1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 SmallVector<Instruction*, 256> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass(&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 bool SimplifyDivRemOfSelect(BinaryOperator &I);
176 Instruction *commonRemTransforms(BinaryOperator &I);
177 Instruction *commonIRemTransforms(BinaryOperator &I);
178 Instruction *commonDivTransforms(BinaryOperator &I);
179 Instruction *commonIDivTransforms(BinaryOperator &I);
180 Instruction *visitUDiv(BinaryOperator &I);
181 Instruction *visitSDiv(BinaryOperator &I);
182 Instruction *visitFDiv(BinaryOperator &I);
183 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
184 Instruction *visitAnd(BinaryOperator &I);
185 Instruction *visitOr (BinaryOperator &I);
186 Instruction *visitXor(BinaryOperator &I);
187 Instruction *visitShl(BinaryOperator &I);
188 Instruction *visitAShr(BinaryOperator &I);
189 Instruction *visitLShr(BinaryOperator &I);
190 Instruction *commonShiftTransforms(BinaryOperator &I);
191 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
193 Instruction *visitFCmpInst(FCmpInst &I);
194 Instruction *visitICmpInst(ICmpInst &I);
195 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
196 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
199 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
200 ConstantInt *DivRHS);
202 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
203 ICmpInst::Predicate Cond, Instruction &I);
204 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
206 Instruction *commonCastTransforms(CastInst &CI);
207 Instruction *commonIntCastTransforms(CastInst &CI);
208 Instruction *commonPointerCastTransforms(CastInst &CI);
209 Instruction *visitTrunc(TruncInst &CI);
210 Instruction *visitZExt(ZExtInst &CI);
211 Instruction *visitSExt(SExtInst &CI);
212 Instruction *visitFPTrunc(FPTruncInst &CI);
213 Instruction *visitFPExt(CastInst &CI);
214 Instruction *visitFPToUI(FPToUIInst &FI);
215 Instruction *visitFPToSI(FPToSIInst &FI);
216 Instruction *visitUIToFP(CastInst &CI);
217 Instruction *visitSIToFP(CastInst &CI);
218 Instruction *visitPtrToInt(CastInst &CI);
219 Instruction *visitIntToPtr(IntToPtrInst &CI);
220 Instruction *visitBitCast(BitCastInst &CI);
221 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
223 Instruction *visitSelectInst(SelectInst &SI);
224 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
225 Instruction *visitCallInst(CallInst &CI);
226 Instruction *visitInvokeInst(InvokeInst &II);
227 Instruction *visitPHINode(PHINode &PN);
228 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
229 Instruction *visitAllocationInst(AllocationInst &AI);
230 Instruction *visitFreeInst(FreeInst &FI);
231 Instruction *visitLoadInst(LoadInst &LI);
232 Instruction *visitStoreInst(StoreInst &SI);
233 Instruction *visitBranchInst(BranchInst &BI);
234 Instruction *visitSwitchInst(SwitchInst &SI);
235 Instruction *visitInsertElementInst(InsertElementInst &IE);
236 Instruction *visitExtractElementInst(ExtractElementInst &EI);
237 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
238 Instruction *visitExtractValueInst(ExtractValueInst &EV);
240 // visitInstruction - Specify what to return for unhandled instructions...
241 Instruction *visitInstruction(Instruction &I) { return 0; }
244 Instruction *visitCallSite(CallSite CS);
245 bool transformConstExprCastCall(CallSite CS);
246 Instruction *transformCallThroughTrampoline(CallSite CS);
247 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
248 bool DoXform = true);
249 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
252 // InsertNewInstBefore - insert an instruction New before instruction Old
253 // in the program. Add the new instruction to the worklist.
255 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
256 assert(New && New->getParent() == 0 &&
257 "New instruction already inserted into a basic block!");
258 BasicBlock *BB = Old.getParent();
259 BB->getInstList().insert(&Old, New); // Insert inst
264 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
265 /// This also adds the cast to the worklist. Finally, this returns the
267 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
269 if (V->getType() == Ty) return V;
271 if (Constant *CV = dyn_cast<Constant>(V))
272 return ConstantExpr::getCast(opc, CV, Ty);
274 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
279 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
280 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
284 // ReplaceInstUsesWith - This method is to be used when an instruction is
285 // found to be dead, replacable with another preexisting expression. Here
286 // we add all uses of I to the worklist, replace all uses of I with the new
287 // value, then return I, so that the inst combiner will know that I was
290 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
291 AddUsersToWorkList(I); // Add all modified instrs to worklist
293 I.replaceAllUsesWith(V);
296 // If we are replacing the instruction with itself, this must be in a
297 // segment of unreachable code, so just clobber the instruction.
298 I.replaceAllUsesWith(UndefValue::get(I.getType()));
303 // UpdateValueUsesWith - This method is to be used when an value is
304 // found to be replacable with another preexisting expression or was
305 // updated. Here we add all uses of I to the worklist, replace all uses of
306 // I with the new value (unless the instruction was just updated), then
307 // return true, so that the inst combiner will know that I was modified.
309 bool UpdateValueUsesWith(Value *Old, Value *New) {
310 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
312 Old->replaceAllUsesWith(New);
313 if (Instruction *I = dyn_cast<Instruction>(Old))
315 if (Instruction *I = dyn_cast<Instruction>(New))
320 // EraseInstFromFunction - When dealing with an instruction that has side
321 // effects or produces a void value, we can't rely on DCE to delete the
322 // instruction. Instead, visit methods should return the value returned by
324 Instruction *EraseInstFromFunction(Instruction &I) {
325 assert(I.use_empty() && "Cannot erase instruction that is used!");
326 AddUsesToWorkList(I);
327 RemoveFromWorkList(&I);
329 return 0; // Don't do anything with FI
332 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
333 APInt &KnownOne, unsigned Depth = 0) const {
334 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
337 bool MaskedValueIsZero(Value *V, const APInt &Mask,
338 unsigned Depth = 0) const {
339 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
341 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
342 return llvm::ComputeNumSignBits(Op, TD, Depth);
346 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
347 /// InsertBefore instruction. This is specialized a bit to avoid inserting
348 /// casts that are known to not do anything...
350 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
351 Value *V, const Type *DestTy,
352 Instruction *InsertBefore);
354 /// SimplifyCommutative - This performs a few simplifications for
355 /// commutative operators.
356 bool SimplifyCommutative(BinaryOperator &I);
358 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
359 /// most-complex to least-complex order.
360 bool SimplifyCompare(CmpInst &I);
362 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
363 /// on the demanded bits.
364 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
365 APInt& KnownZero, APInt& KnownOne,
368 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
369 uint64_t &UndefElts, unsigned Depth = 0);
371 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
372 // PHI node as operand #0, see if we can fold the instruction into the PHI
373 // (which is only possible if all operands to the PHI are constants).
374 Instruction *FoldOpIntoPhi(Instruction &I);
376 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
377 // operator and they all are only used by the PHI, PHI together their
378 // inputs, and do the operation once, to the result of the PHI.
379 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
380 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
383 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
384 ConstantInt *AndRHS, BinaryOperator &TheAnd);
386 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
387 bool isSub, Instruction &I);
388 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
389 bool isSigned, bool Inside, Instruction &IB);
390 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
391 Instruction *MatchBSwap(BinaryOperator &I);
392 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
393 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
394 Instruction *SimplifyMemSet(MemSetInst *MI);
397 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
399 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
401 int &NumCastsRemoved);
402 unsigned GetOrEnforceKnownAlignment(Value *V,
403 unsigned PrefAlign = 0);
408 char InstCombiner::ID = 0;
409 static RegisterPass<InstCombiner>
410 X("instcombine", "Combine redundant instructions");
412 // getComplexity: Assign a complexity or rank value to LLVM Values...
413 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
414 static unsigned getComplexity(Value *V) {
415 if (isa<Instruction>(V)) {
416 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
420 if (isa<Argument>(V)) return 3;
421 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
424 // isOnlyUse - Return true if this instruction will be deleted if we stop using
426 static bool isOnlyUse(Value *V) {
427 return V->hasOneUse() || isa<Constant>(V);
430 // getPromotedType - Return the specified type promoted as it would be to pass
431 // though a va_arg area...
432 static const Type *getPromotedType(const Type *Ty) {
433 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
434 if (ITy->getBitWidth() < 32)
435 return Type::Int32Ty;
440 /// getBitCastOperand - If the specified operand is a CastInst, a constant
441 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
442 /// operand value, otherwise return null.
443 static Value *getBitCastOperand(Value *V) {
444 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
446 return I->getOperand(0);
447 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
448 // GetElementPtrInst?
449 if (GEP->hasAllZeroIndices())
450 return GEP->getOperand(0);
451 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
452 if (CE->getOpcode() == Instruction::BitCast)
453 // BitCast ConstantExp?
454 return CE->getOperand(0);
455 else if (CE->getOpcode() == Instruction::GetElementPtr) {
456 // GetElementPtr ConstantExp?
457 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
459 ConstantInt *CI = dyn_cast<ConstantInt>(I);
460 if (!CI || !CI->isZero())
461 // Any non-zero indices? Not cast-like.
464 // All-zero indices? This is just like casting.
465 return CE->getOperand(0);
471 /// This function is a wrapper around CastInst::isEliminableCastPair. It
472 /// simply extracts arguments and returns what that function returns.
473 static Instruction::CastOps
474 isEliminableCastPair(
475 const CastInst *CI, ///< The first cast instruction
476 unsigned opcode, ///< The opcode of the second cast instruction
477 const Type *DstTy, ///< The target type for the second cast instruction
478 TargetData *TD ///< The target data for pointer size
481 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
482 const Type *MidTy = CI->getType(); // B from above
484 // Get the opcodes of the two Cast instructions
485 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
486 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
488 return Instruction::CastOps(
489 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
490 DstTy, TD->getIntPtrType()));
493 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
494 /// in any code being generated. It does not require codegen if V is simple
495 /// enough or if the cast can be folded into other casts.
496 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
497 const Type *Ty, TargetData *TD) {
498 if (V->getType() == Ty || isa<Constant>(V)) return false;
500 // If this is another cast that can be eliminated, it isn't codegen either.
501 if (const CastInst *CI = dyn_cast<CastInst>(V))
502 if (isEliminableCastPair(CI, opcode, Ty, TD))
507 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
508 /// InsertBefore instruction. This is specialized a bit to avoid inserting
509 /// casts that are known to not do anything...
511 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
512 Value *V, const Type *DestTy,
513 Instruction *InsertBefore) {
514 if (V->getType() == DestTy) return V;
515 if (Constant *C = dyn_cast<Constant>(V))
516 return ConstantExpr::getCast(opcode, C, DestTy);
518 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
521 // SimplifyCommutative - This performs a few simplifications for commutative
524 // 1. Order operands such that they are listed from right (least complex) to
525 // left (most complex). This puts constants before unary operators before
528 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
529 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
531 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
532 bool Changed = false;
533 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
534 Changed = !I.swapOperands();
536 if (!I.isAssociative()) return Changed;
537 Instruction::BinaryOps Opcode = I.getOpcode();
538 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
539 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
540 if (isa<Constant>(I.getOperand(1))) {
541 Constant *Folded = ConstantExpr::get(I.getOpcode(),
542 cast<Constant>(I.getOperand(1)),
543 cast<Constant>(Op->getOperand(1)));
544 I.setOperand(0, Op->getOperand(0));
545 I.setOperand(1, Folded);
547 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
548 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
549 isOnlyUse(Op) && isOnlyUse(Op1)) {
550 Constant *C1 = cast<Constant>(Op->getOperand(1));
551 Constant *C2 = cast<Constant>(Op1->getOperand(1));
553 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
554 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
555 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
559 I.setOperand(0, New);
560 I.setOperand(1, Folded);
567 /// SimplifyCompare - For a CmpInst this function just orders the operands
568 /// so that theyare listed from right (least complex) to left (most complex).
569 /// This puts constants before unary operators before binary operators.
570 bool InstCombiner::SimplifyCompare(CmpInst &I) {
571 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
574 // Compare instructions are not associative so there's nothing else we can do.
578 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
579 // if the LHS is a constant zero (which is the 'negate' form).
581 static inline Value *dyn_castNegVal(Value *V) {
582 if (BinaryOperator::isNeg(V))
583 return BinaryOperator::getNegArgument(V);
585 // Constants can be considered to be negated values if they can be folded.
586 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
587 return ConstantExpr::getNeg(C);
589 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
590 if (C->getType()->getElementType()->isInteger())
591 return ConstantExpr::getNeg(C);
596 static inline Value *dyn_castNotVal(Value *V) {
597 if (BinaryOperator::isNot(V))
598 return BinaryOperator::getNotArgument(V);
600 // Constants can be considered to be not'ed values...
601 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
602 return ConstantInt::get(~C->getValue());
606 // dyn_castFoldableMul - If this value is a multiply that can be folded into
607 // other computations (because it has a constant operand), return the
608 // non-constant operand of the multiply, and set CST to point to the multiplier.
609 // Otherwise, return null.
611 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
612 if (V->hasOneUse() && V->getType()->isInteger())
613 if (Instruction *I = dyn_cast<Instruction>(V)) {
614 if (I->getOpcode() == Instruction::Mul)
615 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
616 return I->getOperand(0);
617 if (I->getOpcode() == Instruction::Shl)
618 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
619 // The multiplier is really 1 << CST.
620 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
621 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
622 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
623 return I->getOperand(0);
629 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
630 /// expression, return it.
631 static User *dyn_castGetElementPtr(Value *V) {
632 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
633 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
634 if (CE->getOpcode() == Instruction::GetElementPtr)
635 return cast<User>(V);
639 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
640 /// opcode value. Otherwise return UserOp1.
641 static unsigned getOpcode(const Value *V) {
642 if (const Instruction *I = dyn_cast<Instruction>(V))
643 return I->getOpcode();
644 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
645 return CE->getOpcode();
646 // Use UserOp1 to mean there's no opcode.
647 return Instruction::UserOp1;
650 /// AddOne - Add one to a ConstantInt
651 static ConstantInt *AddOne(ConstantInt *C) {
652 APInt Val(C->getValue());
653 return ConstantInt::get(++Val);
655 /// SubOne - Subtract one from a ConstantInt
656 static ConstantInt *SubOne(ConstantInt *C) {
657 APInt Val(C->getValue());
658 return ConstantInt::get(--Val);
660 /// Add - Add two ConstantInts together
661 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
662 return ConstantInt::get(C1->getValue() + C2->getValue());
664 /// And - Bitwise AND two ConstantInts together
665 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
666 return ConstantInt::get(C1->getValue() & C2->getValue());
668 /// Subtract - Subtract one ConstantInt from another
669 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
670 return ConstantInt::get(C1->getValue() - C2->getValue());
672 /// Multiply - Multiply two ConstantInts together
673 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
674 return ConstantInt::get(C1->getValue() * C2->getValue());
676 /// MultiplyOverflows - True if the multiply can not be expressed in an int
678 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
679 uint32_t W = C1->getBitWidth();
680 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
689 APInt MulExt = LHSExt * RHSExt;
692 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
693 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
694 return MulExt.slt(Min) || MulExt.sgt(Max);
696 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
700 /// ShrinkDemandedConstant - Check to see if the specified operand of the
701 /// specified instruction is a constant integer. If so, check to see if there
702 /// are any bits set in the constant that are not demanded. If so, shrink the
703 /// constant and return true.
704 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
706 assert(I && "No instruction?");
707 assert(OpNo < I->getNumOperands() && "Operand index too large");
709 // If the operand is not a constant integer, nothing to do.
710 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
711 if (!OpC) return false;
713 // If there are no bits set that aren't demanded, nothing to do.
714 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
715 if ((~Demanded & OpC->getValue()) == 0)
718 // This instruction is producing bits that are not demanded. Shrink the RHS.
719 Demanded &= OpC->getValue();
720 I->setOperand(OpNo, ConstantInt::get(Demanded));
724 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
725 // set of known zero and one bits, compute the maximum and minimum values that
726 // could have the specified known zero and known one bits, returning them in
728 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
729 const APInt& KnownZero,
730 const APInt& KnownOne,
731 APInt& Min, APInt& Max) {
732 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
733 assert(KnownZero.getBitWidth() == BitWidth &&
734 KnownOne.getBitWidth() == BitWidth &&
735 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
736 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
737 APInt UnknownBits = ~(KnownZero|KnownOne);
739 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
740 // bit if it is unknown.
742 Max = KnownOne|UnknownBits;
744 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
746 Max.clear(BitWidth-1);
750 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
751 // a set of known zero and one bits, compute the maximum and minimum values that
752 // could have the specified known zero and known one bits, returning them in
754 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
755 const APInt &KnownZero,
756 const APInt &KnownOne,
757 APInt &Min, APInt &Max) {
758 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
759 assert(KnownZero.getBitWidth() == BitWidth &&
760 KnownOne.getBitWidth() == BitWidth &&
761 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
762 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
763 APInt UnknownBits = ~(KnownZero|KnownOne);
765 // The minimum value is when the unknown bits are all zeros.
767 // The maximum value is when the unknown bits are all ones.
768 Max = KnownOne|UnknownBits;
771 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
772 /// value based on the demanded bits. When this function is called, it is known
773 /// that only the bits set in DemandedMask of the result of V are ever used
774 /// downstream. Consequently, depending on the mask and V, it may be possible
775 /// to replace V with a constant or one of its operands. In such cases, this
776 /// function does the replacement and returns true. In all other cases, it
777 /// returns false after analyzing the expression and setting KnownOne and known
778 /// to be one in the expression. KnownZero contains all the bits that are known
779 /// to be zero in the expression. These are provided to potentially allow the
780 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
781 /// the expression. KnownOne and KnownZero always follow the invariant that
782 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
783 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
784 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
785 /// and KnownOne must all be the same.
786 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
787 APInt& KnownZero, APInt& KnownOne,
789 assert(V != 0 && "Null pointer of Value???");
790 assert(Depth <= 6 && "Limit Search Depth");
791 uint32_t BitWidth = DemandedMask.getBitWidth();
792 const IntegerType *VTy = cast<IntegerType>(V->getType());
793 assert(VTy->getBitWidth() == BitWidth &&
794 KnownZero.getBitWidth() == BitWidth &&
795 KnownOne.getBitWidth() == BitWidth &&
796 "Value *V, DemandedMask, KnownZero and KnownOne \
797 must have same BitWidth");
798 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
799 // We know all of the bits for a constant!
800 KnownOne = CI->getValue() & DemandedMask;
801 KnownZero = ~KnownOne & DemandedMask;
807 if (!V->hasOneUse()) { // Other users may use these bits.
808 if (Depth != 0) { // Not at the root.
809 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
810 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
813 // If this is the root being simplified, allow it to have multiple uses,
814 // just set the DemandedMask to all bits.
815 DemandedMask = APInt::getAllOnesValue(BitWidth);
816 } else if (DemandedMask == 0) { // Not demanding any bits from V.
817 if (V != UndefValue::get(VTy))
818 return UpdateValueUsesWith(V, UndefValue::get(VTy));
820 } else if (Depth == 6) { // Limit search depth.
824 Instruction *I = dyn_cast<Instruction>(V);
825 if (!I) return false; // Only analyze instructions.
827 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
828 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
829 switch (I->getOpcode()) {
831 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
833 case Instruction::And:
834 // If either the LHS or the RHS are Zero, the result is zero.
835 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
836 RHSKnownZero, RHSKnownOne, Depth+1))
838 assert((RHSKnownZero & RHSKnownOne) == 0 &&
839 "Bits known to be one AND zero?");
841 // If something is known zero on the RHS, the bits aren't demanded on the
843 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
844 LHSKnownZero, LHSKnownOne, Depth+1))
846 assert((LHSKnownZero & LHSKnownOne) == 0 &&
847 "Bits known to be one AND zero?");
849 // If all of the demanded bits are known 1 on one side, return the other.
850 // These bits cannot contribute to the result of the 'and'.
851 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
852 (DemandedMask & ~LHSKnownZero))
853 return UpdateValueUsesWith(I, I->getOperand(0));
854 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
855 (DemandedMask & ~RHSKnownZero))
856 return UpdateValueUsesWith(I, I->getOperand(1));
858 // If all of the demanded bits in the inputs are known zeros, return zero.
859 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
860 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
862 // If the RHS is a constant, see if we can simplify it.
863 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
864 return UpdateValueUsesWith(I, I);
866 // Output known-1 bits are only known if set in both the LHS & RHS.
867 RHSKnownOne &= LHSKnownOne;
868 // Output known-0 are known to be clear if zero in either the LHS | RHS.
869 RHSKnownZero |= LHSKnownZero;
871 case Instruction::Or:
872 // If either the LHS or the RHS are One, the result is One.
873 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
874 RHSKnownZero, RHSKnownOne, Depth+1))
876 assert((RHSKnownZero & RHSKnownOne) == 0 &&
877 "Bits known to be one AND zero?");
878 // If something is known one on the RHS, the bits aren't demanded on the
880 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
881 LHSKnownZero, LHSKnownOne, Depth+1))
883 assert((LHSKnownZero & LHSKnownOne) == 0 &&
884 "Bits known to be one AND zero?");
886 // If all of the demanded bits are known zero on one side, return the other.
887 // These bits cannot contribute to the result of the 'or'.
888 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
889 (DemandedMask & ~LHSKnownOne))
890 return UpdateValueUsesWith(I, I->getOperand(0));
891 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
892 (DemandedMask & ~RHSKnownOne))
893 return UpdateValueUsesWith(I, I->getOperand(1));
895 // If all of the potentially set bits on one side are known to be set on
896 // the other side, just use the 'other' side.
897 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
898 (DemandedMask & (~RHSKnownZero)))
899 return UpdateValueUsesWith(I, I->getOperand(0));
900 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
901 (DemandedMask & (~LHSKnownZero)))
902 return UpdateValueUsesWith(I, I->getOperand(1));
904 // If the RHS is a constant, see if we can simplify it.
905 if (ShrinkDemandedConstant(I, 1, DemandedMask))
906 return UpdateValueUsesWith(I, I);
908 // Output known-0 bits are only known if clear in both the LHS & RHS.
909 RHSKnownZero &= LHSKnownZero;
910 // Output known-1 are known to be set if set in either the LHS | RHS.
911 RHSKnownOne |= LHSKnownOne;
913 case Instruction::Xor: {
914 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
915 RHSKnownZero, RHSKnownOne, Depth+1))
917 assert((RHSKnownZero & RHSKnownOne) == 0 &&
918 "Bits known to be one AND zero?");
919 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
920 LHSKnownZero, LHSKnownOne, Depth+1))
922 assert((LHSKnownZero & LHSKnownOne) == 0 &&
923 "Bits known to be one AND zero?");
925 // If all of the demanded bits are known zero on one side, return the other.
926 // These bits cannot contribute to the result of the 'xor'.
927 if ((DemandedMask & RHSKnownZero) == DemandedMask)
928 return UpdateValueUsesWith(I, I->getOperand(0));
929 if ((DemandedMask & LHSKnownZero) == DemandedMask)
930 return UpdateValueUsesWith(I, I->getOperand(1));
932 // Output known-0 bits are known if clear or set in both the LHS & RHS.
933 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
934 (RHSKnownOne & LHSKnownOne);
935 // Output known-1 are known to be set if set in only one of the LHS, RHS.
936 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
937 (RHSKnownOne & LHSKnownZero);
939 // If all of the demanded bits are known to be zero on one side or the
940 // other, turn this into an *inclusive* or.
941 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
942 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
944 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
946 InsertNewInstBefore(Or, *I);
947 return UpdateValueUsesWith(I, Or);
950 // If all of the demanded bits on one side are known, and all of the set
951 // bits on that side are also known to be set on the other side, turn this
952 // into an AND, as we know the bits will be cleared.
953 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
954 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
956 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
957 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
959 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
960 InsertNewInstBefore(And, *I);
961 return UpdateValueUsesWith(I, And);
965 // If the RHS is a constant, see if we can simplify it.
966 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
967 if (ShrinkDemandedConstant(I, 1, DemandedMask))
968 return UpdateValueUsesWith(I, I);
970 RHSKnownZero = KnownZeroOut;
971 RHSKnownOne = KnownOneOut;
974 case Instruction::Select:
975 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
976 RHSKnownZero, RHSKnownOne, Depth+1))
978 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
979 LHSKnownZero, LHSKnownOne, Depth+1))
981 assert((RHSKnownZero & RHSKnownOne) == 0 &&
982 "Bits known to be one AND zero?");
983 assert((LHSKnownZero & LHSKnownOne) == 0 &&
984 "Bits known to be one AND zero?");
986 // If the operands are constants, see if we can simplify them.
987 if (ShrinkDemandedConstant(I, 1, DemandedMask))
988 return UpdateValueUsesWith(I, I);
989 if (ShrinkDemandedConstant(I, 2, DemandedMask))
990 return UpdateValueUsesWith(I, I);
992 // Only known if known in both the LHS and RHS.
993 RHSKnownOne &= LHSKnownOne;
994 RHSKnownZero &= LHSKnownZero;
996 case Instruction::Trunc: {
998 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
999 DemandedMask.zext(truncBf);
1000 RHSKnownZero.zext(truncBf);
1001 RHSKnownOne.zext(truncBf);
1002 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1003 RHSKnownZero, RHSKnownOne, Depth+1))
1005 DemandedMask.trunc(BitWidth);
1006 RHSKnownZero.trunc(BitWidth);
1007 RHSKnownOne.trunc(BitWidth);
1008 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1009 "Bits known to be one AND zero?");
1012 case Instruction::BitCast:
1013 if (!I->getOperand(0)->getType()->isInteger())
1016 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1017 RHSKnownZero, RHSKnownOne, Depth+1))
1019 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1020 "Bits known to be one AND zero?");
1022 case Instruction::ZExt: {
1023 // Compute the bits in the result that are not present in the input.
1024 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1025 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1027 DemandedMask.trunc(SrcBitWidth);
1028 RHSKnownZero.trunc(SrcBitWidth);
1029 RHSKnownOne.trunc(SrcBitWidth);
1030 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1031 RHSKnownZero, RHSKnownOne, Depth+1))
1033 DemandedMask.zext(BitWidth);
1034 RHSKnownZero.zext(BitWidth);
1035 RHSKnownOne.zext(BitWidth);
1036 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1037 "Bits known to be one AND zero?");
1038 // The top bits are known to be zero.
1039 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1042 case Instruction::SExt: {
1043 // Compute the bits in the result that are not present in the input.
1044 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1045 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1047 APInt InputDemandedBits = DemandedMask &
1048 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1050 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1051 // If any of the sign extended bits are demanded, we know that the sign
1053 if ((NewBits & DemandedMask) != 0)
1054 InputDemandedBits.set(SrcBitWidth-1);
1056 InputDemandedBits.trunc(SrcBitWidth);
1057 RHSKnownZero.trunc(SrcBitWidth);
1058 RHSKnownOne.trunc(SrcBitWidth);
1059 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1060 RHSKnownZero, RHSKnownOne, Depth+1))
1062 InputDemandedBits.zext(BitWidth);
1063 RHSKnownZero.zext(BitWidth);
1064 RHSKnownOne.zext(BitWidth);
1065 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1066 "Bits known to be one AND zero?");
1068 // If the sign bit of the input is known set or clear, then we know the
1069 // top bits of the result.
1071 // If the input sign bit is known zero, or if the NewBits are not demanded
1072 // convert this into a zero extension.
1073 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1075 // Convert to ZExt cast
1076 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1077 return UpdateValueUsesWith(I, NewCast);
1078 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1079 RHSKnownOne |= NewBits;
1083 case Instruction::Add: {
1084 // Figure out what the input bits are. If the top bits of the and result
1085 // are not demanded, then the add doesn't demand them from its input
1087 uint32_t NLZ = DemandedMask.countLeadingZeros();
1089 // If there is a constant on the RHS, there are a variety of xformations
1091 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1092 // If null, this should be simplified elsewhere. Some of the xforms here
1093 // won't work if the RHS is zero.
1097 // If the top bit of the output is demanded, demand everything from the
1098 // input. Otherwise, we demand all the input bits except NLZ top bits.
1099 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1101 // Find information about known zero/one bits in the input.
1102 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1103 LHSKnownZero, LHSKnownOne, Depth+1))
1106 // If the RHS of the add has bits set that can't affect the input, reduce
1108 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1109 return UpdateValueUsesWith(I, I);
1111 // Avoid excess work.
1112 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1115 // Turn it into OR if input bits are zero.
1116 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1118 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1120 InsertNewInstBefore(Or, *I);
1121 return UpdateValueUsesWith(I, Or);
1124 // We can say something about the output known-zero and known-one bits,
1125 // depending on potential carries from the input constant and the
1126 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1127 // bits set and the RHS constant is 0x01001, then we know we have a known
1128 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1130 // To compute this, we first compute the potential carry bits. These are
1131 // the bits which may be modified. I'm not aware of a better way to do
1133 const APInt& RHSVal = RHS->getValue();
1134 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1136 // Now that we know which bits have carries, compute the known-1/0 sets.
1138 // Bits are known one if they are known zero in one operand and one in the
1139 // other, and there is no input carry.
1140 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1141 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1143 // Bits are known zero if they are known zero in both operands and there
1144 // is no input carry.
1145 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1147 // If the high-bits of this ADD are not demanded, then it does not demand
1148 // the high bits of its LHS or RHS.
1149 if (DemandedMask[BitWidth-1] == 0) {
1150 // Right fill the mask of bits for this ADD to demand the most
1151 // significant bit and all those below it.
1152 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1153 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1154 LHSKnownZero, LHSKnownOne, Depth+1))
1156 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1157 LHSKnownZero, LHSKnownOne, Depth+1))
1163 case Instruction::Sub:
1164 // If the high-bits of this SUB are not demanded, then it does not demand
1165 // the high bits of its LHS or RHS.
1166 if (DemandedMask[BitWidth-1] == 0) {
1167 // Right fill the mask of bits for this SUB to demand the most
1168 // significant bit and all those below it.
1169 uint32_t NLZ = DemandedMask.countLeadingZeros();
1170 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1171 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1172 LHSKnownZero, LHSKnownOne, Depth+1))
1174 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1175 LHSKnownZero, LHSKnownOne, Depth+1))
1178 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1179 // the known zeros and ones.
1180 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1182 case Instruction::Shl:
1183 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1184 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1185 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1186 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1187 RHSKnownZero, RHSKnownOne, Depth+1))
1189 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1190 "Bits known to be one AND zero?");
1191 RHSKnownZero <<= ShiftAmt;
1192 RHSKnownOne <<= ShiftAmt;
1193 // low bits known zero.
1195 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1198 case Instruction::LShr:
1199 // For a logical shift right
1200 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1201 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1203 // Unsigned shift right.
1204 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1205 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1206 RHSKnownZero, RHSKnownOne, Depth+1))
1208 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1209 "Bits known to be one AND zero?");
1210 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1211 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1213 // Compute the new bits that are at the top now.
1214 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1215 RHSKnownZero |= HighBits; // high bits known zero.
1219 case Instruction::AShr:
1220 // If this is an arithmetic shift right and only the low-bit is set, we can
1221 // always convert this into a logical shr, even if the shift amount is
1222 // variable. The low bit of the shift cannot be an input sign bit unless
1223 // the shift amount is >= the size of the datatype, which is undefined.
1224 if (DemandedMask == 1) {
1225 // Perform the logical shift right.
1226 Value *NewVal = BinaryOperator::CreateLShr(
1227 I->getOperand(0), I->getOperand(1), I->getName());
1228 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1229 return UpdateValueUsesWith(I, NewVal);
1232 // If the sign bit is the only bit demanded by this ashr, then there is no
1233 // need to do it, the shift doesn't change the high bit.
1234 if (DemandedMask.isSignBit())
1235 return UpdateValueUsesWith(I, I->getOperand(0));
1237 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1238 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1240 // Signed shift right.
1241 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1242 // If any of the "high bits" are demanded, we should set the sign bit as
1244 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1245 DemandedMaskIn.set(BitWidth-1);
1246 if (SimplifyDemandedBits(I->getOperand(0),
1248 RHSKnownZero, RHSKnownOne, Depth+1))
1250 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1251 "Bits known to be one AND zero?");
1252 // Compute the new bits that are at the top now.
1253 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1254 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1255 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1257 // Handle the sign bits.
1258 APInt SignBit(APInt::getSignBit(BitWidth));
1259 // Adjust to where it is now in the mask.
1260 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1262 // If the input sign bit is known to be zero, or if none of the top bits
1263 // are demanded, turn this into an unsigned shift right.
1264 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1265 (HighBits & ~DemandedMask) == HighBits) {
1266 // Perform the logical shift right.
1267 Value *NewVal = BinaryOperator::CreateLShr(
1268 I->getOperand(0), SA, I->getName());
1269 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1270 return UpdateValueUsesWith(I, NewVal);
1271 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1272 RHSKnownOne |= HighBits;
1276 case Instruction::SRem:
1277 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1278 APInt RA = Rem->getValue().abs();
1279 if (RA.isPowerOf2()) {
1280 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1281 return UpdateValueUsesWith(I, I->getOperand(0));
1283 APInt LowBits = RA - 1;
1284 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1285 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1286 LHSKnownZero, LHSKnownOne, Depth+1))
1289 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1290 LHSKnownZero |= ~LowBits;
1292 KnownZero |= LHSKnownZero & DemandedMask;
1294 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1298 case Instruction::URem: {
1299 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1300 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1301 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1302 KnownZero2, KnownOne2, Depth+1))
1305 uint32_t Leaders = KnownZero2.countLeadingOnes();
1306 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1307 KnownZero2, KnownOne2, Depth+1))
1310 Leaders = std::max(Leaders,
1311 KnownZero2.countLeadingOnes());
1312 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1315 case Instruction::Call:
1316 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1317 switch (II->getIntrinsicID()) {
1319 case Intrinsic::bswap: {
1320 // If the only bits demanded come from one byte of the bswap result,
1321 // just shift the input byte into position to eliminate the bswap.
1322 unsigned NLZ = DemandedMask.countLeadingZeros();
1323 unsigned NTZ = DemandedMask.countTrailingZeros();
1325 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1326 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1327 // have 14 leading zeros, round to 8.
1330 // If we need exactly one byte, we can do this transformation.
1331 if (BitWidth-NLZ-NTZ == 8) {
1332 unsigned ResultBit = NTZ;
1333 unsigned InputBit = BitWidth-NTZ-8;
1335 // Replace this with either a left or right shift to get the byte into
1337 Instruction *NewVal;
1338 if (InputBit > ResultBit)
1339 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1340 ConstantInt::get(I->getType(), InputBit-ResultBit));
1342 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1343 ConstantInt::get(I->getType(), ResultBit-InputBit));
1344 NewVal->takeName(I);
1345 InsertNewInstBefore(NewVal, *I);
1346 return UpdateValueUsesWith(I, NewVal);
1349 // TODO: Could compute known zero/one bits based on the input.
1354 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1358 // If the client is only demanding bits that we know, return the known
1360 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1361 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1366 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1367 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1368 /// actually used by the caller. This method analyzes which elements of the
1369 /// operand are undef and returns that information in UndefElts.
1371 /// If the information about demanded elements can be used to simplify the
1372 /// operation, the operation is simplified, then the resultant value is
1373 /// returned. This returns null if no change was made.
1374 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1375 uint64_t &UndefElts,
1377 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1378 assert(VWidth <= 64 && "Vector too wide to analyze!");
1379 uint64_t EltMask = ~0ULL >> (64-VWidth);
1380 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1382 if (isa<UndefValue>(V)) {
1383 // If the entire vector is undefined, just return this info.
1384 UndefElts = EltMask;
1386 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1387 UndefElts = EltMask;
1388 return UndefValue::get(V->getType());
1392 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1393 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1394 Constant *Undef = UndefValue::get(EltTy);
1396 std::vector<Constant*> Elts;
1397 for (unsigned i = 0; i != VWidth; ++i)
1398 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1399 Elts.push_back(Undef);
1400 UndefElts |= (1ULL << i);
1401 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1402 Elts.push_back(Undef);
1403 UndefElts |= (1ULL << i);
1404 } else { // Otherwise, defined.
1405 Elts.push_back(CP->getOperand(i));
1408 // If we changed the constant, return it.
1409 Constant *NewCP = ConstantVector::get(Elts);
1410 return NewCP != CP ? NewCP : 0;
1411 } else if (isa<ConstantAggregateZero>(V)) {
1412 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1415 // Check if this is identity. If so, return 0 since we are not simplifying
1417 if (DemandedElts == ((1ULL << VWidth) -1))
1420 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1421 Constant *Zero = Constant::getNullValue(EltTy);
1422 Constant *Undef = UndefValue::get(EltTy);
1423 std::vector<Constant*> Elts;
1424 for (unsigned i = 0; i != VWidth; ++i)
1425 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1426 UndefElts = DemandedElts ^ EltMask;
1427 return ConstantVector::get(Elts);
1430 // Limit search depth.
1434 // If multiple users are using the root value, procede with
1435 // simplification conservatively assuming that all elements
1437 if (!V->hasOneUse()) {
1438 // Quit if we find multiple users of a non-root value though.
1439 // They'll be handled when it's their turn to be visited by
1440 // the main instcombine process.
1442 // TODO: Just compute the UndefElts information recursively.
1445 // Conservatively assume that all elements are needed.
1446 DemandedElts = EltMask;
1449 Instruction *I = dyn_cast<Instruction>(V);
1450 if (!I) return false; // Only analyze instructions.
1452 bool MadeChange = false;
1453 uint64_t UndefElts2;
1455 switch (I->getOpcode()) {
1458 case Instruction::InsertElement: {
1459 // If this is a variable index, we don't know which element it overwrites.
1460 // demand exactly the same input as we produce.
1461 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1463 // Note that we can't propagate undef elt info, because we don't know
1464 // which elt is getting updated.
1465 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1466 UndefElts2, Depth+1);
1467 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1471 // If this is inserting an element that isn't demanded, remove this
1473 unsigned IdxNo = Idx->getZExtValue();
1474 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1475 return AddSoonDeadInstToWorklist(*I, 0);
1477 // Otherwise, the element inserted overwrites whatever was there, so the
1478 // input demanded set is simpler than the output set.
1479 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1480 DemandedElts & ~(1ULL << IdxNo),
1481 UndefElts, Depth+1);
1482 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1484 // The inserted element is defined.
