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 *visitAnd(BinaryOperator &I);
184 Instruction *visitOr (BinaryOperator &I);
185 Instruction *visitXor(BinaryOperator &I);
186 Instruction *visitShl(BinaryOperator &I);
187 Instruction *visitAShr(BinaryOperator &I);
188 Instruction *visitLShr(BinaryOperator &I);
189 Instruction *commonShiftTransforms(BinaryOperator &I);
190 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
192 Instruction *visitFCmpInst(FCmpInst &I);
193 Instruction *visitICmpInst(ICmpInst &I);
194 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
195 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
198 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
199 ConstantInt *DivRHS);
201 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
202 ICmpInst::Predicate Cond, Instruction &I);
203 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
205 Instruction *commonCastTransforms(CastInst &CI);
206 Instruction *commonIntCastTransforms(CastInst &CI);
207 Instruction *commonPointerCastTransforms(CastInst &CI);
208 Instruction *visitTrunc(TruncInst &CI);
209 Instruction *visitZExt(ZExtInst &CI);
210 Instruction *visitSExt(SExtInst &CI);
211 Instruction *visitFPTrunc(FPTruncInst &CI);
212 Instruction *visitFPExt(CastInst &CI);
213 Instruction *visitFPToUI(FPToUIInst &FI);
214 Instruction *visitFPToSI(FPToSIInst &FI);
215 Instruction *visitUIToFP(CastInst &CI);
216 Instruction *visitSIToFP(CastInst &CI);
217 Instruction *visitPtrToInt(CastInst &CI);
218 Instruction *visitIntToPtr(IntToPtrInst &CI);
219 Instruction *visitBitCast(BitCastInst &CI);
220 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
222 Instruction *visitSelectInst(SelectInst &SI);
223 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
224 Instruction *visitCallInst(CallInst &CI);
225 Instruction *visitInvokeInst(InvokeInst &II);
226 Instruction *visitPHINode(PHINode &PN);
227 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
228 Instruction *visitAllocationInst(AllocationInst &AI);
229 Instruction *visitFreeInst(FreeInst &FI);
230 Instruction *visitLoadInst(LoadInst &LI);
231 Instruction *visitStoreInst(StoreInst &SI);
232 Instruction *visitBranchInst(BranchInst &BI);
233 Instruction *visitSwitchInst(SwitchInst &SI);
234 Instruction *visitInsertElementInst(InsertElementInst &IE);
235 Instruction *visitExtractElementInst(ExtractElementInst &EI);
236 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
237 Instruction *visitExtractValueInst(ExtractValueInst &EV);
239 // visitInstruction - Specify what to return for unhandled instructions...
240 Instruction *visitInstruction(Instruction &I) { return 0; }
243 Instruction *visitCallSite(CallSite CS);
244 bool transformConstExprCastCall(CallSite CS);
245 Instruction *transformCallThroughTrampoline(CallSite CS);
246 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
247 bool DoXform = true);
248 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
251 // InsertNewInstBefore - insert an instruction New before instruction Old
252 // in the program. Add the new instruction to the worklist.
254 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
255 assert(New && New->getParent() == 0 &&
256 "New instruction already inserted into a basic block!");
257 BasicBlock *BB = Old.getParent();
258 BB->getInstList().insert(&Old, New); // Insert inst
263 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
264 /// This also adds the cast to the worklist. Finally, this returns the
266 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
268 if (V->getType() == Ty) return V;
270 if (Constant *CV = dyn_cast<Constant>(V))
271 return ConstantExpr::getCast(opc, CV, Ty);
273 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
278 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
279 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
283 // ReplaceInstUsesWith - This method is to be used when an instruction is
284 // found to be dead, replacable with another preexisting expression. Here
285 // we add all uses of I to the worklist, replace all uses of I with the new
286 // value, then return I, so that the inst combiner will know that I was
289 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
290 AddUsersToWorkList(I); // Add all modified instrs to worklist
292 I.replaceAllUsesWith(V);
295 // If we are replacing the instruction with itself, this must be in a
296 // segment of unreachable code, so just clobber the instruction.
297 I.replaceAllUsesWith(UndefValue::get(I.getType()));
302 // UpdateValueUsesWith - This method is to be used when an value is
303 // found to be replacable with another preexisting expression or was
304 // updated. Here we add all uses of I to the worklist, replace all uses of
305 // I with the new value (unless the instruction was just updated), then
306 // return true, so that the inst combiner will know that I was modified.
308 bool UpdateValueUsesWith(Value *Old, Value *New) {
309 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
311 Old->replaceAllUsesWith(New);
312 if (Instruction *I = dyn_cast<Instruction>(Old))
314 if (Instruction *I = dyn_cast<Instruction>(New))
319 // EraseInstFromFunction - When dealing with an instruction that has side
320 // effects or produces a void value, we can't rely on DCE to delete the
321 // instruction. Instead, visit methods should return the value returned by
323 Instruction *EraseInstFromFunction(Instruction &I) {
324 assert(I.use_empty() && "Cannot erase instruction that is used!");
325 AddUsesToWorkList(I);
326 RemoveFromWorkList(&I);
328 return 0; // Don't do anything with FI
331 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
332 APInt &KnownOne, unsigned Depth = 0) const {
333 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
336 bool MaskedValueIsZero(Value *V, const APInt &Mask,
337 unsigned Depth = 0) const {
338 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
340 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
341 return llvm::ComputeNumSignBits(Op, TD, Depth);
345 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
346 /// InsertBefore instruction. This is specialized a bit to avoid inserting
347 /// casts that are known to not do anything...
349 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
350 Value *V, const Type *DestTy,
351 Instruction *InsertBefore);
353 /// SimplifyCommutative - This performs a few simplifications for
354 /// commutative operators.
355 bool SimplifyCommutative(BinaryOperator &I);
357 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
358 /// most-complex to least-complex order.
359 bool SimplifyCompare(CmpInst &I);
361 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
362 /// on the demanded bits.
363 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
364 APInt& KnownZero, APInt& KnownOne,
367 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
368 uint64_t &UndefElts, unsigned Depth = 0);
370 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
371 // PHI node as operand #0, see if we can fold the instruction into the PHI
372 // (which is only possible if all operands to the PHI are constants).
373 Instruction *FoldOpIntoPhi(Instruction &I);
375 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
376 // operator and they all are only used by the PHI, PHI together their
377 // inputs, and do the operation once, to the result of the PHI.
378 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
379 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
382 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
383 ConstantInt *AndRHS, BinaryOperator &TheAnd);
385 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
386 bool isSub, Instruction &I);
387 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
388 bool isSigned, bool Inside, Instruction &IB);
389 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
390 Instruction *MatchBSwap(BinaryOperator &I);
391 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
392 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
393 Instruction *SimplifyMemSet(MemSetInst *MI);
396 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
398 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
400 int &NumCastsRemoved);
401 unsigned GetOrEnforceKnownAlignment(Value *V,
402 unsigned PrefAlign = 0);
407 char InstCombiner::ID = 0;
408 static RegisterPass<InstCombiner>
409 X("instcombine", "Combine redundant instructions");
411 // getComplexity: Assign a complexity or rank value to LLVM Values...
412 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
413 static unsigned getComplexity(Value *V) {
414 if (isa<Instruction>(V)) {
415 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
419 if (isa<Argument>(V)) return 3;
420 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
423 // isOnlyUse - Return true if this instruction will be deleted if we stop using
425 static bool isOnlyUse(Value *V) {
426 return V->hasOneUse() || isa<Constant>(V);
429 // getPromotedType - Return the specified type promoted as it would be to pass
430 // though a va_arg area...
431 static const Type *getPromotedType(const Type *Ty) {
432 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
433 if (ITy->getBitWidth() < 32)
434 return Type::Int32Ty;
439 /// getBitCastOperand - If the specified operand is a CastInst, a constant
440 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
441 /// operand value, otherwise return null.
442 static Value *getBitCastOperand(Value *V) {
443 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
445 return I->getOperand(0);
446 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
447 // GetElementPtrInst?
448 if (GEP->hasAllZeroIndices())
449 return GEP->getOperand(0);
450 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
451 if (CE->getOpcode() == Instruction::BitCast)
452 // BitCast ConstantExp?
453 return CE->getOperand(0);
454 else if (CE->getOpcode() == Instruction::GetElementPtr) {
455 // GetElementPtr ConstantExp?
456 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
458 ConstantInt *CI = dyn_cast<ConstantInt>(I);
459 if (!CI || !CI->isZero())
460 // Any non-zero indices? Not cast-like.
463 // All-zero indices? This is just like casting.
464 return CE->getOperand(0);
470 /// This function is a wrapper around CastInst::isEliminableCastPair. It
471 /// simply extracts arguments and returns what that function returns.
472 static Instruction::CastOps
473 isEliminableCastPair(
474 const CastInst *CI, ///< The first cast instruction
475 unsigned opcode, ///< The opcode of the second cast instruction
476 const Type *DstTy, ///< The target type for the second cast instruction
477 TargetData *TD ///< The target data for pointer size
480 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
481 const Type *MidTy = CI->getType(); // B from above
483 // Get the opcodes of the two Cast instructions
484 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
485 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
487 return Instruction::CastOps(
488 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
489 DstTy, TD->getIntPtrType()));
492 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
493 /// in any code being generated. It does not require codegen if V is simple
494 /// enough or if the cast can be folded into other casts.
495 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
496 const Type *Ty, TargetData *TD) {
497 if (V->getType() == Ty || isa<Constant>(V)) return false;
499 // If this is another cast that can be eliminated, it isn't codegen either.
500 if (const CastInst *CI = dyn_cast<CastInst>(V))
501 if (isEliminableCastPair(CI, opcode, Ty, TD))
506 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
507 /// InsertBefore instruction. This is specialized a bit to avoid inserting
508 /// casts that are known to not do anything...
510 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
511 Value *V, const Type *DestTy,
512 Instruction *InsertBefore) {
513 if (V->getType() == DestTy) return V;
514 if (Constant *C = dyn_cast<Constant>(V))
515 return ConstantExpr::getCast(opcode, C, DestTy);
517 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
520 // SimplifyCommutative - This performs a few simplifications for commutative
523 // 1. Order operands such that they are listed from right (least complex) to
524 // left (most complex). This puts constants before unary operators before
527 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
528 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
530 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
531 bool Changed = false;
532 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
533 Changed = !I.swapOperands();
535 if (!I.isAssociative()) return Changed;
536 Instruction::BinaryOps Opcode = I.getOpcode();
537 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
538 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
539 if (isa<Constant>(I.getOperand(1))) {
540 Constant *Folded = ConstantExpr::get(I.getOpcode(),
541 cast<Constant>(I.getOperand(1)),
542 cast<Constant>(Op->getOperand(1)));
543 I.setOperand(0, Op->getOperand(0));
544 I.setOperand(1, Folded);
546 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
547 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
548 isOnlyUse(Op) && isOnlyUse(Op1)) {
549 Constant *C1 = cast<Constant>(Op->getOperand(1));
550 Constant *C2 = cast<Constant>(Op1->getOperand(1));
552 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
553 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
554 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
558 I.setOperand(0, New);
559 I.setOperand(1, Folded);
566 /// SimplifyCompare - For a CmpInst this function just orders the operands
567 /// so that theyare listed from right (least complex) to left (most complex).
568 /// This puts constants before unary operators before binary operators.
569 bool InstCombiner::SimplifyCompare(CmpInst &I) {
570 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
573 // Compare instructions are not associative so there's nothing else we can do.
577 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
578 // if the LHS is a constant zero (which is the 'negate' form).
580 static inline Value *dyn_castNegVal(Value *V) {
581 if (BinaryOperator::isNeg(V))
582 return BinaryOperator::getNegArgument(V);
584 // Constants can be considered to be negated values if they can be folded.
585 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
586 return ConstantExpr::getNeg(C);
588 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
589 if (C->getType()->getElementType()->isInteger())
590 return ConstantExpr::getNeg(C);
595 static inline Value *dyn_castNotVal(Value *V) {
596 if (BinaryOperator::isNot(V))
597 return BinaryOperator::getNotArgument(V);
599 // Constants can be considered to be not'ed values...
600 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
601 return ConstantInt::get(~C->getValue());
605 // dyn_castFoldableMul - If this value is a multiply that can be folded into
606 // other computations (because it has a constant operand), return the
607 // non-constant operand of the multiply, and set CST to point to the multiplier.
608 // Otherwise, return null.
610 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
611 if (V->hasOneUse() && V->getType()->isInteger())
612 if (Instruction *I = dyn_cast<Instruction>(V)) {
613 if (I->getOpcode() == Instruction::Mul)
614 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
615 return I->getOperand(0);
616 if (I->getOpcode() == Instruction::Shl)
617 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
618 // The multiplier is really 1 << CST.
619 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
620 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
621 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
622 return I->getOperand(0);
628 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
629 /// expression, return it.
630 static User *dyn_castGetElementPtr(Value *V) {
631 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
632 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
633 if (CE->getOpcode() == Instruction::GetElementPtr)
634 return cast<User>(V);
638 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
639 /// opcode value. Otherwise return UserOp1.
640 static unsigned getOpcode(const Value *V) {
641 if (const Instruction *I = dyn_cast<Instruction>(V))
642 return I->getOpcode();
643 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
644 return CE->getOpcode();
645 // Use UserOp1 to mean there's no opcode.
646 return Instruction::UserOp1;
649 /// AddOne - Add one to a ConstantInt
650 static ConstantInt *AddOne(ConstantInt *C) {
651 APInt Val(C->getValue());
652 return ConstantInt::get(++Val);
654 /// SubOne - Subtract one from a ConstantInt
655 static ConstantInt *SubOne(ConstantInt *C) {
656 APInt Val(C->getValue());
657 return ConstantInt::get(--Val);
659 /// Add - Add two ConstantInts together
660 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
661 return ConstantInt::get(C1->getValue() + C2->getValue());
663 /// And - Bitwise AND two ConstantInts together
664 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
665 return ConstantInt::get(C1->getValue() & C2->getValue());
667 /// Subtract - Subtract one ConstantInt from another
668 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
669 return ConstantInt::get(C1->getValue() - C2->getValue());
671 /// Multiply - Multiply two ConstantInts together
672 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
673 return ConstantInt::get(C1->getValue() * C2->getValue());
675 /// MultiplyOverflows - True if the multiply can not be expressed in an int
677 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
678 uint32_t W = C1->getBitWidth();
679 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
688 APInt MulExt = LHSExt * RHSExt;
691 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
692 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
693 return MulExt.slt(Min) || MulExt.sgt(Max);
695 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
699 /// ShrinkDemandedConstant - Check to see if the specified operand of the
700 /// specified instruction is a constant integer. If so, check to see if there
701 /// are any bits set in the constant that are not demanded. If so, shrink the
702 /// constant and return true.
703 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
705 assert(I && "No instruction?");
706 assert(OpNo < I->getNumOperands() && "Operand index too large");
708 // If the operand is not a constant integer, nothing to do.
709 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
710 if (!OpC) return false;
712 // If there are no bits set that aren't demanded, nothing to do.
713 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
714 if ((~Demanded & OpC->getValue()) == 0)
717 // This instruction is producing bits that are not demanded. Shrink the RHS.
718 Demanded &= OpC->getValue();
719 I->setOperand(OpNo, ConstantInt::get(Demanded));
723 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
724 // set of known zero and one bits, compute the maximum and minimum values that
725 // could have the specified known zero and known one bits, returning them in
727 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
728 const APInt& KnownZero,
729 const APInt& KnownOne,
730 APInt& Min, APInt& Max) {
731 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
732 assert(KnownZero.getBitWidth() == BitWidth &&
733 KnownOne.getBitWidth() == BitWidth &&
734 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
735 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
736 APInt UnknownBits = ~(KnownZero|KnownOne);
738 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
739 // bit if it is unknown.
741 Max = KnownOne|UnknownBits;
743 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
745 Max.clear(BitWidth-1);
749 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
750 // a set of known zero and one bits, compute the maximum and minimum values that
751 // could have the specified known zero and known one bits, returning them in
753 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
754 const APInt &KnownZero,
755 const APInt &KnownOne,
756 APInt &Min, APInt &Max) {
757 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
758 assert(KnownZero.getBitWidth() == BitWidth &&
759 KnownOne.getBitWidth() == BitWidth &&
760 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
761 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
762 APInt UnknownBits = ~(KnownZero|KnownOne);
764 // The minimum value is when the unknown bits are all zeros.
766 // The maximum value is when the unknown bits are all ones.
767 Max = KnownOne|UnknownBits;
770 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
771 /// value based on the demanded bits. When this function is called, it is known
772 /// that only the bits set in DemandedMask of the result of V are ever used
773 /// downstream. Consequently, depending on the mask and V, it may be possible
774 /// to replace V with a constant or one of its operands. In such cases, this
775 /// function does the replacement and returns true. In all other cases, it
776 /// returns false after analyzing the expression and setting KnownOne and known
777 /// to be one in the expression. KnownZero contains all the bits that are known
778 /// to be zero in the expression. These are provided to potentially allow the
779 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
780 /// the expression. KnownOne and KnownZero always follow the invariant that
781 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
782 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
783 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
784 /// and KnownOne must all be the same.
785 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
786 APInt& KnownZero, APInt& KnownOne,
788 assert(V != 0 && "Null pointer of Value???");
789 assert(Depth <= 6 && "Limit Search Depth");
790 uint32_t BitWidth = DemandedMask.getBitWidth();
791 const IntegerType *VTy = cast<IntegerType>(V->getType());
792 assert(VTy->getBitWidth() == BitWidth &&
793 KnownZero.getBitWidth() == BitWidth &&
794 KnownOne.getBitWidth() == BitWidth &&
795 "Value *V, DemandedMask, KnownZero and KnownOne \
796 must have same BitWidth");
797 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
798 // We know all of the bits for a constant!
799 KnownOne = CI->getValue() & DemandedMask;
800 KnownZero = ~KnownOne & DemandedMask;
806 if (!V->hasOneUse()) { // Other users may use these bits.
807 if (Depth != 0) { // Not at the root.
808 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
809 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
812 // If this is the root being simplified, allow it to have multiple uses,
813 // just set the DemandedMask to all bits.
814 DemandedMask = APInt::getAllOnesValue(BitWidth);
815 } else if (DemandedMask == 0) { // Not demanding any bits from V.
816 if (V != UndefValue::get(VTy))
817 return UpdateValueUsesWith(V, UndefValue::get(VTy));
819 } else if (Depth == 6) { // Limit search depth.
823 Instruction *I = dyn_cast<Instruction>(V);
824 if (!I) return false; // Only analyze instructions.
826 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
827 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
828 switch (I->getOpcode()) {
830 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
832 case Instruction::And:
833 // If either the LHS or the RHS are Zero, the result is zero.
834 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
835 RHSKnownZero, RHSKnownOne, Depth+1))
837 assert((RHSKnownZero & RHSKnownOne) == 0 &&
838 "Bits known to be one AND zero?");
840 // If something is known zero on the RHS, the bits aren't demanded on the
842 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
843 LHSKnownZero, LHSKnownOne, Depth+1))
845 assert((LHSKnownZero & LHSKnownOne) == 0 &&
846 "Bits known to be one AND zero?");
848 // If all of the demanded bits are known 1 on one side, return the other.
849 // These bits cannot contribute to the result of the 'and'.
850 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
851 (DemandedMask & ~LHSKnownZero))
852 return UpdateValueUsesWith(I, I->getOperand(0));
853 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
854 (DemandedMask & ~RHSKnownZero))
855 return UpdateValueUsesWith(I, I->getOperand(1));
857 // If all of the demanded bits in the inputs are known zeros, return zero.
858 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
859 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
861 // If the RHS is a constant, see if we can simplify it.
862 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
863 return UpdateValueUsesWith(I, I);
865 // Output known-1 bits are only known if set in both the LHS & RHS.
866 RHSKnownOne &= LHSKnownOne;
867 // Output known-0 are known to be clear if zero in either the LHS | RHS.
868 RHSKnownZero |= LHSKnownZero;
870 case Instruction::Or:
871 // If either the LHS or the RHS are One, the result is One.
872 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
873 RHSKnownZero, RHSKnownOne, Depth+1))
875 assert((RHSKnownZero & RHSKnownOne) == 0 &&
876 "Bits known to be one AND zero?");
877 // If something is known one on the RHS, the bits aren't demanded on the
879 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
880 LHSKnownZero, LHSKnownOne, Depth+1))
882 assert((LHSKnownZero & LHSKnownOne) == 0 &&
883 "Bits known to be one AND zero?");
885 // If all of the demanded bits are known zero on one side, return the other.
886 // These bits cannot contribute to the result of the 'or'.
887 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
888 (DemandedMask & ~LHSKnownOne))
889 return UpdateValueUsesWith(I, I->getOperand(0));
890 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
891 (DemandedMask & ~RHSKnownOne))
892 return UpdateValueUsesWith(I, I->getOperand(1));
894 // If all of the potentially set bits on one side are known to be set on
895 // the other side, just use the 'other' side.
896 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
897 (DemandedMask & (~RHSKnownZero)))
898 return UpdateValueUsesWith(I, I->getOperand(0));
899 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
900 (DemandedMask & (~LHSKnownZero)))
901 return UpdateValueUsesWith(I, I->getOperand(1));
903 // If the RHS is a constant, see if we can simplify it.
904 if (ShrinkDemandedConstant(I, 1, DemandedMask))
905 return UpdateValueUsesWith(I, I);
907 // Output known-0 bits are only known if clear in both the LHS & RHS.
908 RHSKnownZero &= LHSKnownZero;
909 // Output known-1 are known to be set if set in either the LHS | RHS.
910 RHSKnownOne |= LHSKnownOne;
912 case Instruction::Xor: {
913 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
914 RHSKnownZero, RHSKnownOne, Depth+1))
916 assert((RHSKnownZero & RHSKnownOne) == 0 &&
917 "Bits known to be one AND zero?");
918 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
919 LHSKnownZero, LHSKnownOne, Depth+1))
921 assert((LHSKnownZero & LHSKnownOne) == 0 &&
922 "Bits known to be one AND zero?");
924 // If all of the demanded bits are known zero on one side, return the other.
925 // These bits cannot contribute to the result of the 'xor'.
926 if ((DemandedMask & RHSKnownZero) == DemandedMask)
927 return UpdateValueUsesWith(I, I->getOperand(0));
928 if ((DemandedMask & LHSKnownZero) == DemandedMask)
929 return UpdateValueUsesWith(I, I->getOperand(1));
931 // Output known-0 bits are known if clear or set in both the LHS & RHS.
932 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
933 (RHSKnownOne & LHSKnownOne);
934 // Output known-1 are known to be set if set in only one of the LHS, RHS.
935 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
936 (RHSKnownOne & LHSKnownZero);
938 // If all of the demanded bits are known to be zero on one side or the
939 // other, turn this into an *inclusive* or.
940 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
941 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
943 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
945 InsertNewInstBefore(Or, *I);
946 return UpdateValueUsesWith(I, Or);
949 // If all of the demanded bits on one side are known, and all of the set
950 // bits on that side are also known to be set on the other side, turn this
951 // into an AND, as we know the bits will be cleared.
952 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
953 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
955 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
956 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
958 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
959 InsertNewInstBefore(And, *I);
960 return UpdateValueUsesWith(I, And);
964 // If the RHS is a constant, see if we can simplify it.
965 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
966 if (ShrinkDemandedConstant(I, 1, DemandedMask))
967 return UpdateValueUsesWith(I, I);
969 RHSKnownZero = KnownZeroOut;
970 RHSKnownOne = KnownOneOut;
973 case Instruction::Select:
974 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
975 RHSKnownZero, RHSKnownOne, Depth+1))
977 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
978 LHSKnownZero, LHSKnownOne, Depth+1))
980 assert((RHSKnownZero & RHSKnownOne) == 0 &&
981 "Bits known to be one AND zero?");
982 assert((LHSKnownZero & LHSKnownOne) == 0 &&
983 "Bits known to be one AND zero?");
985 // If the operands are constants, see if we can simplify them.
986 if (ShrinkDemandedConstant(I, 1, DemandedMask))
987 return UpdateValueUsesWith(I, I);
988 if (ShrinkDemandedConstant(I, 2, DemandedMask))
989 return UpdateValueUsesWith(I, I);
991 // Only known if known in both the LHS and RHS.
992 RHSKnownOne &= LHSKnownOne;
993 RHSKnownZero &= LHSKnownZero;
995 case Instruction::Trunc: {
997 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
998 DemandedMask.zext(truncBf);
999 RHSKnownZero.zext(truncBf);
1000 RHSKnownOne.zext(truncBf);
1001 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1002 RHSKnownZero, RHSKnownOne, Depth+1))
1004 DemandedMask.trunc(BitWidth);
1005 RHSKnownZero.trunc(BitWidth);
1006 RHSKnownOne.trunc(BitWidth);
1007 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1008 "Bits known to be one AND zero?");
1011 case Instruction::BitCast:
1012 if (!I->getOperand(0)->getType()->isInteger())
1015 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1016 RHSKnownZero, RHSKnownOne, Depth+1))
1018 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1019 "Bits known to be one AND zero?");
1021 case Instruction::ZExt: {
1022 // Compute the bits in the result that are not present in the input.
1023 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1024 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1026 DemandedMask.trunc(SrcBitWidth);
1027 RHSKnownZero.trunc(SrcBitWidth);
1028 RHSKnownOne.trunc(SrcBitWidth);
1029 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1030 RHSKnownZero, RHSKnownOne, Depth+1))
1032 DemandedMask.zext(BitWidth);
1033 RHSKnownZero.zext(BitWidth);
1034 RHSKnownOne.zext(BitWidth);
1035 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1036 "Bits known to be one AND zero?");
1037 // The top bits are known to be zero.
