTI = gep_type_begin(GEP);
9600 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
9601 i != e; ++i, ++GTI) {
9602 if (isa<SequentialType>(*GTI)) {
9603 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
9604 if (CI->getOpcode() == Instruction::ZExt ||
9605 CI->getOpcode() == Instruction::SExt) {
9606 const Type *SrcTy = CI->getOperand(0)->getType();
9607 // We can eliminate a cast from i32 to i64 iff the target
9608 // is a 32-bit pointer target.
9609 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9611 *i = CI->getOperand(0);
9615 // If we are using a wider index than needed for this platform, shrink it
9616 // to what we need. If the incoming value needs a cast instruction,
9617 // insert it. This explicit cast can make subsequent optimizations more
9620 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9621 if (Constant *C = dyn_cast<Constant>(Op)) {
9622 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
9625 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9633 if (MadeChange) return &GEP;
9635 // If this GEP instruction doesn't move the pointer, and if the input operand
9636 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9637 // real input to the dest type.
9638 if (GEP.hasAllZeroIndices()) {
9639 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9640 // If the bitcast is of an allocation, and the allocation will be
9641 // converted to match the type of the cast, don't touch this.
9642 if (isa<AllocationInst>(BCI->getOperand(0))) {
9643 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9644 if (Instruction *I = visitBitCast(*BCI)) {
9647 BCI->getParent()->getInstList().insert(BCI, I);
9648 ReplaceInstUsesWith(*BCI, I);
9653 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9657 // Combine Indices - If the source pointer to this getelementptr instruction
9658 // is a getelementptr instruction, combine the indices of the two
9659 // getelementptr instructions into a single instruction.
9661 SmallVector<Value*, 8> SrcGEPOperands;
9662 if (User *Src = dyn_castGetElementPtr(PtrOp))
9663 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9665 if (!SrcGEPOperands.empty()) {
9666 // Note that if our source is a gep chain itself that we wait for that
9667 // chain to be resolved before we perform this transformation. This
9668 // avoids us creating a TON of code in some cases.
9670 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9671 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9672 return 0; // Wait until our source is folded to completion.
9674 SmallVector<Value*, 8> Indices;
9676 // Find out whether the last index in the source GEP is a sequential idx.
9677 bool EndsWithSequential = false;
9678 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9679 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9680 EndsWithSequential = !isa<StructType>(*I);
9682 // Can we combine the two pointer arithmetics offsets?
9683 if (EndsWithSequential) {
9684 // Replace: gep (gep %P, long B), long A, ...
9685 // With: T = long A+B; gep %P, T, ...
9687 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9688 if (SO1 == Constant::getNullValue(SO1->getType())) {
9690 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9693 // If they aren't the same type, convert both to an integer of the
9694 // target's pointer size.
9695 if (SO1->getType() != GO1->getType()) {
9696 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9697 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9698 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9699 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9701 unsigned PS = TD->getPointerSizeInBits();
9702 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9703 // Convert GO1 to SO1's type.
9704 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9706 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9707 // Convert SO1 to GO1's type.
9708 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9710 const Type *PT = TD->getIntPtrType();
9711 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9712 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9716 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9717 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9719 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
9720 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9724 // Recycle the GEP we already have if possible.
9725 if (SrcGEPOperands.size() == 2) {
9726 GEP.setOperand(0, SrcGEPOperands[0]);
9727 GEP.setOperand(1, Sum);
9730 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9731 SrcGEPOperands.end()-1);
9732 Indices.push_back(Sum);
9733 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9735 } else if (isa<Constant>(*GEP.idx_begin()) &&
9736 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9737 SrcGEPOperands.size() != 1) {
9738 // Otherwise we can do the fold if the first index of the GEP is a zero
9739 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9740 SrcGEPOperands.end());
9741 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9744 if (!Indices.empty())
9745 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9746 Indices.end(), GEP.getName());
9748 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9749 // GEP of global variable. If all of the indices for this GEP are
9750 // constants, we can promote this to a constexpr instead of an instruction.
9752 // Scan for nonconstants...
9753 SmallVector<Constant*, 8> Indices;
9754 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9755 for (; I != E && isa<Constant>(*I); ++I)
9756 Indices.push_back(cast<Constant>(*I));
9758 if (I == E) { // If they are all constants...
9759 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9760 &Indices[0],Indices.size());
9762 // Replace all uses of the GEP with the new constexpr...
9763 return ReplaceInstUsesWith(GEP, CE);
9765 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9766 if (!isa<PointerType>(X->getType())) {
9767 // Not interesting. Source pointer must be a cast from pointer.
9768 } else if (HasZeroPointerIndex) {
9769 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9770 // into : GEP [10 x i8]* X, i32 0, ...
9772 // This occurs when the program declares an array extern like "int X[];"
9774 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9775 const PointerType *XTy = cast<PointerType>(X->getType());
9776 if (const ArrayType *XATy =
9777 dyn_cast<ArrayType>(XTy->getElementType()))
9778 if (const ArrayType *CATy =
9779 dyn_cast<ArrayType>(CPTy->getElementType()))
9780 if (CATy->getElementType() == XATy->getElementType()) {
9781 // At this point, we know that the cast source type is a pointer
9782 // to an array of the same type as the destination pointer
9783 // array. Because the array type is never stepped over (there
9784 // is a leading zero) we can fold the cast into this GEP.
9785 GEP.setOperand(0, X);
9788 } else if (GEP.getNumOperands() == 2) {
9789 // Transform things like:
9790 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9791 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9792 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9793 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9794 if (isa<ArrayType>(SrcElTy) &&
9795 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9796 TD->getABITypeSize(ResElTy)) {
9798 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9799 Idx[1] = GEP.getOperand(1);
9800 Value *V = InsertNewInstBefore(
9801 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9802 // V and GEP are both pointer types --> BitCast
9803 return new BitCastInst(V, GEP.getType());
9806 // Transform things like:
9807 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9808 // (where tmp = 8*tmp2) into:
9809 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9811 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9812 uint64_t ArrayEltSize =
9813 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9815 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9816 // allow either a mul, shift, or constant here.