1485 UndefElts &= ~(1ULL << IdxNo);
1488 case Instruction::ShuffleVector: {
1489 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1490 uint64_t LHSVWidth =
1491 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1492 uint64_t LeftDemanded = 0, RightDemanded = 0;
1493 for (unsigned i = 0; i < VWidth; i++) {
1494 if (DemandedElts & (1ULL << i)) {
1495 unsigned MaskVal = Shuffle->getMaskValue(i);
1496 if (MaskVal != -1u) {
1497 assert(MaskVal < LHSVWidth * 2 &&
1498 "shufflevector mask index out of range!");
1499 if (MaskVal < LHSVWidth)
1500 LeftDemanded |= 1ULL << MaskVal;
1502 RightDemanded |= 1ULL << (MaskVal - LHSVWidth);
1507 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1508 UndefElts2, Depth+1);
1509 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1511 uint64_t UndefElts3;
1512 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1513 UndefElts3, Depth+1);
1514 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1516 bool NewUndefElts = false;
1517 for (unsigned i = 0; i < VWidth; i++) {
1518 unsigned MaskVal = Shuffle->getMaskValue(i);
1519 if (MaskVal == -1u) {
1520 uint64_t NewBit = 1ULL << i;
1521 UndefElts |= NewBit;
1522 } else if (MaskVal < LHSVWidth) {
1523 uint64_t NewBit = ((UndefElts2 >> MaskVal) & 1) << i;
1524 NewUndefElts |= NewBit;
1525 UndefElts |= NewBit;
1527 uint64_t NewBit = ((UndefElts3 >> (MaskVal - LHSVWidth)) & 1) << i;
1528 NewUndefElts |= NewBit;
1529 UndefElts |= NewBit;
1534 // Add additional discovered undefs.
1535 std::vector<Constant*> Elts;
1536 for (unsigned i = 0; i < VWidth; ++i) {
1537 if (UndefElts & (1ULL << i))
1538 Elts.push_back(UndefValue::get(Type::Int32Ty));
1540 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1541 Shuffle->getMaskValue(i)));
1543 I->setOperand(2, ConstantVector::get(Elts));
1548 case Instruction::BitCast: {
1549 // Vector->vector casts only.
1550 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1552 unsigned InVWidth = VTy->getNumElements();
1553 uint64_t InputDemandedElts = 0;
1556 if (VWidth == InVWidth) {
1557 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1558 // elements as are demanded of us.
1560 InputDemandedElts = DemandedElts;
1561 } else if (VWidth > InVWidth) {
1565 // If there are more elements in the result than there are in the source,
1566 // then an input element is live if any of the corresponding output
1567 // elements are live.
1568 Ratio = VWidth/InVWidth;
1569 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1570 if (DemandedElts & (1ULL << OutIdx))
1571 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1577 // If there are more elements in the source than there are in the result,
1578 // then an input element is live if the corresponding output element is
1580 Ratio = InVWidth/VWidth;
1581 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1582 if (DemandedElts & (1ULL << InIdx/Ratio))
1583 InputDemandedElts |= 1ULL << InIdx;
1586 // div/rem demand all inputs, because they don't want divide by zero.
1587 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1588 UndefElts2, Depth+1);
1590 I->setOperand(0, TmpV);
1594 UndefElts = UndefElts2;
1595 if (VWidth > InVWidth) {
1596 assert(0 && "Unimp");
1597 // If there are more elements in the result than there are in the source,
1598 // then an output element is undef if the corresponding input element is
1600 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1601 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1602 UndefElts |= 1ULL << OutIdx;
1603 } else if (VWidth < InVWidth) {
1604 assert(0 && "Unimp");
1605 // If there are more elements in the source than there are in the result,
1606 // then a result element is undef if all of the corresponding input
1607 // elements are undef.
1608 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1609 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1610 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1611 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1615 case Instruction::And:
1616 case Instruction::Or:
1617 case Instruction::Xor:
1618 case Instruction::Add:
1619 case Instruction::Sub:
1620 case Instruction::Mul:
1621 // div/rem demand all inputs, because they don't want divide by zero.
1622 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1623 UndefElts, Depth+1);
1624 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1625 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1626 UndefElts2, Depth+1);
1627 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1629 // Output elements are undefined if both are undefined. Consider things
1630 // like undef&0. The result is known zero, not undef.
1631 UndefElts &= UndefElts2;
1634 case Instruction::Call: {
1635 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1637 switch (II->getIntrinsicID()) {
1640 // Binary vector operations that work column-wise. A dest element is a
1641 // function of the corresponding input elements from the two inputs.
1642 case Intrinsic::x86_sse_sub_ss:
1643 case Intrinsic::x86_sse_mul_ss:
1644 case Intrinsic::x86_sse_min_ss:
1645 case Intrinsic::x86_sse_max_ss:
1646 case Intrinsic::x86_sse2_sub_sd:
1647 case Intrinsic::x86_sse2_mul_sd:
1648 case Intrinsic::x86_sse2_min_sd:
1649 case Intrinsic::x86_sse2_max_sd:
1650 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1651 UndefElts, Depth+1);
1652 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1653 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1654 UndefElts2, Depth+1);
1655 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1657 // If only the low elt is demanded and this is a scalarizable intrinsic,
1658 // scalarize it now.
1659 if (DemandedElts == 1) {
1660 switch (II->getIntrinsicID()) {
1662 case Intrinsic::x86_sse_sub_ss:
1663 case Intrinsic::x86_sse_mul_ss:
1664 case Intrinsic::x86_sse2_sub_sd:
1665 case Intrinsic::x86_sse2_mul_sd:
1666 // TODO: Lower MIN/MAX/ABS/etc
1667 Value *LHS = II->getOperand(1);
1668 Value *RHS = II->getOperand(2);
1669 // Extract the element as scalars.
1670 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1671 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1673 switch (II->getIntrinsicID()) {
1674 default: assert(0 && "Case stmts out of sync!");
1675 case Intrinsic::x86_sse_sub_ss:
1676 case Intrinsic::x86_sse2_sub_sd:
1677 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1678 II->getName()), *II);
1680 case Intrinsic::x86_sse_mul_ss:
1681 case Intrinsic::x86_sse2_mul_sd:
1682 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1683 II->getName()), *II);
1688 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1690 InsertNewInstBefore(New, *II);
1691 AddSoonDeadInstToWorklist(*II, 0);
1696 // Output elements are undefined if both are undefined. Consider things
1697 // like undef&0. The result is known zero, not undef.
1698 UndefElts &= UndefElts2;
1704 return MadeChange ? I : 0;
1708 /// AssociativeOpt - Perform an optimization on an associative operator. This
1709 /// function is designed to check a chain of associative operators for a
1710 /// potential to apply a certain optimization. Since the optimization may be
1711 /// applicable if the expression was reassociated, this checks the chain, then
1712 /// reassociates the expression as necessary to expose the optimization
1713 /// opportunity. This makes use of a special Functor, which must define
1714 /// 'shouldApply' and 'apply' methods.
1716 template<typename Functor>
1717 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1718 unsigned Opcode = Root.getOpcode();
1719 Value *LHS = Root.getOperand(0);
1721 // Quick check, see if the immediate LHS matches...
1722 if (F.shouldApply(LHS))
1723 return F.apply(Root);
1725 // Otherwise, if the LHS is not of the same opcode as the root, return.
1726 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1727 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1728 // Should we apply this transform to the RHS?
1729 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1731 // If not to the RHS, check to see if we should apply to the LHS...
1732 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1733 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1737 // If the functor wants to apply the optimization to the RHS of LHSI,
1738 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1740 // Now all of the instructions are in the current basic block, go ahead
1741 // and perform the reassociation.
1742 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1744 // First move the selected RHS to the LHS of the root...
1745 Root.setOperand(0, LHSI->getOperand(1));
1747 // Make what used to be the LHS of the root be the user of the root...
1748 Value *ExtraOperand = TmpLHSI->getOperand(1);
1749 if (&Root == TmpLHSI) {
1750 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1753 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1754 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1755 BasicBlock::iterator ARI = &Root; ++ARI;
1756 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1759 // Now propagate the ExtraOperand down the chain of instructions until we
1761 while (TmpLHSI != LHSI) {
1762 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1763 // Move the instruction to immediately before the chain we are
1764 // constructing to avoid breaking dominance properties.
1765 NextLHSI->moveBefore(ARI);
1768 Value *NextOp = NextLHSI->getOperand(1);
1769 NextLHSI->setOperand(1, ExtraOperand);
1771 ExtraOperand = NextOp;
1774 // Now that the instructions are reassociated, have the functor perform
1775 // the transformation...
1776 return F.apply(Root);
1779 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1786 // AddRHS - Implements: X + X --> X << 1
1789 AddRHS(Value *rhs) : RHS(rhs) {}
1790 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1791 Instruction *apply(BinaryOperator &Add) const {
1792 return BinaryOperator::CreateShl(Add.getOperand(0),
1793 ConstantInt::get(Add.getType(), 1));
1797 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1799 struct AddMaskingAnd {
1801 AddMaskingAnd(Constant *c) : C2(c) {}
1802 bool shouldApply(Value *LHS) const {
1804 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1805 ConstantExpr::getAnd(C1, C2)->isNullValue();
1807 Instruction *apply(BinaryOperator &Add) const {
1808 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1814 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1816 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1817 if (Constant *SOC = dyn_cast<Constant>(SO))
1818 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1820 return IC->InsertNewInstBefore(CastInst::Create(
1821 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1824 // Figure out if the constant is the left or the right argument.
1825 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1826 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1828 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1830 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1831 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1834 Value *Op0 = SO, *Op1 = ConstOperand;
1836 std::swap(Op0, Op1);
1838 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1839 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1840 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1841 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1842 SO->getName()+".cmp");
1844 assert(0 && "Unknown binary instruction type!");
1847 return IC->InsertNewInstBefore(New, I);
1850 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1851 // constant as the other operand, try to fold the binary operator into the
1852 // select arguments. This also works for Cast instructions, which obviously do
1853 // not have a second operand.
1854 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1856 // Don't modify shared select instructions
1857 if (!SI->hasOneUse()) return 0;
1858 Value *TV = SI->getOperand(1);
1859 Value *FV = SI->getOperand(2);
1861 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1862 // Bool selects with constant operands can be folded to logical ops.
1863 if (SI->getType() == Type::Int1Ty) return 0;
1865 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1866 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1868 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1875 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1876 /// node as operand #0, see if we can fold the instruction into the PHI (which
1877 /// is only possible if all operands to the PHI are constants).
1878 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1879 PHINode *PN = cast<PHINode>(I.getOperand(0));
1880 unsigned NumPHIValues = PN->getNumIncomingValues();
1881 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1883 // Check to see if all of the operands of the PHI are constants. If there is
1884 // one non-constant value, remember the BB it is. If there is more than one
1885 // or if *it* is a PHI, bail out.
1886 BasicBlock *NonConstBB = 0;
1887 for (unsigned i = 0; i != NumPHIValues; ++i)
1888 if (!isa<Constant>(PN->getIncomingValue(i))) {
1889 if (NonConstBB) return 0; // More than one non-const value.
1890 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1891 NonConstBB = PN->getIncomingBlock(i);
1893 // If the incoming non-constant value is in I's block, we have an infinite
1895 if (NonConstBB == I.getParent())
1899 // If there is exactly one non-constant value, we can insert a copy of the
1900 // operation in that block. However, if this is a critical edge, we would be
1901 // inserting the computation one some other paths (e.g. inside a loop). Only
1902 // do this if the pred block is unconditionally branching into the phi block.
1904 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1905 if (!BI || !BI->isUnconditional()) return 0;
1908 // Okay, we can do the transformation: create the new PHI node.
1909 PHINode *NewPN = PHINode::Create(I.getType(), "");
1910 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1911 InsertNewInstBefore(NewPN, *PN);
1912 NewPN->takeName(PN);
1914 // Next, add all of the operands to the PHI.
1915 if (I.getNumOperands() == 2) {
1916 Constant *C = cast<Constant>(I.getOperand(1));
1917 for (unsigned i = 0; i != NumPHIValues; ++i) {
1919 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1920 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1921 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1923 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1925 assert(PN->getIncomingBlock(i) == NonConstBB);
1926 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1927 InV = BinaryOperator::Create(BO->getOpcode(),
1928 PN->getIncomingValue(i), C, "phitmp",
1929 NonConstBB->getTerminator());
1930 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1931 InV = CmpInst::Create(CI->getOpcode(),
1933 PN->getIncomingValue(i), C, "phitmp",
1934 NonConstBB->getTerminator());
1936 assert(0 && "Unknown binop!");
1938 AddToWorkList(cast<Instruction>(InV));
1940 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1943 CastInst *CI = cast<CastInst>(&I);
1944 const Type *RetTy = CI->getType();
1945 for (unsigned i = 0; i != NumPHIValues; ++i) {
1947 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1948 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1950 assert(PN->getIncomingBlock(i) == NonConstBB);
1951 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1952 I.getType(), "phitmp",
1953 NonConstBB->getTerminator());
1954 AddToWorkList(cast<Instruction>(InV));
1956 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1959 return ReplaceInstUsesWith(I, NewPN);
1963 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1964 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1965 /// This basically requires proving that the add in the original type would not
1966 /// overflow to change the sign bit or have a carry out.
1967 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1968 // There are different heuristics we can use for this. Here are some simple
1971 // Add has the property that adding any two 2's complement numbers can only
1972 // have one carry bit which can change a sign. As such, if LHS and RHS each
1973 // have at least two sign bits, we know that the addition of the two values will
1974 // sign extend fine.
1975 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1979 // If one of the operands only has one non-zero bit, and if the other operand
1980 // has a known-zero bit in a more significant place than it (not including the
1981 // sign bit) the ripple may go up to and fill the zero, but won't change the
1982 // sign. For example, (X & ~4) + 1.
1990 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1991 bool Changed = SimplifyCommutative(I);
1992 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1994 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1995 // X + undef -> undef
1996 if (isa<UndefValue>(RHS))
1997 return ReplaceInstUsesWith(I, RHS);
2000 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2001 if (RHSC->isNullValue())
2002 return ReplaceInstUsesWith(I, LHS);
2003 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2004 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2005 (I.getType())->getValueAPF()))
2006 return ReplaceInstUsesWith(I, LHS);
2009 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2010 // X + (signbit) --> X ^ signbit
2011 const APInt& Val = CI->getValue();
2012 uint32_t BitWidth = Val.getBitWidth();
2013 if (Val == APInt::getSignBit(BitWidth))
2014 return BinaryOperator::CreateXor(LHS, RHS);
2016 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2017 // (X & 254)+1 -> (X&254)|1
2018 if (!isa<VectorType>(I.getType())) {
2019 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2020 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2021 KnownZero, KnownOne))
2025 // zext(i1) - 1 -> select i1, 0, -1
2026 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2027 if (CI->isAllOnesValue() &&
2028 ZI->getOperand(0)->getType() == Type::Int1Ty)
2029 return SelectInst::Create(ZI->getOperand(0),
2030 Constant::getNullValue(I.getType()),
2031 ConstantInt::getAllOnesValue(I.getType()));
2034 if (isa<PHINode>(LHS))
2035 if (Instruction *NV = FoldOpIntoPhi(I))
2038 ConstantInt *XorRHS = 0;
2040 if (isa<ConstantInt>(RHSC) &&
2041 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2042 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2043 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2045 uint32_t Size = TySizeBits / 2;
2046 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2047 APInt CFF80Val(-C0080Val);
2049 if (TySizeBits > Size) {
2050 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2051 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2052 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2053 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2054 // This is a sign extend if the top bits are known zero.
2055 if (!MaskedValueIsZero(XorLHS,
2056 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2057 Size = 0; // Not a sign ext, but can't be any others either.
2062 C0080Val = APIntOps::lshr(C0080Val, Size);
2063 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2064 } while (Size >= 1);
2066 // FIXME: This shouldn't be necessary. When the backends can handle types
2067 // with funny bit widths then this switch statement should be removed. It
2068 // is just here to get the size of the "middle" type back up to something
2069 // that the back ends can handle.
2070 const Type *MiddleType = 0;
2073 case 32: MiddleType = Type::Int32Ty; break;
2074 case 16: MiddleType = Type::Int16Ty; break;
2075 case 8: MiddleType = Type::Int8Ty; break;
2078 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2079 InsertNewInstBefore(NewTrunc, I);
2080 return new SExtInst(NewTrunc, I.getType(), I.getName());
2085 if (I.getType() == Type::Int1Ty)
2086 return BinaryOperator::CreateXor(LHS, RHS);
2089 if (I.getType()->isInteger()) {
2090 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2092 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2093 if (RHSI->getOpcode() == Instruction::Sub)
2094 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2095 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2097 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2098 if (LHSI->getOpcode() == Instruction::Sub)
2099 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2100 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2105 // -A + -B --> -(A + B)
2106 if (Value *LHSV = dyn_castNegVal(LHS)) {
2107 if (LHS->getType()->isIntOrIntVector()) {
2108 if (Value *RHSV = dyn_castNegVal(RHS)) {
2109 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2110 InsertNewInstBefore(NewAdd, I);
2111 return BinaryOperator::CreateNeg(NewAdd);
2115 return BinaryOperator::CreateSub(RHS, LHSV);
2119 if (!isa<Constant>(RHS))
2120 if (Value *V = dyn_castNegVal(RHS))
2121 return BinaryOperator::CreateSub(LHS, V);
2125 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2126 if (X == RHS) // X*C + X --> X * (C+1)
2127 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2129 // X*C1 + X*C2 --> X * (C1+C2)
2131 if (X == dyn_castFoldableMul(RHS, C1))
2132 return BinaryOperator::CreateMul(X, Add(C1, C2));
2135 // X + X*C --> X * (C+1)
2136 if (dyn_castFoldableMul(RHS, C2) == LHS)
2137 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2139 // X + ~X --> -1 since ~X = -X-1
2140 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2141 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2144 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2145 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2146 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2149 // A+B --> A|B iff A and B have no bits set in common.
2150 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2151 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2152 APInt LHSKnownOne(IT->getBitWidth(), 0);
2153 APInt LHSKnownZero(IT->getBitWidth(), 0);
2154 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2155 if (LHSKnownZero != 0) {
2156 APInt RHSKnownOne(IT->getBitWidth(), 0);
2157 APInt RHSKnownZero(IT->getBitWidth(), 0);
2158 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2160 // No bits in common -> bitwise or.
2161 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2162 return BinaryOperator::CreateOr(LHS, RHS);
2166 // W*X + Y*Z --> W * (X+Z) iff W == Y
2167 if (I.getType()->isIntOrIntVector()) {
2168 Value *W, *X, *Y, *Z;
2169 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2170 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2174 } else if (Y == X) {
2176 } else if (X == Z) {
2183 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2184 LHS->getName()), I);
2185 return BinaryOperator::CreateMul(W, NewAdd);
2190 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2192 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2193 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2195 // (X & FF00) + xx00 -> (X+xx00) & FF00
2196 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2197 Constant *Anded = And(CRHS, C2);
2198 if (Anded == CRHS) {
2199 // See if all bits from the first bit set in the Add RHS up are included
2200 // in the mask. First, get the rightmost bit.
2201 const APInt& AddRHSV = CRHS->getValue();
2203 // Form a mask of all bits from the lowest bit added through the top.
2204 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2206 // See if the and mask includes all of these bits.
2207 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2209 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2210 // Okay, the xform is safe. Insert the new add pronto.
2211 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2212 LHS->getName()), I);
2213 return BinaryOperator::CreateAnd(NewAdd, C2);
2218 // Try to fold constant add into select arguments.
2219 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2220 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2224 // add (cast *A to intptrtype) B ->
2225 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2227 CastInst *CI = dyn_cast<CastInst>(LHS);
2230 CI = dyn_cast<CastInst>(RHS);
2233 if (CI && CI->getType()->isSized() &&
2234 (CI->getType()->getPrimitiveSizeInBits() ==
2235 TD->getIntPtrType()->getPrimitiveSizeInBits())
2236 && isa<PointerType>(CI->getOperand(0)->getType())) {
2238 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2239 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2240 PointerType::get(Type::Int8Ty, AS), I);
2241 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2242 return new PtrToIntInst(I2, CI->getType());
2246 // add (select X 0 (sub n A)) A --> select X A n
2248 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2251 SI = dyn_cast<SelectInst>(RHS);
2254 if (SI && SI->hasOneUse()) {
2255 Value *TV = SI->getTrueValue();
2256 Value *FV = SI->getFalseValue();
2259 // Can we fold the add into the argument of the select?
2260 // We check both true and false select arguments for a matching subtract.
2261 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2262 // Fold the add into the true select value.
2263 return SelectInst::Create(SI->getCondition(), N, A);
2264 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2265 // Fold the add into the false select value.
2266 return SelectInst::Create(SI->getCondition(), A, N);
2270 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2271 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2272 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2273 return ReplaceInstUsesWith(I, LHS);
2275 // Check for (add (sext x), y), see if we can merge this into an
2276 // integer add followed by a sext.
2277 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2278 // (add (sext x), cst) --> (sext (add x, cst'))
2279 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2281 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2282 if (LHSConv->hasOneUse() &&
2283 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2284 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2285 // Insert the new, smaller add.
2286 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2288 InsertNewInstBefore(NewAdd, I);
2289 return new SExtInst(NewAdd, I.getType());
2293 // (add (sext x), (sext y)) --> (sext (add int x, y))
2294 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2295 // Only do this if x/y have the same type, if at last one of them has a
2296 // single use (so we don't increase the number of sexts), and if the
2297 // integer add will not overflow.
2298 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2299 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2300 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2301 RHSConv->getOperand(0))) {
2302 // Insert the new integer add.
2303 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2304 RHSConv->getOperand(0),
2306 InsertNewInstBefore(NewAdd, I);
2307 return new SExtInst(NewAdd, I.getType());
2312 // Check for (add double (sitofp x), y), see if we can merge this into an
2313 // integer add followed by a promotion.
2314 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2315 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2316 // ... if the constant fits in the integer value. This is useful for things
2317 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2318 // requires a constant pool load, and generally allows the add to be better
2320 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2322 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2323 if (LHSConv->hasOneUse() &&
2324 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2325 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2326 // Insert the new integer add.
2327 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2329 InsertNewInstBefore(NewAdd, I);
2330 return new SIToFPInst(NewAdd, I.getType());
2334 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2335 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2336 // Only do this if x/y have the same type, if at last one of them has a
2337 // single use (so we don't increase the number of int->fp conversions),
2338 // and if the integer add will not overflow.
2339 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2340 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2341 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2342 RHSConv->getOperand(0))) {
2343 // Insert the new integer add.
2344 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2345 RHSConv->getOperand(0),
2347 InsertNewInstBefore(NewAdd, I);
2348 return new SIToFPInst(NewAdd, I.getType());
2353 return Changed ? &I : 0;
2356 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2357 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2359 if (Op0 == Op1 && // sub X, X -> 0
2360 !I.getType()->isFPOrFPVector())
2361 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2363 // If this is a 'B = x-(-A)', change to B = x+A...
2364 if (Value *V = dyn_castNegVal(Op1))
2365 return BinaryOperator::CreateAdd(Op0, V);
2367 if (isa<UndefValue>(Op0))
2368 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2369 if (isa<UndefValue>(Op1))
2370 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2372 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2373 // Replace (-1 - A) with (~A)...
2374 if (C->isAllOnesValue())
2375 return BinaryOperator::CreateNot(Op1);
2377 // C - ~X == X + (1+C)
2379 if (match(Op1, m_Not(m_Value(X))))
2380 return BinaryOperator::CreateAdd(X, AddOne(C));
2382 // -(X >>u 31) -> (X >>s 31)
2383 // -(X >>s 31) -> (X >>u 31)
2385 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2386 if (SI->getOpcode() == Instruction::LShr) {
2387 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2388 // Check to see if we are shifting out everything but the sign bit.
2389 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2390 SI->getType()->getPrimitiveSizeInBits()-1) {
2391 // Ok, the transformation is safe. Insert AShr.
2392 return BinaryOperator::Create(Instruction::AShr,
2393 SI->getOperand(0), CU, SI->getName());
2397 else if (SI->getOpcode() == Instruction::AShr) {
2398 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2399 // Check to see if we are shifting out everything but the sign bit.
2400 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2401 SI->getType()->getPrimitiveSizeInBits()-1) {
2402 // Ok, the transformation is safe. Insert LShr.
2403 return BinaryOperator::CreateLShr(
2404 SI->getOperand(0), CU, SI->getName());
2411 // Try to fold constant sub into select arguments.
2412 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2413 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2416 if (isa<PHINode>(Op0))
2417 if (Instruction *NV = FoldOpIntoPhi(I))
2421 if (I.getType() == Type::Int1Ty)
2422 return BinaryOperator::CreateXor(Op0, Op1);
2424 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2425 if (Op1I->getOpcode() == Instruction::Add &&
2426 !Op0->getType()->isFPOrFPVector()) {
2427 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2428 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2429 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2430 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2431 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2432 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2433 // C1-(X+C2) --> (C1-C2)-X
2434 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2435 Op1I->getOperand(0));
2439 if (Op1I->hasOneUse()) {
2440 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2441 // is not used by anyone else...
2443 if (Op1I->getOpcode() == Instruction::Sub &&
2444 !Op1I->getType()->isFPOrFPVector()) {
2445 // Swap the two operands of the subexpr...
2446 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2447 Op1I->setOperand(0, IIOp1);
2448 Op1I->setOperand(1, IIOp0);
2450 // Create the new top level add instruction...
2451 return BinaryOperator::CreateAdd(Op0, Op1);
2454 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2456 if (Op1I->getOpcode() == Instruction::And &&
2457 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2458 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2461 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2462 return BinaryOperator::CreateAnd(Op0, NewNot);
2465 // 0 - (X sdiv C) -> (X sdiv -C)
2466 if (Op1I->getOpcode() == Instruction::SDiv)
2467 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2469 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2470 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2471 ConstantExpr::getNeg(DivRHS));
2473 // X - X*C --> X * (1-C)
2474 ConstantInt *C2 = 0;
2475 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2476 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2477 return BinaryOperator::CreateMul(Op0, CP1);
2480 // X - ((X / Y) * Y) --> X % Y
2481 if (Op1I->getOpcode() == Instruction::Mul)
2482 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2483 if (Op0 == I->getOperand(0) &&
2484 Op1I->getOperand(1) == I->getOperand(1)) {
2485 if (I->getOpcode() == Instruction::SDiv)
2486 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2487 if (I->getOpcode() == Instruction::UDiv)
2488 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2493 if (!Op0->getType()->isFPOrFPVector())
2494 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2495 if (Op0I->getOpcode() == Instruction::Add) {
2496 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2497 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2498 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2499 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2500 } else if (Op0I->getOpcode() == Instruction::Sub) {
2501 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2502 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2507 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2508 if (X == Op1) // X*C - X --> X * (C-1)
2509 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2511 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2512 if (X == dyn_castFoldableMul(Op1, C2))
2513 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2518 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2519 /// comparison only checks the sign bit. If it only checks the sign bit, set
2520 /// TrueIfSigned if the result of the comparison is true when the input value is
2522 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2523 bool &TrueIfSigned) {
2525 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2526 TrueIfSigned = true;
2527 return RHS->isZero();
2528 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2529 TrueIfSigned = true;
2530 return RHS->isAllOnesValue();
2531 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2532 TrueIfSigned = false;
2533 return RHS->isAllOnesValue();
2534 case ICmpInst::ICMP_UGT:
2535 // True if LHS u> RHS and RHS == high-bit-mask - 1
2536 TrueIfSigned = true;
2537 return RHS->getValue() ==
2538 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2539 case ICmpInst::ICMP_UGE:
2540 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2541 TrueIfSigned = true;
2542 return RHS->getValue().isSignBit();
2548 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2549 bool Changed = SimplifyCommutative(I);
2550 Value *Op0 = I.getOperand(0);
2552 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2553 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2555 // Simplify mul instructions with a constant RHS...
2556 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2557 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2559 // ((X << C1)*C2) == (X * (C2 << C1))
2560 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2561 if (SI->getOpcode() == Instruction::Shl)
2562 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2563 return BinaryOperator::CreateMul(SI->getOperand(0),
2564 ConstantExpr::getShl(CI, ShOp));
2567 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2568 if (CI->equalsInt(1)) // X * 1 == X
2569 return ReplaceInstUsesWith(I, Op0);
2570 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2571 return BinaryOperator::CreateNeg(Op0, I.getName());
2573 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2574 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2575 return BinaryOperator::CreateShl(Op0,
2576 ConstantInt::get(Op0->getType(), Val.logBase2()));
2578 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2579 if (Op1F->isNullValue())
2580 return ReplaceInstUsesWith(I, Op1);
2582 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2583 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2584 if (Op1F->isExactlyValue(1.0))
2585 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2586 } else if (isa<VectorType>(Op1->getType())) {
2587 if (isa<ConstantAggregateZero>(Op1))
2588 return ReplaceInstUsesWith(I, Op1);
2590 // As above, vector X*splat(1.0) -> X in all defined cases.
2591 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1))
2592 if (ConstantFP *F = dyn_cast_or_null<ConstantFP>(Op1V->getSplatValue()))
2593 if (F->isExactlyValue(1.0))
2594 return ReplaceInstUsesWith(I, Op0);
2597 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2598 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2599 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2600 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2601 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2603 InsertNewInstBefore(Add, I);
2604 Value *C1C2 = ConstantExpr::getMul(Op1,
2605 cast<Constant>(Op0I->getOperand(1)));
2606 return BinaryOperator::CreateAdd(Add, C1C2);
2610 // Try to fold constant mul into select arguments.
2611 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2612 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2615 if (isa<PHINode>(Op0))
2616 if (Instruction *NV = FoldOpIntoPhi(I))
2620 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2621 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2622 return BinaryOperator::CreateMul(Op0v, Op1v);
2624 if (I.getType() == Type::Int1Ty)
2625 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2627 // If one of the operands of the multiply is a cast from a boolean value, then
2628 // we know the bool is either zero or one, so this is a 'masking' multiply.
2629 // See if we can simplify things based on how the boolean was originally
2631 CastInst *BoolCast = 0;
2632 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2633 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2636 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2637 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2640 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2641 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2642 const Type *SCOpTy = SCIOp0->getType();
2645 // If the icmp is true iff the sign bit of X is set, then convert this
2646 // multiply into a shift/and combination.
2647 if (isa<ConstantInt>(SCIOp1) &&
2648 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2650 // Shift the X value right to turn it into "all signbits".
2651 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2652 SCOpTy->getPrimitiveSizeInBits()-1);
2654 InsertNewInstBefore(
2655 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2656 BoolCast->getOperand(0)->getName()+
2659 // If the multiply type is not the same as the source type, sign extend
2660 // or truncate to the multiply type.
2661 if (I.getType() != V->getType()) {
2662 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2663 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2664 Instruction::CastOps opcode =
2665 (SrcBits == DstBits ? Instruction::BitCast :
2666 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2667 V = InsertCastBefore(opcode, V, I.getType(), I);
2670 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2671 return BinaryOperator::CreateAnd(V, OtherOp);
2676 return Changed ? &I : 0;
2679 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2681 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2682 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2684 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2685 int NonNullOperand = -1;
2686 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2687 if (ST->isNullValue())
2689 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2690 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2691 if (ST->isNullValue())
2694 if (NonNullOperand == -1)
2697 Value *SelectCond = SI->getOperand(0);
2699 // Change the div/rem to use 'Y' instead of the select.
2700 I.setOperand(1, SI->getOperand(NonNullOperand));
2702 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2703 // problem. However, the select, or the condition of the select may have
2704 // multiple uses. Based on our knowledge that the operand must be non-zero,
2705 // propagate the known value for the select into other uses of it, and
2706 // propagate a known value of the condition into its other users.
2708 // If the select and condition only have a single use, don't bother with this,
2710 if (SI->use_empty() && SelectCond->hasOneUse())
2713 // Scan the current block backward, looking for other uses of SI.
2714 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2716 while (BBI != BBFront) {
2718 // If we found a call to a function, we can't assume it will return, so
2719 // information from below it cannot be propagated above it.
2720 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2723 // Replace uses of the select or its condition with the known values.
2724 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2727 *I = SI->getOperand(NonNullOperand);
2729 } else if (*I == SelectCond) {
2730 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2731 ConstantInt::getFalse();
2736 // If we past the instruction, quit looking for it.
2739 if (&*BBI == SelectCond)
2742 // If we ran out of things to eliminate, break out of the loop.
2743 if (SelectCond == 0 && SI == 0)
2751 /// This function implements the transforms on div instructions that work
2752 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2753 /// used by the visitors to those instructions.
2754 /// @brief Transforms common to all three div instructions
2755 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2756 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2758 // undef / X -> 0 for integer.
2759 // undef / X -> undef for FP (the undef could be a snan).
2760 if (isa<UndefValue>(Op0)) {
2761 if (Op0->getType()->isFPOrFPVector())
2762 return ReplaceInstUsesWith(I, Op0);
2763 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2766 // X / undef -> undef
2767 if (isa<UndefValue>(Op1))
2768 return ReplaceInstUsesWith(I, Op1);
2773 /// This function implements the transforms common to both integer division
2774 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2775 /// division instructions.
2776 /// @brief Common integer divide transforms
2777 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2778 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2780 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2782 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2783 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2784 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2785 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2788 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2789 return ReplaceInstUsesWith(I, CI);
2792 if (Instruction *Common = commonDivTransforms(I))
2795 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2796 // This does not apply for fdiv.
2797 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2800 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2802 if (RHS->equalsInt(1))
2803 return ReplaceInstUsesWith(I, Op0);
2805 // (X / C1) / C2 -> X / (C1*C2)
2806 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2807 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2808 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2809 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2810 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2812 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2813 Multiply(RHS, LHSRHS));
2816 if (!RHS->isZero()) { // avoid X udiv 0
2817 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2818 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2820 if (isa<PHINode>(Op0))
2821 if (Instruction *NV = FoldOpIntoPhi(I))
2826 // 0 / X == 0, we don't need to preserve faults!
2827 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2828 if (LHS->equalsInt(0))
2829 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2831 // It can't be division by zero, hence it must be division by one.
2832 if (I.getType() == Type::Int1Ty)
2833 return ReplaceInstUsesWith(I, Op0);
2838 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2839 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2841 // Handle the integer div common cases
2842 if (Instruction *Common = commonIDivTransforms(I))
2845 // X udiv C^2 -> X >> C
2846 // Check to see if this is an unsigned division with an exact power of 2,
2847 // if so, convert to a right shift.
2848 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2849 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2850 return BinaryOperator::CreateLShr(Op0,
2851 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2854 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2855 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2856 if (RHSI->getOpcode() == Instruction::Shl &&
2857 isa<ConstantInt>(RHSI->getOperand(0))) {
2858 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2859 if (C1.isPowerOf2()) {
2860 Value *N = RHSI->getOperand(1);
2861 const Type *NTy = N->getType();
2862 if (uint32_t C2 = C1.logBase2()) {
2863 Constant *C2V = ConstantInt::get(NTy, C2);
2864 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2866 return BinaryOperator::CreateLShr(Op0, N);
2871 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2872 // where C1&C2 are powers of two.
2873 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2874 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2875 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2876 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2877 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2878 // Compute the shift amounts
2879 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2880 // Construct the "on true" case of the select
2881 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2882 Instruction *TSI = BinaryOperator::CreateLShr(
2883 Op0, TC, SI->getName()+".t");
2884 TSI = InsertNewInstBefore(TSI, I);
2886 // Construct the "on false" case of the select
2887 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2888 Instruction *FSI = BinaryOperator::CreateLShr(
2889 Op0, FC, SI->getName()+".f");
2890 FSI = InsertNewInstBefore(FSI, I);
2892 // construct the select instruction and return it.
2893 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2899 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2900 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2902 // Handle the integer div common cases
2903 if (Instruction *Common = commonIDivTransforms(I))
2906 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2908 if (RHS->isAllOnesValue())
2909 return BinaryOperator::CreateNeg(Op0);
2912 if (Value *LHSNeg = dyn_castNegVal(Op0))
2913 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2916 // If the sign bits of both operands are zero (i.e. we can prove they are
2917 // unsigned inputs), turn this into a udiv.
2918 if (I.getType()->isInteger()) {
2919 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2920 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2921 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2922 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2929 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2930 return commonDivTransforms(I);
2933 /// This function implements the transforms on rem instructions that work
2934 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2935 /// is used by the visitors to those instructions.
2936 /// @brief Transforms common to all three rem instructions
2937 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2938 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2940 // 0 % X == 0 for integer, we don't need to preserve faults!
2941 if (Constant *LHS = dyn_cast<Constant>(Op0))
2942 if (LHS->isNullValue())
2943 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2945 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2946 if (I.getType()->isFPOrFPVector())
2947 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2948 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2950 if (isa<UndefValue>(Op1))
2951 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2953 // Handle cases involving: rem X, (select Cond, Y, Z)
2954 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2960 /// This function implements the transforms common to both integer remainder
2961 /// instructions (urem and srem). It is called by the visitors to those integer
2962 /// remainder instructions.
2963 /// @brief Common integer remainder transforms
2964 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2965 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2967 if (Instruction *common = commonRemTransforms(I))
2970 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2971 // X % 0 == undef, we don't need to preserve faults!
2972 if (RHS->equalsInt(0))
2973 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2975 if (RHS->equalsInt(1)) // X % 1 == 0
2976 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2978 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2979 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2980 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2982 } else if (isa<PHINode>(Op0I)) {
2983 if (Instruction *NV = FoldOpIntoPhi(I))
2987 // See if we can fold away this rem instruction.
2988 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2989 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2990 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2991 KnownZero, KnownOne))
2999 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3000 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3002 if (Instruction *common = commonIRemTransforms(I))
3005 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3006 // X urem C^2 -> X and C
3007 // Check to see if this is an unsigned remainder with an exact power of 2,
3008 // if so, convert to a bitwise and.
3009 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3010 if (C->getValue().isPowerOf2())
3011 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3014 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3015 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3016 if (RHSI->getOpcode() == Instruction::Shl &&
3017 isa<ConstantInt>(RHSI->getOperand(0))) {
3018 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3019 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3020 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3022 return BinaryOperator::CreateAnd(Op0, Add);
3027 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3028 // where C1&C2 are powers of two.