1038 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1041 case Instruction::SExt: {
1042 // Compute the bits in the result that are not present in the input.
1043 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1044 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1046 APInt InputDemandedBits = DemandedMask &
1047 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1049 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1050 // If any of the sign extended bits are demanded, we know that the sign
1052 if ((NewBits & DemandedMask) != 0)
1053 InputDemandedBits.set(SrcBitWidth-1);
1055 InputDemandedBits.trunc(SrcBitWidth);
1056 RHSKnownZero.trunc(SrcBitWidth);
1057 RHSKnownOne.trunc(SrcBitWidth);
1058 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1059 RHSKnownZero, RHSKnownOne, Depth+1))
1061 InputDemandedBits.zext(BitWidth);
1062 RHSKnownZero.zext(BitWidth);
1063 RHSKnownOne.zext(BitWidth);
1064 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1065 "Bits known to be one AND zero?");
1067 // If the sign bit of the input is known set or clear, then we know the
1068 // top bits of the result.
1070 // If the input sign bit is known zero, or if the NewBits are not demanded
1071 // convert this into a zero extension.
1072 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1074 // Convert to ZExt cast
1075 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1076 return UpdateValueUsesWith(I, NewCast);
1077 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1078 RHSKnownOne |= NewBits;
1082 case Instruction::Add: {
1083 // Figure out what the input bits are. If the top bits of the and result
1084 // are not demanded, then the add doesn't demand them from its input
1086 uint32_t NLZ = DemandedMask.countLeadingZeros();
1088 // If there is a constant on the RHS, there are a variety of xformations
1090 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1091 // If null, this should be simplified elsewhere. Some of the xforms here
1092 // won't work if the RHS is zero.
1096 // If the top bit of the output is demanded, demand everything from the
1097 // input. Otherwise, we demand all the input bits except NLZ top bits.
1098 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1100 // Find information about known zero/one bits in the input.
1101 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1102 LHSKnownZero, LHSKnownOne, Depth+1))
1105 // If the RHS of the add has bits set that can't affect the input, reduce
1107 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1108 return UpdateValueUsesWith(I, I);
1110 // Avoid excess work.
1111 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1114 // Turn it into OR if input bits are zero.
1115 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1117 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1119 InsertNewInstBefore(Or, *I);
1120 return UpdateValueUsesWith(I, Or);
1123 // We can say something about the output known-zero and known-one bits,
1124 // depending on potential carries from the input constant and the
1125 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1126 // bits set and the RHS constant is 0x01001, then we know we have a known
1127 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1129 // To compute this, we first compute the potential carry bits. These are
1130 // the bits which may be modified. I'm not aware of a better way to do
1132 const APInt& RHSVal = RHS->getValue();
1133 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1135 // Now that we know which bits have carries, compute the known-1/0 sets.
1137 // Bits are known one if they are known zero in one operand and one in the
1138 // other, and there is no input carry.
1139 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1140 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1142 // Bits are known zero if they are known zero in both operands and there
1143 // is no input carry.
1144 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1146 // If the high-bits of this ADD are not demanded, then it does not demand
1147 // the high bits of its LHS or RHS.
1148 if (DemandedMask[BitWidth-1] == 0) {
1149 // Right fill the mask of bits for this ADD to demand the most
1150 // significant bit and all those below it.
1151 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1152 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1153 LHSKnownZero, LHSKnownOne, Depth+1))
1155 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1156 LHSKnownZero, LHSKnownOne, Depth+1))
1162 case Instruction::Sub:
1163 // If the high-bits of this SUB are not demanded, then it does not demand
1164 // the high bits of its LHS or RHS.
1165 if (DemandedMask[BitWidth-1] == 0) {
1166 // Right fill the mask of bits for this SUB to demand the most
1167 // significant bit and all those below it.
1168 uint32_t NLZ = DemandedMask.countLeadingZeros();
1169 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1170 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1171 LHSKnownZero, LHSKnownOne, Depth+1))
1173 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1174 LHSKnownZero, LHSKnownOne, Depth+1))
1177 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1178 // the known zeros and ones.
1179 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1181 case Instruction::Shl:
1182 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1183 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1184 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1185 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1186 RHSKnownZero, RHSKnownOne, Depth+1))
1188 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1189 "Bits known to be one AND zero?");
1190 RHSKnownZero <<= ShiftAmt;
1191 RHSKnownOne <<= ShiftAmt;
1192 // low bits known zero.
1194 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1197 case Instruction::LShr:
1198 // For a logical shift right
1199 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1200 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1202 // Unsigned shift right.
1203 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1204 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1205 RHSKnownZero, RHSKnownOne, Depth+1))
1207 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1208 "Bits known to be one AND zero?");
1209 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1210 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1212 // Compute the new bits that are at the top now.
1213 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1214 RHSKnownZero |= HighBits; // high bits known zero.
1218 case Instruction::AShr:
1219 // If this is an arithmetic shift right and only the low-bit is set, we can
1220 // always convert this into a logical shr, even if the shift amount is
1221 // variable. The low bit of the shift cannot be an input sign bit unless
1222 // the shift amount is >= the size of the datatype, which is undefined.
1223 if (DemandedMask == 1) {
1224 // Perform the logical shift right.
1225 Value *NewVal = BinaryOperator::CreateLShr(
1226 I->getOperand(0), I->getOperand(1), I->getName());
1227 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1228 return UpdateValueUsesWith(I, NewVal);
1231 // If the sign bit is the only bit demanded by this ashr, then there is no
1232 // need to do it, the shift doesn't change the high bit.
1233 if (DemandedMask.isSignBit())
1234 return UpdateValueUsesWith(I, I->getOperand(0));
1236 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1237 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1239 // Signed shift right.
1240 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1241 // If any of the "high bits" are demanded, we should set the sign bit as
1243 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1244 DemandedMaskIn.set(BitWidth-1);
1245 if (SimplifyDemandedBits(I->getOperand(0),
1247 RHSKnownZero, RHSKnownOne, Depth+1))
1249 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1250 "Bits known to be one AND zero?");
1251 // Compute the new bits that are at the top now.
1252 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1253 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1254 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1256 // Handle the sign bits.
1257 APInt SignBit(APInt::getSignBit(BitWidth));
1258 // Adjust to where it is now in the mask.
1259 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1261 // If the input sign bit is known to be zero, or if none of the top bits
1262 // are demanded, turn this into an unsigned shift right.
1263 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1264 (HighBits & ~DemandedMask) == HighBits) {
1265 // Perform the logical shift right.
1266 Value *NewVal = BinaryOperator::CreateLShr(
1267 I->getOperand(0), SA, I->getName());
1268 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1269 return UpdateValueUsesWith(I, NewVal);
1270 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1271 RHSKnownOne |= HighBits;
1275 case Instruction::SRem:
1276 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1277 APInt RA = Rem->getValue().abs();
1278 if (RA.isPowerOf2()) {
1279 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1280 return UpdateValueUsesWith(I, I->getOperand(0));
1282 APInt LowBits = RA - 1;
1283 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1284 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1285 LHSKnownZero, LHSKnownOne, Depth+1))
1288 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1289 LHSKnownZero |= ~LowBits;
1291 KnownZero |= LHSKnownZero & DemandedMask;
1293 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1297 case Instruction::URem: {
1298 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1299 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1300 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1301 KnownZero2, KnownOne2, Depth+1))
1304 uint32_t Leaders = KnownZero2.countLeadingOnes();
1305 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1306 KnownZero2, KnownOne2, Depth+1))
1309 Leaders = std::max(Leaders,
1310 KnownZero2.countLeadingOnes());
1311 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1314 case Instruction::Call:
1315 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1316 switch (II->getIntrinsicID()) {
1318 case Intrinsic::bswap: {
1319 // If the only bits demanded come from one byte of the bswap result,
1320 // just shift the input byte into position to eliminate the bswap.
1321 unsigned NLZ = DemandedMask.countLeadingZeros();
1322 unsigned NTZ = DemandedMask.countTrailingZeros();
1324 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1325 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1326 // have 14 leading zeros, round to 8.
1329 // If we need exactly one byte, we can do this transformation.
1330 if (BitWidth-NLZ-NTZ == 8) {
1331 unsigned ResultBit = NTZ;
1332 unsigned InputBit = BitWidth-NTZ-8;
1334 // Replace this with either a left or right shift to get the byte into
1336 Instruction *NewVal;
1337 if (InputBit > ResultBit)
1338 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1339 ConstantInt::get(I->getType(), InputBit-ResultBit));
1341 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1342 ConstantInt::get(I->getType(), ResultBit-InputBit));
1343 NewVal->takeName(I);
1344 InsertNewInstBefore(NewVal, *I);
1345 return UpdateValueUsesWith(I, NewVal);
1348 // TODO: Could compute known zero/one bits based on the input.
1353 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1357 // If the client is only demanding bits that we know, return the known
1359 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1360 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1365 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1366 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1367 /// actually used by the caller. This method analyzes which elements of the
1368 /// operand are undef and returns that information in UndefElts.
1370 /// If the information about demanded elements can be used to simplify the
1371 /// operation, the operation is simplified, then the resultant value is
1372 /// returned. This returns null if no change was made.
1373 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1374 uint64_t &UndefElts,
1376 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1377 assert(VWidth <= 64 && "Vector too wide to analyze!");
1378 uint64_t EltMask = ~0ULL >> (64-VWidth);
1379 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1381 if (isa<UndefValue>(V)) {
1382 // If the entire vector is undefined, just return this info.
1383 UndefElts = EltMask;
1385 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1386 UndefElts = EltMask;
1387 return UndefValue::get(V->getType());
1391 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1392 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1393 Constant *Undef = UndefValue::get(EltTy);
1395 std::vector<Constant*> Elts;
1396 for (unsigned i = 0; i != VWidth; ++i)
1397 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1398 Elts.push_back(Undef);
1399 UndefElts |= (1ULL << i);
1400 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1401 Elts.push_back(Undef);
1402 UndefElts |= (1ULL << i);
1403 } else { // Otherwise, defined.
1404 Elts.push_back(CP->getOperand(i));
1407 // If we changed the constant, return it.
1408 Constant *NewCP = ConstantVector::get(Elts);
1409 return NewCP != CP ? NewCP : 0;
1410 } else if (isa<ConstantAggregateZero>(V)) {
1411 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1414 // Check if this is identity. If so, return 0 since we are not simplifying
1416 if (DemandedElts == ((1ULL << VWidth) -1))
1419 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1420 Constant *Zero = Constant::getNullValue(EltTy);
1421 Constant *Undef = UndefValue::get(EltTy);
1422 std::vector<Constant*> Elts;
1423 for (unsigned i = 0; i != VWidth; ++i)
1424 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1425 UndefElts = DemandedElts ^ EltMask;
1426 return ConstantVector::get(Elts);
1429 // Limit search depth.
1433 // If multiple users are using the root value, procede with
1434 // simplification conservatively assuming that all elements
1436 if (!V->hasOneUse()) {
1437 // Quit if we find multiple users of a non-root value though.
1438 // They'll be handled when it's their turn to be visited by
1439 // the main instcombine process.
1441 // TODO: Just compute the UndefElts information recursively.
1444 // Conservatively assume that all elements are needed.
1445 DemandedElts = EltMask;
1448 Instruction *I = dyn_cast<Instruction>(V);
1449 if (!I) return false; // Only analyze instructions.
1451 bool MadeChange = false;
1452 uint64_t UndefElts2;
1454 switch (I->getOpcode()) {
1457 case Instruction::InsertElement: {
1458 // If this is a variable index, we don't know which element it overwrites.
1459 // demand exactly the same input as we produce.
1460 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1462 // Note that we can't propagate undef elt info, because we don't know
1463 // which elt is getting updated.
1464 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1465 UndefElts2, Depth+1);
1466 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1470 // If this is inserting an element that isn't demanded, remove this
1472 unsigned IdxNo = Idx->getZExtValue();
1473 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1474 return AddSoonDeadInstToWorklist(*I, 0);
1476 // Otherwise, the element inserted overwrites whatever was there, so the
1477 // input demanded set is simpler than the output set.
1478 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1479 DemandedElts & ~(1ULL << IdxNo),
1480 UndefElts, Depth+1);
1481 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1483 // The inserted element is defined.
1484 UndefElts &= ~(1ULL << IdxNo);
1487 case Instruction::ShuffleVector: {
1488 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1489 uint64_t LHSVWidth =
1490 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1491 uint64_t LeftDemanded = 0, RightDemanded = 0;
1492 for (unsigned i = 0; i < VWidth; i++) {
1493 if (DemandedElts & (1ULL << i)) {
1494 unsigned MaskVal = Shuffle->getMaskValue(i);
1495 if (MaskVal != -1u) {
1496 assert(MaskVal < LHSVWidth * 2 &&
1497 "shufflevector mask index out of range!");
1498 if (MaskVal < LHSVWidth)
1499 LeftDemanded |= 1ULL << MaskVal;
1501 RightDemanded |= 1ULL << (MaskVal - LHSVWidth);
1506 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1507 UndefElts2, Depth+1);
1508 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1510 uint64_t UndefElts3;
1511 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1512 UndefElts3, Depth+1);
1513 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1515 bool NewUndefElts = false;
1516 for (unsigned i = 0; i < VWidth; i++) {
1517 unsigned MaskVal = Shuffle->getMaskValue(i);
1518 if (MaskVal == -1u) {
1519 uint64_t NewBit = 1ULL << i;
1520 UndefElts |= NewBit;
1521 } else if (MaskVal < LHSVWidth) {
1522 uint64_t NewBit = ((UndefElts2 >> MaskVal) & 1) << i;
1523 NewUndefElts |= NewBit;
1524 UndefElts |= NewBit;
1526 uint64_t NewBit = ((UndefElts3 >> (MaskVal - LHSVWidth)) & 1) << i;
1527 NewUndefElts |= NewBit;
1528 UndefElts |= NewBit;
1533 // Add additional discovered undefs.
1534 std::vector<Constant*> Elts;
1535 for (unsigned i = 0; i < VWidth; ++i) {
1536 if (UndefElts & (1ULL << i))
1537 Elts.push_back(UndefValue::get(Type::Int32Ty));
1539 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1540 Shuffle->getMaskValue(i)));
1542 I->setOperand(2, ConstantVector::get(Elts));
1547 case Instruction::BitCast: {
1548 // Vector->vector casts only.
1549 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1551 unsigned InVWidth = VTy->getNumElements();
1552 uint64_t InputDemandedElts = 0;
1555 if (VWidth == InVWidth) {
1556 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1557 // elements as are demanded of us.
1559 InputDemandedElts = DemandedElts;
1560 } else if (VWidth > InVWidth) {
1564 // If there are more elements in the result than there are in the source,
1565 // then an input element is live if any of the corresponding output
1566 // elements are live.
1567 Ratio = VWidth/InVWidth;
1568 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1569 if (DemandedElts & (1ULL << OutIdx))
1570 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1576 // If there are more elements in the source than there are in the result,
1577 // then an input element is live if the corresponding output element is
1579 Ratio = InVWidth/VWidth;
1580 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1581 if (DemandedElts & (1ULL << InIdx/Ratio))
1582 InputDemandedElts |= 1ULL << InIdx;
1585 // div/rem demand all inputs, because they don't want divide by zero.
1586 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1587 UndefElts2, Depth+1);
1589 I->setOperand(0, TmpV);
1593 UndefElts = UndefElts2;
1594 if (VWidth > InVWidth) {
1595 assert(0 && "Unimp");
1596 // If there are more elements in the result than there are in the source,
1597 // then an output element is undef if the corresponding input element is
1599 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1600 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1601 UndefElts |= 1ULL << OutIdx;
1602 } else if (VWidth < InVWidth) {
1603 assert(0 && "Unimp");
1604 // If there are more elements in the source than there are in the result,
1605 // then a result element is undef if all of the corresponding input
1606 // elements are undef.
1607 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1608 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1609 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1610 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1614 case Instruction::And:
1615 case Instruction::Or:
1616 case Instruction::Xor:
1617 case Instruction::Add:
1618 case Instruction::Sub:
1619 case Instruction::Mul:
1620 // div/rem demand all inputs, because they don't want divide by zero.
1621 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1622 UndefElts, Depth+1);
1623 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1624 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1625 UndefElts2, Depth+1);
1626 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1628 // Output elements are undefined if both are undefined. Consider things
1629 // like undef&0. The result is known zero, not undef.
1630 UndefElts &= UndefElts2;
1633 case Instruction::Call: {
1634 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1636 switch (II->getIntrinsicID()) {
1639 // Binary vector operations that work column-wise. A dest element is a
1640 // function of the corresponding input elements from the two inputs.
1641 case Intrinsic::x86_sse_sub_ss:
1642 case Intrinsic::x86_sse_mul_ss:
1643 case Intrinsic::x86_sse_min_ss:
1644 case Intrinsic::x86_sse_max_ss:
1645 case Intrinsic::x86_sse2_sub_sd:
1646 case Intrinsic::x86_sse2_mul_sd:
1647 case Intrinsic::x86_sse2_min_sd:
1648 case Intrinsic::x86_sse2_max_sd:
1649 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1650 UndefElts, Depth+1);
1651 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1652 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1653 UndefElts2, Depth+1);
1654 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1656 // If only the low elt is demanded and this is a scalarizable intrinsic,
1657 // scalarize it now.
1658 if (DemandedElts == 1) {
1659 switch (II->getIntrinsicID()) {
1661 case Intrinsic::x86_sse_sub_ss:
1662 case Intrinsic::x86_sse_mul_ss:
1663 case Intrinsic::x86_sse2_sub_sd:
1664 case Intrinsic::x86_sse2_mul_sd:
1665 // TODO: Lower MIN/MAX/ABS/etc
1666 Value *LHS = II->getOperand(1);
1667 Value *RHS = II->getOperand(2);
1668 // Extract the element as scalars.
1669 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1670 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1672 switch (II->getIntrinsicID()) {
1673 default: assert(0 && "Case stmts out of sync!");
1674 case Intrinsic::x86_sse_sub_ss:
1675 case Intrinsic::x86_sse2_sub_sd:
1676 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1677 II->getName()), *II);
1679 case Intrinsic::x86_sse_mul_ss:
1680 case Intrinsic::x86_sse2_mul_sd:
1681 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1682 II->getName()), *II);
1687 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1689 InsertNewInstBefore(New, *II);
1690 AddSoonDeadInstToWorklist(*II, 0);
1695 // Output elements are undefined if both are undefined. Consider things
1696 // like undef&0. The result is known zero, not undef.
1697 UndefElts &= UndefElts2;
1703 return MadeChange ? I : 0;
1707 /// AssociativeOpt - Perform an optimization on an associative operator. This
1708 /// function is designed to check a chain of associative operators for a
1709 /// potential to apply a certain optimization. Since the optimization may be
1710 /// applicable if the expression was reassociated, this checks the chain, then
1711 /// reassociates the expression as necessary to expose the optimization
1712 /// opportunity. This makes use of a special Functor, which must define
1713 /// 'shouldApply' and 'apply' methods.
1715 template<typename Functor>
1716 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1717 unsigned Opcode = Root.getOpcode();
1718 Value *LHS = Root.getOperand(0);
1720 // Quick check, see if the immediate LHS matches...
1721 if (F.shouldApply(LHS))
1722 return F.apply(Root);
1724 // Otherwise, if the LHS is not of the same opcode as the root, return.
1725 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1726 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1727 // Should we apply this transform to the RHS?
1728 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1730 // If not to the RHS, check to see if we should apply to the LHS...
1731 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1732 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1736 // If the functor wants to apply the optimization to the RHS of LHSI,
1737 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1739 // Now all of the instructions are in the current basic block, go ahead
1740 // and perform the reassociation.
1741 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1743 // First move the selected RHS to the LHS of the root...
1744 Root.setOperand(0, LHSI->getOperand(1));
1746 // Make what used to be the LHS of the root be the user of the root...
1747 Value *ExtraOperand = TmpLHSI->getOperand(1);
1748 if (&Root == TmpLHSI) {
1749 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1752 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1753 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1754 BasicBlock::iterator ARI = &Root; ++ARI;
1755 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1758 // Now propagate the ExtraOperand down the chain of instructions until we
1760 while (TmpLHSI != LHSI) {
1761 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1762 // Move the instruction to immediately before the chain we are
1763 // constructing to avoid breaking dominance properties.
1764 NextLHSI->moveBefore(ARI);
1767 Value *NextOp = NextLHSI->getOperand(1);
1768 NextLHSI->setOperand(1, ExtraOperand);
1770 ExtraOperand = NextOp;
1773 // Now that the instructions are reassociated, have the functor perform
1774 // the transformation...
1775 return F.apply(Root);
1778 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1785 // AddRHS - Implements: X + X --> X << 1
1788 AddRHS(Value *rhs) : RHS(rhs) {}
1789 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1790 Instruction *apply(BinaryOperator &Add) const {
1791 return BinaryOperator::CreateShl(Add.getOperand(0),
1792 ConstantInt::get(Add.getType(), 1));
1796 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1798 struct AddMaskingAnd {
1800 AddMaskingAnd(Constant *c) : C2(c) {}
1801 bool shouldApply(Value *LHS) const {
1803 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1804 ConstantExpr::getAnd(C1, C2)->isNullValue();
1806 Instruction *apply(BinaryOperator &Add) const {
1807 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1813 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1815 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1816 if (Constant *SOC = dyn_cast<Constant>(SO))
1817 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1819 return IC->InsertNewInstBefore(CastInst::Create(
1820 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1823 // Figure out if the constant is the left or the right argument.
1824 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1825 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1827 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1829 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1830 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1833 Value *Op0 = SO, *Op1 = ConstOperand;
1835 std::swap(Op0, Op1);
1837 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1838 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1839 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1840 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1841 SO->getName()+".cmp");
1843 assert(0 && "Unknown binary instruction type!");
1846 return IC->InsertNewInstBefore(New, I);
1849 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1850 // constant as the other operand, try to fold the binary operator into the
1851 // select arguments. This also works for Cast instructions, which obviously do
1852 // not have a second operand.
1853 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1855 // Don't modify shared select instructions
1856 if (!SI->hasOneUse()) return 0;
1857 Value *TV = SI->getOperand(1);
1858 Value *FV = SI->getOperand(2);
1860 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1861 // Bool selects with constant operands can be folded to logical ops.
1862 if (SI->getType() == Type::Int1Ty) return 0;
1864 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1865 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1867 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1874 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1875 /// node as operand #0, see if we can fold the instruction into the PHI (which
1876 /// is only possible if all operands to the PHI are constants).
1877 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1878 PHINode *PN = cast<PHINode>(I.getOperand(0));
1879 unsigned NumPHIValues = PN->getNumIncomingValues();
1880 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1882 // Check to see if all of the operands of the PHI are constants. If there is
1883 // one non-constant value, remember the BB it is. If there is more than one
1884 // or if *it* is a PHI, bail out.
1885 BasicBlock *NonConstBB = 0;
1886 for (unsigned i = 0; i != NumPHIValues; ++i)
1887 if (!isa<Constant>(PN->getIncomingValue(i))) {
1888 if (NonConstBB) return 0; // More than one non-const value.
1889 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1890 NonConstBB = PN->getIncomingBlock(i);
1892 // If the incoming non-constant value is in I's block, we have an infinite
1894 if (NonConstBB == I.getParent())
1898 // If there is exactly one non-constant value, we can insert a copy of the
1899 // operation in that block. However, if this is a critical edge, we would be
1900 // inserting the computation one some other paths (e.g. inside a loop). Only
1901 // do this if the pred block is unconditionally branching into the phi block.
1903 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1904 if (!BI || !BI->isUnconditional()) return 0;
1907 // Okay, we can do the transformation: create the new PHI node.
1908 PHINode *NewPN = PHINode::Create(I.getType(), "");
1909 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1910 InsertNewInstBefore(NewPN, *PN);
1911 NewPN->takeName(PN);
1913 // Next, add all of the operands to the PHI.
1914 if (I.getNumOperands() == 2) {
1915 Constant *C = cast<Constant>(I.getOperand(1));
1916 for (unsigned i = 0; i != NumPHIValues; ++i) {
1918 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1919 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1920 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1922 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1924 assert(PN->getIncomingBlock(i) == NonConstBB);
1925 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1926 InV = BinaryOperator::Create(BO->getOpcode(),
1927 PN->getIncomingValue(i), C, "phitmp",
1928 NonConstBB->getTerminator());
1929 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1930 InV = CmpInst::Create(CI->getOpcode(),
1932 PN->getIncomingValue(i), C, "phitmp",
1933 NonConstBB->getTerminator());
1935 assert(0 && "Unknown binop!");
1937 AddToWorkList(cast<Instruction>(InV));
1939 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1942 CastInst *CI = cast<CastInst>(&I);
1943 const Type *RetTy = CI->getType();
1944 for (unsigned i = 0; i != NumPHIValues; ++i) {
1946 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1947 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1949 assert(PN->getIncomingBlock(i) == NonConstBB);
1950 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1951 I.getType(), "phitmp",
1952 NonConstBB->getTerminator());
1953 AddToWorkList(cast<Instruction>(InV));
1955 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1958 return ReplaceInstUsesWith(I, NewPN);
1962 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1963 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1964 /// This basically requires proving that the add in the original type would not
1965 /// overflow to change the sign bit or have a carry out.
1966 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1967 // There are different heuristics we can use for this. Here are some simple
1970 // Add has the property that adding any two 2's complement numbers can only
1971 // have one carry bit which can change a sign. As such, if LHS and RHS each
1972 // have at least two sign bits, we know that the addition of the two values will
1973 // sign extend fine.