9818 ConstantInt *Scale = 0;
9819 if (ArrayEltSize == 1) {
9820 NewIdx = GEP.getOperand(1);
9821 Scale = ConstantInt::get(NewIdx->getType(), 1);
9822 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9823 NewIdx = ConstantInt::get(CI->getType(), 1);
9825 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9826 if (Inst->getOpcode() == Instruction::Shl &&
9827 isa<ConstantInt>(Inst->getOperand(1))) {
9828 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9829 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9830 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9831 NewIdx = Inst->getOperand(0);
9832 } else if (Inst->getOpcode() == Instruction::Mul &&
9833 isa<ConstantInt>(Inst->getOperand(1))) {
9834 Scale = cast<ConstantInt>(Inst->getOperand(1));
9835 NewIdx = Inst->getOperand(0);
9839 // If the index will be to exactly the right offset with the scale taken
9840 // out, perform the transformation. Note, we don't know whether Scale is
9841 // signed or not. We'll use unsigned version of division/modulo
9842 // operation after making sure Scale doesn't have the sign bit set.
9843 if (Scale && Scale->getSExtValue() >= 0LL &&
9844 Scale->getZExtValue() % ArrayEltSize == 0) {
9845 Scale = ConstantInt::get(Scale->getType(),
9846 Scale->getZExtValue() / ArrayEltSize);
9847 if (Scale->getZExtValue() != 1) {
9848 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9850 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
9851 NewIdx = InsertNewInstBefore(Sc, GEP);
9854 // Insert the new GEP instruction.
9856 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9858 Instruction *NewGEP =
9859 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9860 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9861 // The NewGEP must be pointer typed, so must the old one -> BitCast
9862 return new BitCastInst(NewGEP, GEP.getType());
9871 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9872 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9873 if (AI.isArrayAllocation()) { // Check C != 1
9874 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9876 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9877 AllocationInst *New = 0;
9879 // Create and insert the replacement instruction...
9880 if (isa<MallocInst>(AI))
9881 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9883 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9884 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9887 InsertNewInstBefore(New, AI);
9889 // Scan to the end of the allocation instructions, to skip over a block of
9890 // allocas if possible...
9892 BasicBlock::iterator It = New;
9893 while (isa<AllocationInst>(*It)) ++It;
9895 // Now that I is pointing to the first non-allocation-inst in the block,
9896 // insert our getelementptr instruction...
9898 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9902 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9903 New->getName()+".sub", It);
9905 // Now make everything use the getelementptr instead of the original
9907 return ReplaceInstUsesWith(AI, V);
9908 } else if (isa<UndefValue>(AI.getArraySize())) {
9909 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9913 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9914 // Note that we only do this for alloca's, because malloc should allocate and
9915 // return a unique pointer, even for a zero byte allocation.
9916 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9917 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9918 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9923 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9924 Value *Op = FI.getOperand(0);
9926 // free undef -> unreachable.
9927 if (isa<UndefValue>(Op)) {
9928 // Insert a new store to null because we cannot modify the CFG here.
9929 new StoreInst(ConstantInt::getTrue(),
9930 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9931 return EraseInstFromFunction(FI);
9934 // If we have 'free null' delete the instruction. This can happen in stl code
9935 // when lots of inlining happens.
9936 if (isa<ConstantPointerNull>(Op))
9937 return EraseInstFromFunction(FI);
9939 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9940 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9941 FI.setOperand(0, CI->getOperand(0));
9945 // Change free (gep X, 0,0,0,0) into free(X)
9946 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9947 if (GEPI->hasAllZeroIndices()) {
9948 AddToWorkList(GEPI);
9949 FI.setOperand(0, GEPI->getOperand(0));
9954 // Change free(malloc) into nothing, if the malloc has a single use.
9955 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9956 if (MI->hasOneUse()) {
9957 EraseInstFromFunction(FI);
9958 return EraseInstFromFunction(*MI);
9965 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9966 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9967 const TargetData *TD) {
9968 User *CI = cast<User>(LI.getOperand(0));
9969 Value *CastOp = CI->getOperand(0);
9971 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9972 // Instead of loading constant c string, use corresponding integer value
9973 // directly if string length is small enough.
9974 const std::string &Str = CE->getOperand(0)->getStringValue();
9976 unsigned len = Str.length();
9977 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9978 unsigned numBits = Ty->getPrimitiveSizeInBits();
9979 // Replace LI with immediate integer store.
9980 if ((numBits >> 3) == len + 1) {
9981 APInt StrVal(numBits, 0);
9982 APInt SingleChar(numBits, 0);
9983 if (TD->isLittleEndian()) {
9984 for (signed i = len-1; i >= 0; i--) {
9985 SingleChar = (uint64_t) Str[i];
9986 StrVal = (StrVal << 8) | SingleChar;
9989 for (unsigned i = 0; i < len; i++) {
9990 SingleChar = (uint64_t) Str[i];
9991 StrVal = (StrVal << 8) | SingleChar;
9993 // Append NULL at the end.
9995 StrVal = (StrVal << 8) | SingleChar;
9997 Value *NL = ConstantInt::get(StrVal);
9998 return IC.ReplaceInstUsesWith(LI, NL);
10003 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10004 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10005 const Type *SrcPTy = SrcTy->getElementType();
10007 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10008 isa<VectorType>(DestPTy)) {
10009 // If the source is an array, the code below will not succeed. Check to
10010 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10012 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10013 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10014 if (ASrcTy->getNumElements() != 0) {
10016 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10017 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10018 SrcTy = cast<PointerType>(CastOp->getType());
10019 SrcPTy = SrcTy->getElementType();
10022 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10023 isa<VectorType>(SrcPTy)) &&
10024 // Do not allow turning this into a load of an integer, which is then
10025 // casted to a pointer, this pessimizes pointer analysis a lot.
10026 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10027 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10028 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10030 // Okay, we are casting from one integer or pointer type to another of
10031 // the same size. Instead of casting the pointer before the load, cast
10032 // the result of the loaded value.
10033 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10035 LI.isVolatile()),LI);
10036 // Now cast the result of the load.
10037 return new BitCastInst(NewLoad, LI.getType());
10044 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10045 /// from this value cannot trap. If it is not obviously safe to load from the
10046 /// specified pointer, we do a quick local scan of the basic block containing
10047 /// ScanFrom, to determine if the address is already accessed.
10048 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10049 // If it is an alloca it is always safe to load from.
10050 if (isa<AllocaInst>(V)) return true;
10052 // If it is a global variable it is mostly safe to load from.