3029 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3030 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3031 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3032 // STO == 0 and SFO == 0 handled above.
3033 if ((STO->getValue().isPowerOf2()) &&
3034 (SFO->getValue().isPowerOf2())) {
3035 Value *TrueAnd = InsertNewInstBefore(
3036 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3037 Value *FalseAnd = InsertNewInstBefore(
3038 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3039 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3047 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3048 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3050 // Handle the integer rem common cases
3051 if (Instruction *common = commonIRemTransforms(I))
3054 if (Value *RHSNeg = dyn_castNegVal(Op1))
3055 if (!isa<Constant>(RHSNeg) ||
3056 (isa<ConstantInt>(RHSNeg) &&
3057 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3059 AddUsesToWorkList(I);
3060 I.setOperand(1, RHSNeg);
3064 // If the sign bits of both operands are zero (i.e. we can prove they are
3065 // unsigned inputs), turn this into a urem.
3066 if (I.getType()->isInteger()) {
3067 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3068 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3069 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3070 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3077 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3078 return commonRemTransforms(I);
3081 // isOneBitSet - Return true if there is exactly one bit set in the specified
3083 static bool isOneBitSet(const ConstantInt *CI) {
3084 return CI->getValue().isPowerOf2();
3087 // isHighOnes - Return true if the constant is of the form 1+0+.
3088 // This is the same as lowones(~X).
3089 static bool isHighOnes(const ConstantInt *CI) {
3090 return (~CI->getValue() + 1).isPowerOf2();
3093 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3094 /// are carefully arranged to allow folding of expressions such as:
3096 /// (A < B) | (A > B) --> (A != B)
3098 /// Note that this is only valid if the first and second predicates have the
3099 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3101 /// Three bits are used to represent the condition, as follows:
3106 /// <=> Value Definition
3107 /// 000 0 Always false
3114 /// 111 7 Always true
3116 static unsigned getICmpCode(const ICmpInst *ICI) {
3117 switch (ICI->getPredicate()) {
3119 case ICmpInst::ICMP_UGT: return 1; // 001
3120 case ICmpInst::ICMP_SGT: return 1; // 001
3121 case ICmpInst::ICMP_EQ: return 2; // 010
3122 case ICmpInst::ICMP_UGE: return 3; // 011
3123 case ICmpInst::ICMP_SGE: return 3; // 011
3124 case ICmpInst::ICMP_ULT: return 4; // 100
3125 case ICmpInst::ICMP_SLT: return 4; // 100
3126 case ICmpInst::ICMP_NE: return 5; // 101
3127 case ICmpInst::ICMP_ULE: return 6; // 110
3128 case ICmpInst::ICMP_SLE: return 6; // 110
3131 assert(0 && "Invalid ICmp predicate!");
3136 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3137 /// predicate into a three bit mask. It also returns whether it is an ordered
3138 /// predicate by reference.
3139 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3142 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3143 case FCmpInst::FCMP_UNO: return 0; // 000
3144 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3145 case FCmpInst::FCMP_UGT: return 1; // 001
3146 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3147 case FCmpInst::FCMP_UEQ: return 2; // 010
3148 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3149 case FCmpInst::FCMP_UGE: return 3; // 011
3150 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3151 case FCmpInst::FCMP_ULT: return 4; // 100
3152 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3153 case FCmpInst::FCMP_UNE: return 5; // 101
3154 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3155 case FCmpInst::FCMP_ULE: return 6; // 110
3158 // Not expecting FCMP_FALSE and FCMP_TRUE;
3159 assert(0 && "Unexpected FCmp predicate!");
3164 /// getICmpValue - This is the complement of getICmpCode, which turns an
3165 /// opcode and two operands into either a constant true or false, or a brand
3166 /// new ICmp instruction. The sign is passed in to determine which kind
3167 /// of predicate to use in the new icmp instruction.
3168 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3170 default: assert(0 && "Illegal ICmp code!");
3171 case 0: return ConstantInt::getFalse();
3174 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3176 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3177 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3180 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3182 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3185 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3187 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3188 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3191 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3193 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3194 case 7: return ConstantInt::getTrue();
3198 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3199 /// opcode and two operands into either a FCmp instruction. isordered is passed
3200 /// in to determine which kind of predicate to use in the new fcmp instruction.
3201 static Value *getFCmpValue(bool isordered, unsigned code,
3202 Value *LHS, Value *RHS) {
3204 default: assert(0 && "Illegal FCmp code!");
3207 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3209 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3212 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3214 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3217 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3219 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3222 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3224 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3227 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3229 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3232 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3234 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3237 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3239 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3240 case 7: return ConstantInt::getTrue();
3244 /// PredicatesFoldable - Return true if both predicates match sign or if at
3245 /// least one of them is an equality comparison (which is signless).
3246 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3247 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3248 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3249 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3253 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3254 struct FoldICmpLogical {
3257 ICmpInst::Predicate pred;
3258 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3259 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3260 pred(ICI->getPredicate()) {}
3261 bool shouldApply(Value *V) const {
3262 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3263 if (PredicatesFoldable(pred, ICI->getPredicate()))
3264 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3265 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3268 Instruction *apply(Instruction &Log) const {
3269 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3270 if (ICI->getOperand(0) != LHS) {
3271 assert(ICI->getOperand(1) == LHS);
3272 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3275 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3276 unsigned LHSCode = getICmpCode(ICI);
3277 unsigned RHSCode = getICmpCode(RHSICI);
3279 switch (Log.getOpcode()) {
3280 case Instruction::And: Code = LHSCode & RHSCode; break;
3281 case Instruction::Or: Code = LHSCode | RHSCode; break;
3282 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3283 default: assert(0 && "Illegal logical opcode!"); return 0;
3286 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3287 ICmpInst::isSignedPredicate(ICI->getPredicate());
3289 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3290 if (Instruction *I = dyn_cast<Instruction>(RV))
3292 // Otherwise, it's a constant boolean value...
3293 return IC.ReplaceInstUsesWith(Log, RV);
3296 } // end anonymous namespace
3298 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3299 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3300 // guaranteed to be a binary operator.
3301 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3303 ConstantInt *AndRHS,
3304 BinaryOperator &TheAnd) {
3305 Value *X = Op->getOperand(0);
3306 Constant *Together = 0;
3308 Together = And(AndRHS, OpRHS);
3310 switch (Op->getOpcode()) {
3311 case Instruction::Xor:
3312 if (Op->hasOneUse()) {
3313 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3314 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3315 InsertNewInstBefore(And, TheAnd);
3317 return BinaryOperator::CreateXor(And, Together);
3320 case Instruction::Or:
3321 if (Together == AndRHS) // (X | C) & C --> C
3322 return ReplaceInstUsesWith(TheAnd, AndRHS);
3324 if (Op->hasOneUse() && Together != OpRHS) {
3325 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3326 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3327 InsertNewInstBefore(Or, TheAnd);
3329 return BinaryOperator::CreateAnd(Or, AndRHS);
3332 case Instruction::Add:
3333 if (Op->hasOneUse()) {
3334 // Adding a one to a single bit bit-field should be turned into an XOR
3335 // of the bit. First thing to check is to see if this AND is with a
3336 // single bit constant.
3337 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3339 // If there is only one bit set...
3340 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3341 // Ok, at this point, we know that we are masking the result of the
3342 // ADD down to exactly one bit. If the constant we are adding has
3343 // no bits set below this bit, then we can eliminate the ADD.
3344 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3346 // Check to see if any bits below the one bit set in AndRHSV are set.
3347 if ((AddRHS & (AndRHSV-1)) == 0) {
3348 // If not, the only thing that can effect the output of the AND is
3349 // the bit specified by AndRHSV. If that bit is set, the effect of
3350 // the XOR is to toggle the bit. If it is clear, then the ADD has
3352 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3353 TheAnd.setOperand(0, X);
3356 // Pull the XOR out of the AND.
3357 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3358 InsertNewInstBefore(NewAnd, TheAnd);
3359 NewAnd->takeName(Op);
3360 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3367 case Instruction::Shl: {
3368 // We know that the AND will not produce any of the bits shifted in, so if
3369 // the anded constant includes them, clear them now!
3371 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3372 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3373 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3374 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3376 if (CI->getValue() == ShlMask) {
3377 // Masking out bits that the shift already masks
3378 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3379 } else if (CI != AndRHS) { // Reducing bits set in and.
3380 TheAnd.setOperand(1, CI);
3385 case Instruction::LShr:
3387 // We know that the AND will not produce any of the bits shifted in, so if
3388 // the anded constant includes them, clear them now! This only applies to
3389 // unsigned shifts, because a signed shr may bring in set bits!
3391 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3392 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3393 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3394 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3396 if (CI->getValue() == ShrMask) {
3397 // Masking out bits that the shift already masks.
3398 return ReplaceInstUsesWith(TheAnd, Op);
3399 } else if (CI != AndRHS) {
3400 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3405 case Instruction::AShr:
3407 // See if this is shifting in some sign extension, then masking it out
3409 if (Op->hasOneUse()) {
3410 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3411 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3412 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3413 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3414 if (C == AndRHS) { // Masking out bits shifted in.
3415 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3416 // Make the argument unsigned.
3417 Value *ShVal = Op->getOperand(0);
3418 ShVal = InsertNewInstBefore(
3419 BinaryOperator::CreateLShr(ShVal, OpRHS,
3420 Op->getName()), TheAnd);
3421 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3430 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3431 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3432 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3433 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3434 /// insert new instructions.
3435 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3436 bool isSigned, bool Inside,
3438 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3439 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3440 "Lo is not <= Hi in range emission code!");
3443 if (Lo == Hi) // Trivially false.
3444 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3446 // V >= Min && V < Hi --> V < Hi
3447 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3448 ICmpInst::Predicate pred = (isSigned ?
3449 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3450 return new ICmpInst(pred, V, Hi);
3453 // Emit V-Lo <u Hi-Lo
3454 Constant *NegLo = ConstantExpr::getNeg(Lo);
3455 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3456 InsertNewInstBefore(Add, IB);
3457 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3458 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3461 if (Lo == Hi) // Trivially true.
3462 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3464 // V < Min || V >= Hi -> V > Hi-1
3465 Hi = SubOne(cast<ConstantInt>(Hi));
3466 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3467 ICmpInst::Predicate pred = (isSigned ?
3468 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3469 return new ICmpInst(pred, V, Hi);
3472 // Emit V-Lo >u Hi-1-Lo
3473 // Note that Hi has already had one subtracted from it, above.
3474 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3475 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3476 InsertNewInstBefore(Add, IB);
3477 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3478 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3481 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3482 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3483 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3484 // not, since all 1s are not contiguous.
3485 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3486 const APInt& V = Val->getValue();
3487 uint32_t BitWidth = Val->getType()->getBitWidth();
3488 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3490 // look for the first zero bit after the run of ones
3491 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3492 // look for the first non-zero bit
3493 ME = V.getActiveBits();
3497 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3498 /// where isSub determines whether the operator is a sub. If we can fold one of
3499 /// the following xforms:
3501 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3502 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3503 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3505 /// return (A +/- B).
3507 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3508 ConstantInt *Mask, bool isSub,
3510 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3511 if (!LHSI || LHSI->getNumOperands() != 2 ||
3512 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3514 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3516 switch (LHSI->getOpcode()) {
3518 case Instruction::And:
3519 if (And(N, Mask) == Mask) {
3520 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3521 if ((Mask->getValue().countLeadingZeros() +
3522 Mask->getValue().countPopulation()) ==
3523 Mask->getValue().getBitWidth())
3526 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3527 // part, we don't need any explicit masks to take them out of A. If that
3528 // is all N is, ignore it.
3529 uint32_t MB = 0, ME = 0;
3530 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3531 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3532 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3533 if (MaskedValueIsZero(RHS, Mask))
3538 case Instruction::Or:
3539 case Instruction::Xor:
3540 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3541 if ((Mask->getValue().countLeadingZeros() +
3542 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3543 && And(N, Mask)->isZero())
3550 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3552 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3553 return InsertNewInstBefore(New, I);
3556 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3557 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3558 ICmpInst *LHS, ICmpInst *RHS) {
3560 ConstantInt *LHSCst, *RHSCst;
3561 ICmpInst::Predicate LHSCC, RHSCC;
3563 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3564 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3565 !match(RHS, m_ICmp(RHSCC, m_Specific(Val), m_ConstantInt(RHSCst))))
3568 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3569 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3570 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3571 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3572 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3575 // We can't fold (ugt x, C) & (sgt x, C2).
3576 if (!PredicatesFoldable(LHSCC, RHSCC))
3579 // Ensure that the larger constant is on the RHS.
3580 ICmpInst::Predicate GT;
3581 if (ICmpInst::isSignedPredicate(LHSCC) ||
3582 (ICmpInst::isEquality(LHSCC) &&
3583 ICmpInst::isSignedPredicate(RHSCC)))
3584 GT = ICmpInst::ICMP_SGT;
3586 GT = ICmpInst::ICMP_UGT;
3588 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3589 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3590 std::swap(LHS, RHS);
3591 std::swap(LHSCst, RHSCst);
3592 std::swap(LHSCC, RHSCC);
3595 // At this point, we know we have have two icmp instructions
3596 // comparing a value against two constants and and'ing the result
3597 // together. Because of the above check, we know that we only have
3598 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3599 // (from the FoldICmpLogical check above), that the two constants
3600 // are not equal and that the larger constant is on the RHS
3601 assert(LHSCst != RHSCst && "Compares not folded above?");
3604 default: assert(0 && "Unknown integer condition code!");
3605 case ICmpInst::ICMP_EQ:
3607 default: assert(0 && "Unknown integer condition code!");
3608 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3609 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3610 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3611 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3612 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3613 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3614 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3615 return ReplaceInstUsesWith(I, LHS);
3617 case ICmpInst::ICMP_NE:
3619 default: assert(0 && "Unknown integer condition code!");
3620 case ICmpInst::ICMP_ULT:
3621 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3622 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3623 break; // (X != 13 & X u< 15) -> no change
3624 case ICmpInst::ICMP_SLT:
3625 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3626 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3627 break; // (X != 13 & X s< 15) -> no change
3628 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3629 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3630 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3631 return ReplaceInstUsesWith(I, RHS);
3632 case ICmpInst::ICMP_NE:
3633 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3634 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3635 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3636 Val->getName()+".off");
3637 InsertNewInstBefore(Add, I);
3638 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3639 ConstantInt::get(Add->getType(), 1));
3641 break; // (X != 13 & X != 15) -> no change
3644 case ICmpInst::ICMP_ULT:
3646 default: assert(0 && "Unknown integer condition code!");
3647 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3648 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3649 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3650 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3652 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3653 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3654 return ReplaceInstUsesWith(I, LHS);
3655 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3659 case ICmpInst::ICMP_SLT:
3661 default: assert(0 && "Unknown integer condition code!");
3662 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3663 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3664 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3665 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3667 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3668 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3669 return ReplaceInstUsesWith(I, LHS);
3670 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3674 case ICmpInst::ICMP_UGT:
3676 default: assert(0 && "Unknown integer condition code!");
3677 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3678 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3679 return ReplaceInstUsesWith(I, RHS);
3680 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3682 case ICmpInst::ICMP_NE:
3683 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3684 return new ICmpInst(LHSCC, Val, RHSCst);
3685 break; // (X u> 13 & X != 15) -> no change
3686 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3687 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3688 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3692 case ICmpInst::ICMP_SGT:
3694 default: assert(0 && "Unknown integer condition code!");
3695 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3696 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3697 return ReplaceInstUsesWith(I, RHS);
3698 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3700 case ICmpInst::ICMP_NE:
3701 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3702 return new ICmpInst(LHSCC, Val, RHSCst);
3703 break; // (X s> 13 & X != 15) -> no change
3704 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3705 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3706 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3718 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3719 bool Changed = SimplifyCommutative(I);
3720 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3722 if (isa<UndefValue>(Op1)) // X & undef -> 0
3723 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3727 return ReplaceInstUsesWith(I, Op1);
3729 // See if we can simplify any instructions used by the instruction whose sole
3730 // purpose is to compute bits we don't care about.
3731 if (!isa<VectorType>(I.getType())) {
3732 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3733 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3734 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3735 KnownZero, KnownOne))
3738 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3739 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3740 return ReplaceInstUsesWith(I, I.getOperand(0));
3741 } else if (isa<ConstantAggregateZero>(Op1)) {
3742 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3746 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3747 const APInt& AndRHSMask = AndRHS->getValue();
3748 APInt NotAndRHS(~AndRHSMask);
3750 // Optimize a variety of ((val OP C1) & C2) combinations...
3751 if (isa<BinaryOperator>(Op0)) {
3752 Instruction *Op0I = cast<Instruction>(Op0);
3753 Value *Op0LHS = Op0I->getOperand(0);
3754 Value *Op0RHS = Op0I->getOperand(1);
3755 switch (Op0I->getOpcode()) {
3756 case Instruction::Xor:
3757 case Instruction::Or:
3758 // If the mask is only needed on one incoming arm, push it up.
3759 if (Op0I->hasOneUse()) {
3760 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3761 // Not masking anything out for the LHS, move to RHS.
3762 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3763 Op0RHS->getName()+".masked");
3764 InsertNewInstBefore(NewRHS, I);
3765 return BinaryOperator::Create(
3766 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3768 if (!isa<Constant>(Op0RHS) &&
3769 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3770 // Not masking anything out for the RHS, move to LHS.
3771 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3772 Op0LHS->getName()+".masked");
3773 InsertNewInstBefore(NewLHS, I);
3774 return BinaryOperator::Create(
3775 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3780 case Instruction::Add:
3781 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3782 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3783 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3784 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3785 return BinaryOperator::CreateAnd(V, AndRHS);
3786 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3787 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3790 case Instruction::Sub:
3791 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3792 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3793 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3794 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3795 return BinaryOperator::CreateAnd(V, AndRHS);
3797 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3798 // has 1's for all bits that the subtraction with A might affect.
3799 if (Op0I->hasOneUse()) {
3800 uint32_t BitWidth = AndRHSMask.getBitWidth();
3801 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3802 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3804 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3805 if (!(A && A->isZero()) && // avoid infinite recursion.
3806 MaskedValueIsZero(Op0LHS, Mask)) {
3807 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3808 InsertNewInstBefore(NewNeg, I);
3809 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3814 case Instruction::Shl:
3815 case Instruction::LShr:
3816 // (1 << x) & 1 --> zext(x == 0)
3817 // (1 >> x) & 1 --> zext(x == 0)
3818 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3819 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3820 Constant::getNullValue(I.getType()));
3821 InsertNewInstBefore(NewICmp, I);
3822 return new ZExtInst(NewICmp, I.getType());
3827 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3828 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3830 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3831 // If this is an integer truncation or change from signed-to-unsigned, and
3832 // if the source is an and/or with immediate, transform it. This
3833 // frequently occurs for bitfield accesses.
3834 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3835 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3836 CastOp->getNumOperands() == 2)
3837 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3838 if (CastOp->getOpcode() == Instruction::And) {
3839 // Change: and (cast (and X, C1) to T), C2
3840 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3841 // This will fold the two constants together, which may allow
3842 // other simplifications.
3843 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3844 CastOp->getOperand(0), I.getType(),
3845 CastOp->getName()+".shrunk");
3846 NewCast = InsertNewInstBefore(NewCast, I);
3847 // trunc_or_bitcast(C1)&C2
3848 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3849 C3 = ConstantExpr::getAnd(C3, AndRHS);
3850 return BinaryOperator::CreateAnd(NewCast, C3);
3851 } else if (CastOp->getOpcode() == Instruction::Or) {
3852 // Change: and (cast (or X, C1) to T), C2
3853 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3854 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3855 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3856 return ReplaceInstUsesWith(I, AndRHS);
3862 // Try to fold constant and into select arguments.
3863 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3864 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3866 if (isa<PHINode>(Op0))
3867 if (Instruction *NV = FoldOpIntoPhi(I))
3871 Value *Op0NotVal = dyn_castNotVal(Op0);
3872 Value *Op1NotVal = dyn_castNotVal(Op1);
3874 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3875 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3877 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3878 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3879 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3880 I.getName()+".demorgan");
3881 InsertNewInstBefore(Or, I);
3882 return BinaryOperator::CreateNot(Or);
3886 Value *A = 0, *B = 0, *C = 0, *D = 0;
3887 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3888 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3889 return ReplaceInstUsesWith(I, Op1);
3891 // (A|B) & ~(A&B) -> A^B
3892 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3893 if ((A == C && B == D) || (A == D && B == C))
3894 return BinaryOperator::CreateXor(A, B);
3898 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3899 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3900 return ReplaceInstUsesWith(I, Op0);
3902 // ~(A&B) & (A|B) -> A^B
3903 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3904 if ((A == C && B == D) || (A == D && B == C))
3905 return BinaryOperator::CreateXor(A, B);
3909 if (Op0->hasOneUse() &&
3910 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3911 if (A == Op1) { // (A^B)&A -> A&(A^B)
3912 I.swapOperands(); // Simplify below
3913 std::swap(Op0, Op1);
3914 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3915 cast<BinaryOperator>(Op0)->swapOperands();
3916 I.swapOperands(); // Simplify below
3917 std::swap(Op0, Op1);
3920 if (Op1->hasOneUse() &&
3921 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3922 if (B == Op0) { // B&(A^B) -> B&(B^A)
3923 cast<BinaryOperator>(Op1)->swapOperands();
3926 if (A == Op0) { // A&(A^B) -> A & ~B
3927 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3928 InsertNewInstBefore(NotB, I);
3929 return BinaryOperator::CreateAnd(A, NotB);
3934 { // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3935 // where C is a power of 2
3937 ConstantInt *C1, *C2;
3938 ICmpInst::Predicate LHSCC = ICmpInst::BAD_ICMP_PREDICATE;
3939 ICmpInst::Predicate RHSCC = ICmpInst::BAD_ICMP_PREDICATE;
3940 if (match(&I, m_And(m_ICmp(LHSCC, m_Value(A), m_ConstantInt(C1)),
3941 m_ICmp(RHSCC, m_Value(B), m_ConstantInt(C2)))))
3942 if (C1 == C2 && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3943 C1->getValue().isPowerOf2()) {
3944 Instruction *NewOr = BinaryOperator::CreateOr(A, B);
3945 InsertNewInstBefore(NewOr, I);
3946 return new ICmpInst(LHSCC, NewOr, C1);
3950 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3951 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3952 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3955 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
3956 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
3960 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3961 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3962 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3963 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3964 const Type *SrcTy = Op0C->getOperand(0)->getType();
3965 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3966 // Only do this if the casts both really cause code to be generated.
3967 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3969 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3971 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3972 Op1C->getOperand(0),
3974 InsertNewInstBefore(NewOp, I);
3975 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3979 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3980 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3981 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3982 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3983 SI0->getOperand(1) == SI1->getOperand(1) &&
3984 (SI0->hasOneUse() || SI1->hasOneUse())) {
3985 Instruction *NewOp =
3986 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3988 SI0->getName()), I);
3989 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3990 SI1->getOperand(1));
3994 // If and'ing two fcmp, try combine them into one.
3995 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3996 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3997 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3998 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3999 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4000 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4001 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4002 // If either of the constants are nans, then the whole thing returns
4004 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4005 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4006 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4007 RHS->getOperand(0));
4010 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4011 FCmpInst::Predicate Op0CC, Op1CC;
4012 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4013 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4014 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4015 // Swap RHS operands to match LHS.
4016 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4017 std::swap(Op1LHS, Op1RHS);
4019 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4020 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4022 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4023 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4024 Op1CC == FCmpInst::FCMP_FALSE)
4025 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4026 else if (Op0CC == FCmpInst::FCMP_TRUE)
4027 return ReplaceInstUsesWith(I, Op1);
4028 else if (Op1CC == FCmpInst::FCMP_TRUE)
4029 return ReplaceInstUsesWith(I, Op0);
4032 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4033 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4035 std::swap(Op0, Op1);
4036 std::swap(Op0Pred, Op1Pred);
4037 std::swap(Op0Ordered, Op1Ordered);
4040 // uno && ueq -> uno && (uno || eq) -> ueq
4041 // ord && olt -> ord && (ord && lt) -> olt
4042 if (Op0Ordered == Op1Ordered)
4043 return ReplaceInstUsesWith(I, Op1);
4044 // uno && oeq -> uno && (ord && eq) -> false
4045 // uno && ord -> false
4047 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4048 // ord && ueq -> ord && (uno || eq) -> oeq
4049 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4058 return Changed ? &I : 0;
4061 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4062 /// capable of providing pieces of a bswap. The subexpression provides pieces
4063 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4064 /// the expression came from the corresponding "byte swapped" byte in some other
4065 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4066 /// we know that the expression deposits the low byte of %X into the high byte
4067 /// of the bswap result and that all other bytes are zero. This expression is
4068 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4071 /// This function returns true if the match was unsuccessful and false if so.
4072 /// On entry to the function the "OverallLeftShift" is a signed integer value
4073 /// indicating the number of bytes that the subexpression is later shifted. For
4074 /// example, if the expression is later right shifted by 16 bits, the
4075 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4076 /// byte of ByteValues is actually being set.
4078 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4079 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4080 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4081 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4082 /// always in the local (OverallLeftShift) coordinate space.
4084 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4085 SmallVector<Value*, 8> &ByteValues) {
4086 if (Instruction *I = dyn_cast<Instruction>(V)) {
4087 // If this is an or instruction, it may be an inner node of the bswap.
4088 if (I->getOpcode() == Instruction::Or) {
4089 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4091 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4095 // If this is a logical shift by a constant multiple of 8, recurse with
4096 // OverallLeftShift and ByteMask adjusted.
4097 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4099 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4100 // Ensure the shift amount is defined and of a byte value.
4101 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4104 unsigned ByteShift = ShAmt >> 3;
4105 if (I->getOpcode() == Instruction::Shl) {
4106 // X << 2 -> collect(X, +2)
4107 OverallLeftShift += ByteShift;
4108 ByteMask >>= ByteShift;
4110 // X >>u 2 -> collect(X, -2)
4111 OverallLeftShift -= ByteShift;
4112 ByteMask <<= ByteShift;
4113 ByteMask &= (~0U >> (32-ByteValues.size()));
4116 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4117 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4119 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4123 // If this is a logical 'and' with a mask that clears bytes, clear the
4124 // corresponding bytes in ByteMask.
4125 if (I->getOpcode() == Instruction::And &&
4126 isa<ConstantInt>(I->getOperand(1))) {
4127 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4128 unsigned NumBytes = ByteValues.size();
4129 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4130 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4132 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4133 // If this byte is masked out by a later operation, we don't care what
4135 if ((ByteMask & (1 << i)) == 0)
4138 // If the AndMask is all zeros for this byte, clear the bit.
4139 APInt MaskB = AndMask & Byte;
4141 ByteMask &= ~(1U << i);
4145 // If the AndMask is not all ones for this byte, it's not a bytezap.
4149 // Otherwise, this byte is kept.
4152 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4157 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4158 // the input value to the bswap. Some observations: 1) if more than one byte
4159 // is demanded from this input, then it could not be successfully assembled
4160 // into a byteswap. At least one of the two bytes would not be aligned with
4161 // their ultimate destination.
4162 if (!isPowerOf2_32(ByteMask)) return true;
4163 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4165 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4166 // is demanded, it needs to go into byte 0 of the result. This means that the
4167 // byte needs to be shifted until it lands in the right byte bucket. The
4168 // shift amount depends on the position: if the byte is coming from the high
4169 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4170 // low part, it must be shifted left.
4171 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4172 if (InputByteNo < ByteValues.size()/2) {
4173 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4176 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4180 // If the destination byte value is already defined, the values are or'd
4181 // together, which isn't a bswap (unless it's an or of the same bits).
4182 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4184 ByteValues[DestByteNo] = V;
4188 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4189 /// If so, insert the new bswap intrinsic and return it.
4190 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4191 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4192 if (!ITy || ITy->getBitWidth() % 16 ||
4193 // ByteMask only allows up to 32-byte values.
4194 ITy->getBitWidth() > 32*8)
4195 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4197 /// ByteValues - For each byte of the result, we keep track of which value
4198 /// defines each byte.
4199 SmallVector<Value*, 8> ByteValues;
4200 ByteValues.resize(ITy->getBitWidth()/8);
4202 // Try to find all the pieces corresponding to the bswap.
4203 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4204 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4207 // Check to see if all of the bytes come from the same value.
4208 Value *V = ByteValues[0];
4209 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4211 // Check to make sure that all of the bytes come from the same value.
4212 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4213 if (ByteValues[i] != V)
4215 const Type *Tys[] = { ITy };
4216 Module *M = I.getParent()->getParent()->getParent();
4217 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4218 return CallInst::Create(F, V);
4221 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4222 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4223 /// we can simplify this expression to "cond ? C : D or B".
4224 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4225 Value *C, Value *D) {
4226 // If A is not a select of -1/0, this cannot match.
4228 if (!match(A, m_SelectCst(m_Value(Cond), -1, 0)))
4231 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4232 if (match(D, m_SelectCst(m_Specific(Cond), 0, -1)))
4233 return SelectInst::Create(Cond, C, B);
4234 if (match(D, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4235 return SelectInst::Create(Cond, C, B);
4236 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4237 if (match(B, m_SelectCst(m_Specific(Cond), 0, -1)))
4238 return SelectInst::Create(Cond, C, D);
4239 if (match(B, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4240 return SelectInst::Create(Cond, C, D);
4244 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4245 bool Changed = SimplifyCommutative(I);
4246 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4248 if (isa<UndefValue>(Op1)) // X | undef -> -1
4249 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4253 return ReplaceInstUsesWith(I, Op0);
4255 // See if we can simplify any instructions used by the instruction whose sole
4256 // purpose is to compute bits we don't care about.
4257 if (!isa<VectorType>(I.getType())) {
4258 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4259 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4260 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4261 KnownZero, KnownOne))
4263 } else if (isa<ConstantAggregateZero>(Op1)) {
4264 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4265 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4266 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4267 return ReplaceInstUsesWith(I, I.getOperand(1));
4273 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4274 ConstantInt *C1 = 0; Value *X = 0;
4275 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4276 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4277 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4278 InsertNewInstBefore(Or, I);
4280 return BinaryOperator::CreateAnd(Or,
4281 ConstantInt::get(RHS->getValue() | C1->getValue()));
4284 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4285 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4286 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4287 InsertNewInstBefore(Or, I);
4289 return BinaryOperator::CreateXor(Or,
4290 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4293 // Try to fold constant and into select arguments.
4294 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4295 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4297 if (isa<PHINode>(Op0))
4298 if (Instruction *NV = FoldOpIntoPhi(I))
4302 Value *A = 0, *B = 0;
4303 ConstantInt *C1 = 0, *C2 = 0;
4305 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4306 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4307 return ReplaceInstUsesWith(I, Op1);
4308 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4309 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4310 return ReplaceInstUsesWith(I, Op0);
4312 // (A | B) | C and A | (B | C) -> bswap if possible.
4313 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4314 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4315 match(Op1, m_Or(m_Value(), m_Value())) ||
4316 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4317 match(Op1, m_Shift(m_Value(), m_Value())))) {
4318 if (Instruction *BSwap = MatchBSwap(I))
4322 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4323 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4324 MaskedValueIsZero(Op1, C1->getValue())) {
4325 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4326 InsertNewInstBefore(NOr, I);
4328 return BinaryOperator::CreateXor(NOr, C1);
4331 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4332 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4333 MaskedValueIsZero(Op0, C1->getValue())) {
4334 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4335 InsertNewInstBefore(NOr, I);
4337 return BinaryOperator::CreateXor(NOr, C1);
4341 Value *C = 0, *D = 0;
4342 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4343 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4344 Value *V1 = 0, *V2 = 0, *V3 = 0;
4345 C1 = dyn_cast<ConstantInt>(C);
4346 C2 = dyn_cast<ConstantInt>(D);
4347 if (C1 && C2) { // (A & C1)|(B & C2)
4348 // If we have: ((V + N) & C1) | (V & C2)
4349 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4350 // replace with V+N.
4351 if (C1->getValue() == ~C2->getValue()) {
4352 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4353 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4354 // Add commutes, try both ways.
4355 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4356 return ReplaceInstUsesWith(I, A);
4357 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4358 return ReplaceInstUsesWith(I, A);
4360 // Or commutes, try both ways.
4361 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4362 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4363 // Add commutes, try both ways.
4364 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4365 return ReplaceInstUsesWith(I, B);
4366 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4367 return ReplaceInstUsesWith(I, B);
4370 V1 = 0; V2 = 0; V3 = 0;
4373 // Check to see if we have any common things being and'ed. If so, find the
4374 // terms for V1 & (V2|V3).
4375 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4376 if (A == B) // (A & C)|(A & D) == A & (C|D)
4377 V1 = A, V2 = C, V3 = D;
4378 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4379 V1 = A, V2 = B, V3 = C;
4380 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4381 V1 = C, V2 = A, V3 = D;
4382 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4383 V1 = C, V2 = A, V3 = B;
4387 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4388 return BinaryOperator::CreateAnd(V1, Or);
4392 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4393 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4395 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4397 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4399 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4403 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4404 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4405 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4406 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4407 SI0->getOperand(1) == SI1->getOperand(1) &&
4408 (SI0->hasOneUse() || SI1->hasOneUse())) {
4409 Instruction *NewOp =
4410 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4412 SI0->getName()), I);
4413 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4414 SI1->getOperand(1));
4418 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4419 if (A == Op1) // ~A | A == -1
4420 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4424 // Note, A is still live here!
4425 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4427 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4429 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4430 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4431 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4432 I.getName()+".demorgan"), I);
4433 return BinaryOperator::CreateNot(And);
4437 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4438 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4439 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4443 ConstantInt *LHSCst, *RHSCst;
4444 ICmpInst::Predicate LHSCC, RHSCC;
4445 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4446 if (match(Op0, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) &&
4447 match(RHS, m_ICmp(RHSCC, m_Specific(Val), m_ConstantInt(RHSCst))) &&
4449 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4450 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4451 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4452 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4453 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4455 // We can't fold (ugt x, C) | (sgt x, C2).
4456 PredicatesFoldable(LHSCC, RHSCC)) {
4457 // Ensure that the larger constant is on the RHS.
4458 ICmpInst *LHS = cast<ICmpInst>(Op0);
4460 if (ICmpInst::isEquality(LHSCC) ? ICmpInst::isSignedPredicate(RHSCC)
4461 : ICmpInst::isSignedPredicate(LHSCC))
4462 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4464 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4467 std::swap(LHS, RHS);
4468 std::swap(LHSCst, RHSCst);
4469 std::swap(LHSCC, RHSCC);
4472 // At this point, we know we have have two icmp instructions
4473 // comparing a value against two constants and or'ing the result
4474 // together. Because of the above check, we know that we only have
4475 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4476 // FoldICmpLogical check above), that the two constants are not
4478 assert(LHSCst != RHSCst && "Compares not folded above?");
4481 default: assert(0 && "Unknown integer condition code!");
4482 case ICmpInst::ICMP_EQ:
4484 default: assert(0 && "Unknown integer condition code!");
4485 case ICmpInst::ICMP_EQ:
4486 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4487 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4488 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4489 Val->getName()+".off");
4490 InsertNewInstBefore(Add, I);
4491 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4492 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4494 break; // (X == 13 | X == 15) -> no change
4495 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4496 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4498 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4499 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4500 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4501 return ReplaceInstUsesWith(I, RHS);
4504 case ICmpInst::ICMP_NE:
4506 default: assert(0 && "Unknown integer condition code!");
4507 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4508 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4509 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4510 return ReplaceInstUsesWith(I, LHS);
4511 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4512 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4513 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4514 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4517 case ICmpInst::ICMP_ULT:
4519 default: assert(0 && "Unknown integer condition code!");
4520 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4522 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4523 // If RHSCst is [us]MAXINT, it is always false. Not handling
4524 // this can cause overflow.
4525 if (RHSCst->isMaxValue(false))
4526 return ReplaceInstUsesWith(I, LHS);
4527 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4528 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4530 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4531 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4532 return ReplaceInstUsesWith(I, RHS);
4533 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4537 case ICmpInst::ICMP_SLT:
4539 default: assert(0 && "Unknown integer condition code!");
4540 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4542 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4543 // If RHSCst is [us]MAXINT, it is always false. Not handling
4544 // this can cause overflow.
4545 if (RHSCst->isMaxValue(true))
4546 return ReplaceInstUsesWith(I, LHS);
4547 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4548 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4550 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4551 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4552 return ReplaceInstUsesWith(I, RHS);
4553 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4557 case ICmpInst::ICMP_UGT:
4559 default: assert(0 && "Unknown integer condition code!");
4560 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4561 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4562 return ReplaceInstUsesWith(I, LHS);
4563 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4565 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4566 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4567 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4568 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4572 case ICmpInst::ICMP_SGT:
4574 default: assert(0 && "Unknown integer condition code!");
4575 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4576 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4577 return ReplaceInstUsesWith(I, LHS);
4578 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4580 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4581 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4582 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4583 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4591 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4592 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4593 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4594 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4595 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4596 !isa<ICmpInst>(Op1C->getOperand(0))) {
4597 const Type *SrcTy = Op0C->getOperand(0)->getType();
4598 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4599 // Only do this if the casts both really cause code to be
4601 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4603 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4605 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4606 Op1C->getOperand(0),
4608 InsertNewInstBefore(NewOp, I);
4609 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4616 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4617 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4618 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4619 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4620 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4621 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4622 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4623 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4624 // If either of the constants are nans, then the whole thing returns
4626 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4627 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4629 // Otherwise, no need to compare the two constants, compare the
4631 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4632 RHS->getOperand(0));
4635 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4636 FCmpInst::Predicate Op0CC, Op1CC;
4637 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4638 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4639 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4640 // Swap RHS operands to match LHS.
4641 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4642 std::swap(Op1LHS, Op1RHS);
4644 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4645 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4647 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4648 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4649 Op1CC == FCmpInst::FCMP_TRUE)
4650 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4651 else if (Op0CC == FCmpInst::FCMP_FALSE)
4652 return ReplaceInstUsesWith(I, Op1);
4653 else if (Op1CC == FCmpInst::FCMP_FALSE)
4654 return ReplaceInstUsesWith(I, Op0);
4657 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4658 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4659 if (Op0Ordered == Op1Ordered) {
4660 // If both are ordered or unordered, return a new fcmp with
4661 // or'ed predicates.