1974 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1978 // If one of the operands only has one non-zero bit, and if the other operand
1979 // has a known-zero bit in a more significant place than it (not including the
1980 // sign bit) the ripple may go up to and fill the zero, but won't change the
1981 // sign. For example, (X & ~4) + 1.
1989 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1990 bool Changed = SimplifyCommutative(I);
1991 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1993 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1994 // X + undef -> undef
1995 if (isa<UndefValue>(RHS))
1996 return ReplaceInstUsesWith(I, RHS);
1999 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2000 if (RHSC->isNullValue())
2001 return ReplaceInstUsesWith(I, LHS);
2002 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2003 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2004 (I.getType())->getValueAPF()))
2005 return ReplaceInstUsesWith(I, LHS);
2008 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2009 // X + (signbit) --> X ^ signbit
2010 const APInt& Val = CI->getValue();
2011 uint32_t BitWidth = Val.getBitWidth();
2012 if (Val == APInt::getSignBit(BitWidth))
2013 return BinaryOperator::CreateXor(LHS, RHS);
2015 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2016 // (X & 254)+1 -> (X&254)|1
2017 if (!isa<VectorType>(I.getType())) {
2018 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2019 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2020 KnownZero, KnownOne))
2024 // zext(i1) - 1 -> select i1, 0, -1
2025 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2026 if (CI->isAllOnesValue() &&
2027 ZI->getOperand(0)->getType() == Type::Int1Ty)
2028 return SelectInst::Create(ZI->getOperand(0),
2029 Constant::getNullValue(I.getType()),
2030 ConstantInt::getAllOnesValue(I.getType()));
2033 if (isa<PHINode>(LHS))
2034 if (Instruction *NV = FoldOpIntoPhi(I))
2037 ConstantInt *XorRHS = 0;
2039 if (isa<ConstantInt>(RHSC) &&
2040 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2041 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2042 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2044 uint32_t Size = TySizeBits / 2;
2045 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2046 APInt CFF80Val(-C0080Val);
2048 if (TySizeBits > Size) {
2049 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2050 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2051 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2052 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2053 // This is a sign extend if the top bits are known zero.
2054 if (!MaskedValueIsZero(XorLHS,
2055 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2056 Size = 0; // Not a sign ext, but can't be any others either.
2061 C0080Val = APIntOps::lshr(C0080Val, Size);
2062 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2063 } while (Size >= 1);
2065 // FIXME: This shouldn't be necessary. When the backends can handle types
2066 // with funny bit widths then this switch statement should be removed. It
2067 // is just here to get the size of the "middle" type back up to something
2068 // that the back ends can handle.
2069 const Type *MiddleType = 0;
2072 case 32: MiddleType = Type::Int32Ty; break;
2073 case 16: MiddleType = Type::Int16Ty; break;
2074 case 8: MiddleType = Type::Int8Ty; break;
2077 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2078 InsertNewInstBefore(NewTrunc, I);
2079 return new SExtInst(NewTrunc, I.getType(), I.getName());
2084 if (I.getType() == Type::Int1Ty)
2085 return BinaryOperator::CreateXor(LHS, RHS);
2088 if (I.getType()->isInteger()) {
2089 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2091 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2092 if (RHSI->getOpcode() == Instruction::Sub)
2093 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2094 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2096 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2097 if (LHSI->getOpcode() == Instruction::Sub)
2098 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2099 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2104 // -A + -B --> -(A + B)
2105 if (Value *LHSV = dyn_castNegVal(LHS)) {
2106 if (LHS->getType()->isIntOrIntVector()) {
2107 if (Value *RHSV = dyn_castNegVal(RHS)) {
2108 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2109 InsertNewInstBefore(NewAdd, I);
2110 return BinaryOperator::CreateNeg(NewAdd);
2114 return BinaryOperator::CreateSub(RHS, LHSV);
2118 if (!isa<Constant>(RHS))
2119 if (Value *V = dyn_castNegVal(RHS))
2120 return BinaryOperator::CreateSub(LHS, V);
2124 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2125 if (X == RHS) // X*C + X --> X * (C+1)
2126 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2128 // X*C1 + X*C2 --> X * (C1+C2)
2130 if (X == dyn_castFoldableMul(RHS, C1))
2131 return BinaryOperator::CreateMul(X, Add(C1, C2));
2134 // X + X*C --> X * (C+1)
2135 if (dyn_castFoldableMul(RHS, C2) == LHS)
2136 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2138 // X + ~X --> -1 since ~X = -X-1
2139 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2140 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2143 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2144 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2145 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2148 // A+B --> A|B iff A and B have no bits set in common.
2149 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2150 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2151 APInt LHSKnownOne(IT->getBitWidth(), 0);
2152 APInt LHSKnownZero(IT->getBitWidth(), 0);
2153 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2154 if (LHSKnownZero != 0) {
2155 APInt RHSKnownOne(IT->getBitWidth(), 0);
2156 APInt RHSKnownZero(IT->getBitWidth(), 0);
2157 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2159 // No bits in common -> bitwise or.
2160 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2161 return BinaryOperator::CreateOr(LHS, RHS);
2165 // W*X + Y*Z --> W * (X+Z) iff W == Y
2166 if (I.getType()->isIntOrIntVector()) {
2167 Value *W, *X, *Y, *Z;
2168 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2169 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2173 } else if (Y == X) {
2175 } else if (X == Z) {
2182 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2183 LHS->getName()), I);
2184 return BinaryOperator::CreateMul(W, NewAdd);
2189 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2191 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2192 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2194 // (X & FF00) + xx00 -> (X+xx00) & FF00
2195 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2196 Constant *Anded = And(CRHS, C2);
2197 if (Anded == CRHS) {
2198 // See if all bits from the first bit set in the Add RHS up are included
2199 // in the mask. First, get the rightmost bit.
2200 const APInt& AddRHSV = CRHS->getValue();
2202 // Form a mask of all bits from the lowest bit added through the top.
2203 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2205 // See if the and mask includes all of these bits.
2206 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2208 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2209 // Okay, the xform is safe. Insert the new add pronto.
2210 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2211 LHS->getName()), I);
2212 return BinaryOperator::CreateAnd(NewAdd, C2);
2217 // Try to fold constant add into select arguments.
2218 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2219 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2223 // add (cast *A to intptrtype) B ->
2224 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2226 CastInst *CI = dyn_cast<CastInst>(LHS);
2229 CI = dyn_cast<CastInst>(RHS);
2232 if (CI && CI->getType()->isSized() &&
2233 (CI->getType()->getPrimitiveSizeInBits() ==
2234 TD->getIntPtrType()->getPrimitiveSizeInBits())
2235 && isa<PointerType>(CI->getOperand(0)->getType())) {
2237 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2238 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2239 PointerType::get(Type::Int8Ty, AS), I);
2240 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2241 return new PtrToIntInst(I2, CI->getType());
2245 // add (select X 0 (sub n A)) A --> select X A n
2247 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2250 SI = dyn_cast<SelectInst>(RHS);
2253 if (SI && SI->hasOneUse()) {
2254 Value *TV = SI->getTrueValue();
2255 Value *FV = SI->getFalseValue();
2258 // Can we fold the add into the argument of the select?
2259 // We check both true and false select arguments for a matching subtract.
2260 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2261 // Fold the add into the true select value.
2262 return SelectInst::Create(SI->getCondition(), N, A);
2263 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2264 // Fold the add into the false select value.
2265 return SelectInst::Create(SI->getCondition(), A, N);
2269 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2270 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2271 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2272 return ReplaceInstUsesWith(I, LHS);
2274 // Check for (add (sext x), y), see if we can merge this into an
2275 // integer add followed by a sext.
2276 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2277 // (add (sext x), cst) --> (sext (add x, cst'))
2278 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2280 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2281 if (LHSConv->hasOneUse() &&
2282 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2283 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2284 // Insert the new, smaller add.
2285 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2287 InsertNewInstBefore(NewAdd, I);
2288 return new SExtInst(NewAdd, I.getType());
2292 // (add (sext x), (sext y)) --> (sext (add int x, y))
2293 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2294 // Only do this if x/y have the same type, if at last one of them has a
2295 // single use (so we don't increase the number of sexts), and if the
2296 // integer add will not overflow.
2297 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2298 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2299 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2300 RHSConv->getOperand(0))) {
2301 // Insert the new integer add.
2302 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2303 RHSConv->getOperand(0),
2305 InsertNewInstBefore(NewAdd, I);
2306 return new SExtInst(NewAdd, I.getType());
2311 // Check for (add double (sitofp x), y), see if we can merge this into an
2312 // integer add followed by a promotion.
2313 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2314 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2315 // ... if the constant fits in the integer value. This is useful for things
2316 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2317 // requires a constant pool load, and generally allows the add to be better
2319 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2321 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2322 if (LHSConv->hasOneUse() &&
2323 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2324 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2325 // Insert the new integer add.
2326 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2328 InsertNewInstBefore(NewAdd, I);
2329 return new SIToFPInst(NewAdd, I.getType());
2333 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2334 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2335 // Only do this if x/y have the same type, if at last one of them has a
2336 // single use (so we don't increase the number of int->fp conversions),
2337 // and if the integer add will not overflow.
2338 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2339 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2340 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2341 RHSConv->getOperand(0))) {
2342 // Insert the new integer add.
2343 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2344 RHSConv->getOperand(0),
2346 InsertNewInstBefore(NewAdd, I);
2347 return new SIToFPInst(NewAdd, I.getType());
2352 return Changed ? &I : 0;
2355 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2356 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2358 if (Op0 == Op1 && // sub X, X -> 0
2359 !I.getType()->isFPOrFPVector())
2360 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2362 // If this is a 'B = x-(-A)', change to B = x+A...
2363 if (Value *V = dyn_castNegVal(Op1))
2364 return BinaryOperator::CreateAdd(Op0, V);
2366 if (isa<UndefValue>(Op0))
2367 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2368 if (isa<UndefValue>(Op1))
2369 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2371 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2372 // Replace (-1 - A) with (~A)...
2373 if (C->isAllOnesValue())
2374 return BinaryOperator::CreateNot(Op1);
2376 // C - ~X == X + (1+C)
2378 if (match(Op1, m_Not(m_Value(X))))
2379 return BinaryOperator::CreateAdd(X, AddOne(C));
2381 // -(X >>u 31) -> (X >>s 31)
2382 // -(X >>s 31) -> (X >>u 31)
2384 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2385 if (SI->getOpcode() == Instruction::LShr) {
2386 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2387 // Check to see if we are shifting out everything but the sign bit.
2388 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2389 SI->getType()->getPrimitiveSizeInBits()-1) {
2390 // Ok, the transformation is safe. Insert AShr.
2391 return BinaryOperator::Create(Instruction::AShr,
2392 SI->getOperand(0), CU, SI->getName());
2396 else if (SI->getOpcode() == Instruction::AShr) {
2397 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2398 // Check to see if we are shifting out everything but the sign bit.
2399 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2400 SI->getType()->getPrimitiveSizeInBits()-1) {
2401 // Ok, the transformation is safe. Insert LShr.
2402 return BinaryOperator::CreateLShr(
2403 SI->getOperand(0), CU, SI->getName());
2410 // Try to fold constant sub into select arguments.
2411 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2412 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2415 if (isa<PHINode>(Op0))
2416 if (Instruction *NV = FoldOpIntoPhi(I))
2420 if (I.getType() == Type::Int1Ty)
2421 return BinaryOperator::CreateXor(Op0, Op1);
2423 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2424 if (Op1I->getOpcode() == Instruction::Add &&
2425 !Op0->getType()->isFPOrFPVector()) {
2426 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2427 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2428 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2429 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2430 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2431 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2432 // C1-(X+C2) --> (C1-C2)-X
2433 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2434 Op1I->getOperand(0));
2438 if (Op1I->hasOneUse()) {
2439 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2440 // is not used by anyone else...
2442 if (Op1I->getOpcode() == Instruction::Sub &&
2443 !Op1I->getType()->isFPOrFPVector()) {
2444 // Swap the two operands of the subexpr...
2445 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2446 Op1I->setOperand(0, IIOp1);
2447 Op1I->setOperand(1, IIOp0);
2449 // Create the new top level add instruction...
2450 return BinaryOperator::CreateAdd(Op0, Op1);
2453 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2455 if (Op1I->getOpcode() == Instruction::And &&
2456 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2457 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2460 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2461 return BinaryOperator::CreateAnd(Op0, NewNot);
2464 // 0 - (X sdiv C) -> (X sdiv -C)
2465 if (Op1I->getOpcode() == Instruction::SDiv)
2466 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2468 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2469 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2470 ConstantExpr::getNeg(DivRHS));
2472 // X - X*C --> X * (1-C)
2473 ConstantInt *C2 = 0;
2474 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2475 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2476 return BinaryOperator::CreateMul(Op0, CP1);
2479 // X - ((X / Y) * Y) --> X % Y
2480 if (Op1I->getOpcode() == Instruction::Mul)
2481 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2482 if (Op0 == I->getOperand(0) &&
2483 Op1I->getOperand(1) == I->getOperand(1)) {
2484 if (I->getOpcode() == Instruction::SDiv)
2485 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2486 if (I->getOpcode() == Instruction::UDiv)
2487 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2492 if (!Op0->getType()->isFPOrFPVector())
2493 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2494 if (Op0I->getOpcode() == Instruction::Add) {
2495 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2496 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2497 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2498 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2499 } else if (Op0I->getOpcode() == Instruction::Sub) {
2500 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2501 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2506 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2507 if (X == Op1) // X*C - X --> X * (C-1)
2508 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2510 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2511 if (X == dyn_castFoldableMul(Op1, C2))
2512 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2517 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2518 /// comparison only checks the sign bit. If it only checks the sign bit, set
2519 /// TrueIfSigned if the result of the comparison is true when the input value is
2521 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2522 bool &TrueIfSigned) {
2524 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2525 TrueIfSigned = true;
2526 return RHS->isZero();
2527 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2528 TrueIfSigned = true;
2529 return RHS->isAllOnesValue();
2530 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2531 TrueIfSigned = false;
2532 return RHS->isAllOnesValue();
2533 case ICmpInst::ICMP_UGT:
2534 // True if LHS u> RHS and RHS == high-bit-mask - 1
2535 TrueIfSigned = true;
2536 return RHS->getValue() ==
2537 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2538 case ICmpInst::ICMP_UGE:
2539 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2540 TrueIfSigned = true;
2541 return RHS->getValue().isSignBit();
2547 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2548 bool Changed = SimplifyCommutative(I);
2549 Value *Op0 = I.getOperand(0);
2551 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2552 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2554 // Simplify mul instructions with a constant RHS...
2555 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2556 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2558 // ((X << C1)*C2) == (X * (C2 << C1))
2559 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2560 if (SI->getOpcode() == Instruction::Shl)
2561 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2562 return BinaryOperator::CreateMul(SI->getOperand(0),
2563 ConstantExpr::getShl(CI, ShOp));
2566 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2567 if (CI->equalsInt(1)) // X * 1 == X
2568 return ReplaceInstUsesWith(I, Op0);
2569 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2570 return BinaryOperator::CreateNeg(Op0, I.getName());
2572 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2573 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2574 return BinaryOperator::CreateShl(Op0,
2575 ConstantInt::get(Op0->getType(), Val.logBase2()));
2577 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2578 if (Op1F->isNullValue())
2579 return ReplaceInstUsesWith(I, Op1);
2581 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2582 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2583 if (Op1F->isExactlyValue(1.0))
2584 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2585 } else if (isa<VectorType>(Op1->getType())) {
2586 if (isa<ConstantAggregateZero>(Op1))
2587 return ReplaceInstUsesWith(I, Op1);
2589 // As above, vector X*splat(1.0) -> X in all defined cases.
2590 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1))
2591 if (ConstantFP *F = dyn_cast_or_null<ConstantFP>(Op1V->getSplatValue()))
2592 if (F->isExactlyValue(1.0))
2593 return ReplaceInstUsesWith(I, Op0);
2596 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2597 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2598 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2599 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2600 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2602 InsertNewInstBefore(Add, I);
2603 Value *C1C2 = ConstantExpr::getMul(Op1,
2604 cast<Constant>(Op0I->getOperand(1)));
2605 return BinaryOperator::CreateAdd(Add, C1C2);
2609 // Try to fold constant mul into select arguments.
2610 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2611 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2614 if (isa<PHINode>(Op0))
2615 if (Instruction *NV = FoldOpIntoPhi(I))
2619 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2620 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2621 return BinaryOperator::CreateMul(Op0v, Op1v);
2623 if (I.getType() == Type::Int1Ty)
2624 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2626 // If one of the operands of the multiply is a cast from a boolean value, then
2627 // we know the bool is either zero or one, so this is a 'masking' multiply.
2628 // See if we can simplify things based on how the boolean was originally
2630 CastInst *BoolCast = 0;
2631 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2632 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2635 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2636 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2639 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2640 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2641 const Type *SCOpTy = SCIOp0->getType();
2644 // If the icmp is true iff the sign bit of X is set, then convert this
2645 // multiply into a shift/and combination.
2646 if (isa<ConstantInt>(SCIOp1) &&
2647 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2649 // Shift the X value right to turn it into "all signbits".
2650 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2651 SCOpTy->getPrimitiveSizeInBits()-1);
2653 InsertNewInstBefore(
2654 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2655 BoolCast->getOperand(0)->getName()+
2658 // If the multiply type is not the same as the source type, sign extend
2659 // or truncate to the multiply type.
2660 if (I.getType() != V->getType()) {
2661 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2662 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2663 Instruction::CastOps opcode =
2664 (SrcBits == DstBits ? Instruction::BitCast :
2665 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2666 V = InsertCastBefore(opcode, V, I.getType(), I);
2669 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2670 return BinaryOperator::CreateAnd(V, OtherOp);
2675 return Changed ? &I : 0;
2678 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2680 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2681 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2683 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2684 int NonNullOperand = -1;
2685 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2686 if (ST->isNullValue())
2688 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2689 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2690 if (ST->isNullValue())
2693 if (NonNullOperand == -1)
2696 Value *SelectCond = SI->getOperand(0);
2698 // Change the div/rem to use 'Y' instead of the select.
2699 I.setOperand(1, SI->getOperand(NonNullOperand));
2701 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2702 // problem. However, the select, or the condition of the select may have
2703 // multiple uses. Based on our knowledge that the operand must be non-zero,
2704 // propagate the known value for the select into other uses of it, and
2705 // propagate a known value of the condition into its other users.
2707 // If the select and condition only have a single use, don't bother with this,
2709 if (SI->use_empty() && SelectCond->hasOneUse())
2712 // Scan the current block backward, looking for other uses of SI.
2713 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2715 while (BBI != BBFront) {
2717 // If we found a call to a function, we can't assume it will return, so
2718 // information from below it cannot be propagated above it.
2719 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2722 // Replace uses of the select or its condition with the known values.
2723 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2726 *I = SI->getOperand(NonNullOperand);
2728 } else if (*I == SelectCond) {
2729 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2730 ConstantInt::getFalse();
2735 // If we past the instruction, quit looking for it.
2738 if (&*BBI == SelectCond)
2741 // If we ran out of things to eliminate, break out of the loop.
2742 if (SelectCond == 0 && SI == 0)
2750 /// This function implements the transforms on div instructions that work
2751 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2752 /// used by the visitors to those instructions.
2753 /// @brief Transforms common to all three div instructions
2754 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2755 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2757 // undef / X -> 0 for integer.
2758 // undef / X -> undef for FP (the undef could be a snan).
2759 if (isa<UndefValue>(Op0)) {
2760 if (Op0->getType()->isFPOrFPVector())
2761 return ReplaceInstUsesWith(I, Op0);
2762 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2765 // X / undef -> undef
2766 if (isa<UndefValue>(Op1))
2767 return ReplaceInstUsesWith(I, Op1);
2772 /// This function implements the transforms common to both integer division
2773 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2774 /// division instructions.
2775 /// @brief Common integer divide transforms
2776 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2777 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2779 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2781 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2782 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2783 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2784 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2787 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2788 return ReplaceInstUsesWith(I, CI);
2791 if (Instruction *Common = commonDivTransforms(I))
2794 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2795 // This does not apply for fdiv.
2796 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2799 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2801 if (RHS->equalsInt(1))
2802 return ReplaceInstUsesWith(I, Op0);
2804 // (X / C1) / C2 -> X / (C1*C2)
2805 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2806 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2807 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2808 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2809 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2811 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2812 Multiply(RHS, LHSRHS));
2815 if (!RHS->isZero()) { // avoid X udiv 0
2816 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2817 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2819 if (isa<PHINode>(Op0))
2820 if (Instruction *NV = FoldOpIntoPhi(I))
2825 // 0 / X == 0, we don't need to preserve faults!
2826 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2827 if (LHS->equalsInt(0))
2828 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2830 // It can't be division by zero, hence it must be division by one.
2831 if (I.getType() == Type::Int1Ty)
2832 return ReplaceInstUsesWith(I, Op0);
2837 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2838 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2840 // Handle the integer div common cases
2841 if (Instruction *Common = commonIDivTransforms(I))
2844 // X udiv C^2 -> X >> C
2845 // Check to see if this is an unsigned division with an exact power of 2,
2846 // if so, convert to a right shift.
2847 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2848 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2849 return BinaryOperator::CreateLShr(Op0,
2850 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2853 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2854 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2855 if (RHSI->getOpcode() == Instruction::Shl &&
2856 isa<ConstantInt>(RHSI->getOperand(0))) {
2857 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2858 if (C1.isPowerOf2()) {
2859 Value *N = RHSI->getOperand(1);
2860 const Type *NTy = N->getType();
2861 if (uint32_t C2 = C1.logBase2()) {
2862 Constant *C2V = ConstantInt::get(NTy, C2);
2863 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2865 return BinaryOperator::CreateLShr(Op0, N);
2870 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2871 // where C1&C2 are powers of two.
2872 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2873 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2874 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2875 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2876 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2877 // Compute the shift amounts
2878 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2879 // Construct the "on true" case of the select
2880 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2881 Instruction *TSI = BinaryOperator::CreateLShr(
2882 Op0, TC, SI->getName()+".t");
2883 TSI = InsertNewInstBefore(TSI, I);
2885 // Construct the "on false" case of the select
2886 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2887 Instruction *FSI = BinaryOperator::CreateLShr(
2888 Op0, FC, SI->getName()+".f");
2889 FSI = InsertNewInstBefore(FSI, I);
2891 // construct the select instruction and return it.
2892 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2898 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2899 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2901 // Handle the integer div common cases
2902 if (Instruction *Common = commonIDivTransforms(I))
2905 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2907 if (RHS->isAllOnesValue())
2908 return BinaryOperator::CreateNeg(Op0);
2911 if (Value *LHSNeg = dyn_castNegVal(Op0))
2912 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2915 // If the sign bits of both operands are zero (i.e. we can prove they are
2916 // unsigned inputs), turn this into a udiv.
2917 if (I.getType()->isInteger()) {
2918 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2919 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2920 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2921 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2928 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2929 return commonDivTransforms(I);
2932 /// This function implements the transforms on rem instructions that work
2933 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2934 /// is used by the visitors to those instructions.
2935 /// @brief Transforms common to all three rem instructions
2936 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2937 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2939 // 0 % X == 0 for integer, we don't need to preserve faults!
2940 if (Constant *LHS = dyn_cast<Constant>(Op0))
2941 if (LHS->isNullValue())
2942 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2944 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2945 if (I.getType()->isFPOrFPVector())
2946 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2947 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2949 if (isa<UndefValue>(Op1))
2950 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2952 // Handle cases involving: rem X, (select Cond, Y, Z)
2953 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2959 /// This function implements the transforms common to both integer remainder
2960 /// instructions (urem and srem). It is called by the visitors to those integer
2961 /// remainder instructions.
2962 /// @brief Common integer remainder transforms
2963 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2964 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2966 if (Instruction *common = commonRemTransforms(I))
2969 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2970 // X % 0 == undef, we don't need to preserve faults!
2971 if (RHS->equalsInt(0))
2972 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2974 if (RHS->equalsInt(1)) // X % 1 == 0
2975 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2977 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2978 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2979 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2981 } else if (isa<PHINode>(Op0I)) {
2982 if (Instruction *NV = FoldOpIntoPhi(I))
2986 // See if we can fold away this rem instruction.
2987 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2988 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2989 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2990 KnownZero, KnownOne))
2998 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2999 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3001 if (Instruction *common = commonIRemTransforms(I))
3004 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3005 // X urem C^2 -> X and C
3006 // Check to see if this is an unsigned remainder with an exact power of 2,
3007 // if so, convert to a bitwise and.
3008 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3009 if (C->getValue().isPowerOf2())
3010 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3013 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3014 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3015 if (RHSI->getOpcode() == Instruction::Shl &&
3016 isa<ConstantInt>(RHSI->getOperand(0))) {
3017 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3018 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3019 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3021 return BinaryOperator::CreateAnd(Op0, Add);
3026 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3027 // where C1&C2 are powers of two.
3028 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3029 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3030 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3031 // STO == 0 and SFO == 0 handled above.
3032 if ((STO->getValue().isPowerOf2()) &&
3033 (SFO->getValue().isPowerOf2())) {
3034 Value *TrueAnd = InsertNewInstBefore(
3035 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3036 Value *FalseAnd = InsertNewInstBefore(
3037 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3038 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3046 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3047 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3049 // Handle the integer rem common cases
3050 if (Instruction *common = commonIRemTransforms(I))
3053 if (Value *RHSNeg = dyn_castNegVal(Op1))
3054 if (!isa<Constant>(RHSNeg) ||
3055 (isa<ConstantInt>(RHSNeg) &&
3056 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3058 AddUsesToWorkList(I);
3059 I.setOperand(1, RHSNeg);
3063 // If the sign bits of both operands are zero (i.e. we can prove they are
3064 // unsigned inputs), turn this into a urem.