10053 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10054 // Don't try to evaluate aliases. External weak GV can be null.
10055 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10057 // Otherwise, be a little bit agressive by scanning the local block where we
10058 // want to check to see if the pointer is already being loaded or stored
10059 // from/to. If so, the previous load or store would have already trapped,
10060 // so there is no harm doing an extra load (also, CSE will later eliminate
10061 // the load entirely).
10062 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10067 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10068 if (LI->getOperand(0) == V) return true;
10069 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10070 if (SI->getOperand(1) == V) return true;
10076 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10077 /// until we find the underlying object a pointer is referring to or something
10078 /// we don't understand. Note that the returned pointer may be offset from the
10079 /// input, because we ignore GEP indices.
10080 static Value *GetUnderlyingObject(Value *Ptr) {
10082 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10083 if (CE->getOpcode() == Instruction::BitCast ||
10084 CE->getOpcode() == Instruction::GetElementPtr)
10085 Ptr = CE->getOperand(0);
10088 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10089 Ptr = BCI->getOperand(0);
10090 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10091 Ptr = GEP->getOperand(0);
10098 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10099 Value *Op = LI.getOperand(0);
10101 // Attempt to improve the alignment.
10102 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10104 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10105 LI.getAlignment()))
10106 LI.setAlignment(KnownAlign);
10108 // load (cast X) --> cast (load X) iff safe
10109 if (isa<CastInst>(Op))
10110 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10113 // None of the following transforms are legal for volatile loads.
10114 if (LI.isVolatile()) return 0;
10116 if (&LI.getParent()->front() != &LI) {
10117 BasicBlock::iterator BBI = &LI; --BBI;
10118 // If the instruction immediately before this is a store to the same
10119 // address, do a simple form of store->load forwarding.
10120 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10121 if (SI->getOperand(1) == LI.getOperand(0))
10122 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10123 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10124 if (LIB->getOperand(0) == LI.getOperand(0))
10125 return ReplaceInstUsesWith(LI, LIB);
10128 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10129 const Value *GEPI0 = GEPI->getOperand(0);
10130 // TODO: Consider a target hook for valid address spaces for this xform.
10131 if (isa<ConstantPointerNull>(GEPI0) &&
10132 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10133 // Insert a new store to null instruction before the load to indicate
10134 // that this code is not reachable. We do this instead of inserting
10135 // an unreachable instruction directly because we cannot modify the
10137 new StoreInst(UndefValue::get(LI.getType()),
10138 Constant::getNullValue(Op->getType()), &LI);
10139 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10143 if (Constant *C = dyn_cast<Constant>(Op)) {
10144 // load null/undef -> undef
10145 // TODO: Consider a target hook for valid address spaces for this xform.
10146 if (isa<UndefValue>(C) || (C->isNullValue() &&
10147 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10148 // Insert a new store to null instruction before the load to indicate that
10149 // this code is not reachable. We do this instead of inserting an
10150 // unreachable instruction directly because we cannot modify the CFG.
10151 new StoreInst(UndefValue::get(LI.getType()),
10152 Constant::getNullValue(Op->getType()), &LI);
10153 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10156 // Instcombine load (constant global) into the value loaded.
10157 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10158 if (GV->isConstant() && !GV->isDeclaration())
10159 return ReplaceInstUsesWith(LI, GV->getInitializer());
10161 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10162 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10163 if (CE->getOpcode() == Instruction::GetElementPtr) {
10164 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10165 if (GV->isConstant() && !GV->isDeclaration())
10167 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10168 return ReplaceInstUsesWith(LI, V);
10169 if (CE->getOperand(0)->isNullValue()) {
10170 // Insert a new store to null instruction before the load to indicate
10171 // that this code is not reachable. We do this instead of inserting
10172 // an unreachable instruction directly because we cannot modify the
10174 new StoreInst(UndefValue::get(LI.getType()),
10175 Constant::getNullValue(Op->getType()), &LI);
10176 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10179 } else if (CE->isCast()) {
10180 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10186 // If this load comes from anywhere in a constant global, and if the global
10187 // is all undef or zero, we know what it loads.
10188 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10189 if (GV->isConstant() && GV->hasInitializer()) {
10190 if (GV->getInitializer()->isNullValue())
10191 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10192 else if (isa<UndefValue>(GV->getInitializer()))
10193 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10197 if (Op->hasOneUse()) {
10198 // Change select and PHI nodes to select values instead of addresses: this
10199 // helps alias analysis out a lot, allows many others simplifications, and
10200 // exposes redundancy in the code.
10202 // Note that we cannot do the transformation unless we know that the
10203 // introduced loads cannot trap! Something like this is valid as long as
10204 // the condition is always false: load (select bool %C, int* null, int* %G),
10205 // but it would not be valid if we transformed it to load from null
10206 // unconditionally.
10208 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10209 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10210 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10211 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10212 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10213 SI->getOperand(1)->getName()+".val"), LI);
10214 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10215 SI->getOperand(2)->getName()+".val"), LI);
10216 return SelectInst::Create(SI->getCondition(), V1, V2);
10219 // load (select (cond, null, P)) -> load P
10220 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10221 if (C->isNullValue()) {
10222 LI.setOperand(0, SI->getOperand(2));
10226 // load (select (cond, P, null)) -> load P
10227 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10228 if (C->isNullValue()) {
10229 LI.setOperand(0, SI->getOperand(1));
10237 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10239 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10240 User *CI = cast<User>(SI.getOperand(1));
10241 Value *CastOp = CI->getOperand(0);
10243 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10244 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10245 const Type *SrcPTy = SrcTy->getElementType();
10247 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10248 // If the source is an array, the code below will not succeed. Check to
10249 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10251 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10252 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10253 if (ASrcTy->getNumElements() != 0) {
10255 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10256 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10257 SrcTy = cast<PointerType>(CastOp->getType());
10258 SrcPTy = SrcTy->getElementType();
10261 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10262 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10263 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10265 // Okay, we are casting from one integer or pointer type to another of
10266 // the same size. Instead of casting the pointer before
10267 // the store, cast the value to be stored.