4662 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4664 if (Instruction *I = dyn_cast<Instruction>(RV))
4666 // Otherwise, it's a constant boolean value...
4667 return ReplaceInstUsesWith(I, RV);
4675 return Changed ? &I : 0;
4680 // XorSelf - Implements: X ^ X --> 0
4683 XorSelf(Value *rhs) : RHS(rhs) {}
4684 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4685 Instruction *apply(BinaryOperator &Xor) const {
4692 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4693 bool Changed = SimplifyCommutative(I);
4694 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4696 if (isa<UndefValue>(Op1)) {
4697 if (isa<UndefValue>(Op0))
4698 // Handle undef ^ undef -> 0 special case. This is a common
4700 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4701 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4704 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4705 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4706 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4707 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4710 // See if we can simplify any instructions used by the instruction whose sole
4711 // purpose is to compute bits we don't care about.
4712 if (!isa<VectorType>(I.getType())) {
4713 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4714 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4715 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4716 KnownZero, KnownOne))
4718 } else if (isa<ConstantAggregateZero>(Op1)) {
4719 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4722 // Is this a ~ operation?
4723 if (Value *NotOp = dyn_castNotVal(&I)) {
4724 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4725 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4726 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4727 if (Op0I->getOpcode() == Instruction::And ||
4728 Op0I->getOpcode() == Instruction::Or) {
4729 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4730 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4732 BinaryOperator::CreateNot(Op0I->getOperand(1),
4733 Op0I->getOperand(1)->getName()+".not");
4734 InsertNewInstBefore(NotY, I);
4735 if (Op0I->getOpcode() == Instruction::And)
4736 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4738 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4745 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4746 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4747 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4748 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4749 return new ICmpInst(ICI->getInversePredicate(),
4750 ICI->getOperand(0), ICI->getOperand(1));
4752 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4753 return new FCmpInst(FCI->getInversePredicate(),
4754 FCI->getOperand(0), FCI->getOperand(1));
4757 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4758 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4759 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4760 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4761 Instruction::CastOps Opcode = Op0C->getOpcode();
4762 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4763 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4764 Op0C->getDestTy())) {
4765 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4766 CI->getOpcode(), CI->getInversePredicate(),
4767 CI->getOperand(0), CI->getOperand(1)), I);
4768 NewCI->takeName(CI);
4769 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4776 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4777 // ~(c-X) == X-c-1 == X+(-c-1)
4778 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4779 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4780 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4781 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4782 ConstantInt::get(I.getType(), 1));
4783 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4786 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4787 if (Op0I->getOpcode() == Instruction::Add) {
4788 // ~(X-c) --> (-c-1)-X
4789 if (RHS->isAllOnesValue()) {
4790 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4791 return BinaryOperator::CreateSub(
4792 ConstantExpr::getSub(NegOp0CI,
4793 ConstantInt::get(I.getType(), 1)),
4794 Op0I->getOperand(0));
4795 } else if (RHS->getValue().isSignBit()) {
4796 // (X + C) ^ signbit -> (X + C + signbit)
4797 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4798 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4801 } else if (Op0I->getOpcode() == Instruction::Or) {
4802 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4803 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4804 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4805 // Anything in both C1 and C2 is known to be zero, remove it from
4807 Constant *CommonBits = And(Op0CI, RHS);
4808 NewRHS = ConstantExpr::getAnd(NewRHS,
4809 ConstantExpr::getNot(CommonBits));
4810 AddToWorkList(Op0I);
4811 I.setOperand(0, Op0I->getOperand(0));
4812 I.setOperand(1, NewRHS);
4819 // Try to fold constant and into select arguments.
4820 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4821 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4823 if (isa<PHINode>(Op0))
4824 if (Instruction *NV = FoldOpIntoPhi(I))
4828 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4830 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4832 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4834 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4837 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4840 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4841 if (A == Op0) { // B^(B|A) == (A|B)^B
4842 Op1I->swapOperands();
4844 std::swap(Op0, Op1);
4845 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4846 I.swapOperands(); // Simplified below.
4847 std::swap(Op0, Op1);
4849 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4850 if (Op0 == A) // A^(A^B) == B
4851 return ReplaceInstUsesWith(I, B);
4852 else if (Op0 == B) // A^(B^A) == B
4853 return ReplaceInstUsesWith(I, A);
4854 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4855 if (A == Op0) { // A^(A&B) -> A^(B&A)
4856 Op1I->swapOperands();
4859 if (B == Op0) { // A^(B&A) -> (B&A)^A
4860 I.swapOperands(); // Simplified below.
4861 std::swap(Op0, Op1);
4866 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4869 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4870 if (A == Op1) // (B|A)^B == (A|B)^B
4872 if (B == Op1) { // (A|B)^B == A & ~B
4874 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4875 return BinaryOperator::CreateAnd(A, NotB);
4877 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4878 if (Op1 == A) // (A^B)^A == B
4879 return ReplaceInstUsesWith(I, B);
4880 else if (Op1 == B) // (B^A)^A == B
4881 return ReplaceInstUsesWith(I, A);
4882 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4883 if (A == Op1) // (A&B)^A -> (B&A)^A
4885 if (B == Op1 && // (B&A)^A == ~B & A
4886 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4888 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4889 return BinaryOperator::CreateAnd(N, Op1);
4894 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4895 if (Op0I && Op1I && Op0I->isShift() &&
4896 Op0I->getOpcode() == Op1I->getOpcode() &&
4897 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4898 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4899 Instruction *NewOp =
4900 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4901 Op1I->getOperand(0),
4902 Op0I->getName()), I);
4903 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4904 Op1I->getOperand(1));
4908 Value *A, *B, *C, *D;
4909 // (A & B)^(A | B) -> A ^ B
4910 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4911 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4912 if ((A == C && B == D) || (A == D && B == C))
4913 return BinaryOperator::CreateXor(A, B);
4915 // (A | B)^(A & B) -> A ^ B
4916 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4917 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4918 if ((A == C && B == D) || (A == D && B == C))
4919 return BinaryOperator::CreateXor(A, B);
4923 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4924 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4925 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4926 // (X & Y)^(X & Y) -> (Y^Z) & X
4927 Value *X = 0, *Y = 0, *Z = 0;
4929 X = A, Y = B, Z = D;
4931 X = A, Y = B, Z = C;
4933 X = B, Y = A, Z = D;
4935 X = B, Y = A, Z = C;
4938 Instruction *NewOp =
4939 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4940 return BinaryOperator::CreateAnd(NewOp, X);
4945 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4946 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4947 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4950 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4951 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4952 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4953 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4954 const Type *SrcTy = Op0C->getOperand(0)->getType();
4955 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4956 // Only do this if the casts both really cause code to be generated.
4957 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4959 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4961 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4962 Op1C->getOperand(0),
4964 InsertNewInstBefore(NewOp, I);
4965 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4970 return Changed ? &I : 0;
4973 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4974 /// overflowed for this type.
4975 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4976 ConstantInt *In2, bool IsSigned = false) {
4977 Result = cast<ConstantInt>(Add(In1, In2));
4980 if (In2->getValue().isNegative())
4981 return Result->getValue().sgt(In1->getValue());
4983 return Result->getValue().slt(In1->getValue());
4985 return Result->getValue().ult(In1->getValue());
4988 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
4989 /// overflowed for this type.
4990 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4991 ConstantInt *In2, bool IsSigned = false) {
4992 Result = cast<ConstantInt>(Subtract(In1, In2));
4995 if (In2->getValue().isNegative())
4996 return Result->getValue().slt(In1->getValue());
4998 return Result->getValue().sgt(In1->getValue());
5000 return Result->getValue().ugt(In1->getValue());
5003 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5004 /// code necessary to compute the offset from the base pointer (without adding
5005 /// in the base pointer). Return the result as a signed integer of intptr size.
5006 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5007 TargetData &TD = IC.getTargetData();
5008 gep_type_iterator GTI = gep_type_begin(GEP);
5009 const Type *IntPtrTy = TD.getIntPtrType();
5010 Value *Result = Constant::getNullValue(IntPtrTy);
5012 // Build a mask for high order bits.
5013 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5014 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5016 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5019 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5020 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5021 if (OpC->isZero()) continue;
5023 // Handle a struct index, which adds its field offset to the pointer.
5024 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5025 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5027 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5028 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5030 Result = IC.InsertNewInstBefore(
5031 BinaryOperator::CreateAdd(Result,
5032 ConstantInt::get(IntPtrTy, Size),
5033 GEP->getName()+".offs"), I);
5037 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5038 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5039 Scale = ConstantExpr::getMul(OC, Scale);
5040 if (Constant *RC = dyn_cast<Constant>(Result))
5041 Result = ConstantExpr::getAdd(RC, Scale);
5043 // Emit an add instruction.
5044 Result = IC.InsertNewInstBefore(
5045 BinaryOperator::CreateAdd(Result, Scale,
5046 GEP->getName()+".offs"), I);
5050 // Convert to correct type.
5051 if (Op->getType() != IntPtrTy) {
5052 if (Constant *OpC = dyn_cast<Constant>(Op))
5053 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5055 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5056 Op->getName()+".c"), I);
5059 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5060 if (Constant *OpC = dyn_cast<Constant>(Op))
5061 Op = ConstantExpr::getMul(OpC, Scale);
5062 else // We'll let instcombine(mul) convert this to a shl if possible.
5063 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5064 GEP->getName()+".idx"), I);
5067 // Emit an add instruction.
5068 if (isa<Constant>(Op) && isa<Constant>(Result))
5069 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5070 cast<Constant>(Result));
5072 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5073 GEP->getName()+".offs"), I);
5079 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5080 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5081 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5082 /// complex, and scales are involved. The above expression would also be legal
5083 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5084 /// later form is less amenable to optimization though, and we are allowed to
5085 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5087 /// If we can't emit an optimized form for this expression, this returns null.
5089 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5091 TargetData &TD = IC.getTargetData();
5092 gep_type_iterator GTI = gep_type_begin(GEP);
5094 // Check to see if this gep only has a single variable index. If so, and if
5095 // any constant indices are a multiple of its scale, then we can compute this
5096 // in terms of the scale of the variable index. For example, if the GEP
5097 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5098 // because the expression will cross zero at the same point.
5099 unsigned i, e = GEP->getNumOperands();
5101 for (i = 1; i != e; ++i, ++GTI) {
5102 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5103 // Compute the aggregate offset of constant indices.
5104 if (CI->isZero()) continue;
5106 // Handle a struct index, which adds its field offset to the pointer.
5107 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5108 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5110 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5111 Offset += Size*CI->getSExtValue();
5114 // Found our variable index.
5119 // If there are no variable indices, we must have a constant offset, just
5120 // evaluate it the general way.
5121 if (i == e) return 0;
5123 Value *VariableIdx = GEP->getOperand(i);
5124 // Determine the scale factor of the variable element. For example, this is
5125 // 4 if the variable index is into an array of i32.
5126 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5128 // Verify that there are no other variable indices. If so, emit the hard way.
5129 for (++i, ++GTI; i != e; ++i, ++GTI) {
5130 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5133 // Compute the aggregate offset of constant indices.
5134 if (CI->isZero()) continue;
5136 // Handle a struct index, which adds its field offset to the pointer.
5137 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5138 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5140 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5141 Offset += Size*CI->getSExtValue();
5145 // Okay, we know we have a single variable index, which must be a
5146 // pointer/array/vector index. If there is no offset, life is simple, return
5148 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5150 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5151 // we don't need to bother extending: the extension won't affect where the
5152 // computation crosses zero.
5153 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5154 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5155 VariableIdx->getNameStart(), &I);
5159 // Otherwise, there is an index. The computation we will do will be modulo
5160 // the pointer size, so get it.
5161 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5163 Offset &= PtrSizeMask;
5164 VariableScale &= PtrSizeMask;
5166 // To do this transformation, any constant index must be a multiple of the
5167 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5168 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5169 // multiple of the variable scale.
5170 int64_t NewOffs = Offset / (int64_t)VariableScale;
5171 if (Offset != NewOffs*(int64_t)VariableScale)
5174 // Okay, we can do this evaluation. Start by converting the index to intptr.
5175 const Type *IntPtrTy = TD.getIntPtrType();
5176 if (VariableIdx->getType() != IntPtrTy)
5177 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5179 VariableIdx->getNameStart(), &I);
5180 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5181 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5185 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5186 /// else. At this point we know that the GEP is on the LHS of the comparison.
5187 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5188 ICmpInst::Predicate Cond,
5190 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5192 // Look through bitcasts.
5193 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5194 RHS = BCI->getOperand(0);
5196 Value *PtrBase = GEPLHS->getOperand(0);
5197 if (PtrBase == RHS) {
5198 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5199 // This transformation (ignoring the base and scales) is valid because we
5200 // know pointers can't overflow. See if we can output an optimized form.
5201 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5203 // If not, synthesize the offset the hard way.
5205 Offset = EmitGEPOffset(GEPLHS, I, *this);
5206 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5207 Constant::getNullValue(Offset->getType()));
5208 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5209 // If the base pointers are different, but the indices are the same, just
5210 // compare the base pointer.
5211 if (PtrBase != GEPRHS->getOperand(0)) {
5212 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5213 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5214 GEPRHS->getOperand(0)->getType();
5216 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5217 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5218 IndicesTheSame = false;
5222 // If all indices are the same, just compare the base pointers.
5224 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5225 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5227 // Otherwise, the base pointers are different and the indices are
5228 // different, bail out.
5232 // If one of the GEPs has all zero indices, recurse.
5233 bool AllZeros = true;
5234 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5235 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5236 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5241 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5242 ICmpInst::getSwappedPredicate(Cond), I);
5244 // If the other GEP has all zero indices, recurse.
5246 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5247 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5248 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5253 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5255 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5256 // If the GEPs only differ by one index, compare it.
5257 unsigned NumDifferences = 0; // Keep track of # differences.
5258 unsigned DiffOperand = 0; // The operand that differs.
5259 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5260 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5261 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5262 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5263 // Irreconcilable differences.
5267 if (NumDifferences++) break;
5272 if (NumDifferences == 0) // SAME GEP?
5273 return ReplaceInstUsesWith(I, // No comparison is needed here.
5274 ConstantInt::get(Type::Int1Ty,
5275 ICmpInst::isTrueWhenEqual(Cond)));
5277 else if (NumDifferences == 1) {
5278 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5279 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5280 // Make sure we do a signed comparison here.
5281 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5285 // Only lower this if the icmp is the only user of the GEP or if we expect
5286 // the result to fold to a constant!
5287 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5288 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5289 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5290 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5291 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5292 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5298 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5300 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5303 if (!isa<ConstantFP>(RHSC)) return 0;
5304 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5306 // Get the width of the mantissa. We don't want to hack on conversions that
5307 // might lose information from the integer, e.g. "i64 -> float"
5308 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5309 if (MantissaWidth == -1) return 0; // Unknown.
5311 // Check to see that the input is converted from an integer type that is small
5312 // enough that preserves all bits. TODO: check here for "known" sign bits.
5313 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5314 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5316 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5317 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5321 // If the conversion would lose info, don't hack on this.
5322 if ((int)InputSize > MantissaWidth)
5325 // Otherwise, we can potentially simplify the comparison. We know that it
5326 // will always come through as an integer value and we know the constant is
5327 // not a NAN (it would have been previously simplified).
5328 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5330 ICmpInst::Predicate Pred;
5331 switch (I.getPredicate()) {
5332 default: assert(0 && "Unexpected predicate!");
5333 case FCmpInst::FCMP_UEQ:
5334 case FCmpInst::FCMP_OEQ:
5335 Pred = ICmpInst::ICMP_EQ;
5337 case FCmpInst::FCMP_UGT:
5338 case FCmpInst::FCMP_OGT:
5339 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5341 case FCmpInst::FCMP_UGE:
5342 case FCmpInst::FCMP_OGE:
5343 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5345 case FCmpInst::FCMP_ULT:
5346 case FCmpInst::FCMP_OLT:
5347 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5349 case FCmpInst::FCMP_ULE:
5350 case FCmpInst::FCMP_OLE:
5351 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5353 case FCmpInst::FCMP_UNE:
5354 case FCmpInst::FCMP_ONE:
5355 Pred = ICmpInst::ICMP_NE;
5357 case FCmpInst::FCMP_ORD:
5358 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5359 case FCmpInst::FCMP_UNO:
5360 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5363 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5365 // Now we know that the APFloat is a normal number, zero or inf.
5367 // See if the FP constant is too large for the integer. For example,
5368 // comparing an i8 to 300.0.
5369 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5372 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5373 // and large values.
5374 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5375 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5376 APFloat::rmNearestTiesToEven);
5377 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5378 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5379 Pred == ICmpInst::ICMP_SLE)
5380 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5381 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5384 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5385 // +INF and large values.
5386 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5387 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5388 APFloat::rmNearestTiesToEven);
5389 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5390 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5391 Pred == ICmpInst::ICMP_ULE)
5392 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5393 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5398 // See if the RHS value is < SignedMin.
5399 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5400 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5401 APFloat::rmNearestTiesToEven);
5402 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5403 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5404 Pred == ICmpInst::ICMP_SGE)
5405 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5406 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5410 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5411 // [0, UMAX], but it may still be fractional. See if it is fractional by
5412 // casting the FP value to the integer value and back, checking for equality.
5413 // Don't do this for zero, because -0.0 is not fractional.
5414 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5415 if (!RHS.isZero() &&
5416 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5417 // If we had a comparison against a fractional value, we have to adjust the
5418 // compare predicate and sometimes the value. RHSC is rounded towards zero
5421 default: assert(0 && "Unexpected integer comparison!");
5422 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5423 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5424 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5425 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5426 case ICmpInst::ICMP_ULE:
5427 // (float)int <= 4.4 --> int <= 4
5428 // (float)int <= -4.4 --> false
5429 if (RHS.isNegative())
5430 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5432 case ICmpInst::ICMP_SLE:
5433 // (float)int <= 4.4 --> int <= 4
5434 // (float)int <= -4.4 --> int < -4
5435 if (RHS.isNegative())
5436 Pred = ICmpInst::ICMP_SLT;
5438 case ICmpInst::ICMP_ULT:
5439 // (float)int < -4.4 --> false
5440 // (float)int < 4.4 --> int <= 4
5441 if (RHS.isNegative())
5442 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5443 Pred = ICmpInst::ICMP_ULE;
5445 case ICmpInst::ICMP_SLT:
5446 // (float)int < -4.4 --> int < -4
5447 // (float)int < 4.4 --> int <= 4
5448 if (!RHS.isNegative())
5449 Pred = ICmpInst::ICMP_SLE;
5451 case ICmpInst::ICMP_UGT:
5452 // (float)int > 4.4 --> int > 4
5453 // (float)int > -4.4 --> true
5454 if (RHS.isNegative())
5455 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5457 case ICmpInst::ICMP_SGT:
5458 // (float)int > 4.4 --> int > 4
5459 // (float)int > -4.4 --> int >= -4
5460 if (RHS.isNegative())
5461 Pred = ICmpInst::ICMP_SGE;
5463 case ICmpInst::ICMP_UGE:
5464 // (float)int >= -4.4 --> true
5465 // (float)int >= 4.4 --> int > 4
5466 if (!RHS.isNegative())
5467 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5468 Pred = ICmpInst::ICMP_UGT;
5470 case ICmpInst::ICMP_SGE:
5471 // (float)int >= -4.4 --> int >= -4
5472 // (float)int >= 4.4 --> int > 4
5473 if (!RHS.isNegative())
5474 Pred = ICmpInst::ICMP_SGT;
5479 // Lower this FP comparison into an appropriate integer version of the
5481 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5484 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5485 bool Changed = SimplifyCompare(I);
5486 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5488 // Fold trivial predicates.
5489 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5490 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5491 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5492 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5494 // Simplify 'fcmp pred X, X'
5496 switch (I.getPredicate()) {
5497 default: assert(0 && "Unknown predicate!");
5498 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5499 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5500 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5501 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5502 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5503 case FCmpInst::FCMP_OLT: // True if ordered and less than
5504 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5505 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5507 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5508 case FCmpInst::FCMP_ULT: // True if unordered or less than
5509 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5510 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5511 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5512 I.setPredicate(FCmpInst::FCMP_UNO);
5513 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5516 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5517 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5518 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5519 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5520 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5521 I.setPredicate(FCmpInst::FCMP_ORD);
5522 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5527 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5528 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5530 // Handle fcmp with constant RHS
5531 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5532 // If the constant is a nan, see if we can fold the comparison based on it.
5533 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5534 if (CFP->getValueAPF().isNaN()) {
5535 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5536 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5537 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5538 "Comparison must be either ordered or unordered!");
5539 // True if unordered.
5540 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5544 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5545 switch (LHSI->getOpcode()) {
5546 case Instruction::PHI:
5547 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5548 // block. If in the same block, we're encouraging jump threading. If
5549 // not, we are just pessimizing the code by making an i1 phi.
5550 if (LHSI->getParent() == I.getParent())
5551 if (Instruction *NV = FoldOpIntoPhi(I))
5554 case Instruction::SIToFP:
5555 case Instruction::UIToFP:
5556 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5559 case Instruction::Select:
5560 // If either operand of the select is a constant, we can fold the
5561 // comparison into the select arms, which will cause one to be
5562 // constant folded and the select turned into a bitwise or.
5563 Value *Op1 = 0, *Op2 = 0;
5564 if (LHSI->hasOneUse()) {
5565 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5566 // Fold the known value into the constant operand.
5567 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5568 // Insert a new FCmp of the other select operand.
5569 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5570 LHSI->getOperand(2), RHSC,
5572 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5573 // Fold the known value into the constant operand.
5574 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5575 // Insert a new FCmp of the other select operand.
5576 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5577 LHSI->getOperand(1), RHSC,
5583 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5588 return Changed ? &I : 0;
5591 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5592 bool Changed = SimplifyCompare(I);
5593 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5594 const Type *Ty = Op0->getType();
5598 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5599 I.isTrueWhenEqual()));
5601 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5602 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5604 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5605 // addresses never equal each other! We already know that Op0 != Op1.
5606 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5607 isa<ConstantPointerNull>(Op0)) &&
5608 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5609 isa<ConstantPointerNull>(Op1)))
5610 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5611 !I.isTrueWhenEqual()));
5613 // icmp's with boolean values can always be turned into bitwise operations
5614 if (Ty == Type::Int1Ty) {
5615 switch (I.getPredicate()) {
5616 default: assert(0 && "Invalid icmp instruction!");
5617 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5618 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5619 InsertNewInstBefore(Xor, I);
5620 return BinaryOperator::CreateNot(Xor);
5622 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5623 return BinaryOperator::CreateXor(Op0, Op1);
5625 case ICmpInst::ICMP_UGT:
5626 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5628 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5629 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5630 InsertNewInstBefore(Not, I);
5631 return BinaryOperator::CreateAnd(Not, Op1);
5633 case ICmpInst::ICMP_SGT:
5634 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5636 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5637 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5638 InsertNewInstBefore(Not, I);
5639 return BinaryOperator::CreateAnd(Not, Op0);
5641 case ICmpInst::ICMP_UGE:
5642 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5644 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5645 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5646 InsertNewInstBefore(Not, I);
5647 return BinaryOperator::CreateOr(Not, Op1);
5649 case ICmpInst::ICMP_SGE:
5650 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5652 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5653 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5654 InsertNewInstBefore(Not, I);
5655 return BinaryOperator::CreateOr(Not, Op0);
5660 // See if we are doing a comparison with a constant.
5661 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5664 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5665 if (I.isEquality() && CI->isNullValue() &&
5666 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5667 // (icmp cond A B) if cond is equality
5668 return new ICmpInst(I.getPredicate(), A, B);
5671 // If we have an icmp le or icmp ge instruction, turn it into the
5672 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5673 // them being folded in the code below.
5674 switch (I.getPredicate()) {
5676 case ICmpInst::ICMP_ULE:
5677 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5678 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5679 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5680 case ICmpInst::ICMP_SLE:
5681 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5682 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5683 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5684 case ICmpInst::ICMP_UGE:
5685 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5686 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5687 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5688 case ICmpInst::ICMP_SGE:
5689 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5690 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5691 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5694 // See if we can fold the comparison based on range information we can get
5695 // by checking whether bits are known to be zero or one in the input.
5696 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5697 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5699 // If this comparison is a normal comparison, it demands all
5700 // bits, if it is a sign bit comparison, it only demands the sign bit.
5702 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5704 if (SimplifyDemandedBits(Op0,
5705 isSignBit ? APInt::getSignBit(BitWidth)
5706 : APInt::getAllOnesValue(BitWidth),
5707 KnownZero, KnownOne, 0))
5710 // Given the known and unknown bits, compute a range that the LHS could be
5711 // in. Compute the Min, Max and RHS values based on the known bits. For the
5712 // EQ and NE we use unsigned values.
5713 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5714 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5715 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5717 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5719 // If Min and Max are known to be the same, then SimplifyDemandedBits
5720 // figured out that the LHS is a constant. Just constant fold this now so
5721 // that code below can assume that Min != Max.
5723 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5724 ConstantInt::get(Min),
5727 // Based on the range information we know about the LHS, see if we can
5728 // simplify this comparison. For example, (x&4) < 8 is always true.
5729 const APInt &RHSVal = CI->getValue();
5730 switch (I.getPredicate()) { // LE/GE have been folded already.
5731 default: assert(0 && "Unknown icmp opcode!");
5732 case ICmpInst::ICMP_EQ:
5733 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5734 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5736 case ICmpInst::ICMP_NE:
5737 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5738 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5740 case ICmpInst::ICMP_ULT:
5741 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5742 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5743 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5744 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5745 if (RHSVal == Max) // A <u MAX -> A != MAX
5746 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5747 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5748 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5750 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5751 if (CI->isMinValue(true))
5752 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5753 ConstantInt::getAllOnesValue(Op0->getType()));
5755 case ICmpInst::ICMP_UGT:
5756 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5757 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5758 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5759 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5761 if (RHSVal == Min) // A >u MIN -> A != MIN
5762 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5763 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5764 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5766 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5767 if (CI->isMaxValue(true))
5768 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5769 ConstantInt::getNullValue(Op0->getType()));
5771 case ICmpInst::ICMP_SLT:
5772 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5773 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5774 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5775 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5776 if (RHSVal == Max) // A <s MAX -> A != MAX
5777 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5778 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5779 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5781 case ICmpInst::ICMP_SGT:
5782 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5783 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5784 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5785 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5787 if (RHSVal == Min) // A >s MIN -> A != MIN
5788 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5789 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5790 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5795 // Test if the ICmpInst instruction is used exclusively by a select as
5796 // part of a minimum or maximum operation. If so, refrain from doing
5797 // any other folding. This helps out other analyses which understand
5798 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5799 // and CodeGen. And in this case, at least one of the comparison
5800 // operands has at least one user besides the compare (the select),
5801 // which would often largely negate the benefit of folding anyway.
5803 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5804 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5805 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5808 // See if we are doing a comparison between a constant and an instruction that
5809 // can be folded into the comparison.
5810 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5811 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5812 // instruction, see if that instruction also has constants so that the
5813 // instruction can be folded into the icmp
5814 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5815 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5819 // Handle icmp with constant (but not simple integer constant) RHS
5820 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5821 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5822 switch (LHSI->getOpcode()) {
5823 case Instruction::GetElementPtr:
5824 if (RHSC->isNullValue()) {
5825 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5826 bool isAllZeros = true;
5827 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5828 if (!isa<Constant>(LHSI->getOperand(i)) ||
5829 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5834 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5835 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5839 case Instruction::PHI:
5840 // Only fold icmp into the PHI if the phi and fcmp are in the same
5841 // block. If in the same block, we're encouraging jump threading. If
5842 // not, we are just pessimizing the code by making an i1 phi.
5843 if (LHSI->getParent() == I.getParent())
5844 if (Instruction *NV = FoldOpIntoPhi(I))
5847 case Instruction::Select: {
5848 // If either operand of the select is a constant, we can fold the
5849 // comparison into the select arms, which will cause one to be
5850 // constant folded and the select turned into a bitwise or.
5851 Value *Op1 = 0, *Op2 = 0;
5852 if (LHSI->hasOneUse()) {
5853 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5854 // Fold the known value into the constant operand.
5855 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5856 // Insert a new ICmp of the other select operand.
5857 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5858 LHSI->getOperand(2), RHSC,
5860 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5861 // Fold the known value into the constant operand.
5862 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5863 // Insert a new ICmp of the other select operand.
5864 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5865 LHSI->getOperand(1), RHSC,
5871 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5874 case Instruction::Malloc:
5875 // If we have (malloc != null), and if the malloc has a single use, we
5876 // can assume it is successful and remove the malloc.
5877 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5878 AddToWorkList(LHSI);
5879 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5880 !I.isTrueWhenEqual()));
5886 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5887 if (User *GEP = dyn_castGetElementPtr(Op0))
5888 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5890 if (User *GEP = dyn_castGetElementPtr(Op1))
5891 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5892 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5895 // Test to see if the operands of the icmp are casted versions of other
5896 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5898 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5899 if (isa<PointerType>(Op0->getType()) &&
5900 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5901 // We keep moving the cast from the left operand over to the right
5902 // operand, where it can often be eliminated completely.
5903 Op0 = CI->getOperand(0);
5905 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5906 // so eliminate it as well.
5907 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5908 Op1 = CI2->getOperand(0);
5910 // If Op1 is a constant, we can fold the cast into the constant.
5911 if (Op0->getType() != Op1->getType()) {
5912 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5913 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5915 // Otherwise, cast the RHS right before the icmp
5916 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5919 return new ICmpInst(I.getPredicate(), Op0, Op1);
5923 if (isa<CastInst>(Op0)) {
5924 // Handle the special case of: icmp (cast bool to X), <cst>
5925 // This comes up when you have code like
5928 // For generality, we handle any zero-extension of any operand comparison
5929 // with a constant or another cast from the same type.
5930 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5931 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5935 // See if it's the same type of instruction on the left and right.
5936 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5937 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
5938 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
5939 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
5941 switch (Op0I->getOpcode()) {
5943 case Instruction::Add:
5944 case Instruction::Sub:
5945 case Instruction::Xor:
5946 // a+x icmp eq/ne b+x --> a icmp b
5947 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
5948 Op1I->getOperand(0));
5950 case Instruction::Mul:
5951 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5952 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
5953 // Mask = -1 >> count-trailing-zeros(Cst).
5954 if (!CI->isZero() && !CI->isOne()) {
5955 const APInt &AP = CI->getValue();
5956 ConstantInt *Mask = ConstantInt::get(
5957 APInt::getLowBitsSet(AP.getBitWidth(),
5959 AP.countTrailingZeros()));
5960 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
5962 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
5964 InsertNewInstBefore(And1, I);
5965 InsertNewInstBefore(And2, I);
5966 return new ICmpInst(I.getPredicate(), And1, And2);
5975 // ~x < ~y --> y < x
5977 if (match(Op0, m_Not(m_Value(A))) &&
5978 match(Op1, m_Not(m_Value(B))))
5979 return new ICmpInst(I.getPredicate(), B, A);
5982 if (I.isEquality()) {
5983 Value *A, *B, *C, *D;
5985 // -x == -y --> x == y
5986 if (match(Op0, m_Neg(m_Value(A))) &&
5987 match(Op1, m_Neg(m_Value(B))))
5988 return new ICmpInst(I.getPredicate(), A, B);
5990 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5991 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5992 Value *OtherVal = A == Op1 ? B : A;
5993 return new ICmpInst(I.getPredicate(), OtherVal,
5994 Constant::getNullValue(A->getType()));
5997 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5998 // A^c1 == C^c2 --> A == C^(c1^c2)
5999 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
6000 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
6001 if (Op1->hasOneUse()) {
6002 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6003 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6004 return new ICmpInst(I.getPredicate(), A,
6005 InsertNewInstBefore(Xor, I));
6008 // A^B == A^D -> B == D
6009 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6010 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6011 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6012 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6016 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6017 (A == Op0 || B == Op0)) {
6018 // A == (A^B) -> B == 0
6019 Value *OtherVal = A == Op0 ? B : A;
6020 return new ICmpInst(I.getPredicate(), OtherVal,
6021 Constant::getNullValue(A->getType()));
6023 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
6024 // (A-B) == A -> B == 0
6025 return new ICmpInst(I.getPredicate(), B,
6026 Constant::getNullValue(B->getType()));
6028 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
6029 // A == (A-B) -> B == 0
6030 return new ICmpInst(I.getPredicate(), B,
6031 Constant::getNullValue(B->getType()));
6034 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6035 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6036 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6037 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6038 Value *X = 0, *Y = 0, *Z = 0;
6041 X = B; Y = D; Z = A;
6042 } else if (A == D) {
6043 X = B; Y = C; Z = A;
6044 } else if (B == C) {
6045 X = A; Y = D; Z = B;
6046 } else if (B == D) {
6047 X = A; Y = C; Z = B;
6050 if (X) { // Build (X^Y) & Z
6051 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6052 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6053 I.setOperand(0, Op1);
6054 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6059 return Changed ? &I : 0;
6063 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6064 /// and CmpRHS are both known to be integer constants.
6065 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6066 ConstantInt *DivRHS) {
6067 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6068 const APInt &CmpRHSV = CmpRHS->getValue();
6070 // FIXME: If the operand types don't match the type of the divide
6071 // then don't attempt this transform. The code below doesn't have the
6072 // logic to deal with a signed divide and an unsigned compare (and
6073 // vice versa). This is because (x /s C1) <s C2 produces different
6074 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6075 // (x /u C1) <u C2. Simply casting the operands and result won't
6076 // work. :( The if statement below tests that condition and bails
6078 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6079 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6081 if (DivRHS->isZero())
6082 return 0; // The ProdOV computation fails on divide by zero.
6083 if (DivIsSigned && DivRHS->isAllOnesValue())
6084 return 0; // The overflow computation also screws up here
6085 if (DivRHS->isOne())
6086 return 0; // Not worth bothering, and eliminates some funny cases
6089 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6090 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6091 // C2 (CI). By solving for X we can turn this into a range check
6092 // instead of computing a divide.
6093 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6095 // Determine if the product overflows by seeing if the product is
6096 // not equal to the divide. Make sure we do the same kind of divide
6097 // as in the LHS instruction that we're folding.
6098 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6099 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6101 // Get the ICmp opcode
6102 ICmpInst::Predicate Pred = ICI.getPredicate();
6104 // Figure out the interval that is being checked. For example, a comparison
6105 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6106 // Compute this interval based on the constants involved and the signedness of
6107 // the compare/divide. This computes a half-open interval, keeping track of
6108 // whether either value in the interval overflows. After analysis each
6109 // overflow variable is set to 0 if it's corresponding bound variable is valid
6110 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6111 int LoOverflow = 0, HiOverflow = 0;
6112 ConstantInt *LoBound = 0, *HiBound = 0;
6114 if (!DivIsSigned) { // udiv
6115 // e.g. X/5 op 3 --> [15, 20)
6117 HiOverflow = LoOverflow = ProdOV;
6119 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6120 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6121 if (CmpRHSV == 0) { // (X / pos) op 0
6122 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6123 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6125 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6126 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6127 HiOverflow = LoOverflow = ProdOV;
6129 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6130 } else { // (X / pos) op neg
6131 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6132 HiBound = AddOne(Prod);
6133 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6135 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6136 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6140 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6141 if (CmpRHSV == 0) { // (X / neg) op 0
6142 // e.g. X/-5 op 0 --> [-4, 5)
6143 LoBound = AddOne(DivRHS);
6144 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6145 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6146 HiOverflow = 1; // [INTMIN+1, overflow)
6147 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6149 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6150 // e.g. X/-5 op 3 --> [-19, -14)
6151 HiBound = AddOne(Prod);
6152 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6154 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6155 } else { // (X / neg) op neg
6156 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6157 LoOverflow = HiOverflow = ProdOV;
6159 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6162 // Dividing by a negative swaps the condition. LT <-> GT
6163 Pred = ICmpInst::getSwappedPredicate(Pred);
6166 Value *X = DivI->getOperand(0);
6168 default: assert(0 && "Unhandled icmp opcode!");
6169 case ICmpInst::ICMP_EQ:
6170 if (LoOverflow && HiOverflow)
6171 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6172 else if (HiOverflow)
6173 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6174 ICmpInst::ICMP_UGE, X, LoBound);
6175 else if (LoOverflow)
6176 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6177 ICmpInst::ICMP_ULT, X, HiBound);
6179 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6180 case ICmpInst::ICMP_NE:
6181 if (LoOverflow && HiOverflow)
6182 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6183 else if (HiOverflow)
6184 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6185 ICmpInst::ICMP_ULT, X, LoBound);
6186 else if (LoOverflow)
6187 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6188 ICmpInst::ICMP_UGE, X, HiBound);
6190 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6191 case ICmpInst::ICMP_ULT:
6192 case ICmpInst::ICMP_SLT:
6193 if (LoOverflow == +1) // Low bound is greater than input range.
6194 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6195 if (LoOverflow == -1) // Low bound is less than input range.
6196 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6197 return new ICmpInst(Pred, X, LoBound);
6198 case ICmpInst::ICMP_UGT:
6199 case ICmpInst::ICMP_SGT:
6200 if (HiOverflow == +1) // High bound greater than input range.
6201 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6202 else if (HiOverflow == -1) // High bound less than input range.
6203 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6204 if (Pred == ICmpInst::ICMP_UGT)
6205 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6207 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6212 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6214 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6217 const APInt &RHSV = RHS->getValue();
6219 switch (LHSI->getOpcode()) {
6220 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6221 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6222 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6224 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6225 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6226 Value *CompareVal = LHSI->getOperand(0);
6228 // If the sign bit of the XorCST is not set, there is no change to
6229 // the operation, just stop using the Xor.
6230 if (!XorCST->getValue().isNegative()) {
6231 ICI.setOperand(0, CompareVal);
6232 AddToWorkList(LHSI);
6236 // Was the old condition true if the operand is positive?
6237 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6239 // If so, the new one isn't.
6240 isTrueIfPositive ^= true;
6242 if (isTrueIfPositive)
6243 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6245 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6249 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6250 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6251 LHSI->getOperand(0)->hasOneUse()) {
6252 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6254 // If the LHS is an AND of a truncating cast, we can widen the
6255 // and/compare to be the input width without changing the value
6256 // produced, eliminating a cast.