3065 if (I.getType()->isInteger()) {
3066 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3067 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3068 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3069 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3076 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3077 return commonRemTransforms(I);
3080 // isOneBitSet - Return true if there is exactly one bit set in the specified
3082 static bool isOneBitSet(const ConstantInt *CI) {
3083 return CI->getValue().isPowerOf2();
3086 // isHighOnes - Return true if the constant is of the form 1+0+.
3087 // This is the same as lowones(~X).
3088 static bool isHighOnes(const ConstantInt *CI) {
3089 return (~CI->getValue() + 1).isPowerOf2();
3092 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3093 /// are carefully arranged to allow folding of expressions such as:
3095 /// (A < B) | (A > B) --> (A != B)
3097 /// Note that this is only valid if the first and second predicates have the
3098 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3100 /// Three bits are used to represent the condition, as follows:
3105 /// <=> Value Definition
3106 /// 000 0 Always false
3113 /// 111 7 Always true
3115 static unsigned getICmpCode(const ICmpInst *ICI) {
3116 switch (ICI->getPredicate()) {
3118 case ICmpInst::ICMP_UGT: return 1; // 001
3119 case ICmpInst::ICMP_SGT: return 1; // 001
3120 case ICmpInst::ICMP_EQ: return 2; // 010
3121 case ICmpInst::ICMP_UGE: return 3; // 011
3122 case ICmpInst::ICMP_SGE: return 3; // 011
3123 case ICmpInst::ICMP_ULT: return 4; // 100
3124 case ICmpInst::ICMP_SLT: return 4; // 100
3125 case ICmpInst::ICMP_NE: return 5; // 101
3126 case ICmpInst::ICMP_ULE: return 6; // 110
3127 case ICmpInst::ICMP_SLE: return 6; // 110
3130 assert(0 && "Invalid ICmp predicate!");
3135 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3136 /// predicate into a three bit mask. It also returns whether it is an ordered
3137 /// predicate by reference.
3138 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3141 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3142 case FCmpInst::FCMP_UNO: return 0; // 000
3143 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3144 case FCmpInst::FCMP_UGT: return 1; // 001
3145 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3146 case FCmpInst::FCMP_UEQ: return 2; // 010
3147 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3148 case FCmpInst::FCMP_UGE: return 3; // 011
3149 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3150 case FCmpInst::FCMP_ULT: return 4; // 100
3151 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3152 case FCmpInst::FCMP_UNE: return 5; // 101
3153 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3154 case FCmpInst::FCMP_ULE: return 6; // 110
3157 // Not expecting FCMP_FALSE and FCMP_TRUE;
3158 assert(0 && "Unexpected FCmp predicate!");
3163 /// getICmpValue - This is the complement of getICmpCode, which turns an
3164 /// opcode and two operands into either a constant true or false, or a brand
3165 /// new ICmp instruction. The sign is passed in to determine which kind
3166 /// of predicate to use in the new icmp instruction.
3167 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3169 default: assert(0 && "Illegal ICmp code!");
3170 case 0: return ConstantInt::getFalse();
3173 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3175 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3176 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3179 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3181 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3184 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3186 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3187 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3190 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3192 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3193 case 7: return ConstantInt::getTrue();
3197 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3198 /// opcode and two operands into either a FCmp instruction. isordered is passed
3199 /// in to determine which kind of predicate to use in the new fcmp instruction.
3200 static Value *getFCmpValue(bool isordered, unsigned code,
3201 Value *LHS, Value *RHS) {
3203 default: assert(0 && "Illegal FCmp code!");
3206 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3208 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3211 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3213 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3216 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3218 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3221 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3223 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3226 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3228 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3231 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3233 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3236 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3238 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3239 case 7: return ConstantInt::getTrue();
3244 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3245 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3246 (ICmpInst::isSignedPredicate(p1) &&
3247 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3248 (ICmpInst::isSignedPredicate(p2) &&
3249 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
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 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3557 bool Changed = SimplifyCommutative(I);
3558 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3560 if (isa<UndefValue>(Op1)) // X & undef -> 0
3561 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3565 return ReplaceInstUsesWith(I, Op1);
3567 // See if we can simplify any instructions used by the instruction whose sole
3568 // purpose is to compute bits we don't care about.
3569 if (!isa<VectorType>(I.getType())) {
3570 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3571 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3572 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3573 KnownZero, KnownOne))
3576 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3577 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3578 return ReplaceInstUsesWith(I, I.getOperand(0));
3579 } else if (isa<ConstantAggregateZero>(Op1)) {
3580 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3584 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3585 const APInt& AndRHSMask = AndRHS->getValue();
3586 APInt NotAndRHS(~AndRHSMask);
3588 // Optimize a variety of ((val OP C1) & C2) combinations...
3589 if (isa<BinaryOperator>(Op0)) {
3590 Instruction *Op0I = cast<Instruction>(Op0);
3591 Value *Op0LHS = Op0I->getOperand(0);
3592 Value *Op0RHS = Op0I->getOperand(1);
3593 switch (Op0I->getOpcode()) {
3594 case Instruction::Xor:
3595 case Instruction::Or:
3596 // If the mask is only needed on one incoming arm, push it up.
3597 if (Op0I->hasOneUse()) {
3598 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3599 // Not masking anything out for the LHS, move to RHS.
3600 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3601 Op0RHS->getName()+".masked");
3602 InsertNewInstBefore(NewRHS, I);
3603 return BinaryOperator::Create(
3604 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3606 if (!isa<Constant>(Op0RHS) &&
3607 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3608 // Not masking anything out for the RHS, move to LHS.
3609 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3610 Op0LHS->getName()+".masked");
3611 InsertNewInstBefore(NewLHS, I);
3612 return BinaryOperator::Create(
3613 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3618 case Instruction::Add:
3619 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3620 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3621 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3622 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3623 return BinaryOperator::CreateAnd(V, AndRHS);
3624 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3625 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3628 case Instruction::Sub:
3629 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3630 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3631 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3632 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3633 return BinaryOperator::CreateAnd(V, AndRHS);
3635 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3636 // has 1's for all bits that the subtraction with A might affect.
3637 if (Op0I->hasOneUse()) {
3638 uint32_t BitWidth = AndRHSMask.getBitWidth();
3639 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3640 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3642 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3643 if (!(A && A->isZero()) && // avoid infinite recursion.
3644 MaskedValueIsZero(Op0LHS, Mask)) {
3645 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3646 InsertNewInstBefore(NewNeg, I);
3647 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3652 case Instruction::Shl:
3653 case Instruction::LShr:
3654 // (1 << x) & 1 --> zext(x == 0)
3655 // (1 >> x) & 1 --> zext(x == 0)
3656 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3657 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3658 Constant::getNullValue(I.getType()));
3659 InsertNewInstBefore(NewICmp, I);
3660 return new ZExtInst(NewICmp, I.getType());
3665 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3666 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3668 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3669 // If this is an integer truncation or change from signed-to-unsigned, and
3670 // if the source is an and/or with immediate, transform it. This
3671 // frequently occurs for bitfield accesses.
3672 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3673 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3674 CastOp->getNumOperands() == 2)
3675 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3676 if (CastOp->getOpcode() == Instruction::And) {
3677 // Change: and (cast (and X, C1) to T), C2
3678 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3679 // This will fold the two constants together, which may allow
3680 // other simplifications.
3681 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3682 CastOp->getOperand(0), I.getType(),
3683 CastOp->getName()+".shrunk");
3684 NewCast = InsertNewInstBefore(NewCast, I);
3685 // trunc_or_bitcast(C1)&C2
3686 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3687 C3 = ConstantExpr::getAnd(C3, AndRHS);
3688 return BinaryOperator::CreateAnd(NewCast, C3);
3689 } else if (CastOp->getOpcode() == Instruction::Or) {
3690 // Change: and (cast (or X, C1) to T), C2
3691 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3692 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3693 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3694 return ReplaceInstUsesWith(I, AndRHS);
3700 // Try to fold constant and into select arguments.
3701 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3702 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3704 if (isa<PHINode>(Op0))
3705 if (Instruction *NV = FoldOpIntoPhi(I))
3709 Value *Op0NotVal = dyn_castNotVal(Op0);
3710 Value *Op1NotVal = dyn_castNotVal(Op1);
3712 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3713 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3715 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3716 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3717 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3718 I.getName()+".demorgan");
3719 InsertNewInstBefore(Or, I);
3720 return BinaryOperator::CreateNot(Or);
3724 Value *A = 0, *B = 0, *C = 0, *D = 0;
3725 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3726 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3727 return ReplaceInstUsesWith(I, Op1);
3729 // (A|B) & ~(A&B) -> A^B
3730 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3731 if ((A == C && B == D) || (A == D && B == C))
3732 return BinaryOperator::CreateXor(A, B);
3736 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3737 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3738 return ReplaceInstUsesWith(I, Op0);
3740 // ~(A&B) & (A|B) -> A^B
3741 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3742 if ((A == C && B == D) || (A == D && B == C))
3743 return BinaryOperator::CreateXor(A, B);
3747 if (Op0->hasOneUse() &&
3748 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3749 if (A == Op1) { // (A^B)&A -> A&(A^B)
3750 I.swapOperands(); // Simplify below
3751 std::swap(Op0, Op1);
3752 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3753 cast<BinaryOperator>(Op0)->swapOperands();
3754 I.swapOperands(); // Simplify below
3755 std::swap(Op0, Op1);
3758 if (Op1->hasOneUse() &&
3759 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3760 if (B == Op0) { // B&(A^B) -> B&(B^A)
3761 cast<BinaryOperator>(Op1)->swapOperands();
3764 if (A == Op0) { // A&(A^B) -> A & ~B
3765 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3766 InsertNewInstBefore(NotB, I);
3767 return BinaryOperator::CreateAnd(A, NotB);
3772 { // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3773 // where C is a power of 2
3775 ConstantInt *C1, *C2;
3776 ICmpInst::Predicate LHSCC = ICmpInst::BAD_ICMP_PREDICATE;
3777 ICmpInst::Predicate RHSCC = ICmpInst::BAD_ICMP_PREDICATE;
3778 if (match(&I, m_And(m_ICmp(LHSCC, m_Value(A), m_ConstantInt(C1)),
3779 m_ICmp(RHSCC, m_Value(B), m_ConstantInt(C2)))))
3780 if (C1 == C2 && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3781 C1->getValue().isPowerOf2()) {
3782 Instruction *NewOr = BinaryOperator::CreateOr(A, B);
3783 InsertNewInstBefore(NewOr, I);
3784 return new ICmpInst(LHSCC, NewOr, C1);
3788 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3789 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3790 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3794 ConstantInt *LHSCst, *RHSCst;
3795 ICmpInst::Predicate LHSCC, RHSCC;
3796 if (match(Op0, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))))
3797 if (match(RHS, m_ICmp(RHSCC, m_Specific(Val), m_ConstantInt(RHSCst))))
3798 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3799 if (LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3800 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3801 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3802 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3804 // Don't try to fold ICMP_SLT + ICMP_ULT.
3805 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3806 ICmpInst::isSignedPredicate(LHSCC) ==
3807 ICmpInst::isSignedPredicate(RHSCC))) {
3808 // Ensure that the larger constant is on the RHS.
3809 ICmpInst::Predicate GT;
3810 if (ICmpInst::isSignedPredicate(LHSCC) ||
3811 (ICmpInst::isEquality(LHSCC) &&
3812 ICmpInst::isSignedPredicate(RHSCC)))
3813 GT = ICmpInst::ICMP_SGT;
3815 GT = ICmpInst::ICMP_UGT;
3817 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3818 ICmpInst *LHS = cast<ICmpInst>(Op0);
3819 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3820 std::swap(LHS, RHS);
3821 std::swap(LHSCst, RHSCst);
3822 std::swap(LHSCC, RHSCC);
3825 // At this point, we know we have have two icmp instructions
3826 // comparing a value against two constants and and'ing the result
3827 // together. Because of the above check, we know that we only have
3828 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3829 // (from the FoldICmpLogical check above), that the two constants
3830 // are not equal and that the larger constant is on the RHS
3831 assert(LHSCst != RHSCst && "Compares not folded above?");
3834 default: assert(0 && "Unknown integer condition code!");
3835 case ICmpInst::ICMP_EQ:
3837 default: assert(0 && "Unknown integer condition code!");
3838 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3839 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3840 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3841 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3842 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3843 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3844 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3845 return ReplaceInstUsesWith(I, LHS);
3847 case ICmpInst::ICMP_NE:
3849 default: assert(0 && "Unknown integer condition code!");
3850 case ICmpInst::ICMP_ULT:
3851 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3852 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3853 break; // (X != 13 & X u< 15) -> no change
3854 case ICmpInst::ICMP_SLT:
3855 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3856 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3857 break; // (X != 13 & X s< 15) -> no change
3858 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3859 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3860 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3861 return ReplaceInstUsesWith(I, RHS);
3862 case ICmpInst::ICMP_NE:
3863 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3864 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3865 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3866 Val->getName()+".off");
3867 InsertNewInstBefore(Add, I);
3868 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3869 ConstantInt::get(Add->getType(), 1));
3871 break; // (X != 13 & X != 15) -> no change
3874 case ICmpInst::ICMP_ULT:
3876 default: assert(0 && "Unknown integer condition code!");
3877 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3878 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3879 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3880 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3882 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3883 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3884 return ReplaceInstUsesWith(I, LHS);
3885 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3889 case ICmpInst::ICMP_SLT:
3891 default: assert(0 && "Unknown integer condition code!");
3892 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3893 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3894 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3895 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3897 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3898 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3899 return ReplaceInstUsesWith(I, LHS);
3900 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3904 case ICmpInst::ICMP_UGT:
3906 default: assert(0 && "Unknown integer condition code!");
3907 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3908 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3909 return ReplaceInstUsesWith(I, RHS);
3910 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3912 case ICmpInst::ICMP_NE:
3913 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3914 return new ICmpInst(LHSCC, Val, RHSCst);
3915 break; // (X u> 13 & X != 15) -> no change
3916 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3917 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false,
3919 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3923 case ICmpInst::ICMP_SGT:
3925 default: assert(0 && "Unknown integer condition code!");
3926 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3927 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3928 return ReplaceInstUsesWith(I, RHS);
3929 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3931 case ICmpInst::ICMP_NE:
3932 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3933 return new ICmpInst(LHSCC, Val, RHSCst);
3934 break; // (X s> 13 & X != 15) -> no change
3935 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3936 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true,I);
3937 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3945 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3946 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3947 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3948 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3949 const Type *SrcTy = Op0C->getOperand(0)->getType();
3950 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3951 // Only do this if the casts both really cause code to be generated.
3952 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3954 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3956 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3957 Op1C->getOperand(0),
3959 InsertNewInstBefore(NewOp, I);
3960 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3964 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3965 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3966 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3967 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3968 SI0->getOperand(1) == SI1->getOperand(1) &&
3969 (SI0->hasOneUse() || SI1->hasOneUse())) {
3970 Instruction *NewOp =
3971 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3973 SI0->getName()), I);
3974 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3975 SI1->getOperand(1));
3979 // If and'ing two fcmp, try combine them into one.
3980 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3981 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3982 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3983 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3984 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3985 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3986 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3987 // If either of the constants are nans, then the whole thing returns
3989 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3990 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3991 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3992 RHS->getOperand(0));
3995 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
3996 FCmpInst::Predicate Op0CC, Op1CC;
3997 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
3998 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
3999 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4000 // Swap RHS operands to match LHS.
4001 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4002 std::swap(Op1LHS, Op1RHS);
4004 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4005 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4007 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4008 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4009 Op1CC == FCmpInst::FCMP_FALSE)
4010 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4011 else if (Op0CC == FCmpInst::FCMP_TRUE)
4012 return ReplaceInstUsesWith(I, Op1);
4013 else if (Op1CC == FCmpInst::FCMP_TRUE)
4014 return ReplaceInstUsesWith(I, Op0);
4017 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4018 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4020 std::swap(Op0, Op1);
4021 std::swap(Op0Pred, Op1Pred);
4022 std::swap(Op0Ordered, Op1Ordered);
4025 // uno && ueq -> uno && (uno || eq) -> ueq
4026 // ord && olt -> ord && (ord && lt) -> olt
4027 if (Op0Ordered == Op1Ordered)
4028 return ReplaceInstUsesWith(I, Op1);
4029 // uno && oeq -> uno && (ord && eq) -> false
4030 // uno && ord -> false
4032 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4033 // ord && ueq -> ord && (uno || eq) -> oeq
4034 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4043 return Changed ? &I : 0;
4046 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4047 /// capable of providing pieces of a bswap. The subexpression provides pieces
4048 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4049 /// the expression came from the corresponding "byte swapped" byte in some other
4050 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4051 /// we know that the expression deposits the low byte of %X into the high byte
4052 /// of the bswap result and that all other bytes are zero. This expression is
4053 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4056 /// This function returns true if the match was unsuccessful and false if so.
4057 /// On entry to the function the "OverallLeftShift" is a signed integer value
4058 /// indicating the number of bytes that the subexpression is later shifted. For
4059 /// example, if the expression is later right shifted by 16 bits, the
4060 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4061 /// byte of ByteValues is actually being set.
4063 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4064 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4065 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4066 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4067 /// always in the local (OverallLeftShift) coordinate space.
4069 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4070 SmallVector<Value*, 8> &ByteValues) {
4071 if (Instruction *I = dyn_cast<Instruction>(V)) {
4072 // If this is an or instruction, it may be an inner node of the bswap.
4073 if (I->getOpcode() == Instruction::Or) {
4074 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4076 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4080 // If this is a logical shift by a constant multiple of 8, recurse with
4081 // OverallLeftShift and ByteMask adjusted.
4082 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4084 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4085 // Ensure the shift amount is defined and of a byte value.
4086 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4089 unsigned ByteShift = ShAmt >> 3;
4090 if (I->getOpcode() == Instruction::Shl) {
4091 // X << 2 -> collect(X, +2)
4092 OverallLeftShift += ByteShift;
4093 ByteMask >>= ByteShift;
4095 // X >>u 2 -> collect(X, -2)
4096 OverallLeftShift -= ByteShift;
4097 ByteMask <<= ByteShift;
4098 ByteMask &= (~0U >> (32-ByteValues.size()));
4101 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4102 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4104 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4108 // If this is a logical 'and' with a mask that clears bytes, clear the
4109 // corresponding bytes in ByteMask.
4110 if (I->getOpcode() == Instruction::And &&
4111 isa<ConstantInt>(I->getOperand(1))) {
4112 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4113 unsigned NumBytes = ByteValues.size();
4114 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4115 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4117 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4118 // If this byte is masked out by a later operation, we don't care what
4120 if ((ByteMask & (1 << i)) == 0)
4123 // If the AndMask is all zeros for this byte, clear the bit.
4124 APInt MaskB = AndMask & Byte;
4126 ByteMask &= ~(1U << i);
4130 // If the AndMask is not all ones for this byte, it's not a bytezap.
4134 // Otherwise, this byte is kept.
4137 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4142 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4143 // the input value to the bswap. Some observations: 1) if more than one byte
4144 // is demanded from this input, then it could not be successfully assembled
4145 // into a byteswap. At least one of the two bytes would not be aligned with
4146 // their ultimate destination.
4147 if (!isPowerOf2_32(ByteMask)) return true;
4148 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4150 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4151 // is demanded, it needs to go into byte 0 of the result. This means that the
4152 // byte needs to be shifted until it lands in the right byte bucket. The
4153 // shift amount depends on the position: if the byte is coming from the high
4154 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4155 // low part, it must be shifted left.
4156 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4157 if (InputByteNo < ByteValues.size()/2) {
4158 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4161 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4165 // If the destination byte value is already defined, the values are or'd
4166 // together, which isn't a bswap (unless it's an or of the same bits).
4167 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4169 ByteValues[DestByteNo] = V;
4173 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4174 /// If so, insert the new bswap intrinsic and return it.
4175 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4176 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4177 if (!ITy || ITy->getBitWidth() % 16 ||
4178 // ByteMask only allows up to 32-byte values.
4179 ITy->getBitWidth() > 32*8)
4180 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4182 /// ByteValues - For each byte of the result, we keep track of which value
4183 /// defines each byte.
4184 SmallVector<Value*, 8> ByteValues;
4185 ByteValues.resize(ITy->getBitWidth()/8);
4187 // Try to find all the pieces corresponding to the bswap.
4188 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4189 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4192 // Check to see if all of the bytes come from the same value.
4193 Value *V = ByteValues[0];
4194 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4196 // Check to make sure that all of the bytes come from the same value.
4197 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4198 if (ByteValues[i] != V)
4200 const Type *Tys[] = { ITy };
4201 Module *M = I.getParent()->getParent()->getParent();
4202 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4203 return CallInst::Create(F, V);
4206 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4207 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4208 /// we can simplify this expression to "cond ? C : D or B".
4209 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4210 Value *C, Value *D) {
4211 // If A is not a select of -1/0, this cannot match.
4213 if (!match(A, m_SelectCst(m_Value(Cond), -1, 0)))
4216 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4217 if (match(D, m_SelectCst(m_Specific(Cond), 0, -1)))
4218 return SelectInst::Create(Cond, C, B);
4219 if (match(D, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4220 return SelectInst::Create(Cond, C, B);
4221 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4222 if (match(B, m_SelectCst(m_Specific(Cond), 0, -1)))
4223 return SelectInst::Create(Cond, C, D);
4224 if (match(B, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4225 return SelectInst::Create(Cond, C, D);
4229 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4230 bool Changed = SimplifyCommutative(I);
4231 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4233 if (isa<UndefValue>(Op1)) // X | undef -> -1
4234 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4238 return ReplaceInstUsesWith(I, Op0);
4240 // See if we can simplify any instructions used by the instruction whose sole
4241 // purpose is to compute bits we don't care about.
4242 if (!isa<VectorType>(I.getType())) {
4243 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4244 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4245 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4246 KnownZero, KnownOne))
4248 } else if (isa<ConstantAggregateZero>(Op1)) {
4249 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4250 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4251 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4252 return ReplaceInstUsesWith(I, I.getOperand(1));
4258 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4259 ConstantInt *C1 = 0; Value *X = 0;
4260 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4261 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4262 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4263 InsertNewInstBefore(Or, I);
4265 return BinaryOperator::CreateAnd(Or,
4266 ConstantInt::get(RHS->getValue() | C1->getValue()));
4269 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4270 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4271 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4272 InsertNewInstBefore(Or, I);
4274 return BinaryOperator::CreateXor(Or,
4275 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4278 // Try to fold constant and into select arguments.
4279 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4280 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4282 if (isa<PHINode>(Op0))
4283 if (Instruction *NV = FoldOpIntoPhi(I))
4287 Value *A = 0, *B = 0;
4288 ConstantInt *C1 = 0, *C2 = 0;
4290 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4291 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4292 return ReplaceInstUsesWith(I, Op1);
4293 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4294 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4295 return ReplaceInstUsesWith(I, Op0);
4297 // (A | B) | C and A | (B | C) -> bswap if possible.
4298 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4299 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4300 match(Op1, m_Or(m_Value(), m_Value())) ||
4301 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4302 match(Op1, m_Shift(m_Value(), m_Value())))) {
4303 if (Instruction *BSwap = MatchBSwap(I))
4307 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4308 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4309 MaskedValueIsZero(Op1, C1->getValue())) {
4310 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4311 InsertNewInstBefore(NOr, I);
4313 return BinaryOperator::CreateXor(NOr, C1);
4316 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4317 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4318 MaskedValueIsZero(Op0, C1->getValue())) {
4319 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4320 InsertNewInstBefore(NOr, I);
4322 return BinaryOperator::CreateXor(NOr, C1);
4326 Value *C = 0, *D = 0;
4327 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4328 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4329 Value *V1 = 0, *V2 = 0, *V3 = 0;
4330 C1 = dyn_cast<ConstantInt>(C);
4331 C2 = dyn_cast<ConstantInt>(D);
4332 if (C1 && C2) { // (A & C1)|(B & C2)
4333 // If we have: ((V + N) & C1) | (V & C2)
4334 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4335 // replace with V+N.
4336 if (C1->getValue() == ~C2->getValue()) {
4337 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4338 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4339 // Add commutes, try both ways.
4340 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4341 return ReplaceInstUsesWith(I, A);
4342 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4343 return ReplaceInstUsesWith(I, A);
4345 // Or commutes, try both ways.
4346 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4347 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4348 // Add commutes, try both ways.
4349 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4350 return ReplaceInstUsesWith(I, B);
4351 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4352 return ReplaceInstUsesWith(I, B);
4355 V1 = 0; V2 = 0; V3 = 0;
4358 // Check to see if we have any common things being and'ed. If so, find the
4359 // terms for V1 & (V2|V3).
4360 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4361 if (A == B) // (A & C)|(A & D) == A & (C|D)
4362 V1 = A, V2 = C, V3 = D;
4363 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4364 V1 = A, V2 = B, V3 = C;
4365 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4366 V1 = C, V2 = A, V3 = D;
4367 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4368 V1 = C, V2 = A, V3 = B;
4372 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4373 return BinaryOperator::CreateAnd(V1, Or);
4377 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4378 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4380 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4382 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4384 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4388 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4389 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4390 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4391 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4392 SI0->getOperand(1) == SI1->getOperand(1) &&
4393 (SI0->hasOneUse() || SI1->hasOneUse())) {
4394 Instruction *NewOp =
4395 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4397 SI0->getName()), I);
4398 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4399 SI1->getOperand(1));
4403 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4404 if (A == Op1) // ~A | A == -1
4405 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4409 // Note, A is still live here!