10269 Value *SIOp0 = SI.getOperand(0);
10270 Instruction::CastOps opcode = Instruction::BitCast;
10271 const Type* CastSrcTy = SIOp0->getType();
10272 const Type* CastDstTy = SrcPTy;
10273 if (isa<PointerType>(CastDstTy)) {
10274 if (CastSrcTy->isInteger())
10275 opcode = Instruction::IntToPtr;
10276 } else if (isa<IntegerType>(CastDstTy)) {
10277 if (isa<PointerType>(SIOp0->getType()))
10278 opcode = Instruction::PtrToInt;
10280 if (Constant *C = dyn_cast<Constant>(SIOp0))
10281 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10283 NewCast = IC.InsertNewInstBefore(
10284 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10286 return new StoreInst(NewCast, CastOp);
10293 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10294 Value *Val = SI.getOperand(0);
10295 Value *Ptr = SI.getOperand(1);
10297 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10298 EraseInstFromFunction(SI);
10303 // If the RHS is an alloca with a single use, zapify the store, making the
10305 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10306 if (isa<AllocaInst>(Ptr)) {
10307 EraseInstFromFunction(SI);
10312 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10313 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10314 GEP->getOperand(0)->hasOneUse()) {
10315 EraseInstFromFunction(SI);
10321 // Attempt to improve the alignment.
10322 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10324 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10325 SI.getAlignment()))
10326 SI.setAlignment(KnownAlign);
10328 // Do really simple DSE, to catch cases where there are several consequtive
10329 // stores to the same location, separated by a few arithmetic operations. This
10330 // situation often occurs with bitfield accesses.
10331 BasicBlock::iterator BBI = &SI;
10332 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10336 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10337 // Prev store isn't volatile, and stores to the same location?
10338 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10341 EraseInstFromFunction(*PrevSI);
10347 // If this is a load, we have to stop. However, if the loaded value is from
10348 // the pointer we're loading and is producing the pointer we're storing,
10349 // then *this* store is dead (X = load P; store X -> P).
10350 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10351 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10352 EraseInstFromFunction(SI);
10356 // Otherwise, this is a load from some other location. Stores before it
10357 // may not be dead.
10361 // Don't skip over loads or things that can modify memory.
10362 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10367 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10369 // store X, null -> turns into 'unreachable' in SimplifyCFG
10370 if (isa<ConstantPointerNull>(Ptr)) {
10371 if (!isa<UndefValue>(Val)) {
10372 SI.setOperand(0, UndefValue::get(Val->getType()));
10373 if (Instruction *U = dyn_cast<Instruction>(Val))
10374 AddToWorkList(U); // Dropped a use.
10377 return 0; // Do not modify these!
10380 // store undef, Ptr -> noop
10381 if (isa<UndefValue>(Val)) {
10382 EraseInstFromFunction(SI);
10387 // If the pointer destination is a cast, see if we can fold the cast into the
10389 if (isa<CastInst>(Ptr))
10390 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10394 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10398 // If this store is the last instruction in the basic block, and if the block
10399 // ends with an unconditional branch, try to move it to the successor block.
10401 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10402 if (BI->isUnconditional())
10403 if (SimplifyStoreAtEndOfBlock(SI))
10404 return 0; // xform done!
10409 /// SimplifyStoreAtEndOfBlock - Turn things like:
10410 /// if () { *P = v1; } else { *P = v2 }
10411 /// into a phi node with a store in the successor.
10413 /// Simplify things like:
10414 /// *P = v1; if () { *P = v2; }
10415 /// into a phi node with a store in the successor.
10417 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10418 BasicBlock *StoreBB = SI.getParent();
10420 // Check to see if the successor block has exactly two incoming edges. If
10421 // so, see if the other predecessor contains a store to the same location.
10422 // if so, insert a PHI node (if needed) and move the stores down.
10423 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10425 // Determine whether Dest has exactly two predecessors and, if so, compute
10426 // the other predecessor.
10427 pred_iterator PI = pred_begin(DestBB);
10428 BasicBlock *OtherBB = 0;
10429 if (*PI != StoreBB)
10432 if (PI == pred_end(DestBB))
10435 if (*PI != StoreBB) {
10440 if (++PI != pred_end(DestBB))
10443 // Bail out if all the relevant blocks aren't distinct (this can happen,
10444 // for example, if SI is in an infinite loop)
10445 if (StoreBB == DestBB || OtherBB == DestBB)
10448 // Verify that the other block ends in a branch and is not otherwise empty.
10449 BasicBlock::iterator BBI = OtherBB->getTerminator();
10450 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10451 if (!OtherBr || BBI == OtherBB->begin())
10454 // If the other block ends in an unconditional branch, check for the 'if then
10455 // else' case. there is an instruction before the branch.
10456 StoreInst *OtherStore = 0;
10457 if (OtherBr->isUnconditional()) {
10458 // If this isn't a store, or isn't a store to the same location, bail out.
10460 OtherStore = dyn_cast<StoreInst>(BBI);
10461 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10464 // Otherwise, the other block ended with a conditional branch. If one of the
10465 // destinations is StoreBB, then we have the if/then case.
10466 if (OtherBr->getSuccessor(0) != StoreBB &&
10467 OtherBr->getSuccessor(1) != StoreBB)
10470 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10471 // if/then triangle. See if there is a store to the same ptr as SI that
10472 // lives in OtherBB.
10474 // Check to see if we find the matching store.
10475 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10476 if (OtherStore->getOperand(1) != SI.getOperand(1))
10480 // If we find something that may be using or overwriting the stored
10481 // value, or if we run out of instructions, we can't do the xform.
10482 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
10483 BBI == OtherBB->begin())
10487 // In order to eliminate the store in OtherBr, we have to
10488 // make sure nothing reads or overwrites the stored value in
10490 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10491 // FIXME: This should really be AA driven.
10492 if (I->mayReadFromMemory() || I->mayWriteToMemory())
10497 // Insert a PHI node now if we need it.
10498 Value *MergedVal = OtherStore->getOperand(0);
10499 if (MergedVal != SI.getOperand(0)) {
10500 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10501 PN->reserveOperandSpace(2);
10502 PN->addIncoming(SI.getOperand(0), SI.getParent());
10503 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10504 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10507 // Advance to a place where it is safe to insert the new store and
10509 BBI = DestBB->getFirstNonPHI();
10510 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10511 OtherStore->isVolatile()), *BBI);
10513 // Nuke the old stores.