6257 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6258 // We can do this transformation if either the AND constant does not
6259 // have its sign bit set or if it is an equality comparison.
6260 // Extending a relational comparison when we're checking the sign
6261 // bit would not work.
6262 if (Cast->hasOneUse() &&
6263 (ICI.isEquality() ||
6264 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6266 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6267 APInt NewCST = AndCST->getValue();
6268 NewCST.zext(BitWidth);
6270 NewCI.zext(BitWidth);
6271 Instruction *NewAnd =
6272 BinaryOperator::CreateAnd(Cast->getOperand(0),
6273 ConstantInt::get(NewCST),LHSI->getName());
6274 InsertNewInstBefore(NewAnd, ICI);
6275 return new ICmpInst(ICI.getPredicate(), NewAnd,
6276 ConstantInt::get(NewCI));
6280 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6281 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6282 // happens a LOT in code produced by the C front-end, for bitfield
6284 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6285 if (Shift && !Shift->isShift())
6289 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6290 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6291 const Type *AndTy = AndCST->getType(); // Type of the and.
6293 // We can fold this as long as we can't shift unknown bits
6294 // into the mask. This can only happen with signed shift
6295 // rights, as they sign-extend.
6297 bool CanFold = Shift->isLogicalShift();
6299 // To test for the bad case of the signed shr, see if any
6300 // of the bits shifted in could be tested after the mask.
6301 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6302 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6304 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6305 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6306 AndCST->getValue()) == 0)
6312 if (Shift->getOpcode() == Instruction::Shl)
6313 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6315 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6317 // Check to see if we are shifting out any of the bits being
6319 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6320 // If we shifted bits out, the fold is not going to work out.
6321 // As a special case, check to see if this means that the
6322 // result is always true or false now.
6323 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6324 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6325 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6326 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6328 ICI.setOperand(1, NewCst);
6329 Constant *NewAndCST;
6330 if (Shift->getOpcode() == Instruction::Shl)
6331 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6333 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6334 LHSI->setOperand(1, NewAndCST);
6335 LHSI->setOperand(0, Shift->getOperand(0));
6336 AddToWorkList(Shift); // Shift is dead.
6337 AddUsesToWorkList(ICI);
6343 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6344 // preferable because it allows the C<<Y expression to be hoisted out
6345 // of a loop if Y is invariant and X is not.
6346 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6347 ICI.isEquality() && !Shift->isArithmeticShift() &&
6348 isa<Instruction>(Shift->getOperand(0))) {
6351 if (Shift->getOpcode() == Instruction::LShr) {
6352 NS = BinaryOperator::CreateShl(AndCST,
6353 Shift->getOperand(1), "tmp");
6355 // Insert a logical shift.
6356 NS = BinaryOperator::CreateLShr(AndCST,
6357 Shift->getOperand(1), "tmp");
6359 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6361 // Compute X & (C << Y).
6362 Instruction *NewAnd =
6363 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6364 InsertNewInstBefore(NewAnd, ICI);
6366 ICI.setOperand(0, NewAnd);
6372 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6373 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6376 uint32_t TypeBits = RHSV.getBitWidth();
6378 // Check that the shift amount is in range. If not, don't perform
6379 // undefined shifts. When the shift is visited it will be
6381 if (ShAmt->uge(TypeBits))
6384 if (ICI.isEquality()) {
6385 // If we are comparing against bits always shifted out, the
6386 // comparison cannot succeed.
6388 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6389 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6390 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6391 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6392 return ReplaceInstUsesWith(ICI, Cst);
6395 if (LHSI->hasOneUse()) {
6396 // Otherwise strength reduce the shift into an and.
6397 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6399 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6402 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6403 Mask, LHSI->getName()+".mask");
6404 Value *And = InsertNewInstBefore(AndI, ICI);
6405 return new ICmpInst(ICI.getPredicate(), And,
6406 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6410 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6411 bool TrueIfSigned = false;
6412 if (LHSI->hasOneUse() &&
6413 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6414 // (X << 31) <s 0 --> (X&1) != 0
6415 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6416 (TypeBits-ShAmt->getZExtValue()-1));
6418 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6419 Mask, LHSI->getName()+".mask");
6420 Value *And = InsertNewInstBefore(AndI, ICI);
6422 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6423 And, Constant::getNullValue(And->getType()));
6428 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6429 case Instruction::AShr: {
6430 // Only handle equality comparisons of shift-by-constant.
6431 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6432 if (!ShAmt || !ICI.isEquality()) break;
6434 // Check that the shift amount is in range. If not, don't perform
6435 // undefined shifts. When the shift is visited it will be
6437 uint32_t TypeBits = RHSV.getBitWidth();
6438 if (ShAmt->uge(TypeBits))
6441 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6443 // If we are comparing against bits always shifted out, the
6444 // comparison cannot succeed.
6445 APInt Comp = RHSV << ShAmtVal;
6446 if (LHSI->getOpcode() == Instruction::LShr)
6447 Comp = Comp.lshr(ShAmtVal);
6449 Comp = Comp.ashr(ShAmtVal);
6451 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6452 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6453 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6454 return ReplaceInstUsesWith(ICI, Cst);
6457 // Otherwise, check to see if the bits shifted out are known to be zero.
6458 // If so, we can compare against the unshifted value:
6459 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6460 if (LHSI->hasOneUse() &&
6461 MaskedValueIsZero(LHSI->getOperand(0),
6462 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6463 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6464 ConstantExpr::getShl(RHS, ShAmt));
6467 if (LHSI->hasOneUse()) {
6468 // Otherwise strength reduce the shift into an and.
6469 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6470 Constant *Mask = ConstantInt::get(Val);
6473 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6474 Mask, LHSI->getName()+".mask");
6475 Value *And = InsertNewInstBefore(AndI, ICI);
6476 return new ICmpInst(ICI.getPredicate(), And,
6477 ConstantExpr::getShl(RHS, ShAmt));
6482 case Instruction::SDiv:
6483 case Instruction::UDiv:
6484 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6485 // Fold this div into the comparison, producing a range check.
6486 // Determine, based on the divide type, what the range is being
6487 // checked. If there is an overflow on the low or high side, remember
6488 // it, otherwise compute the range [low, hi) bounding the new value.
6489 // See: InsertRangeTest above for the kinds of replacements possible.
6490 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6491 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6496 case Instruction::Add:
6497 // Fold: icmp pred (add, X, C1), C2
6499 if (!ICI.isEquality()) {
6500 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6502 const APInt &LHSV = LHSC->getValue();
6504 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6507 if (ICI.isSignedPredicate()) {
6508 if (CR.getLower().isSignBit()) {
6509 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6510 ConstantInt::get(CR.getUpper()));
6511 } else if (CR.getUpper().isSignBit()) {
6512 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6513 ConstantInt::get(CR.getLower()));
6516 if (CR.getLower().isMinValue()) {
6517 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6518 ConstantInt::get(CR.getUpper()));
6519 } else if (CR.getUpper().isMinValue()) {
6520 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6521 ConstantInt::get(CR.getLower()));
6528 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6529 if (ICI.isEquality()) {
6530 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6532 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6533 // the second operand is a constant, simplify a bit.
6534 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6535 switch (BO->getOpcode()) {
6536 case Instruction::SRem:
6537 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6538 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6539 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6540 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6541 Instruction *NewRem =
6542 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6544 InsertNewInstBefore(NewRem, ICI);
6545 return new ICmpInst(ICI.getPredicate(), NewRem,
6546 Constant::getNullValue(BO->getType()));
6550 case Instruction::Add:
6551 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6552 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6553 if (BO->hasOneUse())
6554 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6555 Subtract(RHS, BOp1C));
6556 } else if (RHSV == 0) {
6557 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6558 // efficiently invertible, or if the add has just this one use.
6559 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6561 if (Value *NegVal = dyn_castNegVal(BOp1))
6562 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6563 else if (Value *NegVal = dyn_castNegVal(BOp0))
6564 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6565 else if (BO->hasOneUse()) {
6566 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6567 InsertNewInstBefore(Neg, ICI);
6569 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6573 case Instruction::Xor:
6574 // For the xor case, we can xor two constants together, eliminating
6575 // the explicit xor.
6576 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6577 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6578 ConstantExpr::getXor(RHS, BOC));
6581 case Instruction::Sub:
6582 // Replace (([sub|xor] A, B) != 0) with (A != B)
6584 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6588 case Instruction::Or:
6589 // If bits are being or'd in that are not present in the constant we
6590 // are comparing against, then the comparison could never succeed!
6591 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6592 Constant *NotCI = ConstantExpr::getNot(RHS);
6593 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6594 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6599 case Instruction::And:
6600 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6601 // If bits are being compared against that are and'd out, then the
6602 // comparison can never succeed!
6603 if ((RHSV & ~BOC->getValue()) != 0)
6604 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6607 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6608 if (RHS == BOC && RHSV.isPowerOf2())
6609 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6610 ICmpInst::ICMP_NE, LHSI,
6611 Constant::getNullValue(RHS->getType()));
6613 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6614 if (BOC->getValue().isSignBit()) {
6615 Value *X = BO->getOperand(0);
6616 Constant *Zero = Constant::getNullValue(X->getType());
6617 ICmpInst::Predicate pred = isICMP_NE ?
6618 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6619 return new ICmpInst(pred, X, Zero);
6622 // ((X & ~7) == 0) --> X < 8
6623 if (RHSV == 0 && isHighOnes(BOC)) {
6624 Value *X = BO->getOperand(0);
6625 Constant *NegX = ConstantExpr::getNeg(BOC);
6626 ICmpInst::Predicate pred = isICMP_NE ?
6627 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6628 return new ICmpInst(pred, X, NegX);
6633 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6634 // Handle icmp {eq|ne} <intrinsic>, intcst.
6635 if (II->getIntrinsicID() == Intrinsic::bswap) {
6637 ICI.setOperand(0, II->getOperand(1));
6638 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6642 } else { // Not a ICMP_EQ/ICMP_NE
6643 // If the LHS is a cast from an integral value of the same size,
6644 // then since we know the RHS is a constant, try to simlify.
6645 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6646 Value *CastOp = Cast->getOperand(0);
6647 const Type *SrcTy = CastOp->getType();
6648 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6649 if (SrcTy->isInteger() &&
6650 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6651 // If this is an unsigned comparison, try to make the comparison use
6652 // smaller constant values.
6653 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6654 // X u< 128 => X s> -1
6655 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6656 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6657 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6658 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6659 // X u> 127 => X s< 0
6660 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6661 Constant::getNullValue(SrcTy));
6669 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6670 /// We only handle extending casts so far.
6672 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6673 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6674 Value *LHSCIOp = LHSCI->getOperand(0);
6675 const Type *SrcTy = LHSCIOp->getType();
6676 const Type *DestTy = LHSCI->getType();
6679 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6680 // integer type is the same size as the pointer type.
6681 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6682 getTargetData().getPointerSizeInBits() ==
6683 cast<IntegerType>(DestTy)->getBitWidth()) {
6685 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6686 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6687 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6688 RHSOp = RHSC->getOperand(0);
6689 // If the pointer types don't match, insert a bitcast.
6690 if (LHSCIOp->getType() != RHSOp->getType())
6691 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6695 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6698 // The code below only handles extension cast instructions, so far.
6700 if (LHSCI->getOpcode() != Instruction::ZExt &&
6701 LHSCI->getOpcode() != Instruction::SExt)
6704 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6705 bool isSignedCmp = ICI.isSignedPredicate();
6707 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6708 // Not an extension from the same type?
6709 RHSCIOp = CI->getOperand(0);
6710 if (RHSCIOp->getType() != LHSCIOp->getType())
6713 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6714 // and the other is a zext), then we can't handle this.
6715 if (CI->getOpcode() != LHSCI->getOpcode())
6718 // Deal with equality cases early.
6719 if (ICI.isEquality())
6720 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6722 // A signed comparison of sign extended values simplifies into a
6723 // signed comparison.
6724 if (isSignedCmp && isSignedExt)
6725 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6727 // The other three cases all fold into an unsigned comparison.
6728 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6731 // If we aren't dealing with a constant on the RHS, exit early
6732 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6736 // Compute the constant that would happen if we truncated to SrcTy then
6737 // reextended to DestTy.
6738 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6739 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6741 // If the re-extended constant didn't change...
6743 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6744 // For example, we might have:
6745 // %A = sext short %X to uint
6746 // %B = icmp ugt uint %A, 1330
6747 // It is incorrect to transform this into
6748 // %B = icmp ugt short %X, 1330
6749 // because %A may have negative value.
6751 // However, we allow this when the compare is EQ/NE, because they are
6753 if (isSignedExt == isSignedCmp || ICI.isEquality())
6754 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6758 // The re-extended constant changed so the constant cannot be represented
6759 // in the shorter type. Consequently, we cannot emit a simple comparison.
6761 // First, handle some easy cases. We know the result cannot be equal at this
6762 // point so handle the ICI.isEquality() cases
6763 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6764 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6765 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6766 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6768 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6769 // should have been folded away previously and not enter in here.
6772 // We're performing a signed comparison.
6773 if (cast<ConstantInt>(CI)->getValue().isNegative())
6774 Result = ConstantInt::getFalse(); // X < (small) --> false
6776 Result = ConstantInt::getTrue(); // X < (large) --> true
6778 // We're performing an unsigned comparison.
6780 // We're performing an unsigned comp with a sign extended value.
6781 // This is true if the input is >= 0. [aka >s -1]
6782 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6783 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6784 NegOne, ICI.getName()), ICI);
6786 // Unsigned extend & unsigned compare -> always true.
6787 Result = ConstantInt::getTrue();
6791 // Finally, return the value computed.
6792 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6793 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6794 return ReplaceInstUsesWith(ICI, Result);
6796 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6797 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6798 "ICmp should be folded!");
6799 if (Constant *CI = dyn_cast<Constant>(Result))
6800 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6801 return BinaryOperator::CreateNot(Result);
6804 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6805 return commonShiftTransforms(I);
6808 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6809 return commonShiftTransforms(I);
6812 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6813 if (Instruction *R = commonShiftTransforms(I))
6816 Value *Op0 = I.getOperand(0);
6818 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6819 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6820 if (CSI->isAllOnesValue())
6821 return ReplaceInstUsesWith(I, CSI);
6823 // See if we can turn a signed shr into an unsigned shr.
6824 if (!isa<VectorType>(I.getType()) &&
6825 MaskedValueIsZero(Op0,
6826 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6827 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6832 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6833 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6834 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6836 // shl X, 0 == X and shr X, 0 == X
6837 // shl 0, X == 0 and shr 0, X == 0
6838 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6839 Op0 == Constant::getNullValue(Op0->getType()))
6840 return ReplaceInstUsesWith(I, Op0);
6842 if (isa<UndefValue>(Op0)) {
6843 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6844 return ReplaceInstUsesWith(I, Op0);
6845 else // undef << X -> 0, undef >>u X -> 0
6846 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6848 if (isa<UndefValue>(Op1)) {
6849 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6850 return ReplaceInstUsesWith(I, Op0);
6851 else // X << undef, X >>u undef -> 0
6852 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6855 // Try to fold constant and into select arguments.
6856 if (isa<Constant>(Op0))
6857 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6858 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6861 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6862 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6867 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6868 BinaryOperator &I) {
6869 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6871 // See if we can simplify any instructions used by the instruction whose sole
6872 // purpose is to compute bits we don't care about.
6873 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6874 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6875 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6876 KnownZero, KnownOne))
6879 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6880 // of a signed value.
6882 if (Op1->uge(TypeBits)) {
6883 if (I.getOpcode() != Instruction::AShr)
6884 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6886 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6891 // ((X*C1) << C2) == (X * (C1 << C2))
6892 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6893 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6894 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6895 return BinaryOperator::CreateMul(BO->getOperand(0),
6896 ConstantExpr::getShl(BOOp, Op1));
6898 // Try to fold constant and into select arguments.
6899 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6900 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6902 if (isa<PHINode>(Op0))
6903 if (Instruction *NV = FoldOpIntoPhi(I))
6906 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6907 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6908 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6909 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6910 // place. Don't try to do this transformation in this case. Also, we
6911 // require that the input operand is a shift-by-constant so that we have
6912 // confidence that the shifts will get folded together. We could do this
6913 // xform in more cases, but it is unlikely to be profitable.
6914 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6915 isa<ConstantInt>(TrOp->getOperand(1))) {
6916 // Okay, we'll do this xform. Make the shift of shift.
6917 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6918 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6920 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6922 // For logical shifts, the truncation has the effect of making the high
6923 // part of the register be zeros. Emulate this by inserting an AND to
6924 // clear the top bits as needed. This 'and' will usually be zapped by
6925 // other xforms later if dead.
6926 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6927 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6928 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6930 // The mask we constructed says what the trunc would do if occurring
6931 // between the shifts. We want to know the effect *after* the second
6932 // shift. We know that it is a logical shift by a constant, so adjust the
6933 // mask as appropriate.
6934 if (I.getOpcode() == Instruction::Shl)
6935 MaskV <<= Op1->getZExtValue();
6937 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6938 MaskV = MaskV.lshr(Op1->getZExtValue());
6941 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6943 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6945 // Return the value truncated to the interesting size.
6946 return new TruncInst(And, I.getType());
6950 if (Op0->hasOneUse()) {
6951 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6952 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6955 switch (Op0BO->getOpcode()) {
6957 case Instruction::Add:
6958 case Instruction::And:
6959 case Instruction::Or:
6960 case Instruction::Xor: {
6961 // These operators commute.
6962 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6963 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6964 match(Op0BO->getOperand(1),
6965 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6966 Instruction *YS = BinaryOperator::CreateShl(
6967 Op0BO->getOperand(0), Op1,
6969 InsertNewInstBefore(YS, I); // (Y << C)
6971 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6972 Op0BO->getOperand(1)->getName());
6973 InsertNewInstBefore(X, I); // (X + (Y << C))
6974 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6975 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6976 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6979 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6980 Value *Op0BOOp1 = Op0BO->getOperand(1);
6981 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6983 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6984 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6986 Instruction *YS = BinaryOperator::CreateShl(
6987 Op0BO->getOperand(0), Op1,
6989 InsertNewInstBefore(YS, I); // (Y << C)
6991 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6992 V1->getName()+".mask");
6993 InsertNewInstBefore(XM, I); // X & (CC << C)
6995 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7000 case Instruction::Sub: {
7001 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7002 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7003 match(Op0BO->getOperand(0),
7004 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
7005 Instruction *YS = BinaryOperator::CreateShl(
7006 Op0BO->getOperand(1), Op1,
7008 InsertNewInstBefore(YS, I); // (Y << C)
7010 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7011 Op0BO->getOperand(0)->getName());
7012 InsertNewInstBefore(X, I); // (X + (Y << C))
7013 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7014 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7015 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7018 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7019 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7020 match(Op0BO->getOperand(0),
7021 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7022 m_ConstantInt(CC))) && V2 == Op1 &&
7023 cast<BinaryOperator>(Op0BO->getOperand(0))
7024 ->getOperand(0)->hasOneUse()) {
7025 Instruction *YS = BinaryOperator::CreateShl(
7026 Op0BO->getOperand(1), Op1,
7028 InsertNewInstBefore(YS, I); // (Y << C)
7030 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7031 V1->getName()+".mask");
7032 InsertNewInstBefore(XM, I); // X & (CC << C)
7034 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7042 // If the operand is an bitwise operator with a constant RHS, and the
7043 // shift is the only use, we can pull it out of the shift.
7044 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7045 bool isValid = true; // Valid only for And, Or, Xor
7046 bool highBitSet = false; // Transform if high bit of constant set?
7048 switch (Op0BO->getOpcode()) {
7049 default: isValid = false; break; // Do not perform transform!
7050 case Instruction::Add:
7051 isValid = isLeftShift;
7053 case Instruction::Or:
7054 case Instruction::Xor:
7057 case Instruction::And:
7062 // If this is a signed shift right, and the high bit is modified
7063 // by the logical operation, do not perform the transformation.
7064 // The highBitSet boolean indicates the value of the high bit of
7065 // the constant which would cause it to be modified for this
7068 if (isValid && I.getOpcode() == Instruction::AShr)
7069 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7072 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7074 Instruction *NewShift =
7075 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7076 InsertNewInstBefore(NewShift, I);
7077 NewShift->takeName(Op0BO);
7079 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7086 // Find out if this is a shift of a shift by a constant.
7087 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7088 if (ShiftOp && !ShiftOp->isShift())
7091 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7092 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7093 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7094 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7095 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7096 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7097 Value *X = ShiftOp->getOperand(0);
7099 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7100 if (AmtSum > TypeBits)
7103 const IntegerType *Ty = cast<IntegerType>(I.getType());
7105 // Check for (X << c1) << c2 and (X >> c1) >> c2
7106 if (I.getOpcode() == ShiftOp->getOpcode()) {
7107 return BinaryOperator::Create(I.getOpcode(), X,
7108 ConstantInt::get(Ty, AmtSum));
7109 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7110 I.getOpcode() == Instruction::AShr) {
7111 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7112 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7113 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7114 I.getOpcode() == Instruction::LShr) {
7115 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7116 Instruction *Shift =
7117 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7118 InsertNewInstBefore(Shift, I);
7120 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7121 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7124 // Okay, if we get here, one shift must be left, and the other shift must be
7125 // right. See if the amounts are equal.
7126 if (ShiftAmt1 == ShiftAmt2) {
7127 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7128 if (I.getOpcode() == Instruction::Shl) {
7129 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7130 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7132 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7133 if (I.getOpcode() == Instruction::LShr) {
7134 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7135 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7137 // We can simplify ((X << C) >>s C) into a trunc + sext.
7138 // NOTE: we could do this for any C, but that would make 'unusual' integer
7139 // types. For now, just stick to ones well-supported by the code
7141 const Type *SExtType = 0;
7142 switch (Ty->getBitWidth() - ShiftAmt1) {
7149 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7154 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7155 InsertNewInstBefore(NewTrunc, I);
7156 return new SExtInst(NewTrunc, Ty);
7158 // Otherwise, we can't handle it yet.
7159 } else if (ShiftAmt1 < ShiftAmt2) {
7160 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7162 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7163 if (I.getOpcode() == Instruction::Shl) {
7164 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7165 ShiftOp->getOpcode() == Instruction::AShr);
7166 Instruction *Shift =
7167 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7168 InsertNewInstBefore(Shift, I);
7170 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7171 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7174 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7175 if (I.getOpcode() == Instruction::LShr) {
7176 assert(ShiftOp->getOpcode() == Instruction::Shl);
7177 Instruction *Shift =
7178 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7179 InsertNewInstBefore(Shift, I);
7181 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7182 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7185 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7187 assert(ShiftAmt2 < ShiftAmt1);
7188 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7190 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7191 if (I.getOpcode() == Instruction::Shl) {
7192 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7193 ShiftOp->getOpcode() == Instruction::AShr);
7194 Instruction *Shift =
7195 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7196 ConstantInt::get(Ty, ShiftDiff));
7197 InsertNewInstBefore(Shift, I);
7199 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7200 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7203 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7204 if (I.getOpcode() == Instruction::LShr) {
7205 assert(ShiftOp->getOpcode() == Instruction::Shl);
7206 Instruction *Shift =
7207 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7208 InsertNewInstBefore(Shift, I);
7210 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7211 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7214 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7221 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7222 /// expression. If so, decompose it, returning some value X, such that Val is
7225 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7227 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7228 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7229 Offset = CI->getZExtValue();
7231 return ConstantInt::get(Type::Int32Ty, 0);
7232 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7233 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7234 if (I->getOpcode() == Instruction::Shl) {
7235 // This is a value scaled by '1 << the shift amt'.
7236 Scale = 1U << RHS->getZExtValue();
7238 return I->getOperand(0);
7239 } else if (I->getOpcode() == Instruction::Mul) {
7240 // This value is scaled by 'RHS'.
7241 Scale = RHS->getZExtValue();
7243 return I->getOperand(0);
7244 } else if (I->getOpcode() == Instruction::Add) {
7245 // We have X+C. Check to see if we really have (X*C2)+C1,
7246 // where C1 is divisible by C2.
7249 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7250 Offset += RHS->getZExtValue();
7257 // Otherwise, we can't look past this.
7264 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7265 /// try to eliminate the cast by moving the type information into the alloc.
7266 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7267 AllocationInst &AI) {
7268 const PointerType *PTy = cast<PointerType>(CI.getType());
7270 // Remove any uses of AI that are dead.
7271 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7273 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7274 Instruction *User = cast<Instruction>(*UI++);
7275 if (isInstructionTriviallyDead(User)) {
7276 while (UI != E && *UI == User)
7277 ++UI; // If this instruction uses AI more than once, don't break UI.
7280 DOUT << "IC: DCE: " << *User;
7281 EraseInstFromFunction(*User);
7285 // Get the type really allocated and the type casted to.
7286 const Type *AllocElTy = AI.getAllocatedType();
7287 const Type *CastElTy = PTy->getElementType();
7288 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7290 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7291 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7292 if (CastElTyAlign < AllocElTyAlign) return 0;
7294 // If the allocation has multiple uses, only promote it if we are strictly
7295 // increasing the alignment of the resultant allocation. If we keep it the
7296 // same, we open the door to infinite loops of various kinds.
7297 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7299 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7300 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7301 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7303 // See if we can satisfy the modulus by pulling a scale out of the array
7305 unsigned ArraySizeScale;
7307 Value *NumElements = // See if the array size is a decomposable linear expr.
7308 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7310 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7312 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7313 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7315 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7320 // If the allocation size is constant, form a constant mul expression
7321 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7322 if (isa<ConstantInt>(NumElements))
7323 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7324 // otherwise multiply the amount and the number of elements
7325 else if (Scale != 1) {
7326 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7327 Amt = InsertNewInstBefore(Tmp, AI);
7331 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7332 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7333 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7334 Amt = InsertNewInstBefore(Tmp, AI);
7337 AllocationInst *New;
7338 if (isa<MallocInst>(AI))
7339 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7341 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7342 InsertNewInstBefore(New, AI);
7345 // If the allocation has multiple uses, insert a cast and change all things
7346 // that used it to use the new cast. This will also hack on CI, but it will
7348 if (!AI.hasOneUse()) {
7349 AddUsesToWorkList(AI);
7350 // New is the allocation instruction, pointer typed. AI is the original
7351 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7352 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7353 InsertNewInstBefore(NewCast, AI);
7354 AI.replaceAllUsesWith(NewCast);
7356 return ReplaceInstUsesWith(CI, New);
7359 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7360 /// and return it as type Ty without inserting any new casts and without
7361 /// changing the computed value. This is used by code that tries to decide
7362 /// whether promoting or shrinking integer operations to wider or smaller types
7363 /// will allow us to eliminate a truncate or extend.
7365 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7366 /// extension operation if Ty is larger.
7368 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7369 /// should return true if trunc(V) can be computed by computing V in the smaller
7370 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7371 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7372 /// efficiently truncated.
7374 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7375 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7376 /// the final result.
7377 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7379 int &NumCastsRemoved) {
7380 // We can always evaluate constants in another type.
7381 if (isa<ConstantInt>(V))
7384 Instruction *I = dyn_cast<Instruction>(V);
7385 if (!I) return false;
7387 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7389 // If this is an extension or truncate, we can often eliminate it.
7390 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7391 // If this is a cast from the destination type, we can trivially eliminate
7392 // it, and this will remove a cast overall.
7393 if (I->getOperand(0)->getType() == Ty) {
7394 // If the first operand is itself a cast, and is eliminable, do not count
7395 // this as an eliminable cast. We would prefer to eliminate those two
7397 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7403 // We can't extend or shrink something that has multiple uses: doing so would
7404 // require duplicating the instruction in general, which isn't profitable.
7405 if (!I->hasOneUse()) return false;
7407 switch (I->getOpcode()) {
7408 case Instruction::Add:
7409 case Instruction::Sub:
7410 case Instruction::Mul:
7411 case Instruction::And:
7412 case Instruction::Or:
7413 case Instruction::Xor:
7414 // These operators can all arbitrarily be extended or truncated.
7415 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7417 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7420 case Instruction::Shl:
7421 // If we are truncating the result of this SHL, and if it's a shift of a
7422 // constant amount, we can always perform a SHL in a smaller type.
7423 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7424 uint32_t BitWidth = Ty->getBitWidth();
7425 if (BitWidth < OrigTy->getBitWidth() &&
7426 CI->getLimitedValue(BitWidth) < BitWidth)
7427 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7431 case Instruction::LShr:
7432 // If this is a truncate of a logical shr, we can truncate it to a smaller
7433 // lshr iff we know that the bits we would otherwise be shifting in are
7435 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7436 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7437 uint32_t BitWidth = Ty->getBitWidth();
7438 if (BitWidth < OrigBitWidth &&
7439 MaskedValueIsZero(I->getOperand(0),
7440 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7441 CI->getLimitedValue(BitWidth) < BitWidth) {
7442 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7447 case Instruction::ZExt:
7448 case Instruction::SExt:
7449 case Instruction::Trunc:
7450 // If this is the same kind of case as our original (e.g. zext+zext), we
7451 // can safely replace it. Note that replacing it does not reduce the number
7452 // of casts in the input.
7453 if (I->getOpcode() == CastOpc)
7456 case Instruction::Select: {
7457 SelectInst *SI = cast<SelectInst>(I);
7458 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7460 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7463 case Instruction::PHI: {
7464 // We can change a phi if we can change all operands.
7465 PHINode *PN = cast<PHINode>(I);
7466 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7467 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7473 // TODO: Can handle more cases here.
7480 /// EvaluateInDifferentType - Given an expression that
7481 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7482 /// evaluate the expression.
7483 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7485 if (Constant *C = dyn_cast<Constant>(V))
7486 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7488 // Otherwise, it must be an instruction.
7489 Instruction *I = cast<Instruction>(V);
7490 Instruction *Res = 0;
7491 switch (I->getOpcode()) {
7492 case Instruction::Add:
7493 case Instruction::Sub:
7494 case Instruction::Mul:
7495 case Instruction::And:
7496 case Instruction::Or:
7497 case Instruction::Xor:
7498 case Instruction::AShr:
7499 case Instruction::LShr:
7500 case Instruction::Shl: {
7501 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7502 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7503 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7507 case Instruction::Trunc:
7508 case Instruction::ZExt:
7509 case Instruction::SExt:
7510 // If the source type of the cast is the type we're trying for then we can
7511 // just return the source. There's no need to insert it because it is not
7513 if (I->getOperand(0)->getType() == Ty)
7514 return I->getOperand(0);
7516 // Otherwise, must be the same type of cast, so just reinsert a new one.
7517 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7520 case Instruction::Select: {
7521 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7522 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7523 Res = SelectInst::Create(I->getOperand(0), True, False);
7526 case Instruction::PHI: {
7527 PHINode *OPN = cast<PHINode>(I);
7528 PHINode *NPN = PHINode::Create(Ty);
7529 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7530 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7531 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7537 // TODO: Can handle more cases here.
7538 assert(0 && "Unreachable!");
7543 return InsertNewInstBefore(Res, *I);
7546 /// @brief Implement the transforms common to all CastInst visitors.
7547 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7548 Value *Src = CI.getOperand(0);
7550 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7551 // eliminate it now.
7552 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7553 if (Instruction::CastOps opc =
7554 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7555 // The first cast (CSrc) is eliminable so we need to fix up or replace
7556 // the second cast (CI). CSrc will then have a good chance of being dead.
7557 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7561 // If we are casting a select then fold the cast into the select
7562 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7563 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7566 // If we are casting a PHI then fold the cast into the PHI
7567 if (isa<PHINode>(Src))
7568 if (Instruction *NV = FoldOpIntoPhi(CI))
7574 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7575 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7576 Value *Src = CI.getOperand(0);
7578 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7579 // If casting the result of a getelementptr instruction with no offset, turn
7580 // this into a cast of the original pointer!
7581 if (GEP->hasAllZeroIndices()) {
7582 // Changing the cast operand is usually not a good idea but it is safe
7583 // here because the pointer operand is being replaced with another
7584 // pointer operand so the opcode doesn't need to change.
7586 CI.setOperand(0, GEP->getOperand(0));
7590 // If the GEP has a single use, and the base pointer is a bitcast, and the
7591 // GEP computes a constant offset, see if we can convert these three
7592 // instructions into fewer. This typically happens with unions and other
7593 // non-type-safe code.
7594 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7595 if (GEP->hasAllConstantIndices()) {
7596 // We are guaranteed to get a constant from EmitGEPOffset.
7597 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7598 int64_t Offset = OffsetV->getSExtValue();
7600 // Get the base pointer input of the bitcast, and the type it points to.
7601 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7602 const Type *GEPIdxTy =
7603 cast<PointerType>(OrigBase->getType())->getElementType();
7604 if (GEPIdxTy->isSized()) {
7605 SmallVector<Value*, 8> NewIndices;
7607 // Start with the index over the outer type. Note that the type size
7608 // might be zero (even if the offset isn't zero) if the indexed type
7609 // is something like [0 x {int, int}]
7610 const Type *IntPtrTy = TD->getIntPtrType();
7611 int64_t FirstIdx = 0;
7612 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7613 FirstIdx = Offset/TySize;
7616 // Handle silly modulus not returning values values [0..TySize).
7620 assert(Offset >= 0);
7622 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7625 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7627 // Index into the types. If we fail, set OrigBase to null.
7629 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7630 const StructLayout *SL = TD->getStructLayout(STy);
7631 if (Offset < (int64_t)SL->getSizeInBytes()) {
7632 unsigned Elt = SL->getElementContainingOffset(Offset);
7633 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7635 Offset -= SL->getElementOffset(Elt);
7636 GEPIdxTy = STy->getElementType(Elt);
7638 // Otherwise, we can't index into this, bail out.
7642 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7643 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7644 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7645 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7648 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7650 GEPIdxTy = STy->getElementType();
7652 // Otherwise, we can't index into this, bail out.
7658 // If we were able to index down into an element, create the GEP
7659 // and bitcast the result. This eliminates one bitcast, potentially
7661 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7663 NewIndices.end(), "");
7664 InsertNewInstBefore(NGEP, CI);
7665 NGEP->takeName(GEP);
7667 if (isa<BitCastInst>(CI))
7668 return new BitCastInst(NGEP, CI.getType());
7669 assert(isa<PtrToIntInst>(CI));
7670 return new PtrToIntInst(NGEP, CI.getType());
7677 return commonCastTransforms(CI);
7682 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7683 /// integer types. This function implements the common transforms for all those
7685 /// @brief Implement the transforms common to CastInst with integer operands
7686 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7687 if (Instruction *Result = commonCastTransforms(CI))
7690 Value *Src = CI.getOperand(0);
7691 const Type *SrcTy = Src->getType();
7692 const Type *DestTy = CI.getType();
7693 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7694 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7696 // See if we can simplify any instructions used by the LHS whose sole
7697 // purpose is to compute bits we don't care about.
7698 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7699 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7700 KnownZero, KnownOne))
7703 // If the source isn't an instruction or has more than one use then we
7704 // can't do anything more.
7705 Instruction *SrcI = dyn_cast<Instruction>(Src);
7706 if (!SrcI || !Src->hasOneUse())
7709 // Attempt to propagate the cast into the instruction for int->int casts.
7710 int NumCastsRemoved = 0;
7711 if (!isa<BitCastInst>(CI) &&
7712 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7713 CI.getOpcode(), NumCastsRemoved)) {
7714 // If this cast is a truncate, evaluting in a different type always
7715 // eliminates the cast, so it is always a win. If this is a zero-extension,
7716 // we need to do an AND to maintain the clear top-part of the computation,
7717 // so we require that the input have eliminated at least one cast. If this
7718 // is a sign extension, we insert two new casts (to do the extension) so we
7719 // require that two casts have been eliminated.
7721 switch (CI.getOpcode()) {
7723 // All the others use floating point so we shouldn't actually
7724 // get here because of the check above.
7725 assert(0 && "Unknown cast type");
7726 case Instruction::Trunc:
7729 case Instruction::ZExt:
7730 DoXForm = NumCastsRemoved >= 1;
7732 case Instruction::SExt:
7733 DoXForm = NumCastsRemoved >= 2;
7738 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7739 CI.getOpcode() == Instruction::SExt);
7740 assert(Res->getType() == DestTy);
7741 switch (CI.getOpcode()) {
7742 default: assert(0 && "Unknown cast type!");
7743 case Instruction::Trunc:
7744 case Instruction::BitCast:
7745 // Just replace this cast with the result.
7746 return ReplaceInstUsesWith(CI, Res);
7747 case Instruction::ZExt: {
7748 // We need to emit an AND to clear the high bits.
7749 assert(SrcBitSize < DestBitSize && "Not a zext?");
7750 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7752 return BinaryOperator::CreateAnd(Res, C);
7754 case Instruction::SExt:
7755 // We need to emit a cast to truncate, then a cast to sext.
7756 return CastInst::Create(Instruction::SExt,
7757 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7763 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7764 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7766 switch (SrcI->getOpcode()) {
7767 case Instruction::Add:
7768 case Instruction::Mul:
7769 case Instruction::And:
7770 case Instruction::Or:
7771 case Instruction::Xor:
7772 // If we are discarding information, rewrite.
7773 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7774 // Don't insert two casts if they cannot be eliminated. We allow
7775 // two casts to be inserted if the sizes are the same. This could
7776 // only be converting signedness, which is a noop.
7777 if (DestBitSize == SrcBitSize ||
7778 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7779 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7780 Instruction::CastOps opcode = CI.getOpcode();
7781 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7782 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7783 return BinaryOperator::Create(
7784 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7788 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7789 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7790 SrcI->getOpcode() == Instruction::Xor &&
7791 Op1 == ConstantInt::getTrue() &&
7792 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7793 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7794 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7797 case Instruction::SDiv:
7798 case Instruction::UDiv:
7799 case Instruction::SRem:
7800 case Instruction::URem:
7801 // If we are just changing the sign, rewrite.
7802 if (DestBitSize == SrcBitSize) {
7803 // Don't insert two casts if they cannot be eliminated. We allow
7804 // two casts to be inserted if the sizes are the same. This could
7805 // only be converting signedness, which is a noop.