4410 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4412 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4414 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4415 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4416 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4417 I.getName()+".demorgan"), I);
4418 return BinaryOperator::CreateNot(And);
4422 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4423 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4424 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4428 ConstantInt *LHSCst, *RHSCst;
4429 ICmpInst::Predicate LHSCC, RHSCC;
4430 if (match(Op0, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))))
4431 if (match(RHS, m_ICmp(RHSCC, m_Specific(Val), m_ConstantInt(RHSCst))))
4432 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4433 if (LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4434 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4435 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4436 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4437 // We can't fold (ugt x, C) | (sgt x, C2).
4438 PredicatesFoldable(LHSCC, RHSCC)) {
4439 // Ensure that the larger constant is on the RHS.
4440 ICmpInst *LHS = cast<ICmpInst>(Op0);
4442 if (ICmpInst::isEquality(LHSCC) ? ICmpInst::isSignedPredicate(RHSCC)
4443 : ICmpInst::isSignedPredicate(LHSCC))
4444 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4446 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4449 std::swap(LHS, RHS);
4450 std::swap(LHSCst, RHSCst);
4451 std::swap(LHSCC, RHSCC);
4454 // At this point, we know we have have two icmp instructions
4455 // comparing a value against two constants and or'ing the result
4456 // together. Because of the above check, we know that we only have
4457 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4458 // FoldICmpLogical check above), that the two constants are not
4460 assert(LHSCst != RHSCst && "Compares not folded above?");
4463 default: assert(0 && "Unknown integer condition code!");
4464 case ICmpInst::ICMP_EQ:
4466 default: assert(0 && "Unknown integer condition code!");
4467 case ICmpInst::ICMP_EQ:
4468 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4469 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4470 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4471 Val->getName()+".off");
4472 InsertNewInstBefore(Add, I);
4473 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4474 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4476 break; // (X == 13 | X == 15) -> no change
4477 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4478 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4480 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4481 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4482 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4483 return ReplaceInstUsesWith(I, RHS);
4486 case ICmpInst::ICMP_NE:
4488 default: assert(0 && "Unknown integer condition code!");
4489 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4490 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4491 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4492 return ReplaceInstUsesWith(I, LHS);
4493 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4494 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4495 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4496 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4499 case ICmpInst::ICMP_ULT:
4501 default: assert(0 && "Unknown integer condition code!");
4502 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4504 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4505 // If RHSCst is [us]MAXINT, it is always false. Not handling
4506 // this can cause overflow.
4507 if (RHSCst->isMaxValue(false))
4508 return ReplaceInstUsesWith(I, LHS);
4509 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false,
4511 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4513 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4514 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4515 return ReplaceInstUsesWith(I, RHS);
4516 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4520 case ICmpInst::ICMP_SLT:
4522 default: assert(0 && "Unknown integer condition code!");
4523 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4525 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4526 // If RHSCst is [us]MAXINT, it is always false. Not handling
4527 // this can cause overflow.
4528 if (RHSCst->isMaxValue(true))
4529 return ReplaceInstUsesWith(I, LHS);
4530 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true,
4532 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4534 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4535 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4536 return ReplaceInstUsesWith(I, RHS);
4537 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4541 case ICmpInst::ICMP_UGT:
4543 default: assert(0 && "Unknown integer condition code!");
4544 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4545 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4546 return ReplaceInstUsesWith(I, LHS);
4547 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4549 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4550 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4551 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4552 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4556 case ICmpInst::ICMP_SGT:
4558 default: assert(0 && "Unknown integer condition code!");
4559 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4560 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4561 return ReplaceInstUsesWith(I, LHS);
4562 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4564 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4565 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4566 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4567 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4575 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4576 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4577 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4578 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4579 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4580 !isa<ICmpInst>(Op1C->getOperand(0))) {
4581 const Type *SrcTy = Op0C->getOperand(0)->getType();
4582 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4583 // Only do this if the casts both really cause code to be
4585 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4587 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4589 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4590 Op1C->getOperand(0),
4592 InsertNewInstBefore(NewOp, I);
4593 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4600 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4601 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4602 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4603 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4604 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4605 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4606 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4607 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4608 // If either of the constants are nans, then the whole thing returns
4610 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4611 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4613 // Otherwise, no need to compare the two constants, compare the
4615 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4616 RHS->getOperand(0));
4619 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4620 FCmpInst::Predicate Op0CC, Op1CC;
4621 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4622 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4623 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4624 // Swap RHS operands to match LHS.
4625 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4626 std::swap(Op1LHS, Op1RHS);
4628 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4629 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4631 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4632 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4633 Op1CC == FCmpInst::FCMP_TRUE)
4634 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4635 else if (Op0CC == FCmpInst::FCMP_FALSE)
4636 return ReplaceInstUsesWith(I, Op1);
4637 else if (Op1CC == FCmpInst::FCMP_FALSE)
4638 return ReplaceInstUsesWith(I, Op0);
4641 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4642 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4643 if (Op0Ordered == Op1Ordered) {
4644 // If both are ordered or unordered, return a new fcmp with
4645 // or'ed predicates.
4646 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4648 if (Instruction *I = dyn_cast<Instruction>(RV))
4650 // Otherwise, it's a constant boolean value...
4651 return ReplaceInstUsesWith(I, RV);
4659 return Changed ? &I : 0;
4664 // XorSelf - Implements: X ^ X --> 0
4667 XorSelf(Value *rhs) : RHS(rhs) {}
4668 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4669 Instruction *apply(BinaryOperator &Xor) const {
4676 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4677 bool Changed = SimplifyCommutative(I);
4678 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4680 if (isa<UndefValue>(Op1)) {
4681 if (isa<UndefValue>(Op0))
4682 // Handle undef ^ undef -> 0 special case. This is a common
4684 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4685 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4688 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4689 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4690 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4691 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4694 // See if we can simplify any instructions used by the instruction whose sole
4695 // purpose is to compute bits we don't care about.
4696 if (!isa<VectorType>(I.getType())) {
4697 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4698 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4699 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4700 KnownZero, KnownOne))
4702 } else if (isa<ConstantAggregateZero>(Op1)) {
4703 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4706 // Is this a ~ operation?
4707 if (Value *NotOp = dyn_castNotVal(&I)) {
4708 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4709 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4710 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4711 if (Op0I->getOpcode() == Instruction::And ||
4712 Op0I->getOpcode() == Instruction::Or) {
4713 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4714 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4716 BinaryOperator::CreateNot(Op0I->getOperand(1),
4717 Op0I->getOperand(1)->getName()+".not");
4718 InsertNewInstBefore(NotY, I);
4719 if (Op0I->getOpcode() == Instruction::And)
4720 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4722 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4729 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4730 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4731 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4732 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4733 return new ICmpInst(ICI->getInversePredicate(),
4734 ICI->getOperand(0), ICI->getOperand(1));
4736 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4737 return new FCmpInst(FCI->getInversePredicate(),
4738 FCI->getOperand(0), FCI->getOperand(1));
4741 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4742 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4743 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4744 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4745 Instruction::CastOps Opcode = Op0C->getOpcode();
4746 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4747 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4748 Op0C->getDestTy())) {
4749 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4750 CI->getOpcode(), CI->getInversePredicate(),
4751 CI->getOperand(0), CI->getOperand(1)), I);
4752 NewCI->takeName(CI);
4753 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4760 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4761 // ~(c-X) == X-c-1 == X+(-c-1)
4762 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4763 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4764 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4765 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4766 ConstantInt::get(I.getType(), 1));
4767 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4770 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4771 if (Op0I->getOpcode() == Instruction::Add) {
4772 // ~(X-c) --> (-c-1)-X
4773 if (RHS->isAllOnesValue()) {
4774 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4775 return BinaryOperator::CreateSub(
4776 ConstantExpr::getSub(NegOp0CI,
4777 ConstantInt::get(I.getType(), 1)),
4778 Op0I->getOperand(0));
4779 } else if (RHS->getValue().isSignBit()) {
4780 // (X + C) ^ signbit -> (X + C + signbit)
4781 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4782 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4785 } else if (Op0I->getOpcode() == Instruction::Or) {
4786 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4787 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4788 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4789 // Anything in both C1 and C2 is known to be zero, remove it from
4791 Constant *CommonBits = And(Op0CI, RHS);
4792 NewRHS = ConstantExpr::getAnd(NewRHS,
4793 ConstantExpr::getNot(CommonBits));
4794 AddToWorkList(Op0I);
4795 I.setOperand(0, Op0I->getOperand(0));
4796 I.setOperand(1, NewRHS);
4803 // Try to fold constant and into select arguments.
4804 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4805 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4807 if (isa<PHINode>(Op0))
4808 if (Instruction *NV = FoldOpIntoPhi(I))
4812 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4814 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4816 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4818 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4821 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4824 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4825 if (A == Op0) { // B^(B|A) == (A|B)^B
4826 Op1I->swapOperands();
4828 std::swap(Op0, Op1);
4829 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4830 I.swapOperands(); // Simplified below.
4831 std::swap(Op0, Op1);
4833 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4834 if (Op0 == A) // A^(A^B) == B
4835 return ReplaceInstUsesWith(I, B);
4836 else if (Op0 == B) // A^(B^A) == B
4837 return ReplaceInstUsesWith(I, A);
4838 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4839 if (A == Op0) { // A^(A&B) -> A^(B&A)
4840 Op1I->swapOperands();
4843 if (B == Op0) { // A^(B&A) -> (B&A)^A
4844 I.swapOperands(); // Simplified below.
4845 std::swap(Op0, Op1);
4850 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4853 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4854 if (A == Op1) // (B|A)^B == (A|B)^B
4856 if (B == Op1) { // (A|B)^B == A & ~B
4858 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4859 return BinaryOperator::CreateAnd(A, NotB);
4861 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4862 if (Op1 == A) // (A^B)^A == B
4863 return ReplaceInstUsesWith(I, B);
4864 else if (Op1 == B) // (B^A)^A == B
4865 return ReplaceInstUsesWith(I, A);
4866 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4867 if (A == Op1) // (A&B)^A -> (B&A)^A
4869 if (B == Op1 && // (B&A)^A == ~B & A
4870 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4872 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4873 return BinaryOperator::CreateAnd(N, Op1);
4878 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4879 if (Op0I && Op1I && Op0I->isShift() &&
4880 Op0I->getOpcode() == Op1I->getOpcode() &&
4881 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4882 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4883 Instruction *NewOp =
4884 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4885 Op1I->getOperand(0),
4886 Op0I->getName()), I);
4887 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4888 Op1I->getOperand(1));
4892 Value *A, *B, *C, *D;
4893 // (A & B)^(A | B) -> A ^ B
4894 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4895 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4896 if ((A == C && B == D) || (A == D && B == C))
4897 return BinaryOperator::CreateXor(A, B);
4899 // (A | B)^(A & B) -> A ^ B
4900 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4901 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4902 if ((A == C && B == D) || (A == D && B == C))
4903 return BinaryOperator::CreateXor(A, B);
4907 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4908 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4909 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4910 // (X & Y)^(X & Y) -> (Y^Z) & X
4911 Value *X = 0, *Y = 0, *Z = 0;
4913 X = A, Y = B, Z = D;
4915 X = A, Y = B, Z = C;
4917 X = B, Y = A, Z = D;
4919 X = B, Y = A, Z = C;
4922 Instruction *NewOp =
4923 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4924 return BinaryOperator::CreateAnd(NewOp, X);
4929 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4930 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4931 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4934 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4935 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4936 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4937 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4938 const Type *SrcTy = Op0C->getOperand(0)->getType();
4939 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4940 // Only do this if the casts both really cause code to be generated.
4941 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4943 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4945 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4946 Op1C->getOperand(0),
4948 InsertNewInstBefore(NewOp, I);
4949 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4954 return Changed ? &I : 0;
4957 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4958 /// overflowed for this type.
4959 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4960 ConstantInt *In2, bool IsSigned = false) {
4961 Result = cast<ConstantInt>(Add(In1, In2));
4964 if (In2->getValue().isNegative())
4965 return Result->getValue().sgt(In1->getValue());
4967 return Result->getValue().slt(In1->getValue());
4969 return Result->getValue().ult(In1->getValue());
4972 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
4973 /// overflowed for this type.
4974 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4975 ConstantInt *In2, bool IsSigned = false) {
4976 Result = cast<ConstantInt>(Subtract(In1, In2));
4979 if (In2->getValue().isNegative())
4980 return Result->getValue().slt(In1->getValue());
4982 return Result->getValue().sgt(In1->getValue());
4984 return Result->getValue().ugt(In1->getValue());
4987 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4988 /// code necessary to compute the offset from the base pointer (without adding
4989 /// in the base pointer). Return the result as a signed integer of intptr size.
4990 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4991 TargetData &TD = IC.getTargetData();
4992 gep_type_iterator GTI = gep_type_begin(GEP);
4993 const Type *IntPtrTy = TD.getIntPtrType();
4994 Value *Result = Constant::getNullValue(IntPtrTy);
4996 // Build a mask for high order bits.
4997 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4998 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5000 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5003 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5004 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5005 if (OpC->isZero()) continue;
5007 // Handle a struct index, which adds its field offset to the pointer.
5008 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5009 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5011 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5012 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5014 Result = IC.InsertNewInstBefore(
5015 BinaryOperator::CreateAdd(Result,
5016 ConstantInt::get(IntPtrTy, Size),
5017 GEP->getName()+".offs"), I);
5021 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5022 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5023 Scale = ConstantExpr::getMul(OC, Scale);
5024 if (Constant *RC = dyn_cast<Constant>(Result))
5025 Result = ConstantExpr::getAdd(RC, Scale);
5027 // Emit an add instruction.
5028 Result = IC.InsertNewInstBefore(
5029 BinaryOperator::CreateAdd(Result, Scale,
5030 GEP->getName()+".offs"), I);
5034 // Convert to correct type.
5035 if (Op->getType() != IntPtrTy) {
5036 if (Constant *OpC = dyn_cast<Constant>(Op))
5037 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5039 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5040 Op->getName()+".c"), I);
5043 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5044 if (Constant *OpC = dyn_cast<Constant>(Op))
5045 Op = ConstantExpr::getMul(OpC, Scale);
5046 else // We'll let instcombine(mul) convert this to a shl if possible.
5047 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5048 GEP->getName()+".idx"), I);
5051 // Emit an add instruction.
5052 if (isa<Constant>(Op) && isa<Constant>(Result))
5053 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5054 cast<Constant>(Result));
5056 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5057 GEP->getName()+".offs"), I);
5063 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5064 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5065 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5066 /// complex, and scales are involved. The above expression would also be legal
5067 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5068 /// later form is less amenable to optimization though, and we are allowed to
5069 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5071 /// If we can't emit an optimized form for this expression, this returns null.
5073 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5075 TargetData &TD = IC.getTargetData();
5076 gep_type_iterator GTI = gep_type_begin(GEP);
5078 // Check to see if this gep only has a single variable index. If so, and if
5079 // any constant indices are a multiple of its scale, then we can compute this
5080 // in terms of the scale of the variable index. For example, if the GEP
5081 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5082 // because the expression will cross zero at the same point.
5083 unsigned i, e = GEP->getNumOperands();
5085 for (i = 1; i != e; ++i, ++GTI) {
5086 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5087 // Compute the aggregate offset of constant indices.
5088 if (CI->isZero()) continue;
5090 // Handle a struct index, which adds its field offset to the pointer.
5091 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5092 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5094 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5095 Offset += Size*CI->getSExtValue();
5098 // Found our variable index.
5103 // If there are no variable indices, we must have a constant offset, just
5104 // evaluate it the general way.
5105 if (i == e) return 0;
5107 Value *VariableIdx = GEP->getOperand(i);
5108 // Determine the scale factor of the variable element. For example, this is
5109 // 4 if the variable index is into an array of i32.
5110 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5112 // Verify that there are no other variable indices. If so, emit the hard way.
5113 for (++i, ++GTI; i != e; ++i, ++GTI) {
5114 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5117 // Compute the aggregate offset of constant indices.
5118 if (CI->isZero()) continue;
5120 // Handle a struct index, which adds its field offset to the pointer.
5121 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5122 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5124 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5125 Offset += Size*CI->getSExtValue();
5129 // Okay, we know we have a single variable index, which must be a
5130 // pointer/array/vector index. If there is no offset, life is simple, return
5132 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5134 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5135 // we don't need to bother extending: the extension won't affect where the
5136 // computation crosses zero.
5137 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5138 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5139 VariableIdx->getNameStart(), &I);
5143 // Otherwise, there is an index. The computation we will do will be modulo
5144 // the pointer size, so get it.
5145 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5147 Offset &= PtrSizeMask;
5148 VariableScale &= PtrSizeMask;
5150 // To do this transformation, any constant index must be a multiple of the
5151 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5152 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5153 // multiple of the variable scale.
5154 int64_t NewOffs = Offset / (int64_t)VariableScale;
5155 if (Offset != NewOffs*(int64_t)VariableScale)
5158 // Okay, we can do this evaluation. Start by converting the index to intptr.
5159 const Type *IntPtrTy = TD.getIntPtrType();
5160 if (VariableIdx->getType() != IntPtrTy)
5161 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5163 VariableIdx->getNameStart(), &I);
5164 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5165 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5169 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5170 /// else. At this point we know that the GEP is on the LHS of the comparison.
5171 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5172 ICmpInst::Predicate Cond,
5174 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5176 // Look through bitcasts.
5177 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5178 RHS = BCI->getOperand(0);
5180 Value *PtrBase = GEPLHS->getOperand(0);
5181 if (PtrBase == RHS) {
5182 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5183 // This transformation (ignoring the base and scales) is valid because we
5184 // know pointers can't overflow. See if we can output an optimized form.
5185 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5187 // If not, synthesize the offset the hard way.
5189 Offset = EmitGEPOffset(GEPLHS, I, *this);
5190 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5191 Constant::getNullValue(Offset->getType()));
5192 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5193 // If the base pointers are different, but the indices are the same, just
5194 // compare the base pointer.
5195 if (PtrBase != GEPRHS->getOperand(0)) {
5196 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5197 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5198 GEPRHS->getOperand(0)->getType();
5200 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5201 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5202 IndicesTheSame = false;
5206 // If all indices are the same, just compare the base pointers.
5208 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5209 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5211 // Otherwise, the base pointers are different and the indices are
5212 // different, bail out.
5216 // If one of the GEPs has all zero indices, recurse.
5217 bool AllZeros = true;
5218 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5219 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5220 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5225 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5226 ICmpInst::getSwappedPredicate(Cond), I);
5228 // If the other GEP has all zero indices, recurse.
5230 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5231 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5232 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5237 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5239 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5240 // If the GEPs only differ by one index, compare it.
5241 unsigned NumDifferences = 0; // Keep track of # differences.
5242 unsigned DiffOperand = 0; // The operand that differs.
5243 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5244 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5245 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5246 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5247 // Irreconcilable differences.
5251 if (NumDifferences++) break;
5256 if (NumDifferences == 0) // SAME GEP?
5257 return ReplaceInstUsesWith(I, // No comparison is needed here.
5258 ConstantInt::get(Type::Int1Ty,
5259 ICmpInst::isTrueWhenEqual(Cond)));
5261 else if (NumDifferences == 1) {
5262 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5263 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5264 // Make sure we do a signed comparison here.
5265 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5269 // Only lower this if the icmp is the only user of the GEP or if we expect
5270 // the result to fold to a constant!
5271 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5272 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5273 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5274 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5275 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5276 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5282 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5284 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5287 if (!isa<ConstantFP>(RHSC)) return 0;
5288 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5290 // Get the width of the mantissa. We don't want to hack on conversions that
5291 // might lose information from the integer, e.g. "i64 -> float"
5292 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5293 if (MantissaWidth == -1) return 0; // Unknown.
5295 // Check to see that the input is converted from an integer type that is small
5296 // enough that preserves all bits. TODO: check here for "known" sign bits.
5297 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5298 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5300 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5301 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5305 // If the conversion would lose info, don't hack on this.
5306 if ((int)InputSize > MantissaWidth)
5309 // Otherwise, we can potentially simplify the comparison. We know that it
5310 // will always come through as an integer value and we know the constant is
5311 // not a NAN (it would have been previously simplified).
5312 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5314 ICmpInst::Predicate Pred;
5315 switch (I.getPredicate()) {
5316 default: assert(0 && "Unexpected predicate!");
5317 case FCmpInst::FCMP_UEQ:
5318 case FCmpInst::FCMP_OEQ:
5319 Pred = ICmpInst::ICMP_EQ;
5321 case FCmpInst::FCMP_UGT:
5322 case FCmpInst::FCMP_OGT:
5323 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5325 case FCmpInst::FCMP_UGE:
5326 case FCmpInst::FCMP_OGE:
5327 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5329 case FCmpInst::FCMP_ULT:
5330 case FCmpInst::FCMP_OLT:
5331 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5333 case FCmpInst::FCMP_ULE:
5334 case FCmpInst::FCMP_OLE:
5335 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5337 case FCmpInst::FCMP_UNE:
5338 case FCmpInst::FCMP_ONE:
5339 Pred = ICmpInst::ICMP_NE;
5341 case FCmpInst::FCMP_ORD:
5342 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5343 case FCmpInst::FCMP_UNO:
5344 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5347 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5349 // Now we know that the APFloat is a normal number, zero or inf.
5351 // See if the FP constant is too large for the integer. For example,
5352 // comparing an i8 to 300.0.
5353 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5356 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5357 // and large values.
5358 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5359 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5360 APFloat::rmNearestTiesToEven);
5361 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5362 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5363 Pred == ICmpInst::ICMP_SLE)
5364 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5365 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5368 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5369 // +INF and large values.
5370 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5371 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5372 APFloat::rmNearestTiesToEven);
5373 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5374 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5375 Pred == ICmpInst::ICMP_ULE)
5376 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5377 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5382 // See if the RHS value is < SignedMin.
5383 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5384 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5385 APFloat::rmNearestTiesToEven);
5386 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5387 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5388 Pred == ICmpInst::ICMP_SGE)
5389 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5390 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5394 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5395 // [0, UMAX], but it may still be fractional. See if it is fractional by
5396 // casting the FP value to the integer value and back, checking for equality.
5397 // Don't do this for zero, because -0.0 is not fractional.
5398 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5399 if (!RHS.isZero() &&
5400 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5401 // If we had a comparison against a fractional value, we have to adjust the
5402 // compare predicate and sometimes the value. RHSC is rounded towards zero
5405 default: assert(0 && "Unexpected integer comparison!");
5406 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5407 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5408 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5409 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5410 case ICmpInst::ICMP_ULE:
5411 // (float)int <= 4.4 --> int <= 4
5412 // (float)int <= -4.4 --> false
5413 if (RHS.isNegative())
5414 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5416 case ICmpInst::ICMP_SLE:
5417 // (float)int <= 4.4 --> int <= 4
5418 // (float)int <= -4.4 --> int < -4
5419 if (RHS.isNegative())
5420 Pred = ICmpInst::ICMP_SLT;
5422 case ICmpInst::ICMP_ULT:
5423 // (float)int < -4.4 --> false
5424 // (float)int < 4.4 --> int <= 4
5425 if (RHS.isNegative())
5426 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5427 Pred = ICmpInst::ICMP_ULE;
5429 case ICmpInst::ICMP_SLT:
5430 // (float)int < -4.4 --> int < -4
5431 // (float)int < 4.4 --> int <= 4
5432 if (!RHS.isNegative())
5433 Pred = ICmpInst::ICMP_SLE;
5435 case ICmpInst::ICMP_UGT:
5436 // (float)int > 4.4 --> int > 4
5437 // (float)int > -4.4 --> true
5438 if (RHS.isNegative())
5439 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5441 case ICmpInst::ICMP_SGT:
5442 // (float)int > 4.4 --> int > 4
5443 // (float)int > -4.4 --> int >= -4
5444 if (RHS.isNegative())
5445 Pred = ICmpInst::ICMP_SGE;
5447 case ICmpInst::ICMP_UGE:
5448 // (float)int >= -4.4 --> true
5449 // (float)int >= 4.4 --> int > 4
5450 if (!RHS.isNegative())
5451 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5452 Pred = ICmpInst::ICMP_UGT;
5454 case ICmpInst::ICMP_SGE:
5455 // (float)int >= -4.4 --> int >= -4
5456 // (float)int >= 4.4 --> int > 4
5457 if (!RHS.isNegative())
5458 Pred = ICmpInst::ICMP_SGT;
5463 // Lower this FP comparison into an appropriate integer version of the
5465 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5468 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5469 bool Changed = SimplifyCompare(I);
5470 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5472 // Fold trivial predicates.
5473 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5474 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5475 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5476 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5478 // Simplify 'fcmp pred X, X'
5480 switch (I.getPredicate()) {
5481 default: assert(0 && "Unknown predicate!");
5482 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5483 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5484 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5485 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5486 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5487 case FCmpInst::FCMP_OLT: // True if ordered and less than
5488 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5489 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5491 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5492 case FCmpInst::FCMP_ULT: // True if unordered or less than
5493 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5494 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5495 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5496 I.setPredicate(FCmpInst::FCMP_UNO);
5497 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5500 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5501 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5502 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5503 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5504 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5505 I.setPredicate(FCmpInst::FCMP_ORD);
5506 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5511 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5512 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5514 // Handle fcmp with constant RHS
5515 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5516 // If the constant is a nan, see if we can fold the comparison based on it.