10514 EraseInstFromFunction(SI);
10515 EraseInstFromFunction(*OtherStore);
10521 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10522 // Change br (not X), label True, label False to: br X, label False, True
10524 BasicBlock *TrueDest;
10525 BasicBlock *FalseDest;
10526 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10527 !isa<Constant>(X)) {
10528 // Swap Destinations and condition...
10529 BI.setCondition(X);
10530 BI.setSuccessor(0, FalseDest);
10531 BI.setSuccessor(1, TrueDest);
10535 // Cannonicalize fcmp_one -> fcmp_oeq
10536 FCmpInst::Predicate FPred; Value *Y;
10537 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10538 TrueDest, FalseDest)))
10539 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10540 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10541 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10542 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10543 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10544 NewSCC->takeName(I);
10545 // Swap Destinations and condition...
10546 BI.setCondition(NewSCC);
10547 BI.setSuccessor(0, FalseDest);
10548 BI.setSuccessor(1, TrueDest);
10549 RemoveFromWorkList(I);
10550 I->eraseFromParent();
10551 AddToWorkList(NewSCC);
10555 // Cannonicalize icmp_ne -> icmp_eq
10556 ICmpInst::Predicate IPred;
10557 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10558 TrueDest, FalseDest)))
10559 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10560 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10561 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10562 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10563 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10564 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10565 NewSCC->takeName(I);
10566 // Swap Destinations and condition...
10567 BI.setCondition(NewSCC);
10568 BI.setSuccessor(0, FalseDest);
10569 BI.setSuccessor(1, TrueDest);
10570 RemoveFromWorkList(I);
10571 I->eraseFromParent();;
10572 AddToWorkList(NewSCC);
10579 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10580 Value *Cond = SI.getCondition();
10581 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10582 if (I->getOpcode() == Instruction::Add)
10583 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10584 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10585 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10586 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10588 SI.setOperand(0, I->getOperand(0));
10596 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
10597 // See if we are trying to extract a known value. If so, use that instead.
10598 if (Value *Elt = FindInsertedValue(EV.getOperand(0), EV.idx_begin(),
10599 EV.idx_end(), &EV))
10600 return ReplaceInstUsesWith(EV, Elt);
10606 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10607 /// is to leave as a vector operation.
10608 static bool CheapToScalarize(Value *V, bool isConstant) {
10609 if (isa<ConstantAggregateZero>(V))
10611 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10612 if (isConstant) return true;
10613 // If all elts are the same, we can extract.
10614 Constant *Op0 = C->getOperand(0);
10615 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10616 if (C->getOperand(i) != Op0)
10620 Instruction *I = dyn_cast<Instruction>(V);
10621 if (!I) return false;
10623 // Insert element gets simplified to the inserted element or is deleted if
10624 // this is constant idx extract element and its a constant idx insertelt.
10625 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10626 isa<ConstantInt>(I->getOperand(2)))
10628 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10630 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10631 if (BO->hasOneUse() &&
10632 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10633 CheapToScalarize(BO->getOperand(1), isConstant)))
10635 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10636 if (CI->hasOneUse() &&
10637 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10638 CheapToScalarize(CI->getOperand(1), isConstant)))
10644 /// Read and decode a shufflevector mask.
10646 /// It turns undef elements into values that are larger than the number of
10647 /// elements in the input.
10648 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10649 unsigned NElts = SVI->getType()->getNumElements();
10650 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10651 return std::vector<unsigned>(NElts, 0);
10652 if (isa<UndefValue>(SVI->getOperand(2)))
10653 return std::vector<unsigned>(NElts, 2*NElts);
10655 std::vector<unsigned> Result;
10656 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10657 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
10658 if (isa<UndefValue>(*i))
10659 Result.push_back(NElts*2); // undef -> 8
10661 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
10665 /// FindScalarElement - Given a vector and an element number, see if the scalar
10666 /// value is already around as a register, for example if it were inserted then
10667 /// extracted from the vector.
10668 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10669 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10670 const VectorType *PTy = cast<VectorType>(V->getType());
10671 unsigned Width = PTy->getNumElements();
10672 if (EltNo >= Width) // Out of range access.
10673 return UndefValue::get(PTy->getElementType());
10675 if (isa<UndefValue>(V))
10676 return UndefValue::get(PTy->getElementType());
10677 else if (isa<ConstantAggregateZero>(V))
10678 return Constant::getNullValue(PTy->getElementType());
10679 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10680 return CP->getOperand(EltNo);
10681 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10682 // If this is an insert to a variable element, we don't know what it is.
10683 if (!isa<ConstantInt>(III->getOperand(2)))
10685 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10687 // If this is an insert to the element we are looking for, return the
10689 if (EltNo == IIElt)
10690 return III->getOperand(1);
10692 // Otherwise, the insertelement doesn't modify the value, recurse on its
10694 return FindScalarElement(III->getOperand(0), EltNo);
10695 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10696 unsigned InEl = getShuffleMask(SVI)[EltNo];
10698 return FindScalarElement(SVI->getOperand(0), InEl);
10699 else if (InEl < Width*2)
10700 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10702 return UndefValue::get(PTy->getElementType());
10705 // Otherwise, we don't know.
10709 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10710 // If vector val is undef, replace extract with scalar undef.
10711 if (isa<UndefValue>(EI.getOperand(0)))
10712 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10714 // If vector val is constant 0, replace extract with scalar 0.
10715 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10716 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10718 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10719 // If vector val is constant with all elements the same, replace EI with
10720 // that element. When the elements are not identical, we cannot replace yet
10721 // (we do that below, but only when the index is constant).
10722 Constant *op0 = C->getOperand(0);
10723 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10724 if (C->getOperand(i) != op0) {
10729 return ReplaceInstUsesWith(EI, op0);
10732 // If extracting a specified index from the vector, see if we can recursively
10733 // find a previously computed scalar that was inserted into the vector.
10734 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10735 unsigned IndexVal = IdxC->getZExtValue();
10736 unsigned VectorWidth =
10737 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10739 // If this is extracting an invalid index, turn this into undef, to avoid
10740 // crashing the code below.
10741 if (IndexVal >= VectorWidth)
10742 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10744 // This instruction only demands the single element from the input vector.