7806 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7807 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7808 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7810 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7812 return BinaryOperator::Create(
7813 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7818 case Instruction::Shl:
7819 // Allow changing the sign of the source operand. Do not allow
7820 // changing the size of the shift, UNLESS the shift amount is a
7821 // constant. We must not change variable sized shifts to a smaller
7822 // size, because it is undefined to shift more bits out than exist
7824 if (DestBitSize == SrcBitSize ||
7825 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7826 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7827 Instruction::BitCast : Instruction::Trunc);
7828 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7829 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7830 return BinaryOperator::CreateShl(Op0c, Op1c);
7833 case Instruction::AShr:
7834 // If this is a signed shr, and if all bits shifted in are about to be
7835 // truncated off, turn it into an unsigned shr to allow greater
7837 if (DestBitSize < SrcBitSize &&
7838 isa<ConstantInt>(Op1)) {
7839 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7840 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7841 // Insert the new logical shift right.
7842 return BinaryOperator::CreateLShr(Op0, Op1);
7850 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7851 if (Instruction *Result = commonIntCastTransforms(CI))
7854 Value *Src = CI.getOperand(0);
7855 const Type *Ty = CI.getType();
7856 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7857 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7859 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7860 switch (SrcI->getOpcode()) {
7862 case Instruction::LShr:
7863 // We can shrink lshr to something smaller if we know the bits shifted in
7864 // are already zeros.
7865 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7866 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7868 // Get a mask for the bits shifting in.
7869 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7870 Value* SrcIOp0 = SrcI->getOperand(0);
7871 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7872 if (ShAmt >= DestBitWidth) // All zeros.
7873 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7875 // Okay, we can shrink this. Truncate the input, then return a new
7877 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7878 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7880 return BinaryOperator::CreateLShr(V1, V2);
7882 } else { // This is a variable shr.
7884 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7885 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7886 // loop-invariant and CSE'd.
7887 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7888 Value *One = ConstantInt::get(SrcI->getType(), 1);
7890 Value *V = InsertNewInstBefore(
7891 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7893 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7894 SrcI->getOperand(0),
7896 Value *Zero = Constant::getNullValue(V->getType());
7897 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7907 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7908 /// in order to eliminate the icmp.
7909 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7911 // If we are just checking for a icmp eq of a single bit and zext'ing it
7912 // to an integer, then shift the bit to the appropriate place and then
7913 // cast to integer to avoid the comparison.
7914 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7915 const APInt &Op1CV = Op1C->getValue();
7917 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7918 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7919 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7920 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7921 if (!DoXform) return ICI;
7923 Value *In = ICI->getOperand(0);
7924 Value *Sh = ConstantInt::get(In->getType(),
7925 In->getType()->getPrimitiveSizeInBits()-1);
7926 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7927 In->getName()+".lobit"),
7929 if (In->getType() != CI.getType())
7930 In = CastInst::CreateIntegerCast(In, CI.getType(),
7931 false/*ZExt*/, "tmp", &CI);
7933 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7934 Constant *One = ConstantInt::get(In->getType(), 1);
7935 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7936 In->getName()+".not"),
7940 return ReplaceInstUsesWith(CI, In);
7945 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7946 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7947 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7948 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7949 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7950 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7951 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7952 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7953 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7954 // This only works for EQ and NE
7955 ICI->isEquality()) {
7956 // If Op1C some other power of two, convert:
7957 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7958 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7959 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7960 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7962 APInt KnownZeroMask(~KnownZero);
7963 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7964 if (!DoXform) return ICI;
7966 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7967 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7968 // (X&4) == 2 --> false
7969 // (X&4) != 2 --> true
7970 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7971 Res = ConstantExpr::getZExt(Res, CI.getType());
7972 return ReplaceInstUsesWith(CI, Res);
7975 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7976 Value *In = ICI->getOperand(0);
7978 // Perform a logical shr by shiftamt.
7979 // Insert the shift to put the result in the low bit.
7980 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7981 ConstantInt::get(In->getType(), ShiftAmt),
7982 In->getName()+".lobit"), CI);
7985 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7986 Constant *One = ConstantInt::get(In->getType(), 1);
7987 In = BinaryOperator::CreateXor(In, One, "tmp");
7988 InsertNewInstBefore(cast<Instruction>(In), CI);
7991 if (CI.getType() == In->getType())
7992 return ReplaceInstUsesWith(CI, In);
7994 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8002 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8003 // If one of the common conversion will work ..
8004 if (Instruction *Result = commonIntCastTransforms(CI))
8007 Value *Src = CI.getOperand(0);
8009 // If this is a cast of a cast
8010 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8011 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8012 // types and if the sizes are just right we can convert this into a logical
8013 // 'and' which will be much cheaper than the pair of casts.
8014 if (isa<TruncInst>(CSrc)) {
8015 // Get the sizes of the types involved
8016 Value *A = CSrc->getOperand(0);
8017 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8018 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8019 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8020 // If we're actually extending zero bits and the trunc is a no-op
8021 if (MidSize < DstSize && SrcSize == DstSize) {
8022 // Replace both of the casts with an And of the type mask.
8023 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8024 Constant *AndConst = ConstantInt::get(AndValue);
8026 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8027 // Unfortunately, if the type changed, we need to cast it back.
8028 if (And->getType() != CI.getType()) {
8029 And->setName(CSrc->getName()+".mask");
8030 InsertNewInstBefore(And, CI);
8031 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8038 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8039 return transformZExtICmp(ICI, CI);
8041 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8042 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8043 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8044 // of the (zext icmp) will be transformed.
8045 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8046 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8047 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8048 (transformZExtICmp(LHS, CI, false) ||
8049 transformZExtICmp(RHS, CI, false))) {
8050 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8051 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8052 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8059 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8060 if (Instruction *I = commonIntCastTransforms(CI))
8063 Value *Src = CI.getOperand(0);
8065 // Canonicalize sign-extend from i1 to a select.
8066 if (Src->getType() == Type::Int1Ty)
8067 return SelectInst::Create(Src,
8068 ConstantInt::getAllOnesValue(CI.getType()),
8069 Constant::getNullValue(CI.getType()));
8071 // See if the value being truncated is already sign extended. If so, just
8072 // eliminate the trunc/sext pair.
8073 if (getOpcode(Src) == Instruction::Trunc) {
8074 Value *Op = cast<User>(Src)->getOperand(0);
8075 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8076 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8077 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8078 unsigned NumSignBits = ComputeNumSignBits(Op);
8080 if (OpBits == DestBits) {
8081 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8082 // bits, it is already ready.
8083 if (NumSignBits > DestBits-MidBits)
8084 return ReplaceInstUsesWith(CI, Op);
8085 } else if (OpBits < DestBits) {
8086 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8087 // bits, just sext from i32.
8088 if (NumSignBits > OpBits-MidBits)
8089 return new SExtInst(Op, CI.getType(), "tmp");
8091 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8092 // bits, just truncate to i32.
8093 if (NumSignBits > OpBits-MidBits)
8094 return new TruncInst(Op, CI.getType(), "tmp");
8098 // If the input is a shl/ashr pair of a same constant, then this is a sign
8099 // extension from a smaller value. If we could trust arbitrary bitwidth
8100 // integers, we could turn this into a truncate to the smaller bit and then
8101 // use a sext for the whole extension. Since we don't, look deeper and check
8102 // for a truncate. If the source and dest are the same type, eliminate the
8103 // trunc and extend and just do shifts. For example, turn:
8104 // %a = trunc i32 %i to i8
8105 // %b = shl i8 %a, 6
8106 // %c = ashr i8 %b, 6
8107 // %d = sext i8 %c to i32
8109 // %a = shl i32 %i, 30
8110 // %d = ashr i32 %a, 30
8112 ConstantInt *BA = 0, *CA = 0;
8113 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8114 m_ConstantInt(CA))) &&
8115 BA == CA && isa<TruncInst>(A)) {
8116 Value *I = cast<TruncInst>(A)->getOperand(0);
8117 if (I->getType() == CI.getType()) {
8118 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8119 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8120 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8121 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8122 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8124 return BinaryOperator::CreateAShr(I, ShAmtV);
8131 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8132 /// in the specified FP type without changing its value.
8133 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8135 APFloat F = CFP->getValueAPF();
8136 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8138 return ConstantFP::get(F);
8142 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8143 /// through it until we get the source value.
8144 static Value *LookThroughFPExtensions(Value *V) {
8145 if (Instruction *I = dyn_cast<Instruction>(V))
8146 if (I->getOpcode() == Instruction::FPExt)
8147 return LookThroughFPExtensions(I->getOperand(0));
8149 // If this value is a constant, return the constant in the smallest FP type
8150 // that can accurately represent it. This allows us to turn
8151 // (float)((double)X+2.0) into x+2.0f.
8152 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8153 if (CFP->getType() == Type::PPC_FP128Ty)
8154 return V; // No constant folding of this.
8155 // See if the value can be truncated to float and then reextended.
8156 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8158 if (CFP->getType() == Type::DoubleTy)
8159 return V; // Won't shrink.
8160 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8162 // Don't try to shrink to various long double types.
8168 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8169 if (Instruction *I = commonCastTransforms(CI))
8172 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8173 // smaller than the destination type, we can eliminate the truncate by doing
8174 // the add as the smaller type. This applies to add/sub/mul/div as well as
8175 // many builtins (sqrt, etc).
8176 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8177 if (OpI && OpI->hasOneUse()) {
8178 switch (OpI->getOpcode()) {
8180 case Instruction::Add:
8181 case Instruction::Sub:
8182 case Instruction::Mul:
8183 case Instruction::FDiv:
8184 case Instruction::FRem:
8185 const Type *SrcTy = OpI->getType();
8186 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8187 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8188 if (LHSTrunc->getType() != SrcTy &&
8189 RHSTrunc->getType() != SrcTy) {
8190 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8191 // If the source types were both smaller than the destination type of
8192 // the cast, do this xform.
8193 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8194 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8195 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8197 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8199 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8208 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8209 return commonCastTransforms(CI);
8212 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8213 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8215 return commonCastTransforms(FI);
8217 // fptoui(uitofp(X)) --> X
8218 // fptoui(sitofp(X)) --> X
8219 // This is safe if the intermediate type has enough bits in its mantissa to
8220 // accurately represent all values of X. For example, do not do this with
8221 // i64->float->i64. This is also safe for sitofp case, because any negative
8222 // 'X' value would cause an undefined result for the fptoui.
8223 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8224 OpI->getOperand(0)->getType() == FI.getType() &&
8225 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8226 OpI->getType()->getFPMantissaWidth())
8227 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8229 return commonCastTransforms(FI);
8232 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8233 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8235 return commonCastTransforms(FI);
8237 // fptosi(sitofp(X)) --> X
8238 // fptosi(uitofp(X)) --> X
8239 // This is safe if the intermediate type has enough bits in its mantissa to
8240 // accurately represent all values of X. For example, do not do this with
8241 // i64->float->i64. This is also safe for sitofp case, because any negative
8242 // 'X' value would cause an undefined result for the fptoui.
8243 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8244 OpI->getOperand(0)->getType() == FI.getType() &&
8245 (int)FI.getType()->getPrimitiveSizeInBits() <=
8246 OpI->getType()->getFPMantissaWidth())
8247 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8249 return commonCastTransforms(FI);
8252 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8253 return commonCastTransforms(CI);
8256 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8257 return commonCastTransforms(CI);
8260 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8261 return commonPointerCastTransforms(CI);
8264 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8265 if (Instruction *I = commonCastTransforms(CI))
8268 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8269 if (!DestPointee->isSized()) return 0;
8271 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8274 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8275 m_ConstantInt(Cst)))) {
8276 // If the source and destination operands have the same type, see if this
8277 // is a single-index GEP.
8278 if (X->getType() == CI.getType()) {
8279 // Get the size of the pointee type.
8280 uint64_t Size = TD->getABITypeSize(DestPointee);
8282 // Convert the constant to intptr type.
8283 APInt Offset = Cst->getValue();
8284 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8286 // If Offset is evenly divisible by Size, we can do this xform.
8287 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8288 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8289 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8292 // TODO: Could handle other cases, e.g. where add is indexing into field of
8294 } else if (CI.getOperand(0)->hasOneUse() &&
8295 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8296 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8297 // "inttoptr+GEP" instead of "add+intptr".
8299 // Get the size of the pointee type.
8300 uint64_t Size = TD->getABITypeSize(DestPointee);
8302 // Convert the constant to intptr type.
8303 APInt Offset = Cst->getValue();
8304 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8306 // If Offset is evenly divisible by Size, we can do this xform.
8307 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8308 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8310 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8312 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8318 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8319 // If the operands are integer typed then apply the integer transforms,
8320 // otherwise just apply the common ones.
8321 Value *Src = CI.getOperand(0);
8322 const Type *SrcTy = Src->getType();
8323 const Type *DestTy = CI.getType();
8325 if (SrcTy->isInteger() && DestTy->isInteger()) {
8326 if (Instruction *Result = commonIntCastTransforms(CI))
8328 } else if (isa<PointerType>(SrcTy)) {
8329 if (Instruction *I = commonPointerCastTransforms(CI))
8332 if (Instruction *Result = commonCastTransforms(CI))
8337 // Get rid of casts from one type to the same type. These are useless and can
8338 // be replaced by the operand.
8339 if (DestTy == Src->getType())
8340 return ReplaceInstUsesWith(CI, Src);
8342 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8343 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8344 const Type *DstElTy = DstPTy->getElementType();
8345 const Type *SrcElTy = SrcPTy->getElementType();
8347 // If the address spaces don't match, don't eliminate the bitcast, which is
8348 // required for changing types.
8349 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8352 // If we are casting a malloc or alloca to a pointer to a type of the same
8353 // size, rewrite the allocation instruction to allocate the "right" type.
8354 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8355 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8358 // If the source and destination are pointers, and this cast is equivalent
8359 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8360 // This can enhance SROA and other transforms that want type-safe pointers.
8361 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8362 unsigned NumZeros = 0;
8363 while (SrcElTy != DstElTy &&
8364 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8365 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8366 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8370 // If we found a path from the src to dest, create the getelementptr now.
8371 if (SrcElTy == DstElTy) {
8372 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8373 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8374 ((Instruction*) NULL));
8378 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8379 if (SVI->hasOneUse()) {
8380 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8381 // a bitconvert to a vector with the same # elts.
8382 if (isa<VectorType>(DestTy) &&
8383 cast<VectorType>(DestTy)->getNumElements() ==
8384 SVI->getType()->getNumElements() &&
8385 SVI->getType()->getNumElements() ==
8386 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8388 // If either of the operands is a cast from CI.getType(), then
8389 // evaluating the shuffle in the casted destination's type will allow
8390 // us to eliminate at least one cast.
8391 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8392 Tmp->getOperand(0)->getType() == DestTy) ||
8393 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8394 Tmp->getOperand(0)->getType() == DestTy)) {
8395 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8396 SVI->getOperand(0), DestTy, &CI);
8397 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8398 SVI->getOperand(1), DestTy, &CI);
8399 // Return a new shuffle vector. Use the same element ID's, as we
8400 // know the vector types match #elts.
8401 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8409 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8411 /// %D = select %cond, %C, %A
8413 /// %C = select %cond, %B, 0
8416 /// Assuming that the specified instruction is an operand to the select, return
8417 /// a bitmask indicating which operands of this instruction are foldable if they
8418 /// equal the other incoming value of the select.
8420 static unsigned GetSelectFoldableOperands(Instruction *I) {
8421 switch (I->getOpcode()) {
8422 case Instruction::Add:
8423 case Instruction::Mul:
8424 case Instruction::And:
8425 case Instruction::Or:
8426 case Instruction::Xor:
8427 return 3; // Can fold through either operand.
8428 case Instruction::Sub: // Can only fold on the amount subtracted.
8429 case Instruction::Shl: // Can only fold on the shift amount.
8430 case Instruction::LShr:
8431 case Instruction::AShr:
8434 return 0; // Cannot fold
8438 /// GetSelectFoldableConstant - For the same transformation as the previous
8439 /// function, return the identity constant that goes into the select.
8440 static Constant *GetSelectFoldableConstant(Instruction *I) {
8441 switch (I->getOpcode()) {
8442 default: assert(0 && "This cannot happen!"); abort();
8443 case Instruction::Add:
8444 case Instruction::Sub:
8445 case Instruction::Or:
8446 case Instruction::Xor:
8447 case Instruction::Shl:
8448 case Instruction::LShr:
8449 case Instruction::AShr:
8450 return Constant::getNullValue(I->getType());
8451 case Instruction::And:
8452 return Constant::getAllOnesValue(I->getType());
8453 case Instruction::Mul:
8454 return ConstantInt::get(I->getType(), 1);
8458 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8459 /// have the same opcode and only one use each. Try to simplify this.
8460 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8462 if (TI->getNumOperands() == 1) {
8463 // If this is a non-volatile load or a cast from the same type,
8466 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8469 return 0; // unknown unary op.
8472 // Fold this by inserting a select from the input values.
8473 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8474 FI->getOperand(0), SI.getName()+".v");
8475 InsertNewInstBefore(NewSI, SI);
8476 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8480 // Only handle binary operators here.
8481 if (!isa<BinaryOperator>(TI))
8484 // Figure out if the operations have any operands in common.
8485 Value *MatchOp, *OtherOpT, *OtherOpF;
8487 if (TI->getOperand(0) == FI->getOperand(0)) {
8488 MatchOp = TI->getOperand(0);
8489 OtherOpT = TI->getOperand(1);
8490 OtherOpF = FI->getOperand(1);
8491 MatchIsOpZero = true;
8492 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8493 MatchOp = TI->getOperand(1);
8494 OtherOpT = TI->getOperand(0);
8495 OtherOpF = FI->getOperand(0);
8496 MatchIsOpZero = false;
8497 } else if (!TI->isCommutative()) {
8499 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8500 MatchOp = TI->getOperand(0);
8501 OtherOpT = TI->getOperand(1);
8502 OtherOpF = FI->getOperand(0);
8503 MatchIsOpZero = true;
8504 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8505 MatchOp = TI->getOperand(1);
8506 OtherOpT = TI->getOperand(0);
8507 OtherOpF = FI->getOperand(1);
8508 MatchIsOpZero = true;
8513 // If we reach here, they do have operations in common.
8514 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8515 OtherOpF, SI.getName()+".v");
8516 InsertNewInstBefore(NewSI, SI);
8518 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8520 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8522 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8524 assert(0 && "Shouldn't get here");
8528 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8529 /// ICmpInst as its first operand.
8531 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8533 bool Changed = false;
8534 ICmpInst::Predicate Pred = ICI->getPredicate();
8535 Value *CmpLHS = ICI->getOperand(0);
8536 Value *CmpRHS = ICI->getOperand(1);
8537 Value *TrueVal = SI.getTrueValue();
8538 Value *FalseVal = SI.getFalseValue();
8540 // Check cases where the comparison is with a constant that
8541 // can be adjusted to fit the min/max idiom. We may edit ICI in
8542 // place here, so make sure the select is the only user.
8543 if (ICI->hasOneUse())
8544 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8547 case ICmpInst::ICMP_ULT:
8548 case ICmpInst::ICMP_SLT: {
8549 // X < MIN ? T : F --> F
8550 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8551 return ReplaceInstUsesWith(SI, FalseVal);
8552 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8553 Constant *AdjustedRHS = SubOne(CI);
8554 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8555 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8556 Pred = ICmpInst::getSwappedPredicate(Pred);
8557 CmpRHS = AdjustedRHS;
8558 std::swap(FalseVal, TrueVal);
8559 ICI->setPredicate(Pred);
8560 ICI->setOperand(1, CmpRHS);
8561 SI.setOperand(1, TrueVal);
8562 SI.setOperand(2, FalseVal);
8567 case ICmpInst::ICMP_UGT:
8568 case ICmpInst::ICMP_SGT: {
8569 // X > MAX ? T : F --> F
8570 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8571 return ReplaceInstUsesWith(SI, FalseVal);
8572 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8573 Constant *AdjustedRHS = AddOne(CI);
8574 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8575 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8576 Pred = ICmpInst::getSwappedPredicate(Pred);
8577 CmpRHS = AdjustedRHS;
8578 std::swap(FalseVal, TrueVal);
8579 ICI->setPredicate(Pred);
8580 ICI->setOperand(1, CmpRHS);
8581 SI.setOperand(1, TrueVal);
8582 SI.setOperand(2, FalseVal);
8589 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8590 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8591 CmpInst::Predicate Pred = ICI->getPredicate();
8592 if (match(TrueVal, m_ConstantInt(0)) &&
8593 match(FalseVal, m_ConstantInt(-1)))
8594 Pred = CmpInst::getInversePredicate(Pred);
8595 else if (!match(TrueVal, m_ConstantInt(-1)) ||
8596 !match(FalseVal, m_ConstantInt(0)))
8597 Pred = CmpInst::BAD_ICMP_PREDICATE;
8598 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8599 // If we are just checking for a icmp eq of a single bit and zext'ing it
8600 // to an integer, then shift the bit to the appropriate place and then
8601 // cast to integer to avoid the comparison.
8602 const APInt &Op1CV = CI->getValue();
8604 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8605 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8606 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8607 (Pred == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8608 Value *In = ICI->getOperand(0);
8609 Value *Sh = ConstantInt::get(In->getType(),
8610 In->getType()->getPrimitiveSizeInBits()-1);
8611 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8612 In->getName()+".lobit"),
8614 if (In->getType() != SI.getType())
8615 In = CastInst::CreateIntegerCast(In, SI.getType(),
8616 true/*SExt*/, "tmp", ICI);
8618 if (Pred == ICmpInst::ICMP_SGT)
8619 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8620 In->getName()+".not"), *ICI);
8622 return ReplaceInstUsesWith(SI, In);
8627 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8628 // Transform (X == Y) ? X : Y -> Y
8629 if (Pred == ICmpInst::ICMP_EQ)
8630 return ReplaceInstUsesWith(SI, FalseVal);
8631 // Transform (X != Y) ? X : Y -> X
8632 if (Pred == ICmpInst::ICMP_NE)
8633 return ReplaceInstUsesWith(SI, TrueVal);
8634 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8636 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8637 // Transform (X == Y) ? Y : X -> X
8638 if (Pred == ICmpInst::ICMP_EQ)
8639 return ReplaceInstUsesWith(SI, FalseVal);
8640 // Transform (X != Y) ? Y : X -> Y
8641 if (Pred == ICmpInst::ICMP_NE)
8642 return ReplaceInstUsesWith(SI, TrueVal);
8643 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8646 /// NOTE: if we wanted to, this is where to detect integer ABS
8648 return Changed ? &SI : 0;
8651 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8652 Value *CondVal = SI.getCondition();
8653 Value *TrueVal = SI.getTrueValue();
8654 Value *FalseVal = SI.getFalseValue();
8656 // select true, X, Y -> X
8657 // select false, X, Y -> Y
8658 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8659 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8661 // select C, X, X -> X
8662 if (TrueVal == FalseVal)
8663 return ReplaceInstUsesWith(SI, TrueVal);
8665 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8666 return ReplaceInstUsesWith(SI, FalseVal);
8667 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8668 return ReplaceInstUsesWith(SI, TrueVal);
8669 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8670 if (isa<Constant>(TrueVal))
8671 return ReplaceInstUsesWith(SI, TrueVal);
8673 return ReplaceInstUsesWith(SI, FalseVal);
8676 if (SI.getType() == Type::Int1Ty) {
8677 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8678 if (C->getZExtValue()) {
8679 // Change: A = select B, true, C --> A = or B, C
8680 return BinaryOperator::CreateOr(CondVal, FalseVal);
8682 // Change: A = select B, false, C --> A = and !B, C
8684 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8685 "not."+CondVal->getName()), SI);
8686 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8688 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8689 if (C->getZExtValue() == false) {
8690 // Change: A = select B, C, false --> A = and B, C
8691 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8693 // Change: A = select B, C, true --> A = or !B, C
8695 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8696 "not."+CondVal->getName()), SI);
8697 return BinaryOperator::CreateOr(NotCond, TrueVal);
8701 // select a, b, a -> a&b
8702 // select a, a, b -> a|b
8703 if (CondVal == TrueVal)
8704 return BinaryOperator::CreateOr(CondVal, FalseVal);
8705 else if (CondVal == FalseVal)
8706 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8709 // Selecting between two integer constants?
8710 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8711 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8712 // select C, 1, 0 -> zext C to int
8713 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8714 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8715 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8716 // select C, 0, 1 -> zext !C to int
8718 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8719 "not."+CondVal->getName()), SI);
8720 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8723 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8725 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8727 // (x <s 0) ? -1 : 0 -> ashr x, 31
8728 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8729 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8730 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8731 // The comparison constant and the result are not neccessarily the
8732 // same width. Make an all-ones value by inserting a AShr.
8733 Value *X = IC->getOperand(0);
8734 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8735 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8736 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8738 InsertNewInstBefore(SRA, SI);
8740 // Finally, convert to the type of the select RHS. We figure out
8741 // if this requires a SExt, Trunc or BitCast based on the sizes.
8742 Instruction::CastOps opc = Instruction::BitCast;
8743 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8744 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8745 if (SRASize < SISize)
8746 opc = Instruction::SExt;
8747 else if (SRASize > SISize)
8748 opc = Instruction::Trunc;
8749 return CastInst::Create(opc, SRA, SI.getType());
8754 // If one of the constants is zero (we know they can't both be) and we
8755 // have an icmp instruction with zero, and we have an 'and' with the
8756 // non-constant value, eliminate this whole mess. This corresponds to
8757 // cases like this: ((X & 27) ? 27 : 0)
8758 if (TrueValC->isZero() || FalseValC->isZero())
8759 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8760 cast<Constant>(IC->getOperand(1))->isNullValue())
8761 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8762 if (ICA->getOpcode() == Instruction::And &&
8763 isa<ConstantInt>(ICA->getOperand(1)) &&
8764 (ICA->getOperand(1) == TrueValC ||
8765 ICA->getOperand(1) == FalseValC) &&
8766 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8767 // Okay, now we know that everything is set up, we just don't
8768 // know whether we have a icmp_ne or icmp_eq and whether the
8769 // true or false val is the zero.
8770 bool ShouldNotVal = !TrueValC->isZero();
8771 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8774 V = InsertNewInstBefore(BinaryOperator::Create(
8775 Instruction::Xor, V, ICA->getOperand(1)), SI);
8776 return ReplaceInstUsesWith(SI, V);
8781 // See if we are selecting two values based on a comparison of the two values.
8782 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8783 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8784 // Transform (X == Y) ? X : Y -> Y
8785 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8786 // This is not safe in general for floating point:
8787 // consider X== -0, Y== +0.
8788 // It becomes safe if either operand is a nonzero constant.
8789 ConstantFP *CFPt, *CFPf;
8790 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8791 !CFPt->getValueAPF().isZero()) ||
8792 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8793 !CFPf->getValueAPF().isZero()))
8794 return ReplaceInstUsesWith(SI, FalseVal);
8796 // Transform (X != Y) ? X : Y -> X
8797 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8798 return ReplaceInstUsesWith(SI, TrueVal);
8799 // NOTE: if we wanted to, this is where to detect MIN/MAX
8801 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8802 // Transform (X == Y) ? Y : X -> X
8803 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8804 // This is not safe in general for floating point:
8805 // consider X== -0, Y== +0.
8806 // It becomes safe if either operand is a nonzero constant.
8807 ConstantFP *CFPt, *CFPf;
8808 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8809 !CFPt->getValueAPF().isZero()) ||
8810 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8811 !CFPf->getValueAPF().isZero()))
8812 return ReplaceInstUsesWith(SI, FalseVal);
8814 // Transform (X != Y) ? Y : X -> Y
8815 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8816 return ReplaceInstUsesWith(SI, TrueVal);
8817 // NOTE: if we wanted to, this is where to detect MIN/MAX
8819 // NOTE: if we wanted to, this is where to detect ABS
8822 // See if we are selecting two values based on a comparison of the two values.
8823 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8824 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8827 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8828 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8829 if (TI->hasOneUse() && FI->hasOneUse()) {
8830 Instruction *AddOp = 0, *SubOp = 0;
8832 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8833 if (TI->getOpcode() == FI->getOpcode())
8834 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8837 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8838 // even legal for FP.
8839 if (TI->getOpcode() == Instruction::Sub &&
8840 FI->getOpcode() == Instruction::Add) {
8841 AddOp = FI; SubOp = TI;
8842 } else if (FI->getOpcode() == Instruction::Sub &&
8843 TI->getOpcode() == Instruction::Add) {
8844 AddOp = TI; SubOp = FI;
8848 Value *OtherAddOp = 0;
8849 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8850 OtherAddOp = AddOp->getOperand(1);
8851 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8852 OtherAddOp = AddOp->getOperand(0);
8856 // So at this point we know we have (Y -> OtherAddOp):
8857 // select C, (add X, Y), (sub X, Z)
8858 Value *NegVal; // Compute -Z
8859 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8860 NegVal = ConstantExpr::getNeg(C);
8862 NegVal = InsertNewInstBefore(
8863 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8866 Value *NewTrueOp = OtherAddOp;
8867 Value *NewFalseOp = NegVal;
8869 std::swap(NewTrueOp, NewFalseOp);
8870 Instruction *NewSel =
8871 SelectInst::Create(CondVal, NewTrueOp,
8872 NewFalseOp, SI.getName() + ".p");
8874 NewSel = InsertNewInstBefore(NewSel, SI);
8875 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8880 // See if we can fold the select into one of our operands.
8881 if (SI.getType()->isInteger()) {
8882 // See the comment above GetSelectFoldableOperands for a description of the
8883 // transformation we are doing here.
8884 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8885 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8886 !isa<Constant>(FalseVal))
8887 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8888 unsigned OpToFold = 0;
8889 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8891 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8896 Constant *C = GetSelectFoldableConstant(TVI);
8897 Instruction *NewSel =
8898 SelectInst::Create(SI.getCondition(),
8899 TVI->getOperand(2-OpToFold), C);
8900 InsertNewInstBefore(NewSel, SI);
8901 NewSel->takeName(TVI);
8902 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8903 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8905 assert(0 && "Unknown instruction!!");
8910 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8911 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8912 !isa<Constant>(TrueVal))
8913 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8914 unsigned OpToFold = 0;
8915 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8917 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8922 Constant *C = GetSelectFoldableConstant(FVI);
8923 Instruction *NewSel =
8924 SelectInst::Create(SI.getCondition(), C,
8925 FVI->getOperand(2-OpToFold));
8926 InsertNewInstBefore(NewSel, SI);
8927 NewSel->takeName(FVI);
8928 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8929 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8931 assert(0 && "Unknown instruction!!");
8936 if (BinaryOperator::isNot(CondVal)) {
8937 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8938 SI.setOperand(1, FalseVal);
8939 SI.setOperand(2, TrueVal);
8946 /// EnforceKnownAlignment - If the specified pointer points to an object that
8947 /// we control, modify the object's alignment to PrefAlign. This isn't
8948 /// often possible though. If alignment is important, a more reliable approach
8949 /// is to simply align all global variables and allocation instructions to
8950 /// their preferred alignment from the beginning.
8952 static unsigned EnforceKnownAlignment(Value *V,
8953 unsigned Align, unsigned PrefAlign) {
8955 User *U = dyn_cast<User>(V);
8956 if (!U) return Align;
8958 switch (getOpcode(U)) {
8960 case Instruction::BitCast:
8961 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8962 case Instruction::GetElementPtr: {
8963 // If all indexes are zero, it is just the alignment of the base pointer.
8964 bool AllZeroOperands = true;
8965 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8966 if (!isa<Constant>(*i) ||
8967 !cast<Constant>(*i)->isNullValue()) {
8968 AllZeroOperands = false;
8972 if (AllZeroOperands) {
8973 // Treat this like a bitcast.
8974 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8980 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8981 // If there is a large requested alignment and we can, bump up the alignment
8983 if (!GV->isDeclaration()) {
8984 GV->setAlignment(PrefAlign);
8987 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8988 // If there is a requested alignment and if this is an alloca, round up. We
8989 // don't do this for malloc, because some systems can't respect the request.
8990 if (isa<AllocaInst>(AI)) {
8991 AI->setAlignment(PrefAlign);
8999 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9000 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9001 /// and it is more than the alignment of the ultimate object, see if we can
9002 /// increase the alignment of the ultimate object, making this check succeed.
9003 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9004 unsigned PrefAlign) {
9005 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9006 sizeof(PrefAlign) * CHAR_BIT;
9007 APInt Mask = APInt::getAllOnesValue(BitWidth);
9008 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9009 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9010 unsigned TrailZ = KnownZero.countTrailingOnes();
9011 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9013 if (PrefAlign > Align)
9014 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9016 // We don't need to make any adjustment.
9020 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9021 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9022 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9023 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9024 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9026 if (CopyAlign < MinAlign) {
9027 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9031 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9033 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9034 if (MemOpLength == 0) return 0;
9036 // Source and destination pointer types are always "i8*" for intrinsic. See
9037 // if the size is something we can handle with a single primitive load/store.
9038 // A single load+store correctly handles overlapping memory in the memmove
9040 unsigned Size = MemOpLength->getZExtValue();
9041 if (Size == 0) return MI; // Delete this mem transfer.
9043 if (Size > 8 || (Size&(Size-1)))
9044 return 0; // If not 1/2/4/8 bytes, exit.
9046 // Use an integer load+store unless we can find something better.
9047 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9049 // Memcpy forces the use of i8* for the source and destination. That means
9050 // that if you're using memcpy to move one double around, you'll get a cast
9051 // from double* to i8*. We'd much rather use a double load+store rather than
9052 // an i64 load+store, here because this improves the odds that the source or
9053 // dest address will be promotable. See if we can find a better type than the
9054 // integer datatype.
9055 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9056 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9057 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9058 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9059 // down through these levels if so.
9060 while (!SrcETy->isSingleValueType()) {
9061 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9062 if (STy->getNumElements() == 1)
9063 SrcETy = STy->getElementType(0);
9066 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9067 if (ATy->getNumElements() == 1)
9068 SrcETy = ATy->getElementType();
9075 if (SrcETy->isSingleValueType())
9076 NewPtrTy = PointerType::getUnqual(SrcETy);
9081 // If the memcpy/memmove provides better alignment info than we can
9083 SrcAlign = std::max(SrcAlign, CopyAlign);
9084 DstAlign = std::max(DstAlign, CopyAlign);
9086 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9087 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9088 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9089 InsertNewInstBefore(L, *MI);
9090 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9092 // Set the size of the copy to 0, it will be deleted on the next iteration.
9093 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9097 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9098 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9099 if (MI->getAlignment()->getZExtValue() < Alignment) {
9100 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9104 // Extract the length and alignment and fill if they are constant.
9105 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9106 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9107 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9109 uint64_t Len = LenC->getZExtValue();
9110 Alignment = MI->getAlignment()->getZExtValue();
9112 // If the length is zero, this is a no-op
9113 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9115 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9116 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9117 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9119 Value *Dest = MI->getDest();
9120 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9122 // Alignment 0 is identity for alignment 1 for memset, but not store.
9123 if (Alignment == 0) Alignment = 1;
9125 // Extract the fill value and store.
9126 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9127 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9130 // Set the size of the copy to 0, it will be deleted on the next iteration.
9131 MI->setLength(Constant::getNullValue(LenC->getType()));
9139 /// visitCallInst - CallInst simplification. This mostly only handles folding
9140 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9141 /// the heavy lifting.
9143 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9144 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9145 if (!II) return visitCallSite(&CI);
9147 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9149 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9150 bool Changed = false;
9152 // memmove/cpy/set of zero bytes is a noop.
9153 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9154 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9156 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9157 if (CI->getZExtValue() == 1) {
9158 // Replace the instruction with just byte operations. We would
9159 // transform other cases to loads/stores, but we don't know if
9160 // alignment is sufficient.
9164 // If we have a memmove and the source operation is a constant global,
9165 // then the source and dest pointers can't alias, so we can change this
9166 // into a call to memcpy.
9167 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9168 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9169 if (GVSrc->isConstant()) {
9170 Module *M = CI.getParent()->getParent()->getParent();
9171 Intrinsic::ID MemCpyID;
9172 if (CI.getOperand(3)->getType() == Type::Int32Ty)
9173 MemCpyID = Intrinsic::memcpy_i32;
9175 MemCpyID = Intrinsic::memcpy_i64;
9176 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
9180 // memmove(x,x,size) -> noop.
9181 if (MMI->getSource() == MMI->getDest())
9182 return EraseInstFromFunction(CI);
9185 // If we can determine a pointer alignment that is bigger than currently
9186 // set, update the alignment.
9187 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9188 if (Instruction *I = SimplifyMemTransfer(MI))
9190 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9191 if (Instruction *I = SimplifyMemSet(MSI))
9195 if (Changed) return II;
9198 switch (II->getIntrinsicID()) {
9200 case Intrinsic::bswap:
9201 // bswap(bswap(x)) -> x
9202 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9203 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9204 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9206 case Intrinsic::ppc_altivec_lvx:
9207 case Intrinsic::ppc_altivec_lvxl:
9208 case Intrinsic::x86_sse_loadu_ps:
9209 case Intrinsic::x86_sse2_loadu_pd:
9210 case Intrinsic::x86_sse2_loadu_dq:
9211 // Turn PPC lvx -> load if the pointer is known aligned.
9212 // Turn X86 loadups -> load if the pointer is known aligned.
9213 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9214 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9215 PointerType::getUnqual(II->getType()),
9217 return new LoadInst(Ptr);
9220 case Intrinsic::ppc_altivec_stvx:
9221 case Intrinsic::ppc_altivec_stvxl:
9222 // Turn stvx -> store if the pointer is known aligned.
9223 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9224 const Type *OpPtrTy =
9225 PointerType::getUnqual(II->getOperand(1)->getType());
9226 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9227 return new StoreInst(II->getOperand(1), Ptr);
9230 case Intrinsic::x86_sse_storeu_ps:
9231 case Intrinsic::x86_sse2_storeu_pd:
9232 case Intrinsic::x86_sse2_storeu_dq:
9233 // Turn X86 storeu -> store if the pointer is known aligned.
9234 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9235 const Type *OpPtrTy =
9236 PointerType::getUnqual(II->getOperand(2)->getType());
9237 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9238 return new StoreInst(II->getOperand(2), Ptr);
9242 case Intrinsic::x86_sse_cvttss2si: {
9243 // These intrinsics only demands the 0th element of its input vector. If
9244 // we can simplify the input based on that, do so now.