5517 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5518 if (CFP->getValueAPF().isNaN()) {
5519 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5520 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5521 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5522 "Comparison must be either ordered or unordered!");
5523 // True if unordered.
5524 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5528 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5529 switch (LHSI->getOpcode()) {
5530 case Instruction::PHI:
5531 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5532 // block. If in the same block, we're encouraging jump threading. If
5533 // not, we are just pessimizing the code by making an i1 phi.
5534 if (LHSI->getParent() == I.getParent())
5535 if (Instruction *NV = FoldOpIntoPhi(I))
5538 case Instruction::SIToFP:
5539 case Instruction::UIToFP:
5540 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5543 case Instruction::Select:
5544 // If either operand of the select is a constant, we can fold the
5545 // comparison into the select arms, which will cause one to be
5546 // constant folded and the select turned into a bitwise or.
5547 Value *Op1 = 0, *Op2 = 0;
5548 if (LHSI->hasOneUse()) {
5549 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5550 // Fold the known value into the constant operand.
5551 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5552 // Insert a new FCmp of the other select operand.
5553 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5554 LHSI->getOperand(2), RHSC,
5556 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5557 // Fold the known value into the constant operand.
5558 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5559 // Insert a new FCmp of the other select operand.
5560 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5561 LHSI->getOperand(1), RHSC,
5567 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5572 return Changed ? &I : 0;
5575 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5576 bool Changed = SimplifyCompare(I);
5577 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5578 const Type *Ty = Op0->getType();
5582 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5583 I.isTrueWhenEqual()));
5585 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5586 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5588 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5589 // addresses never equal each other! We already know that Op0 != Op1.
5590 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5591 isa<ConstantPointerNull>(Op0)) &&
5592 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5593 isa<ConstantPointerNull>(Op1)))
5594 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5595 !I.isTrueWhenEqual()));
5597 // icmp's with boolean values can always be turned into bitwise operations
5598 if (Ty == Type::Int1Ty) {
5599 switch (I.getPredicate()) {
5600 default: assert(0 && "Invalid icmp instruction!");
5601 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5602 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5603 InsertNewInstBefore(Xor, I);
5604 return BinaryOperator::CreateNot(Xor);
5606 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5607 return BinaryOperator::CreateXor(Op0, Op1);
5609 case ICmpInst::ICMP_UGT:
5610 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5612 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5613 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5614 InsertNewInstBefore(Not, I);
5615 return BinaryOperator::CreateAnd(Not, Op1);
5617 case ICmpInst::ICMP_SGT:
5618 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5620 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5621 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5622 InsertNewInstBefore(Not, I);
5623 return BinaryOperator::CreateAnd(Not, Op0);
5625 case ICmpInst::ICMP_UGE:
5626 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5628 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5629 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5630 InsertNewInstBefore(Not, I);
5631 return BinaryOperator::CreateOr(Not, Op1);
5633 case ICmpInst::ICMP_SGE:
5634 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5636 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5637 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5638 InsertNewInstBefore(Not, I);
5639 return BinaryOperator::CreateOr(Not, Op0);
5644 // See if we are doing a comparison with a constant.
5645 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5648 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5649 if (I.isEquality() && CI->isNullValue() &&
5650 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5651 // (icmp cond A B) if cond is equality
5652 return new ICmpInst(I.getPredicate(), A, B);
5655 // If we have an icmp le or icmp ge instruction, turn it into the
5656 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5657 // them being folded in the code below.
5658 switch (I.getPredicate()) {
5660 case ICmpInst::ICMP_ULE:
5661 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5662 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5663 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5664 case ICmpInst::ICMP_SLE:
5665 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5666 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5667 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5668 case ICmpInst::ICMP_UGE:
5669 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5670 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5671 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5672 case ICmpInst::ICMP_SGE:
5673 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5674 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5675 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5678 // See if we can fold the comparison based on range information we can get
5679 // by checking whether bits are known to be zero or one in the input.
5680 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5681 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5683 // If this comparison is a normal comparison, it demands all
5684 // bits, if it is a sign bit comparison, it only demands the sign bit.
5686 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5688 if (SimplifyDemandedBits(Op0,
5689 isSignBit ? APInt::getSignBit(BitWidth)
5690 : APInt::getAllOnesValue(BitWidth),
5691 KnownZero, KnownOne, 0))
5694 // Given the known and unknown bits, compute a range that the LHS could be
5695 // in. Compute the Min, Max and RHS values based on the known bits. For the
5696 // EQ and NE we use unsigned values.
5697 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5698 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5699 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5701 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5703 // If Min and Max are known to be the same, then SimplifyDemandedBits
5704 // figured out that the LHS is a constant. Just constant fold this now so
5705 // that code below can assume that Min != Max.
5707 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5708 ConstantInt::get(Min),
5711 // Based on the range information we know about the LHS, see if we can
5712 // simplify this comparison. For example, (x&4) < 8 is always true.
5713 const APInt &RHSVal = CI->getValue();
5714 switch (I.getPredicate()) { // LE/GE have been folded already.
5715 default: assert(0 && "Unknown icmp opcode!");
5716 case ICmpInst::ICMP_EQ:
5717 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5718 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5720 case ICmpInst::ICMP_NE:
5721 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5722 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5724 case ICmpInst::ICMP_ULT:
5725 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5726 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5727 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5728 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5729 if (RHSVal == Max) // A <u MAX -> A != MAX
5730 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5731 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5732 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5734 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5735 if (CI->isMinValue(true))
5736 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5737 ConstantInt::getAllOnesValue(Op0->getType()));
5739 case ICmpInst::ICMP_UGT:
5740 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5741 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5742 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5743 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5745 if (RHSVal == Min) // A >u MIN -> A != MIN
5746 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5747 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5748 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5750 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5751 if (CI->isMaxValue(true))
5752 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5753 ConstantInt::getNullValue(Op0->getType()));
5755 case ICmpInst::ICMP_SLT:
5756 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5757 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5758 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5759 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5760 if (RHSVal == Max) // A <s MAX -> A != MAX
5761 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5762 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5763 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5765 case ICmpInst::ICMP_SGT:
5766 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5767 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5768 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5769 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5771 if (RHSVal == Min) // A >s MIN -> A != MIN
5772 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5773 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5774 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5779 // Test if the ICmpInst instruction is used exclusively by a select as
5780 // part of a minimum or maximum operation. If so, refrain from doing
5781 // any other folding. This helps out other analyses which understand
5782 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5783 // and CodeGen. And in this case, at least one of the comparison
5784 // operands has at least one user besides the compare (the select),
5785 // which would often largely negate the benefit of folding anyway.
5787 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5788 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5789 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5792 // See if we are doing a comparison between a constant and an instruction that
5793 // can be folded into the comparison.
5794 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5795 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5796 // instruction, see if that instruction also has constants so that the
5797 // instruction can be folded into the icmp
5798 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5799 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5803 // Handle icmp with constant (but not simple integer constant) RHS
5804 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5805 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5806 switch (LHSI->getOpcode()) {
5807 case Instruction::GetElementPtr:
5808 if (RHSC->isNullValue()) {
5809 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5810 bool isAllZeros = true;
5811 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5812 if (!isa<Constant>(LHSI->getOperand(i)) ||
5813 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5818 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5819 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5823 case Instruction::PHI:
5824 // Only fold icmp into the PHI if the phi and fcmp are in the same
5825 // block. If in the same block, we're encouraging jump threading. If
5826 // not, we are just pessimizing the code by making an i1 phi.
5827 if (LHSI->getParent() == I.getParent())
5828 if (Instruction *NV = FoldOpIntoPhi(I))
5831 case Instruction::Select: {
5832 // If either operand of the select is a constant, we can fold the
5833 // comparison into the select arms, which will cause one to be
5834 // constant folded and the select turned into a bitwise or.
5835 Value *Op1 = 0, *Op2 = 0;
5836 if (LHSI->hasOneUse()) {
5837 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5838 // Fold the known value into the constant operand.
5839 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5840 // Insert a new ICmp of the other select operand.
5841 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5842 LHSI->getOperand(2), RHSC,
5844 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5845 // Fold the known value into the constant operand.
5846 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5847 // Insert a new ICmp of the other select operand.
5848 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5849 LHSI->getOperand(1), RHSC,
5855 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5858 case Instruction::Malloc:
5859 // If we have (malloc != null), and if the malloc has a single use, we
5860 // can assume it is successful and remove the malloc.
5861 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5862 AddToWorkList(LHSI);
5863 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5864 !I.isTrueWhenEqual()));
5870 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5871 if (User *GEP = dyn_castGetElementPtr(Op0))
5872 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5874 if (User *GEP = dyn_castGetElementPtr(Op1))
5875 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5876 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5879 // Test to see if the operands of the icmp are casted versions of other
5880 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5882 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5883 if (isa<PointerType>(Op0->getType()) &&
5884 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5885 // We keep moving the cast from the left operand over to the right
5886 // operand, where it can often be eliminated completely.
5887 Op0 = CI->getOperand(0);
5889 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5890 // so eliminate it as well.
5891 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5892 Op1 = CI2->getOperand(0);
5894 // If Op1 is a constant, we can fold the cast into the constant.
5895 if (Op0->getType() != Op1->getType()) {
5896 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5897 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5899 // Otherwise, cast the RHS right before the icmp
5900 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5903 return new ICmpInst(I.getPredicate(), Op0, Op1);
5907 if (isa<CastInst>(Op0)) {
5908 // Handle the special case of: icmp (cast bool to X), <cst>
5909 // This comes up when you have code like
5912 // For generality, we handle any zero-extension of any operand comparison
5913 // with a constant or another cast from the same type.
5914 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5915 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5919 // See if it's the same type of instruction on the left and right.
5920 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5921 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
5922 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
5923 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
5925 switch (Op0I->getOpcode()) {
5927 case Instruction::Add:
5928 case Instruction::Sub:
5929 case Instruction::Xor:
5930 // a+x icmp eq/ne b+x --> a icmp b
5931 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
5932 Op1I->getOperand(0));
5934 case Instruction::Mul:
5935 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5936 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
5937 // Mask = -1 >> count-trailing-zeros(Cst).
5938 if (!CI->isZero() && !CI->isOne()) {
5939 const APInt &AP = CI->getValue();
5940 ConstantInt *Mask = ConstantInt::get(
5941 APInt::getLowBitsSet(AP.getBitWidth(),
5943 AP.countTrailingZeros()));
5944 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
5946 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
5948 InsertNewInstBefore(And1, I);
5949 InsertNewInstBefore(And2, I);
5950 return new ICmpInst(I.getPredicate(), And1, And2);
5959 // ~x < ~y --> y < x
5961 if (match(Op0, m_Not(m_Value(A))) &&
5962 match(Op1, m_Not(m_Value(B))))
5963 return new ICmpInst(I.getPredicate(), B, A);
5966 if (I.isEquality()) {
5967 Value *A, *B, *C, *D;
5969 // -x == -y --> x == y
5970 if (match(Op0, m_Neg(m_Value(A))) &&
5971 match(Op1, m_Neg(m_Value(B))))
5972 return new ICmpInst(I.getPredicate(), A, B);
5974 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5975 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5976 Value *OtherVal = A == Op1 ? B : A;
5977 return new ICmpInst(I.getPredicate(), OtherVal,
5978 Constant::getNullValue(A->getType()));
5981 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5982 // A^c1 == C^c2 --> A == C^(c1^c2)
5983 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5984 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5985 if (Op1->hasOneUse()) {
5986 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5987 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5988 return new ICmpInst(I.getPredicate(), A,
5989 InsertNewInstBefore(Xor, I));
5992 // A^B == A^D -> B == D
5993 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5994 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5995 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5996 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6000 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6001 (A == Op0 || B == Op0)) {
6002 // A == (A^B) -> B == 0
6003 Value *OtherVal = A == Op0 ? B : A;
6004 return new ICmpInst(I.getPredicate(), OtherVal,
6005 Constant::getNullValue(A->getType()));
6007 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
6008 // (A-B) == A -> B == 0
6009 return new ICmpInst(I.getPredicate(), B,
6010 Constant::getNullValue(B->getType()));
6012 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
6013 // A == (A-B) -> B == 0
6014 return new ICmpInst(I.getPredicate(), B,
6015 Constant::getNullValue(B->getType()));
6018 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6019 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6020 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6021 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6022 Value *X = 0, *Y = 0, *Z = 0;
6025 X = B; Y = D; Z = A;
6026 } else if (A == D) {
6027 X = B; Y = C; Z = A;
6028 } else if (B == C) {
6029 X = A; Y = D; Z = B;
6030 } else if (B == D) {
6031 X = A; Y = C; Z = B;
6034 if (X) { // Build (X^Y) & Z
6035 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6036 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6037 I.setOperand(0, Op1);
6038 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6043 return Changed ? &I : 0;
6047 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6048 /// and CmpRHS are both known to be integer constants.
6049 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6050 ConstantInt *DivRHS) {
6051 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6052 const APInt &CmpRHSV = CmpRHS->getValue();
6054 // FIXME: If the operand types don't match the type of the divide
6055 // then don't attempt this transform. The code below doesn't have the
6056 // logic to deal with a signed divide and an unsigned compare (and
6057 // vice versa). This is because (x /s C1) <s C2 produces different
6058 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6059 // (x /u C1) <u C2. Simply casting the operands and result won't
6060 // work. :( The if statement below tests that condition and bails
6062 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6063 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6065 if (DivRHS->isZero())
6066 return 0; // The ProdOV computation fails on divide by zero.
6067 if (DivIsSigned && DivRHS->isAllOnesValue())
6068 return 0; // The overflow computation also screws up here
6069 if (DivRHS->isOne())
6070 return 0; // Not worth bothering, and eliminates some funny cases
6073 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6074 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6075 // C2 (CI). By solving for X we can turn this into a range check
6076 // instead of computing a divide.
6077 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6079 // Determine if the product overflows by seeing if the product is
6080 // not equal to the divide. Make sure we do the same kind of divide
6081 // as in the LHS instruction that we're folding.
6082 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6083 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6085 // Get the ICmp opcode
6086 ICmpInst::Predicate Pred = ICI.getPredicate();
6088 // Figure out the interval that is being checked. For example, a comparison
6089 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6090 // Compute this interval based on the constants involved and the signedness of
6091 // the compare/divide. This computes a half-open interval, keeping track of
6092 // whether either value in the interval overflows. After analysis each
6093 // overflow variable is set to 0 if it's corresponding bound variable is valid
6094 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6095 int LoOverflow = 0, HiOverflow = 0;
6096 ConstantInt *LoBound = 0, *HiBound = 0;
6098 if (!DivIsSigned) { // udiv
6099 // e.g. X/5 op 3 --> [15, 20)
6101 HiOverflow = LoOverflow = ProdOV;
6103 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6104 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6105 if (CmpRHSV == 0) { // (X / pos) op 0
6106 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6107 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6109 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6110 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6111 HiOverflow = LoOverflow = ProdOV;
6113 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6114 } else { // (X / pos) op neg
6115 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6116 HiBound = AddOne(Prod);
6117 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6119 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6120 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6124 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6125 if (CmpRHSV == 0) { // (X / neg) op 0
6126 // e.g. X/-5 op 0 --> [-4, 5)
6127 LoBound = AddOne(DivRHS);
6128 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6129 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6130 HiOverflow = 1; // [INTMIN+1, overflow)
6131 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6133 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6134 // e.g. X/-5 op 3 --> [-19, -14)
6135 HiBound = AddOne(Prod);
6136 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6138 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6139 } else { // (X / neg) op neg
6140 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6141 LoOverflow = HiOverflow = ProdOV;
6143 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6146 // Dividing by a negative swaps the condition. LT <-> GT
6147 Pred = ICmpInst::getSwappedPredicate(Pred);
6150 Value *X = DivI->getOperand(0);
6152 default: assert(0 && "Unhandled icmp opcode!");
6153 case ICmpInst::ICMP_EQ:
6154 if (LoOverflow && HiOverflow)
6155 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6156 else if (HiOverflow)
6157 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6158 ICmpInst::ICMP_UGE, X, LoBound);
6159 else if (LoOverflow)
6160 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6161 ICmpInst::ICMP_ULT, X, HiBound);
6163 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6164 case ICmpInst::ICMP_NE:
6165 if (LoOverflow && HiOverflow)
6166 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6167 else if (HiOverflow)
6168 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6169 ICmpInst::ICMP_ULT, X, LoBound);
6170 else if (LoOverflow)
6171 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6172 ICmpInst::ICMP_UGE, X, HiBound);
6174 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6175 case ICmpInst::ICMP_ULT:
6176 case ICmpInst::ICMP_SLT:
6177 if (LoOverflow == +1) // Low bound is greater than input range.
6178 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6179 if (LoOverflow == -1) // Low bound is less than input range.
6180 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6181 return new ICmpInst(Pred, X, LoBound);
6182 case ICmpInst::ICMP_UGT:
6183 case ICmpInst::ICMP_SGT:
6184 if (HiOverflow == +1) // High bound greater than input range.
6185 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6186 else if (HiOverflow == -1) // High bound less than input range.
6187 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6188 if (Pred == ICmpInst::ICMP_UGT)
6189 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6191 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6196 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6198 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6201 const APInt &RHSV = RHS->getValue();
6203 switch (LHSI->getOpcode()) {
6204 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6205 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6206 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6208 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6209 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6210 Value *CompareVal = LHSI->getOperand(0);
6212 // If the sign bit of the XorCST is not set, there is no change to
6213 // the operation, just stop using the Xor.
6214 if (!XorCST->getValue().isNegative()) {
6215 ICI.setOperand(0, CompareVal);
6216 AddToWorkList(LHSI);
6220 // Was the old condition true if the operand is positive?
6221 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6223 // If so, the new one isn't.
6224 isTrueIfPositive ^= true;
6226 if (isTrueIfPositive)
6227 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6229 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6233 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6234 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6235 LHSI->getOperand(0)->hasOneUse()) {
6236 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6238 // If the LHS is an AND of a truncating cast, we can widen the
6239 // and/compare to be the input width without changing the value
6240 // produced, eliminating a cast.
6241 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6242 // We can do this transformation if either the AND constant does not
6243 // have its sign bit set or if it is an equality comparison.
6244 // Extending a relational comparison when we're checking the sign
6245 // bit would not work.
6246 if (Cast->hasOneUse() &&
6247 (ICI.isEquality() ||
6248 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6250 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6251 APInt NewCST = AndCST->getValue();
6252 NewCST.zext(BitWidth);
6254 NewCI.zext(BitWidth);
6255 Instruction *NewAnd =
6256 BinaryOperator::CreateAnd(Cast->getOperand(0),
6257 ConstantInt::get(NewCST),LHSI->getName());
6258 InsertNewInstBefore(NewAnd, ICI);
6259 return new ICmpInst(ICI.getPredicate(), NewAnd,
6260 ConstantInt::get(NewCI));
6264 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6265 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6266 // happens a LOT in code produced by the C front-end, for bitfield
6268 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6269 if (Shift && !Shift->isShift())
6273 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6274 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6275 const Type *AndTy = AndCST->getType(); // Type of the and.
6277 // We can fold this as long as we can't shift unknown bits
6278 // into the mask. This can only happen with signed shift
6279 // rights, as they sign-extend.
6281 bool CanFold = Shift->isLogicalShift();
6283 // To test for the bad case of the signed shr, see if any
6284 // of the bits shifted in could be tested after the mask.
6285 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6286 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6288 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6289 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6290 AndCST->getValue()) == 0)
6296 if (Shift->getOpcode() == Instruction::Shl)
6297 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6299 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6301 // Check to see if we are shifting out any of the bits being
6303 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6304 // If we shifted bits out, the fold is not going to work out.
6305 // As a special case, check to see if this means that the
6306 // result is always true or false now.
6307 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6308 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6309 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6310 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6312 ICI.setOperand(1, NewCst);
6313 Constant *NewAndCST;
6314 if (Shift->getOpcode() == Instruction::Shl)
6315 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6317 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6318 LHSI->setOperand(1, NewAndCST);
6319 LHSI->setOperand(0, Shift->getOperand(0));
6320 AddToWorkList(Shift); // Shift is dead.
6321 AddUsesToWorkList(ICI);
6327 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6328 // preferable because it allows the C<<Y expression to be hoisted out
6329 // of a loop if Y is invariant and X is not.
6330 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6331 ICI.isEquality() && !Shift->isArithmeticShift() &&
6332 isa<Instruction>(Shift->getOperand(0))) {
6335 if (Shift->getOpcode() == Instruction::LShr) {
6336 NS = BinaryOperator::CreateShl(AndCST,
6337 Shift->getOperand(1), "tmp");
6339 // Insert a logical shift.
6340 NS = BinaryOperator::CreateLShr(AndCST,
6341 Shift->getOperand(1), "tmp");
6343 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6345 // Compute X & (C << Y).
6346 Instruction *NewAnd =
6347 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6348 InsertNewInstBefore(NewAnd, ICI);
6350 ICI.setOperand(0, NewAnd);
6356 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6357 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6360 uint32_t TypeBits = RHSV.getBitWidth();
6362 // Check that the shift amount is in range. If not, don't perform
6363 // undefined shifts. When the shift is visited it will be
6365 if (ShAmt->uge(TypeBits))
6368 if (ICI.isEquality()) {
6369 // If we are comparing against bits always shifted out, the
6370 // comparison cannot succeed.
6372 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6373 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6374 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6375 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6376 return ReplaceInstUsesWith(ICI, Cst);
6379 if (LHSI->hasOneUse()) {
6380 // Otherwise strength reduce the shift into an and.
6381 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6383 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6386 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6387 Mask, LHSI->getName()+".mask");
6388 Value *And = InsertNewInstBefore(AndI, ICI);
6389 return new ICmpInst(ICI.getPredicate(), And,
6390 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6394 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6395 bool TrueIfSigned = false;
6396 if (LHSI->hasOneUse() &&
6397 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6398 // (X << 31) <s 0 --> (X&1) != 0
6399 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6400 (TypeBits-ShAmt->getZExtValue()-1));
6402 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6403 Mask, LHSI->getName()+".mask");
6404 Value *And = InsertNewInstBefore(AndI, ICI);
6406 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6407 And, Constant::getNullValue(And->getType()));
6412 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6413 case Instruction::AShr: {
6414 // Only handle equality comparisons of shift-by-constant.
6415 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6416 if (!ShAmt || !ICI.isEquality()) break;
6418 // Check that the shift amount is in range. If not, don't perform
6419 // undefined shifts. When the shift is visited it will be
6421 uint32_t TypeBits = RHSV.getBitWidth();
6422 if (ShAmt->uge(TypeBits))
6425 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6427 // If we are comparing against bits always shifted out, the
6428 // comparison cannot succeed.
6429 APInt Comp = RHSV << ShAmtVal;
6430 if (LHSI->getOpcode() == Instruction::LShr)
6431 Comp = Comp.lshr(ShAmtVal);
6433 Comp = Comp.ashr(ShAmtVal);
6435 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6436 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6437 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6438 return ReplaceInstUsesWith(ICI, Cst);
6441 // Otherwise, check to see if the bits shifted out are known to be zero.
6442 // If so, we can compare against the unshifted value:
6443 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6444 if (LHSI->hasOneUse() &&
6445 MaskedValueIsZero(LHSI->getOperand(0),
6446 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6447 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6448 ConstantExpr::getShl(RHS, ShAmt));
6451 if (LHSI->hasOneUse()) {
6452 // Otherwise strength reduce the shift into an and.
6453 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6454 Constant *Mask = ConstantInt::get(Val);
6457 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6458 Mask, LHSI->getName()+".mask");
6459 Value *And = InsertNewInstBefore(AndI, ICI);
6460 return new ICmpInst(ICI.getPredicate(), And,
6461 ConstantExpr::getShl(RHS, ShAmt));
6466 case Instruction::SDiv:
6467 case Instruction::UDiv:
6468 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6469 // Fold this div into the comparison, producing a range check.
6470 // Determine, based on the divide type, what the range is being
6471 // checked. If there is an overflow on the low or high side, remember
6472 // it, otherwise compute the range [low, hi) bounding the new value.
6473 // See: InsertRangeTest above for the kinds of replacements possible.
6474 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6475 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6480 case Instruction::Add:
6481 // Fold: icmp pred (add, X, C1), C2
6483 if (!ICI.isEquality()) {
6484 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6486 const APInt &LHSV = LHSC->getValue();
6488 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6491 if (ICI.isSignedPredicate()) {
6492 if (CR.getLower().isSignBit()) {
6493 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6494 ConstantInt::get(CR.getUpper()));
6495 } else if (CR.getUpper().isSignBit()) {
6496 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6497 ConstantInt::get(CR.getLower()));
6500 if (CR.getLower().isMinValue()) {
6501 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6502 ConstantInt::get(CR.getUpper()));
6503 } else if (CR.getUpper().isMinValue()) {
6504 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6505 ConstantInt::get(CR.getLower()));
6512 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6513 if (ICI.isEquality()) {
6514 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6516 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6517 // the second operand is a constant, simplify a bit.
6518 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6519 switch (BO->getOpcode()) {
6520 case Instruction::SRem:
6521 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6522 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6523 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6524 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6525 Instruction *NewRem =
6526 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6528 InsertNewInstBefore(NewRem, ICI);
6529 return new ICmpInst(ICI.getPredicate(), NewRem,
6530 Constant::getNullValue(BO->getType()));
6534 case Instruction::Add:
6535 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6536 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6537 if (BO->hasOneUse())
6538 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6539 Subtract(RHS, BOp1C));
6540 } else if (RHSV == 0) {
6541 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6542 // efficiently invertible, or if the add has just this one use.