10745 // If the input vector has a single use, simplify it based on this use
10747 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10748 uint64_t UndefElts;
10749 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10752 EI.setOperand(0, V);
10757 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10758 return ReplaceInstUsesWith(EI, Elt);
10760 // If the this extractelement is directly using a bitcast from a vector of
10761 // the same number of elements, see if we can find the source element from
10762 // it. In this case, we will end up needing to bitcast the scalars.
10763 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10764 if (const VectorType *VT =
10765 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10766 if (VT->getNumElements() == VectorWidth)
10767 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10768 return new BitCastInst(Elt, EI.getType());
10772 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10773 if (I->hasOneUse()) {
10774 // Push extractelement into predecessor operation if legal and
10775 // profitable to do so
10776 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10777 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10778 if (CheapToScalarize(BO, isConstantElt)) {
10779 ExtractElementInst *newEI0 =
10780 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10781 EI.getName()+".lhs");
10782 ExtractElementInst *newEI1 =
10783 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10784 EI.getName()+".rhs");
10785 InsertNewInstBefore(newEI0, EI);
10786 InsertNewInstBefore(newEI1, EI);
10787 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
10789 } else if (isa<LoadInst>(I)) {
10791 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10792 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10793 PointerType::get(EI.getType(), AS),EI);
10794 GetElementPtrInst *GEP =
10795 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
10796 InsertNewInstBefore(GEP, EI);
10797 return new LoadInst(GEP);
10800 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10801 // Extracting the inserted element?
10802 if (IE->getOperand(2) == EI.getOperand(1))
10803 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10804 // If the inserted and extracted elements are constants, they must not
10805 // be the same value, extract from the pre-inserted value instead.
10806 if (isa<Constant>(IE->getOperand(2)) &&
10807 isa<Constant>(EI.getOperand(1))) {
10808 AddUsesToWorkList(EI);
10809 EI.setOperand(0, IE->getOperand(0));
10812 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10813 // If this is extracting an element from a shufflevector, figure out where
10814 // it came from and extract from the appropriate input element instead.
10815 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10816 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10818 if (SrcIdx < SVI->getType()->getNumElements())
10819 Src = SVI->getOperand(0);
10820 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10821 SrcIdx -= SVI->getType()->getNumElements();
10822 Src = SVI->getOperand(1);
10824 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10826 return new ExtractElementInst(Src, SrcIdx);
10833 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10834 /// elements from either LHS or RHS, return the shuffle mask and true.
10835 /// Otherwise, return false.
10836 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10837 std::vector<Constant*> &Mask) {
10838 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10839 "Invalid CollectSingleShuffleElements");
10840 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10842 if (isa<UndefValue>(V)) {
10843 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10845 } else if (V == LHS) {
10846 for (unsigned i = 0; i != NumElts; ++i)
10847 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10849 } else if (V == RHS) {
10850 for (unsigned i = 0; i != NumElts; ++i)
10851 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10853 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10854 // If this is an insert of an extract from some other vector, include it.
10855 Value *VecOp = IEI->getOperand(0);
10856 Value *ScalarOp = IEI->getOperand(1);
10857 Value *IdxOp = IEI->getOperand(2);
10859 if (!isa<ConstantInt>(IdxOp))
10861 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10863 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10864 // Okay, we can handle this if the vector we are insertinting into is
10865 // transitively ok.
10866 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10867 // If so, update the mask to reflect the inserted undef.
10868 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10871 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10872 if (isa<ConstantInt>(EI->getOperand(1)) &&
10873 EI->getOperand(0)->getType() == V->getType()) {
10874 unsigned ExtractedIdx =
10875 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10877 // This must be extracting from either LHS or RHS.
10878 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10879 // Okay, we can handle this if the vector we are insertinting into is
10880 // transitively ok.
10881 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10882 // If so, update the mask to reflect the inserted value.
10883 if (EI->getOperand(0) == LHS) {
10884 Mask[InsertedIdx & (NumElts-1)] =
10885 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10887 assert(EI->getOperand(0) == RHS);
10888 Mask[InsertedIdx & (NumElts-1)] =
10889 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10898 // TODO: Handle shufflevector here!
10903 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10904 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10905 /// that computes V and the LHS value of the shuffle.
10906 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10908 assert(isa<VectorType>(V->getType()) &&
10909 (RHS == 0 || V->getType() == RHS->getType()) &&
10910 "Invalid shuffle!");
10911 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10913 if (isa<UndefValue>(V)) {
10914 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10916 } else if (isa<ConstantAggregateZero>(V)) {
10917 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10919 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10920 // If this is an insert of an extract from some other vector, include it.
10921 Value *VecOp = IEI->getOperand(0);
10922 Value *ScalarOp = IEI->getOperand(1);
10923 Value *IdxOp = IEI->getOperand(2);
10925 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10926 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10927 EI->getOperand(0)->getType() == V->getType()) {
10928 unsigned ExtractedIdx =
10929 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10930 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10932 // Either the extracted from or inserted into vector must be RHSVec,
10933 // otherwise we'd end up with a shuffle of three inputs.
10934 if (EI->getOperand(0) == RHS || RHS == 0) {
10935 RHS = EI->getOperand(0);
10936 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10937 Mask[InsertedIdx & (NumElts-1)] =
10938 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10942 if (VecOp == RHS) {
10943 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10944 // Everything but the extracted element is replaced with the RHS.
10945 for (unsigned i = 0; i != NumElts; ++i) {
10946 if (i != InsertedIdx)
10947 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10952 // If this insertelement is a chain that comes from exactly these two
10953 // vectors, return the vector and the effective shuffle.
10954 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10955 return EI->getOperand(0);
10960 // TODO: Handle shufflevector here!
10962 // Otherwise, can't do anything fancy. Return an identity vector.
10963 for (unsigned i = 0; i != NumElts; ++i)
10964 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10968 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10969 Value *VecOp = IE.getOperand(0);
10970 Value *ScalarOp = IE.getOperand(1);
10971 Value *IdxOp = IE.getOperand(2);
10973 // Inserting an undef or into an undefined place, remove this.
10974 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10975 ReplaceInstUsesWith(IE, VecOp);
10977 // If the inserted element was extracted from some other vector, and if the
10978 // indexes are constant, try to turn this into a shufflevector operation.