9246 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9248 II->setOperand(1, V);
9254 case Intrinsic::ppc_altivec_vperm:
9255 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9256 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9257 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9259 // Check that all of the elements are integer constants or undefs.
9260 bool AllEltsOk = true;
9261 for (unsigned i = 0; i != 16; ++i) {
9262 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9263 !isa<UndefValue>(Mask->getOperand(i))) {
9270 // Cast the input vectors to byte vectors.
9271 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9272 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9273 Value *Result = UndefValue::get(Op0->getType());
9275 // Only extract each element once.
9276 Value *ExtractedElts[32];
9277 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9279 for (unsigned i = 0; i != 16; ++i) {
9280 if (isa<UndefValue>(Mask->getOperand(i)))
9282 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9283 Idx &= 31; // Match the hardware behavior.
9285 if (ExtractedElts[Idx] == 0) {
9287 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9288 InsertNewInstBefore(Elt, CI);
9289 ExtractedElts[Idx] = Elt;
9292 // Insert this value into the result vector.
9293 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9295 InsertNewInstBefore(cast<Instruction>(Result), CI);
9297 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9302 case Intrinsic::stackrestore: {
9303 // If the save is right next to the restore, remove the restore. This can
9304 // happen when variable allocas are DCE'd.
9305 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9306 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9307 BasicBlock::iterator BI = SS;
9309 return EraseInstFromFunction(CI);
9313 // Scan down this block to see if there is another stack restore in the
9314 // same block without an intervening call/alloca.
9315 BasicBlock::iterator BI = II;
9316 TerminatorInst *TI = II->getParent()->getTerminator();
9317 bool CannotRemove = false;
9318 for (++BI; &*BI != TI; ++BI) {
9319 if (isa<AllocaInst>(BI)) {
9320 CannotRemove = true;
9323 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9324 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9325 // If there is a stackrestore below this one, remove this one.
9326 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9327 return EraseInstFromFunction(CI);
9328 // Otherwise, ignore the intrinsic.
9330 // If we found a non-intrinsic call, we can't remove the stack
9332 CannotRemove = true;
9338 // If the stack restore is in a return/unwind block and if there are no
9339 // allocas or calls between the restore and the return, nuke the restore.
9340 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9341 return EraseInstFromFunction(CI);
9346 return visitCallSite(II);
9349 // InvokeInst simplification
9351 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9352 return visitCallSite(&II);
9355 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9356 /// passed through the varargs area, we can eliminate the use of the cast.
9357 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9358 const CastInst * const CI,
9359 const TargetData * const TD,
9361 if (!CI->isLosslessCast())
9364 // The size of ByVal arguments is derived from the type, so we
9365 // can't change to a type with a different size. If the size were
9366 // passed explicitly we could avoid this check.
9367 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9371 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9372 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9373 if (!SrcTy->isSized() || !DstTy->isSized())
9375 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9380 // visitCallSite - Improvements for call and invoke instructions.
9382 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9383 bool Changed = false;
9385 // If the callee is a constexpr cast of a function, attempt to move the cast
9386 // to the arguments of the call/invoke.
9387 if (transformConstExprCastCall(CS)) return 0;
9389 Value *Callee = CS.getCalledValue();
9391 if (Function *CalleeF = dyn_cast<Function>(Callee))
9392 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9393 Instruction *OldCall = CS.getInstruction();
9394 // If the call and callee calling conventions don't match, this call must
9395 // be unreachable, as the call is undefined.
9396 new StoreInst(ConstantInt::getTrue(),
9397 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9399 if (!OldCall->use_empty())
9400 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9401 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9402 return EraseInstFromFunction(*OldCall);
9406 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9407 // This instruction is not reachable, just remove it. We insert a store to
9408 // undef so that we know that this code is not reachable, despite the fact
9409 // that we can't modify the CFG here.
9410 new StoreInst(ConstantInt::getTrue(),
9411 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9412 CS.getInstruction());
9414 if (!CS.getInstruction()->use_empty())
9415 CS.getInstruction()->
9416 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9418 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9419 // Don't break the CFG, insert a dummy cond branch.
9420 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9421 ConstantInt::getTrue(), II);
9423 return EraseInstFromFunction(*CS.getInstruction());
9426 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9427 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9428 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9429 return transformCallThroughTrampoline(CS);
9431 const PointerType *PTy = cast<PointerType>(Callee->getType());
9432 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9433 if (FTy->isVarArg()) {
9434 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9435 // See if we can optimize any arguments passed through the varargs area of
9437 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9438 E = CS.arg_end(); I != E; ++I, ++ix) {
9439 CastInst *CI = dyn_cast<CastInst>(*I);
9440 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9441 *I = CI->getOperand(0);
9447 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9448 // Inline asm calls cannot throw - mark them 'nounwind'.
9449 CS.setDoesNotThrow();
9453 return Changed ? CS.getInstruction() : 0;
9456 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9457 // attempt to move the cast to the arguments of the call/invoke.
9459 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9460 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9461 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9462 if (CE->getOpcode() != Instruction::BitCast ||
9463 !isa<Function>(CE->getOperand(0)))
9465 Function *Callee = cast<Function>(CE->getOperand(0));
9466 Instruction *Caller = CS.getInstruction();
9467 const AttrListPtr &CallerPAL = CS.getAttributes();
9469 // Okay, this is a cast from a function to a different type. Unless doing so
9470 // would cause a type conversion of one of our arguments, change this call to
9471 // be a direct call with arguments casted to the appropriate types.
9473 const FunctionType *FT = Callee->getFunctionType();
9474 const Type *OldRetTy = Caller->getType();
9475 const Type *NewRetTy = FT->getReturnType();
9477 if (isa<StructType>(NewRetTy))
9478 return false; // TODO: Handle multiple return values.
9480 // Check to see if we are changing the return type...
9481 if (OldRetTy != NewRetTy) {
9482 if (Callee->isDeclaration() &&
9483 // Conversion is ok if changing from one pointer type to another or from
9484 // a pointer to an integer of the same size.
9485 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9486 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9487 return false; // Cannot transform this return value.
9489 if (!Caller->use_empty() &&
9490 // void -> non-void is handled specially
9491 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9492 return false; // Cannot transform this return value.
9494 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9495 Attributes RAttrs = CallerPAL.getRetAttributes();
9496 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9497 return false; // Attribute not compatible with transformed value.
9500 // If the callsite is an invoke instruction, and the return value is used by
9501 // a PHI node in a successor, we cannot change the return type of the call
9502 // because there is no place to put the cast instruction (without breaking
9503 // the critical edge). Bail out in this case.
9504 if (!Caller->use_empty())
9505 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9506 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9508 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9509 if (PN->getParent() == II->getNormalDest() ||
9510 PN->getParent() == II->getUnwindDest())
9514 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9515 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9517 CallSite::arg_iterator AI = CS.arg_begin();
9518 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9519 const Type *ParamTy = FT->getParamType(i);
9520 const Type *ActTy = (*AI)->getType();
9522 if (!CastInst::isCastable(ActTy, ParamTy))
9523 return false; // Cannot transform this parameter value.
9525 if (CallerPAL.getParamAttributes(i + 1)
9526 & Attribute::typeIncompatible(ParamTy))
9527 return false; // Attribute not compatible with transformed value.
9529 // Converting from one pointer type to another or between a pointer and an
9530 // integer of the same size is safe even if we do not have a body.
9531 bool isConvertible = ActTy == ParamTy ||
9532 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9533 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9534 if (Callee->isDeclaration() && !isConvertible) return false;
9537 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9538 Callee->isDeclaration())
9539 return false; // Do not delete arguments unless we have a function body.
9541 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9542 !CallerPAL.isEmpty())
9543 // In this case we have more arguments than the new function type, but we
9544 // won't be dropping them. Check that these extra arguments have attributes
9545 // that are compatible with being a vararg call argument.
9546 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9547 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9549 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9550 if (PAttrs & Attribute::VarArgsIncompatible)
9554 // Okay, we decided that this is a safe thing to do: go ahead and start
9555 // inserting cast instructions as necessary...
9556 std::vector<Value*> Args;
9557 Args.reserve(NumActualArgs);
9558 SmallVector<AttributeWithIndex, 8> attrVec;
9559 attrVec.reserve(NumCommonArgs);
9561 // Get any return attributes.
9562 Attributes RAttrs = CallerPAL.getRetAttributes();
9564 // If the return value is not being used, the type may not be compatible
9565 // with the existing attributes. Wipe out any problematic attributes.
9566 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9568 // Add the new return attributes.
9570 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9572 AI = CS.arg_begin();
9573 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9574 const Type *ParamTy = FT->getParamType(i);
9575 if ((*AI)->getType() == ParamTy) {
9576 Args.push_back(*AI);
9578 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9579 false, ParamTy, false);
9580 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9581 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9584 // Add any parameter attributes.
9585 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9586 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9589 // If the function takes more arguments than the call was taking, add them
9591 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9592 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9594 // If we are removing arguments to the function, emit an obnoxious warning...
9595 if (FT->getNumParams() < NumActualArgs) {
9596 if (!FT->isVarArg()) {
9597 cerr << "WARNING: While resolving call to function '"
9598 << Callee->getName() << "' arguments were dropped!\n";
9600 // Add all of the arguments in their promoted form to the arg list...
9601 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9602 const Type *PTy = getPromotedType((*AI)->getType());
9603 if (PTy != (*AI)->getType()) {
9604 // Must promote to pass through va_arg area!
9605 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9607 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9608 InsertNewInstBefore(Cast, *Caller);
9609 Args.push_back(Cast);
9611 Args.push_back(*AI);
9614 // Add any parameter attributes.
9615 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9616 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9621 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9622 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9624 if (NewRetTy == Type::VoidTy)
9625 Caller->setName(""); // Void type should not have a name.
9627 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9630 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9631 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9632 Args.begin(), Args.end(),
9633 Caller->getName(), Caller);
9634 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9635 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9637 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9638 Caller->getName(), Caller);
9639 CallInst *CI = cast<CallInst>(Caller);
9640 if (CI->isTailCall())
9641 cast<CallInst>(NC)->setTailCall();
9642 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9643 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9646 // Insert a cast of the return type as necessary.
9648 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9649 if (NV->getType() != Type::VoidTy) {
9650 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9652 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9654 // If this is an invoke instruction, we should insert it after the first
9655 // non-phi, instruction in the normal successor block.
9656 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9657 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9658 InsertNewInstBefore(NC, *I);
9660 // Otherwise, it's a call, just insert cast right after the call instr
9661 InsertNewInstBefore(NC, *Caller);
9663 AddUsersToWorkList(*Caller);
9665 NV = UndefValue::get(Caller->getType());
9669 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9670 Caller->replaceAllUsesWith(NV);
9671 Caller->eraseFromParent();
9672 RemoveFromWorkList(Caller);
9676 // transformCallThroughTrampoline - Turn a call to a function created by the
9677 // init_trampoline intrinsic into a direct call to the underlying function.
9679 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9680 Value *Callee = CS.getCalledValue();
9681 const PointerType *PTy = cast<PointerType>(Callee->getType());
9682 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9683 const AttrListPtr &Attrs = CS.getAttributes();
9685 // If the call already has the 'nest' attribute somewhere then give up -
9686 // otherwise 'nest' would occur twice after splicing in the chain.
9687 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9690 IntrinsicInst *Tramp =
9691 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9693 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9694 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9695 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9697 const AttrListPtr &NestAttrs = NestF->getAttributes();
9698 if (!NestAttrs.isEmpty()) {
9699 unsigned NestIdx = 1;
9700 const Type *NestTy = 0;
9701 Attributes NestAttr = Attribute::None;
9703 // Look for a parameter marked with the 'nest' attribute.
9704 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9705 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9706 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9707 // Record the parameter type and any other attributes.
9709 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9714 Instruction *Caller = CS.getInstruction();
9715 std::vector<Value*> NewArgs;
9716 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9718 SmallVector<AttributeWithIndex, 8> NewAttrs;
9719 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9721 // Insert the nest argument into the call argument list, which may
9722 // mean appending it. Likewise for attributes.
9724 // Add any result attributes.
9725 if (Attributes Attr = Attrs.getRetAttributes())
9726 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9730 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9732 if (Idx == NestIdx) {
9733 // Add the chain argument and attributes.
9734 Value *NestVal = Tramp->getOperand(3);
9735 if (NestVal->getType() != NestTy)
9736 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9737 NewArgs.push_back(NestVal);
9738 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9744 // Add the original argument and attributes.
9745 NewArgs.push_back(*I);
9746 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9748 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9754 // Add any function attributes.
9755 if (Attributes Attr = Attrs.getFnAttributes())
9756 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
9758 // The trampoline may have been bitcast to a bogus type (FTy).
9759 // Handle this by synthesizing a new function type, equal to FTy
9760 // with the chain parameter inserted.
9762 std::vector<const Type*> NewTypes;
9763 NewTypes.reserve(FTy->getNumParams()+1);
9765 // Insert the chain's type into the list of parameter types, which may
9766 // mean appending it.
9769 FunctionType::param_iterator I = FTy->param_begin(),
9770 E = FTy->param_end();
9774 // Add the chain's type.
9775 NewTypes.push_back(NestTy);
9780 // Add the original type.
9781 NewTypes.push_back(*I);
9787 // Replace the trampoline call with a direct call. Let the generic
9788 // code sort out any function type mismatches.
9789 FunctionType *NewFTy =
9790 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9791 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9792 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9793 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
9795 Instruction *NewCaller;
9796 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9797 NewCaller = InvokeInst::Create(NewCallee,
9798 II->getNormalDest(), II->getUnwindDest(),
9799 NewArgs.begin(), NewArgs.end(),
9800 Caller->getName(), Caller);
9801 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9802 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
9804 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9805 Caller->getName(), Caller);
9806 if (cast<CallInst>(Caller)->isTailCall())
9807 cast<CallInst>(NewCaller)->setTailCall();
9808 cast<CallInst>(NewCaller)->
9809 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9810 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
9812 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9813 Caller->replaceAllUsesWith(NewCaller);
9814 Caller->eraseFromParent();
9815 RemoveFromWorkList(Caller);
9820 // Replace the trampoline call with a direct call. Since there is no 'nest'
9821 // parameter, there is no need to adjust the argument list. Let the generic
9822 // code sort out any function type mismatches.
9823 Constant *NewCallee =
9824 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9825 CS.setCalledFunction(NewCallee);
9826 return CS.getInstruction();
9829 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9830 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9831 /// and a single binop.
9832 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9833 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9834 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9835 isa<CmpInst>(FirstInst));
9836 unsigned Opc = FirstInst->getOpcode();
9837 Value *LHSVal = FirstInst->getOperand(0);
9838 Value *RHSVal = FirstInst->getOperand(1);
9840 const Type *LHSType = LHSVal->getType();
9841 const Type *RHSType = RHSVal->getType();
9843 // Scan to see if all operands are the same opcode, all have one use, and all
9844 // kill their operands (i.e. the operands have one use).
9845 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9846 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9847 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9848 // Verify type of the LHS matches so we don't fold cmp's of different
9849 // types or GEP's with different index types.
9850 I->getOperand(0)->getType() != LHSType ||
9851 I->getOperand(1)->getType() != RHSType)
9854 // If they are CmpInst instructions, check their predicates
9855 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9856 if (cast<CmpInst>(I)->getPredicate() !=
9857 cast<CmpInst>(FirstInst)->getPredicate())
9860 // Keep track of which operand needs a phi node.
9861 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9862 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9865 // Otherwise, this is safe to transform, determine if it is profitable.
9867 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9868 // Indexes are often folded into load/store instructions, so we don't want to
9869 // hide them behind a phi.
9870 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9873 Value *InLHS = FirstInst->getOperand(0);
9874 Value *InRHS = FirstInst->getOperand(1);
9875 PHINode *NewLHS = 0, *NewRHS = 0;
9877 NewLHS = PHINode::Create(LHSType,
9878 FirstInst->getOperand(0)->getName() + ".pn");
9879 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9880 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9881 InsertNewInstBefore(NewLHS, PN);
9886 NewRHS = PHINode::Create(RHSType,
9887 FirstInst->getOperand(1)->getName() + ".pn");
9888 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9889 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9890 InsertNewInstBefore(NewRHS, PN);
9894 // Add all operands to the new PHIs.
9895 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9897 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9898 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9901 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9902 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9906 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9907 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9908 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9909 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9912 assert(isa<GetElementPtrInst>(FirstInst));
9913 return GetElementPtrInst::Create(LHSVal, RHSVal);
9917 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9918 /// of the block that defines it. This means that it must be obvious the value
9919 /// of the load is not changed from the point of the load to the end of the
9922 /// Finally, it is safe, but not profitable, to sink a load targetting a
9923 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9925 static bool isSafeToSinkLoad(LoadInst *L) {
9926 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9928 for (++BBI; BBI != E; ++BBI)
9929 if (BBI->mayWriteToMemory())
9932 // Check for non-address taken alloca. If not address-taken already, it isn't
9933 // profitable to do this xform.
9934 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9935 bool isAddressTaken = false;
9936 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9938 if (isa<LoadInst>(UI)) continue;
9939 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9940 // If storing TO the alloca, then the address isn't taken.
9941 if (SI->getOperand(1) == AI) continue;
9943 isAddressTaken = true;
9947 if (!isAddressTaken)
9955 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9956 // operator and they all are only used by the PHI, PHI together their
9957 // inputs, and do the operation once, to the result of the PHI.
9958 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9959 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9961 // Scan the instruction, looking for input operations that can be folded away.
9962 // If all input operands to the phi are the same instruction (e.g. a cast from
9963 // the same type or "+42") we can pull the operation through the PHI, reducing
9964 // code size and simplifying code.
9965 Constant *ConstantOp = 0;
9966 const Type *CastSrcTy = 0;
9967 bool isVolatile = false;
9968 if (isa<CastInst>(FirstInst)) {
9969 CastSrcTy = FirstInst->getOperand(0)->getType();
9970 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9971 // Can fold binop, compare or shift here if the RHS is a constant,
9972 // otherwise call FoldPHIArgBinOpIntoPHI.
9973 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9974 if (ConstantOp == 0)
9975 return FoldPHIArgBinOpIntoPHI(PN);
9976 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9977 isVolatile = LI->isVolatile();
9978 // We can't sink the load if the loaded value could be modified between the
9979 // load and the PHI.
9980 if (LI->getParent() != PN.getIncomingBlock(0) ||
9981 !isSafeToSinkLoad(LI))
9984 // If the PHI is of volatile loads and the load block has multiple
9985 // successors, sinking it would remove a load of the volatile value from
9986 // the path through the other successor.
9988 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9991 } else if (isa<GetElementPtrInst>(FirstInst)) {
9992 if (FirstInst->getNumOperands() == 2)
9993 return FoldPHIArgBinOpIntoPHI(PN);
9994 // Can't handle general GEPs yet.
9997 return 0; // Cannot fold this operation.
10000 // Check to see if all arguments are the same operation.
10001 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10002 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10003 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10004 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10007 if (I->getOperand(0)->getType() != CastSrcTy)
10008 return 0; // Cast operation must match.
10009 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10010 // We can't sink the load if the loaded value could be modified between
10011 // the load and the PHI.
10012 if (LI->isVolatile() != isVolatile ||
10013 LI->getParent() != PN.getIncomingBlock(i) ||
10014 !isSafeToSinkLoad(LI))
10017 // If the PHI is of volatile loads and the load block has multiple
10018 // successors, sinking it would remove a load of the volatile value from
10019 // the path through the other successor.
10021 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10025 } else if (I->getOperand(1) != ConstantOp) {
10030 // Okay, they are all the same operation. Create a new PHI node of the
10031 // correct type, and PHI together all of the LHS's of the instructions.
10032 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10033 PN.getName()+".in");
10034 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10036 Value *InVal = FirstInst->getOperand(0);
10037 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10039 // Add all operands to the new PHI.
10040 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10041 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10042 if (NewInVal != InVal)
10044 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10049 // The new PHI unions all of the same values together. This is really
10050 // common, so we handle it intelligently here for compile-time speed.
10054 InsertNewInstBefore(NewPN, PN);
10058 // Insert and return the new operation.
10059 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10060 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10061 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10062 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10063 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10064 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10065 PhiVal, ConstantOp);
10066 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10068 // If this was a volatile load that we are merging, make sure to loop through
10069 // and mark all the input loads as non-volatile. If we don't do this, we will
10070 // insert a new volatile load and the old ones will not be deletable.
10072 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10073 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10075 return new LoadInst(PhiVal, "", isVolatile);
10078 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10080 static bool DeadPHICycle(PHINode *PN,
10081 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10082 if (PN->use_empty()) return true;
10083 if (!PN->hasOneUse()) return false;
10085 // Remember this node, and if we find the cycle, return.
10086 if (!PotentiallyDeadPHIs.insert(PN))
10089 // Don't scan crazily complex things.
10090 if (PotentiallyDeadPHIs.size() == 16)
10093 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10094 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10099 /// PHIsEqualValue - Return true if this phi node is always equal to
10100 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10101 /// z = some value; x = phi (y, z); y = phi (x, z)
10102 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10103 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10104 // See if we already saw this PHI node.
10105 if (!ValueEqualPHIs.insert(PN))
10108 // Don't scan crazily complex things.
10109 if (ValueEqualPHIs.size() == 16)
10112 // Scan the operands to see if they are either phi nodes or are equal to
10114 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10115 Value *Op = PN->getIncomingValue(i);
10116 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10117 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10119 } else if (Op != NonPhiInVal)
10127 // PHINode simplification
10129 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10130 // If LCSSA is around, don't mess with Phi nodes
10131 if (MustPreserveLCSSA) return 0;
10133 if (Value *V = PN.hasConstantValue())
10134 return ReplaceInstUsesWith(PN, V);
10136 // If all PHI operands are the same operation, pull them through the PHI,
10137 // reducing code size.
10138 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10139 PN.getIncomingValue(0)->hasOneUse())
10140 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10143 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10144 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10145 // PHI)... break the cycle.
10146 if (PN.hasOneUse()) {
10147 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10148 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10149 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10150 PotentiallyDeadPHIs.insert(&PN);
10151 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10152 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10155 // If this phi has a single use, and if that use just computes a value for
10156 // the next iteration of a loop, delete the phi. This occurs with unused
10157 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10158 // common case here is good because the only other things that catch this
10159 // are induction variable analysis (sometimes) and ADCE, which is only run
10161 if (PHIUser->hasOneUse() &&
10162 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10163 PHIUser->use_back() == &PN) {
10164 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10168 // We sometimes end up with phi cycles that non-obviously end up being the
10169 // same value, for example:
10170 // z = some value; x = phi (y, z); y = phi (x, z)
10171 // where the phi nodes don't necessarily need to be in the same block. Do a
10172 // quick check to see if the PHI node only contains a single non-phi value, if
10173 // so, scan to see if the phi cycle is actually equal to that value.
10175 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10176 // Scan for the first non-phi operand.
10177 while (InValNo != NumOperandVals &&
10178 isa<PHINode>(PN.getIncomingValue(InValNo)))
10181 if (InValNo != NumOperandVals) {
10182 Value *NonPhiInVal = PN.getOperand(InValNo);
10184 // Scan the rest of the operands to see if there are any conflicts, if so
10185 // there is no need to recursively scan other phis.
10186 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10187 Value *OpVal = PN.getIncomingValue(InValNo);
10188 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10192 // If we scanned over all operands, then we have one unique value plus
10193 // phi values. Scan PHI nodes to see if they all merge in each other or
10195 if (InValNo == NumOperandVals) {
10196 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10197 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10198 return ReplaceInstUsesWith(PN, NonPhiInVal);
10205 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10206 Instruction *InsertPoint,
10207 InstCombiner *IC) {
10208 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10209 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10210 // We must cast correctly to the pointer type. Ensure that we
10211 // sign extend the integer value if it is smaller as this is
10212 // used for address computation.
10213 Instruction::CastOps opcode =
10214 (VTySize < PtrSize ? Instruction::SExt :
10215 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10216 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10220 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10221 Value *PtrOp = GEP.getOperand(0);
10222 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10223 // If so, eliminate the noop.
10224 if (GEP.getNumOperands() == 1)
10225 return ReplaceInstUsesWith(GEP, PtrOp);
10227 if (isa<UndefValue>(GEP.getOperand(0)))
10228 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10230 bool HasZeroPointerIndex = false;
10231 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10232 HasZeroPointerIndex = C->isNullValue();
10234 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10235 return ReplaceInstUsesWith(GEP, PtrOp);
10237 // Eliminate unneeded casts for indices.
10238 bool MadeChange = false;
10240 gep_type_iterator GTI = gep_type_begin(GEP);
10241 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10242 i != e; ++i, ++GTI) {
10243 if (isa<SequentialType>(*GTI)) {
10244 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10245 if (CI->getOpcode() == Instruction::ZExt ||
10246 CI->getOpcode() == Instruction::SExt) {
10247 const Type *SrcTy = CI->getOperand(0)->getType();
10248 // We can eliminate a cast from i32 to i64 iff the target
10249 // is a 32-bit pointer target.
10250 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10252 *i = CI->getOperand(0);
10256 // If we are using a wider index than needed for this platform, shrink it
10257 // to what we need. If narrower, sign-extend it to what we need.
10258 // If the incoming value needs a cast instruction,
10259 // insert it. This explicit cast can make subsequent optimizations more
10262 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10263 if (Constant *C = dyn_cast<Constant>(Op)) {
10264 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10267 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10272 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10273 if (Constant *C = dyn_cast<Constant>(Op)) {
10274 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10277 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10285 if (MadeChange) return &GEP;
10287 // If this GEP instruction doesn't move the pointer, and if the input operand
10288 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10289 // real input to the dest type.
10290 if (GEP.hasAllZeroIndices()) {
10291 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10292 // If the bitcast is of an allocation, and the allocation will be
10293 // converted to match the type of the cast, don't touch this.
10294 if (isa<AllocationInst>(BCI->getOperand(0))) {
10295 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10296 if (Instruction *I = visitBitCast(*BCI)) {
10299 BCI->getParent()->getInstList().insert(BCI, I);
10300 ReplaceInstUsesWith(*BCI, I);
10305 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10309 // Combine Indices - If the source pointer to this getelementptr instruction
10310 // is a getelementptr instruction, combine the indices of the two
10311 // getelementptr instructions into a single instruction.
10313 SmallVector<Value*, 8> SrcGEPOperands;
10314 if (User *Src = dyn_castGetElementPtr(PtrOp))
10315 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10317 if (!SrcGEPOperands.empty()) {
10318 // Note that if our source is a gep chain itself that we wait for that
10319 // chain to be resolved before we perform this transformation. This
10320 // avoids us creating a TON of code in some cases.
10322 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10323 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10324 return 0; // Wait until our source is folded to completion.
10326 SmallVector<Value*, 8> Indices;
10328 // Find out whether the last index in the source GEP is a sequential idx.
10329 bool EndsWithSequential = false;
10330 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10331 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10332 EndsWithSequential = !isa<StructType>(*I);
10334 // Can we combine the two pointer arithmetics offsets?
10335 if (EndsWithSequential) {
10336 // Replace: gep (gep %P, long B), long A, ...
10337 // With: T = long A+B; gep %P, T, ...
10339 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10340 if (SO1 == Constant::getNullValue(SO1->getType())) {
10342 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10345 // If they aren't the same type, convert both to an integer of the
10346 // target's pointer size.
10347 if (SO1->getType() != GO1->getType()) {
10348 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10349 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10350 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10351 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10353 unsigned PS = TD->getPointerSizeInBits();
10354 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10355 // Convert GO1 to SO1's type.
10356 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10358 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10359 // Convert SO1 to GO1's type.
10360 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10362 const Type *PT = TD->getIntPtrType();
10363 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10364 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10368 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10369 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10371 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10372 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10376 // Recycle the GEP we already have if possible.
10377 if (SrcGEPOperands.size() == 2) {
10378 GEP.setOperand(0, SrcGEPOperands[0]);
10379 GEP.setOperand(1, Sum);
10382 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10383 SrcGEPOperands.end()-1);
10384 Indices.push_back(Sum);
10385 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10387 } else if (isa<Constant>(*GEP.idx_begin()) &&
10388 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10389 SrcGEPOperands.size() != 1) {
10390 // Otherwise we can do the fold if the first index of the GEP is a zero
10391 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10392 SrcGEPOperands.end());
10393 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10396 if (!Indices.empty())
10397 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10398 Indices.end(), GEP.getName());
10400 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10401 // GEP of global variable. If all of the indices for this GEP are
10402 // constants, we can promote this to a constexpr instead of an instruction.
10404 // Scan for nonconstants...
10405 SmallVector<Constant*, 8> Indices;
10406 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10407 for (; I != E && isa<Constant>(*I); ++I)
10408 Indices.push_back(cast<Constant>(*I));
10410 if (I == E) { // If they are all constants...
10411 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10412 &Indices[0],Indices.size());
10414 // Replace all uses of the GEP with the new constexpr...
10415 return ReplaceInstUsesWith(GEP, CE);
10417 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10418 if (!isa<PointerType>(X->getType())) {
10419 // Not interesting. Source pointer must be a cast from pointer.
10420 } else if (HasZeroPointerIndex) {
10421 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10422 // into : GEP [10 x i8]* X, i32 0, ...
10424 // This occurs when the program declares an array extern like "int X[];"
10426 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10427 const PointerType *XTy = cast<PointerType>(X->getType());
10428 if (const ArrayType *XATy =
10429 dyn_cast<ArrayType>(XTy->getElementType()))
10430 if (const ArrayType *CATy =
10431 dyn_cast<ArrayType>(CPTy->getElementType()))
10432 if (CATy->getElementType() == XATy->getElementType()) {
10433 // At this point, we know that the cast source type is a pointer
10434 // to an array of the same type as the destination pointer
10435 // array. Because the array type is never stepped over (there
10436 // is a leading zero) we can fold the cast into this GEP.
10437 GEP.setOperand(0, X);
10440 } else if (GEP.getNumOperands() == 2) {
10441 // Transform things like:
10442 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10443 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10444 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10445 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10446 if (isa<ArrayType>(SrcElTy) &&
10447 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10448 TD->getABITypeSize(ResElTy)) {
10450 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10451 Idx[1] = GEP.getOperand(1);
10452 Value *V = InsertNewInstBefore(
10453 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10454 // V and GEP are both pointer types --> BitCast
10455 return new BitCastInst(V, GEP.getType());
10458 // Transform things like:
10459 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10460 // (where tmp = 8*tmp2) into:
10461 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10463 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10464 uint64_t ArrayEltSize =
10465 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10467 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10468 // allow either a mul, shift, or constant here.
10470 ConstantInt *Scale = 0;
10471 if (ArrayEltSize == 1) {
10472 NewIdx = GEP.getOperand(1);
10473 Scale = ConstantInt::get(NewIdx->getType(), 1);
10474 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10475 NewIdx = ConstantInt::get(CI->getType(), 1);
10477 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10478 if (Inst->getOpcode() == Instruction::Shl &&
10479 isa<ConstantInt>(Inst->getOperand(1))) {
10480 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10481 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10482 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10483 NewIdx = Inst->getOperand(0);
10484 } else if (Inst->getOpcode() == Instruction::Mul &&
10485 isa<ConstantInt>(Inst->getOperand(1))) {
10486 Scale = cast<ConstantInt>(Inst->getOperand(1));
10487 NewIdx = Inst->getOperand(0);
10491 // If the index will be to exactly the right offset with the scale taken
10492 // out, perform the transformation. Note, we don't know whether Scale is
10493 // signed or not. We'll use unsigned version of division/modulo
10494 // operation after making sure Scale doesn't have the sign bit set.
10495 if (Scale && Scale->getSExtValue() >= 0LL &&
10496 Scale->getZExtValue() % ArrayEltSize == 0) {
10497 Scale = ConstantInt::get(Scale->getType(),
10498 Scale->getZExtValue() / ArrayEltSize);
10499 if (Scale->getZExtValue() != 1) {
10500 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10502 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10503 NewIdx = InsertNewInstBefore(Sc, GEP);
10506 // Insert the new GEP instruction.
10508 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10510 Instruction *NewGEP =
10511 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10512 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10513 // The NewGEP must be pointer typed, so must the old one -> BitCast
10514 return new BitCastInst(NewGEP, GEP.getType());
10523 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10524 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10525 if (AI.isArrayAllocation()) { // Check C != 1
10526 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10527 const Type *NewTy =
10528 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10529 AllocationInst *New = 0;
10531 // Create and insert the replacement instruction...
10532 if (isa<MallocInst>(AI))
10533 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10535 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10536 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10539 InsertNewInstBefore(New, AI);
10541 // Scan to the end of the allocation instructions, to skip over a block of
10542 // allocas if possible...
10544 BasicBlock::iterator It = New;
10545 while (isa<AllocationInst>(*It)) ++It;
10547 // Now that I is pointing to the first non-allocation-inst in the block,
10548 // insert our getelementptr instruction...
10550 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10554 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10555 New->getName()+".sub", It);
10557 // Now make everything use the getelementptr instead of the original
10559 return ReplaceInstUsesWith(AI, V);
10560 } else if (isa<UndefValue>(AI.getArraySize())) {
10561 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10565 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10566 // Note that we only do this for alloca's, because malloc should allocate and
10567 // return a unique pointer, even for a zero byte allocation.
10568 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10569 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10570 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10575 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10576 Value *Op = FI.getOperand(0);
10578 // free undef -> unreachable.
10579 if (isa<UndefValue>(Op)) {
10580 // Insert a new store to null because we cannot modify the CFG here.
10581 new StoreInst(ConstantInt::getTrue(),
10582 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10583 return EraseInstFromFunction(FI);
10586 // If we have 'free null' delete the instruction. This can happen in stl code
10587 // when lots of inlining happens.
10588 if (isa<ConstantPointerNull>(Op))
10589 return EraseInstFromFunction(FI);
10591 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10592 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10593 FI.setOperand(0, CI->getOperand(0));
10597 // Change free (gep X, 0,0,0,0) into free(X)
10598 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10599 if (GEPI->hasAllZeroIndices()) {
10600 AddToWorkList(GEPI);
10601 FI.setOperand(0, GEPI->getOperand(0));
10606 // Change free(malloc) into nothing, if the malloc has a single use.
10607 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10608 if (MI->hasOneUse()) {
10609 EraseInstFromFunction(FI);
10610 return EraseInstFromFunction(*MI);
10617 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10618 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10619 const TargetData *TD) {
10620 User *CI = cast<User>(LI.getOperand(0));
10621 Value *CastOp = CI->getOperand(0);
10623 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10624 // Instead of loading constant c string, use corresponding integer value
10625 // directly if string length is small enough.
10627 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10628 unsigned len = Str.length();
10629 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10630 unsigned numBits = Ty->getPrimitiveSizeInBits();
10631 // Replace LI with immediate integer store.
10632 if ((numBits >> 3) == len + 1) {
10633 APInt StrVal(numBits, 0);
10634 APInt SingleChar(numBits, 0);
10635 if (TD->isLittleEndian()) {
10636 for (signed i = len-1; i >= 0; i--) {
10637 SingleChar = (uint64_t) Str[i];
10638 StrVal = (StrVal << 8) | SingleChar;
10641 for (unsigned i = 0; i < len; i++) {
10642 SingleChar = (uint64_t) Str[i];
10643 StrVal = (StrVal << 8) | SingleChar;
10645 // Append NULL at the end.
10647 StrVal = (StrVal << 8) | SingleChar;
10649 Value *NL = ConstantInt::get(StrVal);
10650 return IC.ReplaceInstUsesWith(LI, NL);
10655 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10656 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10657 const Type *SrcPTy = SrcTy->getElementType();
10659 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10660 isa<VectorType>(DestPTy)) {
10661 // If the source is an array, the code below will not succeed. Check to
10662 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10664 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10665 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10666 if (ASrcTy->getNumElements() != 0) {
10668 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10669 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10670 SrcTy = cast<PointerType>(CastOp->getType());
10671 SrcPTy = SrcTy->getElementType();
10674 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10675 isa<VectorType>(SrcPTy)) &&
10676 // Do not allow turning this into a load of an integer, which is then
10677 // casted to a pointer, this pessimizes pointer analysis a lot.
10678 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10679 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10680 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10682 // Okay, we are casting from one integer or pointer type to another of
10683 // the same size. Instead of casting the pointer before the load, cast
10684 // the result of the loaded value.
10685 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10687 LI.isVolatile()),LI);
10688 // Now cast the result of the load.
10689 return new BitCastInst(NewLoad, LI.getType());
10696 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10697 /// from this value cannot trap. If it is not obviously safe to load from the
10698 /// specified pointer, we do a quick local scan of the basic block containing
10699 /// ScanFrom, to determine if the address is already accessed.
10700 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10701 // If it is an alloca it is always safe to load from.
10702 if (isa<AllocaInst>(V)) return true;
10704 // If it is a global variable it is mostly safe to load from.
10705 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10706 // Don't try to evaluate aliases. External weak GV can be null.
10707 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10709 // Otherwise, be a little bit agressive by scanning the local block where we
10710 // want to check to see if the pointer is already being loaded or stored
10711 // from/to. If so, the previous load or store would have already trapped,
10712 // so there is no harm doing an extra load (also, CSE will later eliminate
10713 // the load entirely).
10714 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10719 // If we see a free or a call (which might do a free) the pointer could be
10721 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10724 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10725 if (LI->getOperand(0) == V) return true;
10726 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10727 if (SI->getOperand(1) == V) return true;
10734 /// equivalentAddressValues - Test if A and B will obviously have the same
10735 /// value. This includes recognizing that %t0 and %t1 will have the same
10736 /// value in code like this:
10737 /// %t0 = getelementptr @a, 0, 3
10738 /// store i32 0, i32* %t0
10739 /// %t1 = getelementptr @a, 0, 3
10740 /// %t2 = load i32* %t1
10742 static bool equivalentAddressValues(Value *A, Value *B) {
10743 // Test if the values are trivially equivalent.
10744 if (A == B) return true;
10746 // Test if the values come form identical arithmetic instructions.
10747 if (isa<BinaryOperator>(A) ||
10748 isa<CastInst>(A) ||
10750 isa<GetElementPtrInst>(A))
10751 if (Instruction *BI = dyn_cast<Instruction>(B))
10752 if (cast<Instruction>(A)->isIdenticalTo(BI))
10755 // Otherwise they may not be equivalent.