6543 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6545 if (Value *NegVal = dyn_castNegVal(BOp1))
6546 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6547 else if (Value *NegVal = dyn_castNegVal(BOp0))
6548 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6549 else if (BO->hasOneUse()) {
6550 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6551 InsertNewInstBefore(Neg, ICI);
6553 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6557 case Instruction::Xor:
6558 // For the xor case, we can xor two constants together, eliminating
6559 // the explicit xor.
6560 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6561 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6562 ConstantExpr::getXor(RHS, BOC));
6565 case Instruction::Sub:
6566 // Replace (([sub|xor] A, B) != 0) with (A != B)
6568 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6572 case Instruction::Or:
6573 // If bits are being or'd in that are not present in the constant we
6574 // are comparing against, then the comparison could never succeed!
6575 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6576 Constant *NotCI = ConstantExpr::getNot(RHS);
6577 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6578 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6583 case Instruction::And:
6584 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6585 // If bits are being compared against that are and'd out, then the
6586 // comparison can never succeed!
6587 if ((RHSV & ~BOC->getValue()) != 0)
6588 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6591 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6592 if (RHS == BOC && RHSV.isPowerOf2())
6593 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6594 ICmpInst::ICMP_NE, LHSI,
6595 Constant::getNullValue(RHS->getType()));
6597 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6598 if (BOC->getValue().isSignBit()) {
6599 Value *X = BO->getOperand(0);
6600 Constant *Zero = Constant::getNullValue(X->getType());
6601 ICmpInst::Predicate pred = isICMP_NE ?
6602 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6603 return new ICmpInst(pred, X, Zero);
6606 // ((X & ~7) == 0) --> X < 8
6607 if (RHSV == 0 && isHighOnes(BOC)) {
6608 Value *X = BO->getOperand(0);
6609 Constant *NegX = ConstantExpr::getNeg(BOC);
6610 ICmpInst::Predicate pred = isICMP_NE ?
6611 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6612 return new ICmpInst(pred, X, NegX);
6617 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6618 // Handle icmp {eq|ne} <intrinsic>, intcst.
6619 if (II->getIntrinsicID() == Intrinsic::bswap) {
6621 ICI.setOperand(0, II->getOperand(1));
6622 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6626 } else { // Not a ICMP_EQ/ICMP_NE
6627 // If the LHS is a cast from an integral value of the same size,
6628 // then since we know the RHS is a constant, try to simlify.
6629 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6630 Value *CastOp = Cast->getOperand(0);
6631 const Type *SrcTy = CastOp->getType();
6632 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6633 if (SrcTy->isInteger() &&
6634 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6635 // If this is an unsigned comparison, try to make the comparison use
6636 // smaller constant values.
6637 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6638 // X u< 128 => X s> -1
6639 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6640 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6641 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6642 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6643 // X u> 127 => X s< 0
6644 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6645 Constant::getNullValue(SrcTy));
6653 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6654 /// We only handle extending casts so far.
6656 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6657 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6658 Value *LHSCIOp = LHSCI->getOperand(0);
6659 const Type *SrcTy = LHSCIOp->getType();
6660 const Type *DestTy = LHSCI->getType();
6663 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6664 // integer type is the same size as the pointer type.
6665 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6666 getTargetData().getPointerSizeInBits() ==
6667 cast<IntegerType>(DestTy)->getBitWidth()) {
6669 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6670 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6671 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6672 RHSOp = RHSC->getOperand(0);
6673 // If the pointer types don't match, insert a bitcast.
6674 if (LHSCIOp->getType() != RHSOp->getType())
6675 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6679 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6682 // The code below only handles extension cast instructions, so far.
6684 if (LHSCI->getOpcode() != Instruction::ZExt &&
6685 LHSCI->getOpcode() != Instruction::SExt)
6688 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6689 bool isSignedCmp = ICI.isSignedPredicate();
6691 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6692 // Not an extension from the same type?
6693 RHSCIOp = CI->getOperand(0);
6694 if (RHSCIOp->getType() != LHSCIOp->getType())
6697 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6698 // and the other is a zext), then we can't handle this.
6699 if (CI->getOpcode() != LHSCI->getOpcode())
6702 // Deal with equality cases early.
6703 if (ICI.isEquality())
6704 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6706 // A signed comparison of sign extended values simplifies into a
6707 // signed comparison.
6708 if (isSignedCmp && isSignedExt)
6709 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6711 // The other three cases all fold into an unsigned comparison.
6712 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6715 // If we aren't dealing with a constant on the RHS, exit early
6716 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6720 // Compute the constant that would happen if we truncated to SrcTy then
6721 // reextended to DestTy.
6722 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6723 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6725 // If the re-extended constant didn't change...
6727 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6728 // For example, we might have:
6729 // %A = sext short %X to uint
6730 // %B = icmp ugt uint %A, 1330
6731 // It is incorrect to transform this into
6732 // %B = icmp ugt short %X, 1330
6733 // because %A may have negative value.
6735 // However, we allow this when the compare is EQ/NE, because they are
6737 if (isSignedExt == isSignedCmp || ICI.isEquality())
6738 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6742 // The re-extended constant changed so the constant cannot be represented
6743 // in the shorter type. Consequently, we cannot emit a simple comparison.
6745 // First, handle some easy cases. We know the result cannot be equal at this
6746 // point so handle the ICI.isEquality() cases
6747 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6748 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6749 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6750 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6752 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6753 // should have been folded away previously and not enter in here.
6756 // We're performing a signed comparison.
6757 if (cast<ConstantInt>(CI)->getValue().isNegative())
6758 Result = ConstantInt::getFalse(); // X < (small) --> false
6760 Result = ConstantInt::getTrue(); // X < (large) --> true
6762 // We're performing an unsigned comparison.
6764 // We're performing an unsigned comp with a sign extended value.
6765 // This is true if the input is >= 0. [aka >s -1]
6766 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6767 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6768 NegOne, ICI.getName()), ICI);
6770 // Unsigned extend & unsigned compare -> always true.
6771 Result = ConstantInt::getTrue();
6775 // Finally, return the value computed.
6776 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6777 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6778 return ReplaceInstUsesWith(ICI, Result);
6780 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6781 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6782 "ICmp should be folded!");
6783 if (Constant *CI = dyn_cast<Constant>(Result))
6784 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6785 return BinaryOperator::CreateNot(Result);
6788 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6789 return commonShiftTransforms(I);
6792 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6793 return commonShiftTransforms(I);
6796 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6797 if (Instruction *R = commonShiftTransforms(I))
6800 Value *Op0 = I.getOperand(0);
6802 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6803 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6804 if (CSI->isAllOnesValue())
6805 return ReplaceInstUsesWith(I, CSI);
6807 // See if we can turn a signed shr into an unsigned shr.
6808 if (!isa<VectorType>(I.getType()) &&
6809 MaskedValueIsZero(Op0,
6810 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6811 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6816 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6817 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6818 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6820 // shl X, 0 == X and shr X, 0 == X
6821 // shl 0, X == 0 and shr 0, X == 0
6822 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6823 Op0 == Constant::getNullValue(Op0->getType()))
6824 return ReplaceInstUsesWith(I, Op0);
6826 if (isa<UndefValue>(Op0)) {
6827 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6828 return ReplaceInstUsesWith(I, Op0);
6829 else // undef << X -> 0, undef >>u X -> 0
6830 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6832 if (isa<UndefValue>(Op1)) {
6833 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6834 return ReplaceInstUsesWith(I, Op0);
6835 else // X << undef, X >>u undef -> 0
6836 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6839 // Try to fold constant and into select arguments.
6840 if (isa<Constant>(Op0))
6841 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6842 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6845 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6846 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6851 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6852 BinaryOperator &I) {
6853 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6855 // See if we can simplify any instructions used by the instruction whose sole
6856 // purpose is to compute bits we don't care about.
6857 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6858 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6859 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6860 KnownZero, KnownOne))
6863 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6864 // of a signed value.
6866 if (Op1->uge(TypeBits)) {
6867 if (I.getOpcode() != Instruction::AShr)
6868 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6870 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6875 // ((X*C1) << C2) == (X * (C1 << C2))
6876 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6877 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6878 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6879 return BinaryOperator::CreateMul(BO->getOperand(0),
6880 ConstantExpr::getShl(BOOp, Op1));
6882 // Try to fold constant and into select arguments.
6883 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6884 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6886 if (isa<PHINode>(Op0))
6887 if (Instruction *NV = FoldOpIntoPhi(I))
6890 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6891 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6892 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6893 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6894 // place. Don't try to do this transformation in this case. Also, we
6895 // require that the input operand is a shift-by-constant so that we have
6896 // confidence that the shifts will get folded together. We could do this
6897 // xform in more cases, but it is unlikely to be profitable.
6898 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6899 isa<ConstantInt>(TrOp->getOperand(1))) {
6900 // Okay, we'll do this xform. Make the shift of shift.
6901 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6902 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6904 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6906 // For logical shifts, the truncation has the effect of making the high
6907 // part of the register be zeros. Emulate this by inserting an AND to
6908 // clear the top bits as needed. This 'and' will usually be zapped by
6909 // other xforms later if dead.
6910 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6911 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6912 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6914 // The mask we constructed says what the trunc would do if occurring
6915 // between the shifts. We want to know the effect *after* the second
6916 // shift. We know that it is a logical shift by a constant, so adjust the
6917 // mask as appropriate.
6918 if (I.getOpcode() == Instruction::Shl)
6919 MaskV <<= Op1->getZExtValue();
6921 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6922 MaskV = MaskV.lshr(Op1->getZExtValue());
6925 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6927 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6929 // Return the value truncated to the interesting size.
6930 return new TruncInst(And, I.getType());
6934 if (Op0->hasOneUse()) {
6935 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6936 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6939 switch (Op0BO->getOpcode()) {
6941 case Instruction::Add:
6942 case Instruction::And:
6943 case Instruction::Or:
6944 case Instruction::Xor: {
6945 // These operators commute.
6946 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6947 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6948 match(Op0BO->getOperand(1),
6949 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6950 Instruction *YS = BinaryOperator::CreateShl(
6951 Op0BO->getOperand(0), Op1,
6953 InsertNewInstBefore(YS, I); // (Y << C)
6955 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6956 Op0BO->getOperand(1)->getName());
6957 InsertNewInstBefore(X, I); // (X + (Y << C))
6958 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6959 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6960 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6963 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6964 Value *Op0BOOp1 = Op0BO->getOperand(1);
6965 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6967 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6968 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6970 Instruction *YS = BinaryOperator::CreateShl(
6971 Op0BO->getOperand(0), Op1,
6973 InsertNewInstBefore(YS, I); // (Y << C)
6975 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6976 V1->getName()+".mask");
6977 InsertNewInstBefore(XM, I); // X & (CC << C)
6979 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6984 case Instruction::Sub: {
6985 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6986 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6987 match(Op0BO->getOperand(0),
6988 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6989 Instruction *YS = BinaryOperator::CreateShl(
6990 Op0BO->getOperand(1), Op1,
6992 InsertNewInstBefore(YS, I); // (Y << C)
6994 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6995 Op0BO->getOperand(0)->getName());
6996 InsertNewInstBefore(X, I); // (X + (Y << C))
6997 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6998 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6999 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7002 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7003 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7004 match(Op0BO->getOperand(0),
7005 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7006 m_ConstantInt(CC))) && V2 == Op1 &&
7007 cast<BinaryOperator>(Op0BO->getOperand(0))
7008 ->getOperand(0)->hasOneUse()) {
7009 Instruction *YS = BinaryOperator::CreateShl(
7010 Op0BO->getOperand(1), Op1,
7012 InsertNewInstBefore(YS, I); // (Y << C)
7014 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7015 V1->getName()+".mask");
7016 InsertNewInstBefore(XM, I); // X & (CC << C)
7018 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7026 // If the operand is an bitwise operator with a constant RHS, and the
7027 // shift is the only use, we can pull it out of the shift.
7028 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7029 bool isValid = true; // Valid only for And, Or, Xor
7030 bool highBitSet = false; // Transform if high bit of constant set?
7032 switch (Op0BO->getOpcode()) {
7033 default: isValid = false; break; // Do not perform transform!
7034 case Instruction::Add:
7035 isValid = isLeftShift;
7037 case Instruction::Or:
7038 case Instruction::Xor:
7041 case Instruction::And:
7046 // If this is a signed shift right, and the high bit is modified
7047 // by the logical operation, do not perform the transformation.
7048 // The highBitSet boolean indicates the value of the high bit of
7049 // the constant which would cause it to be modified for this
7052 if (isValid && I.getOpcode() == Instruction::AShr)
7053 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7056 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7058 Instruction *NewShift =
7059 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7060 InsertNewInstBefore(NewShift, I);
7061 NewShift->takeName(Op0BO);
7063 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7070 // Find out if this is a shift of a shift by a constant.
7071 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7072 if (ShiftOp && !ShiftOp->isShift())
7075 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7076 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7077 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7078 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7079 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7080 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7081 Value *X = ShiftOp->getOperand(0);
7083 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7084 if (AmtSum > TypeBits)
7087 const IntegerType *Ty = cast<IntegerType>(I.getType());
7089 // Check for (X << c1) << c2 and (X >> c1) >> c2
7090 if (I.getOpcode() == ShiftOp->getOpcode()) {
7091 return BinaryOperator::Create(I.getOpcode(), X,
7092 ConstantInt::get(Ty, AmtSum));
7093 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7094 I.getOpcode() == Instruction::AShr) {
7095 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7096 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7097 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7098 I.getOpcode() == Instruction::LShr) {
7099 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7100 Instruction *Shift =
7101 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7102 InsertNewInstBefore(Shift, I);
7104 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7105 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7108 // Okay, if we get here, one shift must be left, and the other shift must be
7109 // right. See if the amounts are equal.
7110 if (ShiftAmt1 == ShiftAmt2) {
7111 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7112 if (I.getOpcode() == Instruction::Shl) {
7113 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7114 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7116 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7117 if (I.getOpcode() == Instruction::LShr) {
7118 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7119 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7121 // We can simplify ((X << C) >>s C) into a trunc + sext.
7122 // NOTE: we could do this for any C, but that would make 'unusual' integer
7123 // types. For now, just stick to ones well-supported by the code
7125 const Type *SExtType = 0;
7126 switch (Ty->getBitWidth() - ShiftAmt1) {
7133 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7138 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7139 InsertNewInstBefore(NewTrunc, I);
7140 return new SExtInst(NewTrunc, Ty);
7142 // Otherwise, we can't handle it yet.
7143 } else if (ShiftAmt1 < ShiftAmt2) {
7144 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7146 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7147 if (I.getOpcode() == Instruction::Shl) {
7148 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7149 ShiftOp->getOpcode() == Instruction::AShr);
7150 Instruction *Shift =
7151 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7152 InsertNewInstBefore(Shift, I);
7154 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7155 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7158 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7159 if (I.getOpcode() == Instruction::LShr) {
7160 assert(ShiftOp->getOpcode() == Instruction::Shl);
7161 Instruction *Shift =
7162 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7163 InsertNewInstBefore(Shift, I);
7165 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7166 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7169 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7171 assert(ShiftAmt2 < ShiftAmt1);
7172 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7174 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7175 if (I.getOpcode() == Instruction::Shl) {
7176 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7177 ShiftOp->getOpcode() == Instruction::AShr);
7178 Instruction *Shift =
7179 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7180 ConstantInt::get(Ty, ShiftDiff));
7181 InsertNewInstBefore(Shift, I);
7183 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7184 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7187 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7188 if (I.getOpcode() == Instruction::LShr) {
7189 assert(ShiftOp->getOpcode() == Instruction::Shl);
7190 Instruction *Shift =
7191 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7192 InsertNewInstBefore(Shift, I);
7194 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7195 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7198 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7205 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7206 /// expression. If so, decompose it, returning some value X, such that Val is
7209 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7211 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7212 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7213 Offset = CI->getZExtValue();
7215 return ConstantInt::get(Type::Int32Ty, 0);
7216 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7217 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7218 if (I->getOpcode() == Instruction::Shl) {
7219 // This is a value scaled by '1 << the shift amt'.
7220 Scale = 1U << RHS->getZExtValue();
7222 return I->getOperand(0);
7223 } else if (I->getOpcode() == Instruction::Mul) {
7224 // This value is scaled by 'RHS'.
7225 Scale = RHS->getZExtValue();
7227 return I->getOperand(0);
7228 } else if (I->getOpcode() == Instruction::Add) {
7229 // We have X+C. Check to see if we really have (X*C2)+C1,
7230 // where C1 is divisible by C2.
7233 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7234 Offset += RHS->getZExtValue();
7241 // Otherwise, we can't look past this.
7248 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7249 /// try to eliminate the cast by moving the type information into the alloc.
7250 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7251 AllocationInst &AI) {
7252 const PointerType *PTy = cast<PointerType>(CI.getType());
7254 // Remove any uses of AI that are dead.
7255 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7257 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7258 Instruction *User = cast<Instruction>(*UI++);
7259 if (isInstructionTriviallyDead(User)) {
7260 while (UI != E && *UI == User)
7261 ++UI; // If this instruction uses AI more than once, don't break UI.
7264 DOUT << "IC: DCE: " << *User;
7265 EraseInstFromFunction(*User);
7269 // Get the type really allocated and the type casted to.
7270 const Type *AllocElTy = AI.getAllocatedType();
7271 const Type *CastElTy = PTy->getElementType();
7272 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7274 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7275 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7276 if (CastElTyAlign < AllocElTyAlign) return 0;
7278 // If the allocation has multiple uses, only promote it if we are strictly
7279 // increasing the alignment of the resultant allocation. If we keep it the
7280 // same, we open the door to infinite loops of various kinds.
7281 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7283 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7284 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7285 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7287 // See if we can satisfy the modulus by pulling a scale out of the array
7289 unsigned ArraySizeScale;
7291 Value *NumElements = // See if the array size is a decomposable linear expr.
7292 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7294 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7296 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7297 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7299 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7304 // If the allocation size is constant, form a constant mul expression
7305 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7306 if (isa<ConstantInt>(NumElements))
7307 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7308 // otherwise multiply the amount and the number of elements
7309 else if (Scale != 1) {
7310 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7311 Amt = InsertNewInstBefore(Tmp, AI);
7315 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7316 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7317 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7318 Amt = InsertNewInstBefore(Tmp, AI);
7321 AllocationInst *New;
7322 if (isa<MallocInst>(AI))
7323 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7325 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7326 InsertNewInstBefore(New, AI);
7329 // If the allocation has multiple uses, insert a cast and change all things
7330 // that used it to use the new cast. This will also hack on CI, but it will
7332 if (!AI.hasOneUse()) {
7333 AddUsesToWorkList(AI);
7334 // New is the allocation instruction, pointer typed. AI is the original
7335 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7336 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7337 InsertNewInstBefore(NewCast, AI);
7338 AI.replaceAllUsesWith(NewCast);
7340 return ReplaceInstUsesWith(CI, New);
7343 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7344 /// and return it as type Ty without inserting any new casts and without
7345 /// changing the computed value. This is used by code that tries to decide
7346 /// whether promoting or shrinking integer operations to wider or smaller types
7347 /// will allow us to eliminate a truncate or extend.
7349 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7350 /// extension operation if Ty is larger.
7352 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7353 /// should return true if trunc(V) can be computed by computing V in the smaller
7354 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7355 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7356 /// efficiently truncated.
7358 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7359 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7360 /// the final result.
7361 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7363 int &NumCastsRemoved) {
7364 // We can always evaluate constants in another type.
7365 if (isa<ConstantInt>(V))
7368 Instruction *I = dyn_cast<Instruction>(V);
7369 if (!I) return false;
7371 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7373 // If this is an extension or truncate, we can often eliminate it.
7374 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7375 // If this is a cast from the destination type, we can trivially eliminate
7376 // it, and this will remove a cast overall.
7377 if (I->getOperand(0)->getType() == Ty) {
7378 // If the first operand is itself a cast, and is eliminable, do not count
7379 // this as an eliminable cast. We would prefer to eliminate those two
7381 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7387 // We can't extend or shrink something that has multiple uses: doing so would
7388 // require duplicating the instruction in general, which isn't profitable.
7389 if (!I->hasOneUse()) return false;
7391 switch (I->getOpcode()) {
7392 case Instruction::Add:
7393 case Instruction::Sub:
7394 case Instruction::Mul:
7395 case Instruction::And:
7396 case Instruction::Or:
7397 case Instruction::Xor:
7398 // These operators can all arbitrarily be extended or truncated.
7399 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7401 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7404 case Instruction::Shl:
7405 // If we are truncating the result of this SHL, and if it's a shift of a
7406 // constant amount, we can always perform a SHL in a smaller type.
7407 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7408 uint32_t BitWidth = Ty->getBitWidth();
7409 if (BitWidth < OrigTy->getBitWidth() &&
7410 CI->getLimitedValue(BitWidth) < BitWidth)
7411 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7415 case Instruction::LShr:
7416 // If this is a truncate of a logical shr, we can truncate it to a smaller
7417 // lshr iff we know that the bits we would otherwise be shifting in are
7419 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7420 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7421 uint32_t BitWidth = Ty->getBitWidth();
7422 if (BitWidth < OrigBitWidth &&
7423 MaskedValueIsZero(I->getOperand(0),
7424 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7425 CI->getLimitedValue(BitWidth) < BitWidth) {
7426 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7431 case Instruction::ZExt:
7432 case Instruction::SExt:
7433 case Instruction::Trunc:
7434 // If this is the same kind of case as our original (e.g. zext+zext), we
7435 // can safely replace it. Note that replacing it does not reduce the number
7436 // of casts in the input.
7437 if (I->getOpcode() == CastOpc)
7440 case Instruction::Select: {
7441 SelectInst *SI = cast<SelectInst>(I);
7442 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7444 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7447 case Instruction::PHI: {
7448 // We can change a phi if we can change all operands.
7449 PHINode *PN = cast<PHINode>(I);
7450 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7451 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7457 // TODO: Can handle more cases here.
7464 /// EvaluateInDifferentType - Given an expression that
7465 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7466 /// evaluate the expression.
7467 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7469 if (Constant *C = dyn_cast<Constant>(V))
7470 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7472 // Otherwise, it must be an instruction.
7473 Instruction *I = cast<Instruction>(V);
7474 Instruction *Res = 0;
7475 switch (I->getOpcode()) {
7476 case Instruction::Add:
7477 case Instruction::Sub:
7478 case Instruction::Mul:
7479 case Instruction::And:
7480 case Instruction::Or:
7481 case Instruction::Xor:
7482 case Instruction::AShr:
7483 case Instruction::LShr:
7484 case Instruction::Shl: {
7485 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7486 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7487 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7491 case Instruction::Trunc:
7492 case Instruction::ZExt:
7493 case Instruction::SExt:
7494 // If the source type of the cast is the type we're trying for then we can
7495 // just return the source. There's no need to insert it because it is not
7497 if (I->getOperand(0)->getType() == Ty)
7498 return I->getOperand(0);
7500 // Otherwise, must be the same type of cast, so just reinsert a new one.
7501 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7504 case Instruction::Select: {
7505 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7506 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7507 Res = SelectInst::Create(I->getOperand(0), True, False);
7510 case Instruction::PHI: {
7511 PHINode *OPN = cast<PHINode>(I);
7512 PHINode *NPN = PHINode::Create(Ty);
7513 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7514 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7515 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7521 // TODO: Can handle more cases here.
7522 assert(0 && "Unreachable!");
7527 return InsertNewInstBefore(Res, *I);
7530 /// @brief Implement the transforms common to all CastInst visitors.
7531 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7532 Value *Src = CI.getOperand(0);
7534 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7535 // eliminate it now.
7536 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7537 if (Instruction::CastOps opc =
7538 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7539 // The first cast (CSrc) is eliminable so we need to fix up or replace
7540 // the second cast (CI). CSrc will then have a good chance of being dead.
7541 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7545 // If we are casting a select then fold the cast into the select
7546 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7547 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7550 // If we are casting a PHI then fold the cast into the PHI
7551 if (isa<PHINode>(Src))
7552 if (Instruction *NV = FoldOpIntoPhi(CI))
7558 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7559 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7560 Value *Src = CI.getOperand(0);
7562 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7563 // If casting the result of a getelementptr instruction with no offset, turn
7564 // this into a cast of the original pointer!
7565 if (GEP->hasAllZeroIndices()) {
7566 // Changing the cast operand is usually not a good idea but it is safe
7567 // here because the pointer operand is being replaced with another
7568 // pointer operand so the opcode doesn't need to change.
7570 CI.setOperand(0, GEP->getOperand(0));
7574 // If the GEP has a single use, and the base pointer is a bitcast, and the
7575 // GEP computes a constant offset, see if we can convert these three
7576 // instructions into fewer. This typically happens with unions and other
7577 // non-type-safe code.
7578 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7579 if (GEP->hasAllConstantIndices()) {
7580 // We are guaranteed to get a constant from EmitGEPOffset.
7581 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7582 int64_t Offset = OffsetV->getSExtValue();
7584 // Get the base pointer input of the bitcast, and the type it points to.
7585 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7586 const Type *GEPIdxTy =
7587 cast<PointerType>(OrigBase->getType())->getElementType();
7588 if (GEPIdxTy->isSized()) {
7589 SmallVector<Value*, 8> NewIndices;
7591 // Start with the index over the outer type. Note that the type size
7592 // might be zero (even if the offset isn't zero) if the indexed type
7593 // is something like [0 x {int, int}]
7594 const Type *IntPtrTy = TD->getIntPtrType();
7595 int64_t FirstIdx = 0;
7596 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7597 FirstIdx = Offset/TySize;
7600 // Handle silly modulus not returning values values [0..TySize).