10979 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10980 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10981 EI->getOperand(0)->getType() == IE.getType()) {
10982 unsigned NumVectorElts = IE.getType()->getNumElements();
10983 unsigned ExtractedIdx =
10984 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10985 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10987 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10988 return ReplaceInstUsesWith(IE, VecOp);
10990 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10991 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10993 // If we are extracting a value from a vector, then inserting it right
10994 // back into the same place, just use the input vector.
10995 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10996 return ReplaceInstUsesWith(IE, VecOp);
10998 // We could theoretically do this for ANY input. However, doing so could
10999 // turn chains of insertelement instructions into a chain of shufflevector
11000 // instructions, and right now we do not merge shufflevectors. As such,
11001 // only do this in a situation where it is clear that there is benefit.
11002 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11003 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11004 // the values of VecOp, except then one read from EIOp0.
11005 // Build a new shuffle mask.
11006 std::vector<Constant*> Mask;
11007 if (isa<UndefValue>(VecOp))
11008 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11010 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11011 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11014 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11015 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11016 ConstantVector::get(Mask));
11019 // If this insertelement isn't used by some other insertelement, turn it
11020 // (and any insertelements it points to), into one big shuffle.
11021 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11022 std::vector<Constant*> Mask;
11024 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11025 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11026 // We now have a shuffle of LHS, RHS, Mask.
11027 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11036 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11037 Value *LHS = SVI.getOperand(0);
11038 Value *RHS = SVI.getOperand(1);
11039 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11041 bool MadeChange = false;
11043 // Undefined shuffle mask -> undefined value.
11044 if (isa<UndefValue>(SVI.getOperand(2)))
11045 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11047 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11048 // the undef, change them to undefs.
11049 if (isa<UndefValue>(SVI.getOperand(1))) {
11050 // Scan to see if there are any references to the RHS. If so, replace them
11051 // with undef element refs and set MadeChange to true.
11052 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11053 if (Mask[i] >= e && Mask[i] != 2*e) {
11060 // Remap any references to RHS to use LHS.
11061 std::vector<Constant*> Elts;
11062 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11063 if (Mask[i] == 2*e)
11064 Elts.push_back(UndefValue::get(Type::Int32Ty));
11066 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11068 SVI.setOperand(2, ConstantVector::get(Elts));
11072 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11073 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11074 if (LHS == RHS || isa<UndefValue>(LHS)) {
11075 if (isa<UndefValue>(LHS) && LHS == RHS) {
11076 // shuffle(undef,undef,mask) -> undef.
11077 return ReplaceInstUsesWith(SVI, LHS);
11080 // Remap any references to RHS to use LHS.
11081 std::vector<Constant*> Elts;
11082 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11083 if (Mask[i] >= 2*e)
11084 Elts.push_back(UndefValue::get(Type::Int32Ty));
11086 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11087 (Mask[i] < e && isa<UndefValue>(LHS)))
11088 Mask[i] = 2*e; // Turn into undef.
11090 Mask[i] &= (e-1); // Force to LHS.
11091 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11094 SVI.setOperand(0, SVI.getOperand(1));
11095 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11096 SVI.setOperand(2, ConstantVector::get(Elts));
11097 LHS = SVI.getOperand(0);
11098 RHS = SVI.getOperand(1);
11102 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11103 bool isLHSID = true, isRHSID = true;
11105 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11106 if (Mask[i] >= e*2) continue; // Ignore undef values.
11107 // Is this an identity shuffle of the LHS value?
11108 isLHSID &= (Mask[i] == i);
11110 // Is this an identity shuffle of the RHS value?
11111 isRHSID &= (Mask[i]-e == i);
11114 // Eliminate identity shuffles.
11115 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11116 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11118 // If the LHS is a shufflevector itself, see if we can combine it with this
11119 // one without producing an unusual shuffle. Here we are really conservative:
11120 // we are absolutely afraid of producing a shuffle mask not in the input
11121 // program, because the code gen may not be smart enough to turn a merged
11122 // shuffle into two specific shuffles: it may produce worse code. As such,
11123 // we only merge two shuffles if the result is one of the two input shuffle
11124 // masks. In this case, merging the shuffles just removes one instruction,
11125 // which we know is safe. This is good for things like turning:
11126 // (splat(splat)) -> splat.
11127 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11128 if (isa<UndefValue>(RHS)) {
11129 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11131 std::vector<unsigned> NewMask;
11132 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11133 if (Mask[i] >= 2*e)
11134 NewMask.push_back(2*e);
11136 NewMask.push_back(LHSMask[Mask[i]]);
11138 // If the result mask is equal to the src shuffle or this shuffle mask, do
11139 // the replacement.
11140 if (NewMask == LHSMask || NewMask == Mask) {
11141 std::vector<Constant*> Elts;
11142 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11143 if (NewMask[i] >= e*2) {
11144 Elts.push_back(UndefValue::get(Type::Int32Ty));
11146 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11149 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11150 LHSSVI->getOperand(1),
11151 ConstantVector::get(Elts));
11156 return MadeChange ? &SVI : 0;
11162 /// TryToSinkInstruction - Try to move the specified instruction from its
11163 /// current block into the beginning of DestBlock, which can only happen if it's
11164 /// safe to move the instruction past all of the instructions between it and the
11165 /// end of its block.
11166 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11167 assert(I->hasOneUse() && "Invariants didn't hold!");
11169 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11170 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11173 // Do not sink alloca instructions out of the entry block.
11174 if (isa<AllocaInst>(I) && I->getParent() ==
11175 &DestBlock->getParent()->getEntryBlock())
11178 // We can only sink load instructions if there is nothing between the load and
11179 // the end of block that could change the value.
11180 if (I->mayReadFromMemory()) {
11181 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11183 if (Scan->mayWriteToMemory())
11187 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11189 I->moveBefore(InsertPos);
11195 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11196 /// all reachable code to the worklist.
11198 /// This has a couple of tricks to make the code faster and more powerful. In
11199 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11200 /// them to the worklist (this significantly speeds up instcombine on code where
11201 /// many instructions are dead or constant). Additionally, if we find a branch
11202 /// whose condition is a known constant, we only visit the reachable successors.
11204 static void AddReachableCodeToWorklist(BasicBlock *BB,
11205 SmallPtrSet<BasicBlock*, 64> &Visited,
11207 const TargetData *TD) {
11208 std::vector<BasicBlock*> Worklist;
11209 Worklist.push_back(BB);
11211 while (!Worklist.empty()) {
11212 BB = Worklist.back();
11213 Worklist.pop_back();
11215 // We have now visited this block! If we've already been here, ignore it.