10759 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10760 Value *Op = LI.getOperand(0);
10762 // Attempt to improve the alignment.
10763 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10765 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10766 LI.getAlignment()))
10767 LI.setAlignment(KnownAlign);
10769 // load (cast X) --> cast (load X) iff safe
10770 if (isa<CastInst>(Op))
10771 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10774 // None of the following transforms are legal for volatile loads.
10775 if (LI.isVolatile()) return 0;
10777 // Do really simple store-to-load forwarding and load CSE, to catch cases
10778 // where there are several consequtive memory accesses to the same location,
10779 // separated by a few arithmetic operations.
10780 BasicBlock::iterator BBI = &LI;
10781 for (unsigned ScanInsts = 6; BBI != LI.getParent()->begin() && ScanInsts;
10785 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10786 if (equivalentAddressValues(SI->getOperand(1), LI.getOperand(0)))
10787 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10788 } else if (LoadInst *LIB = dyn_cast<LoadInst>(BBI)) {
10789 if (equivalentAddressValues(LIB->getOperand(0), LI.getOperand(0)))
10790 return ReplaceInstUsesWith(LI, LIB);
10793 // Don't skip over things that can modify memory.
10794 if (BBI->mayWriteToMemory())
10798 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10799 const Value *GEPI0 = GEPI->getOperand(0);
10800 // TODO: Consider a target hook for valid address spaces for this xform.
10801 if (isa<ConstantPointerNull>(GEPI0) &&
10802 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10803 // Insert a new store to null instruction before the load to indicate
10804 // that this code is not reachable. We do this instead of inserting
10805 // an unreachable instruction directly because we cannot modify the
10807 new StoreInst(UndefValue::get(LI.getType()),
10808 Constant::getNullValue(Op->getType()), &LI);
10809 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10813 if (Constant *C = dyn_cast<Constant>(Op)) {
10814 // load null/undef -> undef
10815 // TODO: Consider a target hook for valid address spaces for this xform.
10816 if (isa<UndefValue>(C) || (C->isNullValue() &&
10817 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10818 // Insert a new store to null instruction before the load to indicate that
10819 // this code is not reachable. We do this instead of inserting an
10820 // unreachable instruction directly because we cannot modify the CFG.
10821 new StoreInst(UndefValue::get(LI.getType()),
10822 Constant::getNullValue(Op->getType()), &LI);
10823 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10826 // Instcombine load (constant global) into the value loaded.
10827 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10828 if (GV->isConstant() && !GV->isDeclaration())
10829 return ReplaceInstUsesWith(LI, GV->getInitializer());
10831 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10832 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10833 if (CE->getOpcode() == Instruction::GetElementPtr) {
10834 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10835 if (GV->isConstant() && !GV->isDeclaration())
10837 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10838 return ReplaceInstUsesWith(LI, V);
10839 if (CE->getOperand(0)->isNullValue()) {
10840 // Insert a new store to null instruction before the load to indicate
10841 // that this code is not reachable. We do this instead of inserting
10842 // an unreachable instruction directly because we cannot modify the
10844 new StoreInst(UndefValue::get(LI.getType()),
10845 Constant::getNullValue(Op->getType()), &LI);
10846 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10849 } else if (CE->isCast()) {
10850 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10856 // If this load comes from anywhere in a constant global, and if the global
10857 // is all undef or zero, we know what it loads.
10858 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
10859 if (GV->isConstant() && GV->hasInitializer()) {
10860 if (GV->getInitializer()->isNullValue())
10861 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10862 else if (isa<UndefValue>(GV->getInitializer()))
10863 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10867 if (Op->hasOneUse()) {
10868 // Change select and PHI nodes to select values instead of addresses: this
10869 // helps alias analysis out a lot, allows many others simplifications, and
10870 // exposes redundancy in the code.
10872 // Note that we cannot do the transformation unless we know that the
10873 // introduced loads cannot trap! Something like this is valid as long as
10874 // the condition is always false: load (select bool %C, int* null, int* %G),
10875 // but it would not be valid if we transformed it to load from null
10876 // unconditionally.
10878 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10879 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10880 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10881 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10882 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10883 SI->getOperand(1)->getName()+".val"), LI);
10884 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10885 SI->getOperand(2)->getName()+".val"), LI);
10886 return SelectInst::Create(SI->getCondition(), V1, V2);
10889 // load (select (cond, null, P)) -> load P
10890 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10891 if (C->isNullValue()) {
10892 LI.setOperand(0, SI->getOperand(2));
10896 // load (select (cond, P, null)) -> load P
10897 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10898 if (C->isNullValue()) {
10899 LI.setOperand(0, SI->getOperand(1));
10907 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10909 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10910 User *CI = cast<User>(SI.getOperand(1));
10911 Value *CastOp = CI->getOperand(0);
10913 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10914 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10915 const Type *SrcPTy = SrcTy->getElementType();
10917 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10918 // If the source is an array, the code below will not succeed. Check to
10919 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10921 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10922 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10923 if (ASrcTy->getNumElements() != 0) {
10925 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10926 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10927 SrcTy = cast<PointerType>(CastOp->getType());
10928 SrcPTy = SrcTy->getElementType();
10931 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10932 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10933 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10935 // Okay, we are casting from one integer or pointer type to another of
10936 // the same size. Instead of casting the pointer before
10937 // the store, cast the value to be stored.
10939 Value *SIOp0 = SI.getOperand(0);
10940 Instruction::CastOps opcode = Instruction::BitCast;
10941 const Type* CastSrcTy = SIOp0->getType();
10942 const Type* CastDstTy = SrcPTy;
10943 if (isa<PointerType>(CastDstTy)) {
10944 if (CastSrcTy->isInteger())
10945 opcode = Instruction::IntToPtr;
10946 } else if (isa<IntegerType>(CastDstTy)) {
10947 if (isa<PointerType>(SIOp0->getType()))
10948 opcode = Instruction::PtrToInt;
10950 if (Constant *C = dyn_cast<Constant>(SIOp0))
10951 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10953 NewCast = IC.InsertNewInstBefore(
10954 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10956 return new StoreInst(NewCast, CastOp);
10963 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10964 Value *Val = SI.getOperand(0);
10965 Value *Ptr = SI.getOperand(1);
10967 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10968 EraseInstFromFunction(SI);
10973 // If the RHS is an alloca with a single use, zapify the store, making the
10975 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10976 if (isa<AllocaInst>(Ptr)) {
10977 EraseInstFromFunction(SI);
10982 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10983 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10984 GEP->getOperand(0)->hasOneUse()) {
10985 EraseInstFromFunction(SI);
10991 // Attempt to improve the alignment.
10992 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10994 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10995 SI.getAlignment()))
10996 SI.setAlignment(KnownAlign);
10998 // Do really simple DSE, to catch cases where there are several consequtive
10999 // stores to the same location, separated by a few arithmetic operations. This
11000 // situation often occurs with bitfield accesses.
11001 BasicBlock::iterator BBI = &SI;
11002 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11006 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11007 // Prev store isn't volatile, and stores to the same location?
11008 if (!PrevSI->isVolatile() && equivalentAddressValues(PrevSI->getOperand(1),
11009 SI.getOperand(1))) {
11012 EraseInstFromFunction(*PrevSI);
11018 // If this is a load, we have to stop. However, if the loaded value is from
11019 // the pointer we're loading and is producing the pointer we're storing,
11020 // then *this* store is dead (X = load P; store X -> P).
11021 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11022 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11023 !SI.isVolatile()) {
11024 EraseInstFromFunction(SI);
11028 // Otherwise, this is a load from some other location. Stores before it
11029 // may not be dead.
11033 // Don't skip over loads or things that can modify memory.
11034 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11039 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11041 // store X, null -> turns into 'unreachable' in SimplifyCFG
11042 if (isa<ConstantPointerNull>(Ptr)) {
11043 if (!isa<UndefValue>(Val)) {
11044 SI.setOperand(0, UndefValue::get(Val->getType()));
11045 if (Instruction *U = dyn_cast<Instruction>(Val))
11046 AddToWorkList(U); // Dropped a use.
11049 return 0; // Do not modify these!
11052 // store undef, Ptr -> noop
11053 if (isa<UndefValue>(Val)) {
11054 EraseInstFromFunction(SI);
11059 // If the pointer destination is a cast, see if we can fold the cast into the
11061 if (isa<CastInst>(Ptr))
11062 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11064 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11066 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11070 // If this store is the last instruction in the basic block, and if the block
11071 // ends with an unconditional branch, try to move it to the successor block.
11073 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11074 if (BI->isUnconditional())
11075 if (SimplifyStoreAtEndOfBlock(SI))
11076 return 0; // xform done!
11081 /// SimplifyStoreAtEndOfBlock - Turn things like:
11082 /// if () { *P = v1; } else { *P = v2 }
11083 /// into a phi node with a store in the successor.
11085 /// Simplify things like:
11086 /// *P = v1; if () { *P = v2; }
11087 /// into a phi node with a store in the successor.
11089 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11090 BasicBlock *StoreBB = SI.getParent();
11092 // Check to see if the successor block has exactly two incoming edges. If
11093 // so, see if the other predecessor contains a store to the same location.
11094 // if so, insert a PHI node (if needed) and move the stores down.
11095 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11097 // Determine whether Dest has exactly two predecessors and, if so, compute
11098 // the other predecessor.
11099 pred_iterator PI = pred_begin(DestBB);
11100 BasicBlock *OtherBB = 0;
11101 if (*PI != StoreBB)
11104 if (PI == pred_end(DestBB))
11107 if (*PI != StoreBB) {
11112 if (++PI != pred_end(DestBB))
11115 // Bail out if all the relevant blocks aren't distinct (this can happen,
11116 // for example, if SI is in an infinite loop)
11117 if (StoreBB == DestBB || OtherBB == DestBB)
11120 // Verify that the other block ends in a branch and is not otherwise empty.
11121 BasicBlock::iterator BBI = OtherBB->getTerminator();
11122 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11123 if (!OtherBr || BBI == OtherBB->begin())
11126 // If the other block ends in an unconditional branch, check for the 'if then
11127 // else' case. there is an instruction before the branch.
11128 StoreInst *OtherStore = 0;
11129 if (OtherBr->isUnconditional()) {
11130 // If this isn't a store, or isn't a store to the same location, bail out.
11132 OtherStore = dyn_cast<StoreInst>(BBI);
11133 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11136 // Otherwise, the other block ended with a conditional branch. If one of the
11137 // destinations is StoreBB, then we have the if/then case.
11138 if (OtherBr->getSuccessor(0) != StoreBB &&
11139 OtherBr->getSuccessor(1) != StoreBB)
11142 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11143 // if/then triangle. See if there is a store to the same ptr as SI that
11144 // lives in OtherBB.
11146 // Check to see if we find the matching store.
11147 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11148 if (OtherStore->getOperand(1) != SI.getOperand(1))
11152 // If we find something that may be using or overwriting the stored
11153 // value, or if we run out of instructions, we can't do the xform.
11154 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11155 BBI == OtherBB->begin())
11159 // In order to eliminate the store in OtherBr, we have to
11160 // make sure nothing reads or overwrites the stored value in
11162 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11163 // FIXME: This should really be AA driven.
11164 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11169 // Insert a PHI node now if we need it.
11170 Value *MergedVal = OtherStore->getOperand(0);
11171 if (MergedVal != SI.getOperand(0)) {
11172 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11173 PN->reserveOperandSpace(2);
11174 PN->addIncoming(SI.getOperand(0), SI.getParent());
11175 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11176 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11179 // Advance to a place where it is safe to insert the new store and
11181 BBI = DestBB->getFirstNonPHI();
11182 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11183 OtherStore->isVolatile()), *BBI);
11185 // Nuke the old stores.
11186 EraseInstFromFunction(SI);
11187 EraseInstFromFunction(*OtherStore);
11193 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11194 // Change br (not X), label True, label False to: br X, label False, True
11196 BasicBlock *TrueDest;
11197 BasicBlock *FalseDest;
11198 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11199 !isa<Constant>(X)) {
11200 // Swap Destinations and condition...
11201 BI.setCondition(X);
11202 BI.setSuccessor(0, FalseDest);
11203 BI.setSuccessor(1, TrueDest);
11207 // Cannonicalize fcmp_one -> fcmp_oeq
11208 FCmpInst::Predicate FPred; Value *Y;
11209 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11210 TrueDest, FalseDest)))
11211 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11212 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11213 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11214 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11215 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11216 NewSCC->takeName(I);
11217 // Swap Destinations and condition...
11218 BI.setCondition(NewSCC);
11219 BI.setSuccessor(0, FalseDest);
11220 BI.setSuccessor(1, TrueDest);
11221 RemoveFromWorkList(I);
11222 I->eraseFromParent();
11223 AddToWorkList(NewSCC);
11227 // Cannonicalize icmp_ne -> icmp_eq
11228 ICmpInst::Predicate IPred;
11229 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11230 TrueDest, FalseDest)))
11231 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11232 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11233 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11234 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11235 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11236 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11237 NewSCC->takeName(I);
11238 // Swap Destinations and condition...
11239 BI.setCondition(NewSCC);
11240 BI.setSuccessor(0, FalseDest);
11241 BI.setSuccessor(1, TrueDest);
11242 RemoveFromWorkList(I);
11243 I->eraseFromParent();;
11244 AddToWorkList(NewSCC);
11251 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11252 Value *Cond = SI.getCondition();
11253 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11254 if (I->getOpcode() == Instruction::Add)
11255 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11256 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11257 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11258 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11260 SI.setOperand(0, I->getOperand(0));
11268 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11269 Value *Agg = EV.getAggregateOperand();
11271 if (!EV.hasIndices())
11272 return ReplaceInstUsesWith(EV, Agg);
11274 if (Constant *C = dyn_cast<Constant>(Agg)) {
11275 if (isa<UndefValue>(C))
11276 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11278 if (isa<ConstantAggregateZero>(C))
11279 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11281 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11282 // Extract the element indexed by the first index out of the constant
11283 Value *V = C->getOperand(*EV.idx_begin());
11284 if (EV.getNumIndices() > 1)
11285 // Extract the remaining indices out of the constant indexed by the
11287 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11289 return ReplaceInstUsesWith(EV, V);
11291 return 0; // Can't handle other constants
11293 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11294 // We're extracting from an insertvalue instruction, compare the indices
11295 const unsigned *exti, *exte, *insi, *inse;
11296 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11297 exte = EV.idx_end(), inse = IV->idx_end();
11298 exti != exte && insi != inse;
11300 if (*insi != *exti)
11301 // The insert and extract both reference distinctly different elements.
11302 // This means the extract is not influenced by the insert, and we can
11303 // replace the aggregate operand of the extract with the aggregate
11304 // operand of the insert. i.e., replace
11305 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11306 // %E = extractvalue { i32, { i32 } } %I, 0
11308 // %E = extractvalue { i32, { i32 } } %A, 0
11309 return ExtractValueInst::Create(IV->getAggregateOperand(),
11310 EV.idx_begin(), EV.idx_end());
11312 if (exti == exte && insi == inse)
11313 // Both iterators are at the end: Index lists are identical. Replace
11314 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11315 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11317 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11318 if (exti == exte) {
11319 // The extract list is a prefix of the insert list. i.e. replace
11320 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11321 // %E = extractvalue { i32, { i32 } } %I, 1
11323 // %X = extractvalue { i32, { i32 } } %A, 1
11324 // %E = insertvalue { i32 } %X, i32 42, 0
11325 // by switching the order of the insert and extract (though the
11326 // insertvalue should be left in, since it may have other uses).
11327 Value *NewEV = InsertNewInstBefore(
11328 ExtractValueInst::Create(IV->getAggregateOperand(),
11329 EV.idx_begin(), EV.idx_end()),
11331 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11335 // The insert list is a prefix of the extract list
11336 // We can simply remove the common indices from the extract and make it
11337 // operate on the inserted value instead of the insertvalue result.
11339 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11340 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11342 // %E extractvalue { i32 } { i32 42 }, 0
11343 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11346 // Can't simplify extracts from other values. Note that nested extracts are
11347 // already simplified implicitely by the above (extract ( extract (insert) )
11348 // will be translated into extract ( insert ( extract ) ) first and then just
11349 // the value inserted, if appropriate).
11353 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11354 /// is to leave as a vector operation.
11355 static bool CheapToScalarize(Value *V, bool isConstant) {
11356 if (isa<ConstantAggregateZero>(V))
11358 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11359 if (isConstant) return true;
11360 // If all elts are the same, we can extract.
11361 Constant *Op0 = C->getOperand(0);
11362 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11363 if (C->getOperand(i) != Op0)
11367 Instruction *I = dyn_cast<Instruction>(V);
11368 if (!I) return false;
11370 // Insert element gets simplified to the inserted element or is deleted if
11371 // this is constant idx extract element and its a constant idx insertelt.
11372 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11373 isa<ConstantInt>(I->getOperand(2)))
11375 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11377 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11378 if (BO->hasOneUse() &&
11379 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11380 CheapToScalarize(BO->getOperand(1), isConstant)))
11382 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11383 if (CI->hasOneUse() &&
11384 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11385 CheapToScalarize(CI->getOperand(1), isConstant)))
11391 /// Read and decode a shufflevector mask.
11393 /// It turns undef elements into values that are larger than the number of
11394 /// elements in the input.
11395 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11396 unsigned NElts = SVI->getType()->getNumElements();
11397 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11398 return std::vector<unsigned>(NElts, 0);
11399 if (isa<UndefValue>(SVI->getOperand(2)))
11400 return std::vector<unsigned>(NElts, 2*NElts);
11402 std::vector<unsigned> Result;
11403 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11404 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11405 if (isa<UndefValue>(*i))
11406 Result.push_back(NElts*2); // undef -> 8
11408 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11412 /// FindScalarElement - Given a vector and an element number, see if the scalar
11413 /// value is already around as a register, for example if it were inserted then
11414 /// extracted from the vector.
11415 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11416 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11417 const VectorType *PTy = cast<VectorType>(V->getType());
11418 unsigned Width = PTy->getNumElements();
11419 if (EltNo >= Width) // Out of range access.
11420 return UndefValue::get(PTy->getElementType());
11422 if (isa<UndefValue>(V))
11423 return UndefValue::get(PTy->getElementType());
11424 else if (isa<ConstantAggregateZero>(V))
11425 return Constant::getNullValue(PTy->getElementType());
11426 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11427 return CP->getOperand(EltNo);
11428 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11429 // If this is an insert to a variable element, we don't know what it is.
11430 if (!isa<ConstantInt>(III->getOperand(2)))
11432 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11434 // If this is an insert to the element we are looking for, return the
11436 if (EltNo == IIElt)
11437 return III->getOperand(1);
11439 // Otherwise, the insertelement doesn't modify the value, recurse on its
11441 return FindScalarElement(III->getOperand(0), EltNo);
11442 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11443 unsigned LHSWidth =
11444 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11445 unsigned InEl = getShuffleMask(SVI)[EltNo];
11446 if (InEl < LHSWidth)
11447 return FindScalarElement(SVI->getOperand(0), InEl);
11448 else if (InEl < LHSWidth*2)
11449 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11451 return UndefValue::get(PTy->getElementType());
11454 // Otherwise, we don't know.
11458 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11459 // If vector val is undef, replace extract with scalar undef.
11460 if (isa<UndefValue>(EI.getOperand(0)))
11461 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11463 // If vector val is constant 0, replace extract with scalar 0.
11464 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11465 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11467 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11468 // If vector val is constant with all elements the same, replace EI with
11469 // that element. When the elements are not identical, we cannot replace yet
11470 // (we do that below, but only when the index is constant).
11471 Constant *op0 = C->getOperand(0);
11472 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11473 if (C->getOperand(i) != op0) {
11478 return ReplaceInstUsesWith(EI, op0);
11481 // If extracting a specified index from the vector, see if we can recursively
11482 // find a previously computed scalar that was inserted into the vector.
11483 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11484 unsigned IndexVal = IdxC->getZExtValue();
11485 unsigned VectorWidth =
11486 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11488 // If this is extracting an invalid index, turn this into undef, to avoid
11489 // crashing the code below.
11490 if (IndexVal >= VectorWidth)
11491 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11493 // This instruction only demands the single element from the input vector.
11494 // If the input vector has a single use, simplify it based on this use
11496 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11497 uint64_t UndefElts;
11498 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11501 EI.setOperand(0, V);
11506 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11507 return ReplaceInstUsesWith(EI, Elt);
11509 // If the this extractelement is directly using a bitcast from a vector of
11510 // the same number of elements, see if we can find the source element from
11511 // it. In this case, we will end up needing to bitcast the scalars.
11512 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11513 if (const VectorType *VT =
11514 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11515 if (VT->getNumElements() == VectorWidth)
11516 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11517 return new BitCastInst(Elt, EI.getType());
11521 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11522 if (I->hasOneUse()) {
11523 // Push extractelement into predecessor operation if legal and
11524 // profitable to do so
11525 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11526 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11527 if (CheapToScalarize(BO, isConstantElt)) {
11528 ExtractElementInst *newEI0 =
11529 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11530 EI.getName()+".lhs");
11531 ExtractElementInst *newEI1 =
11532 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11533 EI.getName()+".rhs");
11534 InsertNewInstBefore(newEI0, EI);
11535 InsertNewInstBefore(newEI1, EI);
11536 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11538 } else if (isa<LoadInst>(I)) {
11540 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11541 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11542 PointerType::get(EI.getType(), AS),EI);
11543 GetElementPtrInst *GEP =
11544 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11545 InsertNewInstBefore(GEP, EI);
11546 return new LoadInst(GEP);
11549 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11550 // Extracting the inserted element?
11551 if (IE->getOperand(2) == EI.getOperand(1))
11552 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11553 // If the inserted and extracted elements are constants, they must not
11554 // be the same value, extract from the pre-inserted value instead.
11555 if (isa<Constant>(IE->getOperand(2)) &&
11556 isa<Constant>(EI.getOperand(1))) {
11557 AddUsesToWorkList(EI);
11558 EI.setOperand(0, IE->getOperand(0));
11561 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11562 // If this is extracting an element from a shufflevector, figure out where
11563 // it came from and extract from the appropriate input element instead.
11564 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11565 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11567 unsigned LHSWidth =
11568 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11570 if (SrcIdx < LHSWidth)
11571 Src = SVI->getOperand(0);
11572 else if (SrcIdx < LHSWidth*2) {
11573 SrcIdx -= LHSWidth;
11574 Src = SVI->getOperand(1);
11576 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11578 return new ExtractElementInst(Src, SrcIdx);
11585 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11586 /// elements from either LHS or RHS, return the shuffle mask and true.
11587 /// Otherwise, return false.
11588 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11589 std::vector<Constant*> &Mask) {
11590 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11591 "Invalid CollectSingleShuffleElements");
11592 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11594 if (isa<UndefValue>(V)) {
11595 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11597 } else if (V == LHS) {
11598 for (unsigned i = 0; i != NumElts; ++i)
11599 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11601 } else if (V == RHS) {
11602 for (unsigned i = 0; i != NumElts; ++i)
11603 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11605 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11606 // If this is an insert of an extract from some other vector, include it.
11607 Value *VecOp = IEI->getOperand(0);
11608 Value *ScalarOp = IEI->getOperand(1);
11609 Value *IdxOp = IEI->getOperand(2);
11611 if (!isa<ConstantInt>(IdxOp))
11613 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11615 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11616 // Okay, we can handle this if the vector we are insertinting into is
11617 // transitively ok.
11618 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11619 // If so, update the mask to reflect the inserted undef.
11620 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11623 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11624 if (isa<ConstantInt>(EI->getOperand(1)) &&
11625 EI->getOperand(0)->getType() == V->getType()) {
11626 unsigned ExtractedIdx =
11627 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11629 // This must be extracting from either LHS or RHS.
11630 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11631 // Okay, we can handle this if the vector we are insertinting into is
11632 // transitively ok.
11633 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11634 // If so, update the mask to reflect the inserted value.
11635 if (EI->getOperand(0) == LHS) {
11636 Mask[InsertedIdx % NumElts] =
11637 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11639 assert(EI->getOperand(0) == RHS);
11640 Mask[InsertedIdx % NumElts] =
11641 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11650 // TODO: Handle shufflevector here!
11655 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11656 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11657 /// that computes V and the LHS value of the shuffle.
11658 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11660 assert(isa<VectorType>(V->getType()) &&
11661 (RHS == 0 || V->getType() == RHS->getType()) &&
11662 "Invalid shuffle!");
11663 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11665 if (isa<UndefValue>(V)) {
11666 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11668 } else if (isa<ConstantAggregateZero>(V)) {
11669 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11671 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11672 // If this is an insert of an extract from some other vector, include it.
11673 Value *VecOp = IEI->getOperand(0);
11674 Value *ScalarOp = IEI->getOperand(1);
11675 Value *IdxOp = IEI->getOperand(2);
11677 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11678 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11679 EI->getOperand(0)->getType() == V->getType()) {
11680 unsigned ExtractedIdx =
11681 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11682 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11684 // Either the extracted from or inserted into vector must be RHSVec,
11685 // otherwise we'd end up with a shuffle of three inputs.
11686 if (EI->getOperand(0) == RHS || RHS == 0) {
11687 RHS = EI->getOperand(0);
11688 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11689 Mask[InsertedIdx % NumElts] =
11690 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11694 if (VecOp == RHS) {
11695 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11696 // Everything but the extracted element is replaced with the RHS.
11697 for (unsigned i = 0; i != NumElts; ++i) {
11698 if (i != InsertedIdx)
11699 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11704 // If this insertelement is a chain that comes from exactly these two
11705 // vectors, return the vector and the effective shuffle.
11706 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11707 return EI->getOperand(0);
11712 // TODO: Handle shufflevector here!
11714 // Otherwise, can't do anything fancy. Return an identity vector.
11715 for (unsigned i = 0; i != NumElts; ++i)
11716 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11720 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11721 Value *VecOp = IE.getOperand(0);
11722 Value *ScalarOp = IE.getOperand(1);
11723 Value *IdxOp = IE.getOperand(2);
11725 // Inserting an undef or into an undefined place, remove this.
11726 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11727 ReplaceInstUsesWith(IE, VecOp);
11729 // If the inserted element was extracted from some other vector, and if the
11730 // indexes are constant, try to turn this into a shufflevector operation.
11731 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11732 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11733 EI->getOperand(0)->getType() == IE.getType()) {
11734 unsigned NumVectorElts = IE.getType()->getNumElements();
11735 unsigned ExtractedIdx =
11736 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11737 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11739 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11740 return ReplaceInstUsesWith(IE, VecOp);
11742 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11743 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11745 // If we are extracting a value from a vector, then inserting it right
11746 // back into the same place, just use the input vector.
11747 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11748 return ReplaceInstUsesWith(IE, VecOp);
11750 // We could theoretically do this for ANY input. However, doing so could
11751 // turn chains of insertelement instructions into a chain of shufflevector
11752 // instructions, and right now we do not merge shufflevectors. As such,
11753 // only do this in a situation where it is clear that there is benefit.
11754 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11755 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11756 // the values of VecOp, except then one read from EIOp0.
11757 // Build a new shuffle mask.
11758 std::vector<Constant*> Mask;
11759 if (isa<UndefValue>(VecOp))
11760 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11762 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11763 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11766 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11767 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11768 ConstantVector::get(Mask));
11771 // If this insertelement isn't used by some other insertelement, turn it
11772 // (and any insertelements it points to), into one big shuffle.
11773 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11774 std::vector<Constant*> Mask;
11776 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11777 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11778 // We now have a shuffle of LHS, RHS, Mask.
11779 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11788 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11789 Value *LHS = SVI.getOperand(0);
11790 Value *RHS = SVI.getOperand(1);
11791 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11793 bool MadeChange = false;
11795 // Undefined shuffle mask -> undefined value.
11796 if (isa<UndefValue>(SVI.getOperand(2)))
11797 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11799 uint64_t UndefElts;
11800 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11802 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
11805 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
11806 if (VWidth <= 64 &&
11807 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
11808 LHS = SVI.getOperand(0);
11809 RHS = SVI.getOperand(1);
11813 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11814 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11815 if (LHS == RHS || isa<UndefValue>(LHS)) {
11816 if (isa<UndefValue>(LHS) && LHS == RHS) {
11817 // shuffle(undef,undef,mask) -> undef.
11818 return ReplaceInstUsesWith(SVI, LHS);
11821 // Remap any references to RHS to use LHS.
11822 std::vector<Constant*> Elts;
11823 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11824 if (Mask[i] >= 2*e)
11825 Elts.push_back(UndefValue::get(Type::Int32Ty));
11827 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11828 (Mask[i] < e && isa<UndefValue>(LHS))) {
11829 Mask[i] = 2*e; // Turn into undef.
11830 Elts.push_back(UndefValue::get(Type::Int32Ty));
11832 Mask[i] = Mask[i] % e; // Force to LHS.
11833 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11837 SVI.setOperand(0, SVI.getOperand(1));
11838 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11839 SVI.setOperand(2, ConstantVector::get(Elts));
11840 LHS = SVI.getOperand(0);
11841 RHS = SVI.getOperand(1);
11845 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11846 bool isLHSID = true, isRHSID = true;
11848 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11849 if (Mask[i] >= e*2) continue; // Ignore undef values.
11850 // Is this an identity shuffle of the LHS value?
11851 isLHSID &= (Mask[i] == i);
11853 // Is this an identity shuffle of the RHS value?
11854 isRHSID &= (Mask[i]-e == i);
11857 // Eliminate identity shuffles.
11858 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11859 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11861 // If the LHS is a shufflevector itself, see if we can combine it with this
11862 // one without producing an unusual shuffle. Here we are really conservative:
11863 // we are absolutely afraid of producing a shuffle mask not in the input
11864 // program, because the code gen may not be smart enough to turn a merged
11865 // shuffle into two specific shuffles: it may produce worse code. As such,
11866 // we only merge two shuffles if the result is one of the two input shuffle
11867 // masks. In this case, merging the shuffles just removes one instruction,
11868 // which we know is safe. This is good for things like turning:
11869 // (splat(splat)) -> splat.
11870 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11871 if (isa<UndefValue>(RHS)) {
11872 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11874 std::vector<unsigned> NewMask;
11875 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11876 if (Mask[i] >= 2*e)
11877 NewMask.push_back(2*e);
11879 NewMask.push_back(LHSMask[Mask[i]]);
11881 // If the result mask is equal to the src shuffle or this shuffle mask, do
11882 // the replacement.
11883 if (NewMask == LHSMask || NewMask == Mask) {
11884 std::vector<Constant*> Elts;
11885 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11886 if (NewMask[i] >= e*2) {
11887 Elts.push_back(UndefValue::get(Type::Int32Ty));
11889 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11892 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11893 LHSSVI->getOperand(1),
11894 ConstantVector::get(Elts));
11899 return MadeChange ? &SVI : 0;
11905 /// TryToSinkInstruction - Try to move the specified instruction from its
11906 /// current block into the beginning of DestBlock, which can only happen if it's
11907 /// safe to move the instruction past all of the instructions between it and the
11908 /// end of its block.
11909 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11910 assert(I->hasOneUse() && "Invariants didn't hold!");
11912 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11913 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11916 // Do not sink alloca instructions out of the entry block.
11917 if (isa<AllocaInst>(I) && I->getParent() ==
11918 &DestBlock->getParent()->getEntryBlock())
11921 // We can only sink load instructions if there is nothing between the load and
11922 // the end of block that could change the value.
11923 if (I->mayReadFromMemory()) {
11924 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11926 if (Scan->mayWriteToMemory())
11930 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11932 I->moveBefore(InsertPos);
11938 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11939 /// all reachable code to the worklist.
11941 /// This has a couple of tricks to make the code faster and more powerful. In
11942 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11943 /// them to the worklist (this significantly speeds up instcombine on code where
11944 /// many instructions are dead or constant). Additionally, if we find a branch
11945 /// whose condition is a known constant, we only visit the reachable successors.
11947 static void AddReachableCodeToWorklist(BasicBlock *BB,
11948 SmallPtrSet<BasicBlock*, 64> &Visited,
11950 const TargetData *TD) {
11951 SmallVector<BasicBlock*, 256> Worklist;
11952 Worklist.push_back(BB);
11954 while (!Worklist.empty()) {
11955 BB = Worklist.back();
11956 Worklist.pop_back();
11958 // We have now visited this block! If we've already been here, ignore it.
11959 if (!Visited.insert(BB)) continue;
11961 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11962 Instruction *Inst = BBI++;
11964 // DCE instruction if trivially dead.
11965 if (isInstructionTriviallyDead(Inst)) {
11967 DOUT << "IC: DCE: " << *Inst;
11968 Inst->eraseFromParent();
11972 // ConstantProp instruction if trivially constant.
11973 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11974 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11975 Inst->replaceAllUsesWith(C);
11977 Inst->eraseFromParent();
11981 IC.AddToWorkList(Inst);
11984 // Recursively visit successors. If this is a branch or switch on a
11985 // constant, only visit the reachable successor.
11986 TerminatorInst *TI = BB->getTerminator();
11987 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11988 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11989 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11990 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11991 Worklist.push_back(ReachableBB);
11994 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11995 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11996 // See if this is an explicit destination.
11997 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11998 if (SI->getCaseValue(i) == Cond) {
11999 BasicBlock *ReachableBB = SI->getSuccessor(i);
12000 Worklist.push_back(ReachableBB);
12004 // Otherwise it is the default destination.
12005 Worklist.push_back(SI->getSuccessor(0));
12010 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12011 Worklist.push_back(TI->getSuccessor(i));
12015 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12016 bool Changed = false;
12017 TD = &getAnalysis<TargetData>();
12019 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12020 << F.getNameStr() << "\n");
12023 // Do a depth-first traversal of the function, populate the worklist with
12024 // the reachable instructions. Ignore blocks that are not reachable. Keep
12025 // track of which blocks we visit.
12026 SmallPtrSet<BasicBlock*, 64> Visited;
12027 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12029 // Do a quick scan over the function. If we find any blocks that are
12030 // unreachable, remove any instructions inside of them. This prevents
12031 // the instcombine code from having to deal with some bad special cases.
12032 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12033 if (!Visited.count(BB)) {
12034 Instruction *Term = BB->getTerminator();
12035 while (Term != BB->begin()) { // Remove instrs bottom-up
12036 BasicBlock::iterator I = Term; --I;
12038 DOUT << "IC: DCE: " << *I;
12041 if (!I->use_empty())
12042 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12043 I->eraseFromParent();
12048 while (!Worklist.empty()) {
12049 Instruction *I = RemoveOneFromWorkList();
12050 if (I == 0) continue; // skip null values.
12052 // Check to see if we can DCE the instruction.
12053 if (isInstructionTriviallyDead(I)) {
12054 // Add operands to the worklist.
12055 if (I->getNumOperands() < 4)
12056 AddUsesToWorkList(*I);
12059 DOUT << "IC: DCE: " << *I;
12061 I->eraseFromParent();
12062 RemoveFromWorkList(I);
12066 // Instruction isn't dead, see if we can constant propagate it.
12067 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12068 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12070 // Add operands to the worklist.
12071 AddUsesToWorkList(*I);
12072 ReplaceInstUsesWith(*I, C);
12075 I->eraseFromParent();
12076 RemoveFromWorkList(I);
12080 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12081 // See if we can constant fold its operands.
12082 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
12083 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
12084 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12090 // See if we can trivially sink this instruction to a successor basic block.
12091 if (I->hasOneUse()) {
12092 BasicBlock *BB = I->getParent();
12093 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12094 if (UserParent != BB) {
12095 bool UserIsSuccessor = false;
12096 // See if the user is one of our successors.
12097 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12098 if (*SI == UserParent) {
12099 UserIsSuccessor = true;
12103 // If the user is one of our immediate successors, and if that successor
12104 // only has us as a predecessors (we'd have to split the critical edge
12105 // otherwise), we can keep going.
12106 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12107 next(pred_begin(UserParent)) == pred_end(UserParent))
12108 // Okay, the CFG is simple enough, try to sink this instruction.
12109 Changed |= TryToSinkInstruction(I, UserParent);
12113 // Now that we have an instruction, try combining it to simplify it...
12117 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12118 if (Instruction *Result = visit(*I)) {
12120 // Should we replace the old instruction with a new one?
12122 DOUT << "IC: Old = " << *I
12123 << " New = " << *Result;
12125 // Everything uses the new instruction now.
12126 I->replaceAllUsesWith(Result);
12128 // Push the new instruction and any users onto the worklist.
12129 AddToWorkList(Result);
12130 AddUsersToWorkList(*Result);
12132 // Move the name to the new instruction first.
12133 Result->takeName(I);
12135 // Insert the new instruction into the basic block...
12136 BasicBlock *InstParent = I->getParent();
12137 BasicBlock::iterator InsertPos = I;
12139 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12140 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12143 InstParent->getInstList().insert(InsertPos, Result);
12145 // Make sure that we reprocess all operands now that we reduced their
12147 AddUsesToWorkList(*I);
12149 // Instructions can end up on the worklist more than once. Make sure
12150 // we do not process an instruction that has been deleted.
12151 RemoveFromWorkList(I);
12153 // Erase the old instruction.
12154 InstParent->getInstList().erase(I);
12157 DOUT << "IC: Mod = " << OrigI
12158 << " New = " << *I;
12161 // If the instruction was modified, it's possible that it is now dead.
12162 // if so, remove it.
12163 if (isInstructionTriviallyDead(I)) {
12164 // Make sure we process all operands now that we are reducing their
12166 AddUsesToWorkList(*I);
12168 // Instructions may end up in the worklist more than once. Erase all
12169 // occurrences of this instruction.
12170 RemoveFromWorkList(I);
12171 I->eraseFromParent();
12174 AddUsersToWorkList(*I);
12181 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12183 // Do an explicit clear, this shrinks the map if needed.
12184 WorklistMap.clear();
12189 bool InstCombiner::runOnFunction(Function &F) {
12190 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12192 bool EverMadeChange = false;
12194 // Iterate while there is work to do.
12195 unsigned Iteration = 0;
12196 while (DoOneIteration(F, Iteration++))
12197 EverMadeChange = true;
12198 return EverMadeChange;
12201 FunctionPass *llvm::createInstructionCombiningPass() {
12202 return new InstCombiner();