11216 if (!Visited.insert(BB)) continue;
11218 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11219 Instruction *Inst = BBI++;
11221 // DCE instruction if trivially dead.
11222 if (isInstructionTriviallyDead(Inst)) {
11224 DOUT << "IC: DCE: " << *Inst;
11225 Inst->eraseFromParent();
11229 // ConstantProp instruction if trivially constant.
11230 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11231 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11232 Inst->replaceAllUsesWith(C);
11234 Inst->eraseFromParent();
11238 IC.AddToWorkList(Inst);
11241 // Recursively visit successors. If this is a branch or switch on a
11242 // constant, only visit the reachable successor.
11243 TerminatorInst *TI = BB->getTerminator();
11244 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11245 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11246 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11247 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11248 Worklist.push_back(ReachableBB);
11251 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11252 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11253 // See if this is an explicit destination.
11254 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11255 if (SI->getCaseValue(i) == Cond) {
11256 BasicBlock *ReachableBB = SI->getSuccessor(i);
11257 Worklist.push_back(ReachableBB);
11261 // Otherwise it is the default destination.
11262 Worklist.push_back(SI->getSuccessor(0));
11267 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11268 Worklist.push_back(TI->getSuccessor(i));
11272 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11273 bool Changed = false;
11274 TD = &getAnalysis<TargetData>();
11276 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11277 << F.getNameStr() << "\n");
11280 // Do a depth-first traversal of the function, populate the worklist with
11281 // the reachable instructions. Ignore blocks that are not reachable. Keep
11282 // track of which blocks we visit.
11283 SmallPtrSet<BasicBlock*, 64> Visited;
11284 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11286 // Do a quick scan over the function. If we find any blocks that are
11287 // unreachable, remove any instructions inside of them. This prevents
11288 // the instcombine code from having to deal with some bad special cases.
11289 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11290 if (!Visited.count(BB)) {
11291 Instruction *Term = BB->getTerminator();
11292 while (Term != BB->begin()) { // Remove instrs bottom-up
11293 BasicBlock::iterator I = Term; --I;
11295 DOUT << "IC: DCE: " << *I;
11298 if (!I->use_empty())
11299 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11300 I->eraseFromParent();
11305 while (!Worklist.empty()) {
11306 Instruction *I = RemoveOneFromWorkList();
11307 if (I == 0) continue; // skip null values.
11309 // Check to see if we can DCE the instruction.
11310 if (isInstructionTriviallyDead(I)) {
11311 // Add operands to the worklist.
11312 if (I->getNumOperands() < 4)
11313 AddUsesToWorkList(*I);
11316 DOUT << "IC: DCE: " << *I;
11318 I->eraseFromParent();
11319 RemoveFromWorkList(I);
11323 // Instruction isn't dead, see if we can constant propagate it.
11324 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11325 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11327 // Add operands to the worklist.
11328 AddUsesToWorkList(*I);
11329 ReplaceInstUsesWith(*I, C);
11332 I->eraseFromParent();
11333 RemoveFromWorkList(I);
11337 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
11338 // See if we can constant fold its operands.
11339 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
11340 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
11341 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
11347 // See if we can trivially sink this instruction to a successor basic block.
11348 // FIXME: Remove GetResultInst test when first class support for aggregates
11350 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11351 BasicBlock *BB = I->getParent();
11352 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11353 if (UserParent != BB) {
11354 bool UserIsSuccessor = false;
11355 // See if the user is one of our successors.
11356 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11357 if (*SI == UserParent) {
11358 UserIsSuccessor = true;
11362 // If the user is one of our immediate successors, and if that successor
11363 // only has us as a predecessors (we'd have to split the critical edge
11364 // otherwise), we can keep going.
11365 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11366 next(pred_begin(UserParent)) == pred_end(UserParent))
11367 // Okay, the CFG is simple enough, try to sink this instruction.
11368 Changed |= TryToSinkInstruction(I, UserParent);
11372 // Now that we have an instruction, try combining it to simplify it...
11376 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11377 if (Instruction *Result = visit(*I)) {
11379 // Should we replace the old instruction with a new one?
11381 DOUT << "IC: Old = " << *I
11382 << " New = " << *Result;
11384 // Everything uses the new instruction now.
11385 I->replaceAllUsesWith(Result);
11387 // Push the new instruction and any users onto the worklist.
11388 AddToWorkList(Result);
11389 AddUsersToWorkList(*Result);
11391 // Move the name to the new instruction first.
11392 Result->takeName(I);
11394 // Insert the new instruction into the basic block...
11395 BasicBlock *InstParent = I->getParent();
11396 BasicBlock::iterator InsertPos = I;
11398 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11399 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11402 InstParent->getInstList().insert(InsertPos, Result);
11404 // Make sure that we reprocess all operands now that we reduced their
11406 AddUsesToWorkList(*I);
11408 // Instructions can end up on the worklist more than once. Make sure
11409 // we do not process an instruction that has been deleted.
11410 RemoveFromWorkList(I);
11412 // Erase the old instruction.
11413 InstParent->getInstList().erase(I);
11416 DOUT << "IC: Mod = " << OrigI
11417 << " New = " << *I;
11420 // If the instruction was modified, it's possible that it is now dead.
11421 // if so, remove it.
11422 if (isInstructionTriviallyDead(I)) {
11423 // Make sure we process all operands now that we are reducing their
11425 AddUsesToWorkList(*I);
11427 // Instructions may end up in the worklist more than once. Erase all
11428 // occurrences of this instruction.
11429 RemoveFromWorkList(I);
11430 I->eraseFromParent();
11433 AddUsersToWorkList(*I);
11440 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11442 // Do an explicit clear, this shrinks the map if needed.
11443 WorklistMap.clear();
11448 bool InstCombiner::runOnFunction(Function &F) {
11449 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11451 bool EverMadeChange = false;
11453 // Iterate while there is work to do.
11454 unsigned Iteration = 0;
11455 while (DoOneIteration(F, Iteration++))
11456 EverMadeChange = true;
11457 return EverMadeChange;
11460 FunctionPass *llvm::createInstructionCombiningPass() {
11461 return new InstCombiner();