7604 assert(Offset >= 0);
7606 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7609 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7611 // Index into the types. If we fail, set OrigBase to null.
7613 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7614 const StructLayout *SL = TD->getStructLayout(STy);
7615 if (Offset < (int64_t)SL->getSizeInBytes()) {
7616 unsigned Elt = SL->getElementContainingOffset(Offset);
7617 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7619 Offset -= SL->getElementOffset(Elt);
7620 GEPIdxTy = STy->getElementType(Elt);
7622 // Otherwise, we can't index into this, bail out.
7626 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7627 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7628 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7629 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7632 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7634 GEPIdxTy = STy->getElementType();
7636 // Otherwise, we can't index into this, bail out.
7642 // If we were able to index down into an element, create the GEP
7643 // and bitcast the result. This eliminates one bitcast, potentially
7645 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7647 NewIndices.end(), "");
7648 InsertNewInstBefore(NGEP, CI);
7649 NGEP->takeName(GEP);
7651 if (isa<BitCastInst>(CI))
7652 return new BitCastInst(NGEP, CI.getType());
7653 assert(isa<PtrToIntInst>(CI));
7654 return new PtrToIntInst(NGEP, CI.getType());
7661 return commonCastTransforms(CI);
7666 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7667 /// integer types. This function implements the common transforms for all those
7669 /// @brief Implement the transforms common to CastInst with integer operands
7670 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7671 if (Instruction *Result = commonCastTransforms(CI))
7674 Value *Src = CI.getOperand(0);
7675 const Type *SrcTy = Src->getType();
7676 const Type *DestTy = CI.getType();
7677 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7678 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7680 // See if we can simplify any instructions used by the LHS whose sole
7681 // purpose is to compute bits we don't care about.
7682 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7683 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7684 KnownZero, KnownOne))
7687 // If the source isn't an instruction or has more than one use then we
7688 // can't do anything more.
7689 Instruction *SrcI = dyn_cast<Instruction>(Src);
7690 if (!SrcI || !Src->hasOneUse())
7693 // Attempt to propagate the cast into the instruction for int->int casts.
7694 int NumCastsRemoved = 0;
7695 if (!isa<BitCastInst>(CI) &&
7696 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7697 CI.getOpcode(), NumCastsRemoved)) {
7698 // If this cast is a truncate, evaluting in a different type always
7699 // eliminates the cast, so it is always a win. If this is a zero-extension,
7700 // we need to do an AND to maintain the clear top-part of the computation,
7701 // so we require that the input have eliminated at least one cast. If this
7702 // is a sign extension, we insert two new casts (to do the extension) so we
7703 // require that two casts have been eliminated.
7705 switch (CI.getOpcode()) {
7707 // All the others use floating point so we shouldn't actually
7708 // get here because of the check above.
7709 assert(0 && "Unknown cast type");
7710 case Instruction::Trunc:
7713 case Instruction::ZExt:
7714 DoXForm = NumCastsRemoved >= 1;
7716 case Instruction::SExt:
7717 DoXForm = NumCastsRemoved >= 2;
7722 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7723 CI.getOpcode() == Instruction::SExt);
7724 assert(Res->getType() == DestTy);
7725 switch (CI.getOpcode()) {
7726 default: assert(0 && "Unknown cast type!");
7727 case Instruction::Trunc:
7728 case Instruction::BitCast:
7729 // Just replace this cast with the result.
7730 return ReplaceInstUsesWith(CI, Res);
7731 case Instruction::ZExt: {
7732 // We need to emit an AND to clear the high bits.
7733 assert(SrcBitSize < DestBitSize && "Not a zext?");
7734 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7736 return BinaryOperator::CreateAnd(Res, C);
7738 case Instruction::SExt:
7739 // We need to emit a cast to truncate, then a cast to sext.
7740 return CastInst::Create(Instruction::SExt,
7741 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7747 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7748 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7750 switch (SrcI->getOpcode()) {
7751 case Instruction::Add:
7752 case Instruction::Mul:
7753 case Instruction::And:
7754 case Instruction::Or:
7755 case Instruction::Xor:
7756 // If we are discarding information, rewrite.
7757 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7758 // Don't insert two casts if they cannot be eliminated. We allow
7759 // two casts to be inserted if the sizes are the same. This could
7760 // only be converting signedness, which is a noop.
7761 if (DestBitSize == SrcBitSize ||
7762 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7763 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7764 Instruction::CastOps opcode = CI.getOpcode();
7765 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7766 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7767 return BinaryOperator::Create(
7768 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7772 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7773 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7774 SrcI->getOpcode() == Instruction::Xor &&
7775 Op1 == ConstantInt::getTrue() &&
7776 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7777 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7778 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7781 case Instruction::SDiv:
7782 case Instruction::UDiv:
7783 case Instruction::SRem:
7784 case Instruction::URem:
7785 // If we are just changing the sign, rewrite.
7786 if (DestBitSize == SrcBitSize) {
7787 // Don't insert two casts if they cannot be eliminated. We allow
7788 // two casts to be inserted if the sizes are the same. This could
7789 // only be converting signedness, which is a noop.
7790 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7791 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7792 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7794 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7796 return BinaryOperator::Create(
7797 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7802 case Instruction::Shl:
7803 // Allow changing the sign of the source operand. Do not allow
7804 // changing the size of the shift, UNLESS the shift amount is a
7805 // constant. We must not change variable sized shifts to a smaller
7806 // size, because it is undefined to shift more bits out than exist
7808 if (DestBitSize == SrcBitSize ||
7809 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7810 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7811 Instruction::BitCast : Instruction::Trunc);
7812 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7813 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7814 return BinaryOperator::CreateShl(Op0c, Op1c);
7817 case Instruction::AShr:
7818 // If this is a signed shr, and if all bits shifted in are about to be
7819 // truncated off, turn it into an unsigned shr to allow greater
7821 if (DestBitSize < SrcBitSize &&
7822 isa<ConstantInt>(Op1)) {
7823 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7824 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7825 // Insert the new logical shift right.
7826 return BinaryOperator::CreateLShr(Op0, Op1);
7834 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7835 if (Instruction *Result = commonIntCastTransforms(CI))
7838 Value *Src = CI.getOperand(0);
7839 const Type *Ty = CI.getType();
7840 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7841 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7843 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7844 switch (SrcI->getOpcode()) {
7846 case Instruction::LShr:
7847 // We can shrink lshr to something smaller if we know the bits shifted in
7848 // are already zeros.
7849 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7850 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7852 // Get a mask for the bits shifting in.
7853 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7854 Value* SrcIOp0 = SrcI->getOperand(0);
7855 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7856 if (ShAmt >= DestBitWidth) // All zeros.
7857 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7859 // Okay, we can shrink this. Truncate the input, then return a new
7861 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7862 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7864 return BinaryOperator::CreateLShr(V1, V2);
7866 } else { // This is a variable shr.
7868 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7869 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7870 // loop-invariant and CSE'd.
7871 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7872 Value *One = ConstantInt::get(SrcI->getType(), 1);
7874 Value *V = InsertNewInstBefore(
7875 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7877 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7878 SrcI->getOperand(0),
7880 Value *Zero = Constant::getNullValue(V->getType());
7881 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7891 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7892 /// in order to eliminate the icmp.
7893 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7895 // If we are just checking for a icmp eq of a single bit and zext'ing it
7896 // to an integer, then shift the bit to the appropriate place and then
7897 // cast to integer to avoid the comparison.
7898 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7899 const APInt &Op1CV = Op1C->getValue();
7901 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7902 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7903 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7904 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7905 if (!DoXform) return ICI;
7907 Value *In = ICI->getOperand(0);
7908 Value *Sh = ConstantInt::get(In->getType(),
7909 In->getType()->getPrimitiveSizeInBits()-1);
7910 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7911 In->getName()+".lobit"),
7913 if (In->getType() != CI.getType())
7914 In = CastInst::CreateIntegerCast(In, CI.getType(),
7915 false/*ZExt*/, "tmp", &CI);
7917 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7918 Constant *One = ConstantInt::get(In->getType(), 1);
7919 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7920 In->getName()+".not"),
7924 return ReplaceInstUsesWith(CI, In);
7929 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7930 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7931 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7932 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7933 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7934 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7935 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7936 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7937 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7938 // This only works for EQ and NE
7939 ICI->isEquality()) {
7940 // If Op1C some other power of two, convert:
7941 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7942 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7943 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7944 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7946 APInt KnownZeroMask(~KnownZero);
7947 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7948 if (!DoXform) return ICI;
7950 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7951 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7952 // (X&4) == 2 --> false
7953 // (X&4) != 2 --> true
7954 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7955 Res = ConstantExpr::getZExt(Res, CI.getType());
7956 return ReplaceInstUsesWith(CI, Res);
7959 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7960 Value *In = ICI->getOperand(0);
7962 // Perform a logical shr by shiftamt.
7963 // Insert the shift to put the result in the low bit.
7964 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7965 ConstantInt::get(In->getType(), ShiftAmt),
7966 In->getName()+".lobit"), CI);
7969 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7970 Constant *One = ConstantInt::get(In->getType(), 1);
7971 In = BinaryOperator::CreateXor(In, One, "tmp");
7972 InsertNewInstBefore(cast<Instruction>(In), CI);
7975 if (CI.getType() == In->getType())
7976 return ReplaceInstUsesWith(CI, In);
7978 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7986 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7987 // If one of the common conversion will work ..
7988 if (Instruction *Result = commonIntCastTransforms(CI))
7991 Value *Src = CI.getOperand(0);
7993 // If this is a cast of a cast
7994 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7995 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7996 // types and if the sizes are just right we can convert this into a logical
7997 // 'and' which will be much cheaper than the pair of casts.
7998 if (isa<TruncInst>(CSrc)) {
7999 // Get the sizes of the types involved
8000 Value *A = CSrc->getOperand(0);
8001 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8002 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8003 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8004 // If we're actually extending zero bits and the trunc is a no-op
8005 if (MidSize < DstSize && SrcSize == DstSize) {
8006 // Replace both of the casts with an And of the type mask.
8007 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8008 Constant *AndConst = ConstantInt::get(AndValue);
8010 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8011 // Unfortunately, if the type changed, we need to cast it back.
8012 if (And->getType() != CI.getType()) {
8013 And->setName(CSrc->getName()+".mask");
8014 InsertNewInstBefore(And, CI);
8015 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8022 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8023 return transformZExtICmp(ICI, CI);
8025 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8026 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8027 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8028 // of the (zext icmp) will be transformed.
8029 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8030 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8031 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8032 (transformZExtICmp(LHS, CI, false) ||
8033 transformZExtICmp(RHS, CI, false))) {
8034 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8035 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8036 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8043 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8044 if (Instruction *I = commonIntCastTransforms(CI))
8047 Value *Src = CI.getOperand(0);
8049 // Canonicalize sign-extend from i1 to a select.
8050 if (Src->getType() == Type::Int1Ty)
8051 return SelectInst::Create(Src,
8052 ConstantInt::getAllOnesValue(CI.getType()),
8053 Constant::getNullValue(CI.getType()));
8055 // See if the value being truncated is already sign extended. If so, just
8056 // eliminate the trunc/sext pair.
8057 if (getOpcode(Src) == Instruction::Trunc) {
8058 Value *Op = cast<User>(Src)->getOperand(0);
8059 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8060 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8061 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8062 unsigned NumSignBits = ComputeNumSignBits(Op);
8064 if (OpBits == DestBits) {
8065 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8066 // bits, it is already ready.
8067 if (NumSignBits > DestBits-MidBits)
8068 return ReplaceInstUsesWith(CI, Op);
8069 } else if (OpBits < DestBits) {
8070 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8071 // bits, just sext from i32.
8072 if (NumSignBits > OpBits-MidBits)
8073 return new SExtInst(Op, CI.getType(), "tmp");
8075 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8076 // bits, just truncate to i32.
8077 if (NumSignBits > OpBits-MidBits)
8078 return new TruncInst(Op, CI.getType(), "tmp");
8082 // If the input is a shl/ashr pair of a same constant, then this is a sign
8083 // extension from a smaller value. If we could trust arbitrary bitwidth
8084 // integers, we could turn this into a truncate to the smaller bit and then
8085 // use a sext for the whole extension. Since we don't, look deeper and check
8086 // for a truncate. If the source and dest are the same type, eliminate the
8087 // trunc and extend and just do shifts. For example, turn:
8088 // %a = trunc i32 %i to i8
8089 // %b = shl i8 %a, 6
8090 // %c = ashr i8 %b, 6
8091 // %d = sext i8 %c to i32
8093 // %a = shl i32 %i, 30
8094 // %d = ashr i32 %a, 30
8096 ConstantInt *BA = 0, *CA = 0;
8097 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8098 m_ConstantInt(CA))) &&
8099 BA == CA && isa<TruncInst>(A)) {
8100 Value *I = cast<TruncInst>(A)->getOperand(0);
8101 if (I->getType() == CI.getType()) {
8102 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8103 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8104 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8105 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8106 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8108 return BinaryOperator::CreateAShr(I, ShAmtV);
8115 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8116 /// in the specified FP type without changing its value.
8117 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8119 APFloat F = CFP->getValueAPF();
8120 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8122 return ConstantFP::get(F);
8126 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8127 /// through it until we get the source value.
8128 static Value *LookThroughFPExtensions(Value *V) {
8129 if (Instruction *I = dyn_cast<Instruction>(V))
8130 if (I->getOpcode() == Instruction::FPExt)
8131 return LookThroughFPExtensions(I->getOperand(0));
8133 // If this value is a constant, return the constant in the smallest FP type
8134 // that can accurately represent it. This allows us to turn
8135 // (float)((double)X+2.0) into x+2.0f.
8136 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8137 if (CFP->getType() == Type::PPC_FP128Ty)
8138 return V; // No constant folding of this.
8139 // See if the value can be truncated to float and then reextended.
8140 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8142 if (CFP->getType() == Type::DoubleTy)
8143 return V; // Won't shrink.
8144 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8146 // Don't try to shrink to various long double types.
8152 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8153 if (Instruction *I = commonCastTransforms(CI))
8156 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8157 // smaller than the destination type, we can eliminate the truncate by doing
8158 // the add as the smaller type. This applies to add/sub/mul/div as well as
8159 // many builtins (sqrt, etc).
8160 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8161 if (OpI && OpI->hasOneUse()) {
8162 switch (OpI->getOpcode()) {
8164 case Instruction::Add:
8165 case Instruction::Sub:
8166 case Instruction::Mul:
8167 case Instruction::FDiv:
8168 case Instruction::FRem:
8169 const Type *SrcTy = OpI->getType();
8170 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8171 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8172 if (LHSTrunc->getType() != SrcTy &&
8173 RHSTrunc->getType() != SrcTy) {
8174 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8175 // If the source types were both smaller than the destination type of
8176 // the cast, do this xform.
8177 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8178 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8179 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8181 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8183 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8192 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8193 return commonCastTransforms(CI);
8196 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8197 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8199 return commonCastTransforms(FI);
8201 // fptoui(uitofp(X)) --> X
8202 // fptoui(sitofp(X)) --> X
8203 // This is safe if the intermediate type has enough bits in its mantissa to
8204 // accurately represent all values of X. For example, do not do this with
8205 // i64->float->i64. This is also safe for sitofp case, because any negative
8206 // 'X' value would cause an undefined result for the fptoui.
8207 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8208 OpI->getOperand(0)->getType() == FI.getType() &&
8209 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8210 OpI->getType()->getFPMantissaWidth())
8211 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8213 return commonCastTransforms(FI);
8216 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8217 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8219 return commonCastTransforms(FI);
8221 // fptosi(sitofp(X)) --> X
8222 // fptosi(uitofp(X)) --> X
8223 // This is safe if the intermediate type has enough bits in its mantissa to
8224 // accurately represent all values of X. For example, do not do this with
8225 // i64->float->i64. This is also safe for sitofp case, because any negative
8226 // 'X' value would cause an undefined result for the fptoui.
8227 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8228 OpI->getOperand(0)->getType() == FI.getType() &&
8229 (int)FI.getType()->getPrimitiveSizeInBits() <=
8230 OpI->getType()->getFPMantissaWidth())
8231 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8233 return commonCastTransforms(FI);
8236 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8237 return commonCastTransforms(CI);
8240 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8241 return commonCastTransforms(CI);
8244 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8245 return commonPointerCastTransforms(CI);
8248 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8249 if (Instruction *I = commonCastTransforms(CI))
8252 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8253 if (!DestPointee->isSized()) return 0;
8255 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8258 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8259 m_ConstantInt(Cst)))) {
8260 // If the source and destination operands have the same type, see if this
8261 // is a single-index GEP.
8262 if (X->getType() == CI.getType()) {
8263 // Get the size of the pointee type.
8264 uint64_t Size = TD->getABITypeSize(DestPointee);
8266 // Convert the constant to intptr type.
8267 APInt Offset = Cst->getValue();
8268 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8270 // If Offset is evenly divisible by Size, we can do this xform.
8271 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8272 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8273 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8276 // TODO: Could handle other cases, e.g. where add is indexing into field of
8278 } else if (CI.getOperand(0)->hasOneUse() &&
8279 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8280 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8281 // "inttoptr+GEP" instead of "add+intptr".
8283 // Get the size of the pointee type.
8284 uint64_t Size = TD->getABITypeSize(DestPointee);
8286 // Convert the constant to intptr type.
8287 APInt Offset = Cst->getValue();
8288 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8290 // If Offset is evenly divisible by Size, we can do this xform.
8291 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8292 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8294 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8296 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8302 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8303 // If the operands are integer typed then apply the integer transforms,
8304 // otherwise just apply the common ones.
8305 Value *Src = CI.getOperand(0);
8306 const Type *SrcTy = Src->getType();
8307 const Type *DestTy = CI.getType();
8309 if (SrcTy->isInteger() && DestTy->isInteger()) {
8310 if (Instruction *Result = commonIntCastTransforms(CI))
8312 } else if (isa<PointerType>(SrcTy)) {
8313 if (Instruction *I = commonPointerCastTransforms(CI))
8316 if (Instruction *Result = commonCastTransforms(CI))
8321 // Get rid of casts from one type to the same type. These are useless and can
8322 // be replaced by the operand.
8323 if (DestTy == Src->getType())
8324 return ReplaceInstUsesWith(CI, Src);
8326 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8327 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8328 const Type *DstElTy = DstPTy->getElementType();
8329 const Type *SrcElTy = SrcPTy->getElementType();
8331 // If the address spaces don't match, don't eliminate the bitcast, which is
8332 // required for changing types.
8333 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8336 // If we are casting a malloc or alloca to a pointer to a type of the same
8337 // size, rewrite the allocation instruction to allocate the "right" type.
8338 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8339 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8342 // If the source and destination are pointers, and this cast is equivalent
8343 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8344 // This can enhance SROA and other transforms that want type-safe pointers.
8345 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8346 unsigned NumZeros = 0;
8347 while (SrcElTy != DstElTy &&
8348 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8349 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8350 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8354 // If we found a path from the src to dest, create the getelementptr now.
8355 if (SrcElTy == DstElTy) {
8356 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8357 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8358 ((Instruction*) NULL));
8362 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8363 if (SVI->hasOneUse()) {
8364 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8365 // a bitconvert to a vector with the same # elts.
8366 if (isa<VectorType>(DestTy) &&
8367 cast<VectorType>(DestTy)->getNumElements() ==
8368 SVI->getType()->getNumElements() &&
8369 SVI->getType()->getNumElements() ==
8370 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8372 // If either of the operands is a cast from CI.getType(), then
8373 // evaluating the shuffle in the casted destination's type will allow
8374 // us to eliminate at least one cast.
8375 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8376 Tmp->getOperand(0)->getType() == DestTy) ||
8377 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8378 Tmp->getOperand(0)->getType() == DestTy)) {
8379 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8380 SVI->getOperand(0), DestTy, &CI);
8381 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8382 SVI->getOperand(1), DestTy, &CI);
8383 // Return a new shuffle vector. Use the same element ID's, as we
8384 // know the vector types match #elts.
8385 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8393 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8395 /// %D = select %cond, %C, %A
8397 /// %C = select %cond, %B, 0
8400 /// Assuming that the specified instruction is an operand to the select, return
8401 /// a bitmask indicating which operands of this instruction are foldable if they
8402 /// equal the other incoming value of the select.
8404 static unsigned GetSelectFoldableOperands(Instruction *I) {
8405 switch (I->getOpcode()) {
8406 case Instruction::Add:
8407 case Instruction::Mul:
8408 case Instruction::And:
8409 case Instruction::Or:
8410 case Instruction::Xor:
8411 return 3; // Can fold through either operand.
8412 case Instruction::Sub: // Can only fold on the amount subtracted.
8413 case Instruction::Shl: // Can only fold on the shift amount.
8414 case Instruction::LShr:
8415 case Instruction::AShr:
8418 return 0; // Cannot fold
8422 /// GetSelectFoldableConstant - For the same transformation as the previous
8423 /// function, return the identity constant that goes into the select.
8424 static Constant *GetSelectFoldableConstant(Instruction *I) {
8425 switch (I->getOpcode()) {
8426 default: assert(0 && "This cannot happen!"); abort();
8427 case Instruction::Add:
8428 case Instruction::Sub:
8429 case Instruction::Or:
8430 case Instruction::Xor:
8431 case Instruction::Shl:
8432 case Instruction::LShr:
8433 case Instruction::AShr:
8434 return Constant::getNullValue(I->getType());
8435 case Instruction::And:
8436 return Constant::getAllOnesValue(I->getType());
8437 case Instruction::Mul:
8438 return ConstantInt::get(I->getType(), 1);
8442 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8443 /// have the same opcode and only one use each. Try to simplify this.
8444 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8446 if (TI->getNumOperands() == 1) {
8447 // If this is a non-volatile load or a cast from the same type,
8450 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8453 return 0; // unknown unary op.
8456 // Fold this by inserting a select from the input values.
8457 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8458 FI->getOperand(0), SI.getName()+".v");
8459 InsertNewInstBefore(NewSI, SI);
8460 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8464 // Only handle binary operators here.
8465 if (!isa<BinaryOperator>(TI))
8468 // Figure out if the operations have any operands in common.
8469 Value *MatchOp, *OtherOpT, *OtherOpF;
8471 if (TI->getOperand(0) == FI->getOperand(0)) {
8472 MatchOp = TI->getOperand(0);
8473 OtherOpT = TI->getOperand(1);
8474 OtherOpF = FI->getOperand(1);
8475 MatchIsOpZero = true;
8476 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8477 MatchOp = TI->getOperand(1);
8478 OtherOpT = TI->getOperand(0);
8479 OtherOpF = FI->getOperand(0);
8480 MatchIsOpZero = false;
8481 } else if (!TI->isCommutative()) {
8483 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8484 MatchOp = TI->getOperand(0);
8485 OtherOpT = TI->getOperand(1);
8486 OtherOpF = FI->getOperand(0);
8487 MatchIsOpZero = true;
8488 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8489 MatchOp = TI->getOperand(1);
8490 OtherOpT = TI->getOperand(0);
8491 OtherOpF = FI->getOperand(1);
8492 MatchIsOpZero = true;
8497 // If we reach here, they do have operations in common.
8498 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8499 OtherOpF, SI.getName()+".v");
8500 InsertNewInstBefore(NewSI, SI);
8502 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8504 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8506 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8508 assert(0 && "Shouldn't get here");
8512 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8513 /// ICmpInst as its first operand.
8515 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8517 bool Changed = false;
8518 ICmpInst::Predicate Pred = ICI->getPredicate();
8519 Value *CmpLHS = ICI->getOperand(0);
8520 Value *CmpRHS = ICI->getOperand(1);
8521 Value *TrueVal = SI.getTrueValue();
8522 Value *FalseVal = SI.getFalseValue();
8524 // Check cases where the comparison is with a constant that
8525 // can be adjusted to fit the min/max idiom. We may edit ICI in
8526 // place here, so make sure the select is the only user.
8527 if (ICI->hasOneUse())
8528 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8531 case ICmpInst::ICMP_ULT:
8532 case ICmpInst::ICMP_SLT: {
8533 // X < MIN ? T : F --> F
8534 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8535 return ReplaceInstUsesWith(SI, FalseVal);
8536 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8537 Constant *AdjustedRHS = SubOne(CI);
8538 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8539 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8540 Pred = ICmpInst::getSwappedPredicate(Pred);
8541 CmpRHS = AdjustedRHS;
8542 std::swap(FalseVal, TrueVal);
8543 ICI->setPredicate(Pred);
8544 ICI->setOperand(1, CmpRHS);
8545 SI.setOperand(1, TrueVal);
8546 SI.setOperand(2, FalseVal);
8551 case ICmpInst::ICMP_UGT:
8552 case ICmpInst::ICMP_SGT: {
8553 // X > MAX ? T : F --> F
8554 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8555 return ReplaceInstUsesWith(SI, FalseVal);
8556 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8557 Constant *AdjustedRHS = AddOne(CI);
8558 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8559 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8560 Pred = ICmpInst::getSwappedPredicate(Pred);
8561 CmpRHS = AdjustedRHS;
8562 std::swap(FalseVal, TrueVal);
8563 ICI->setPredicate(Pred);
8564 ICI->setOperand(1, CmpRHS);
8565 SI.setOperand(1, TrueVal);
8566 SI.setOperand(2, FalseVal);
8573 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8574 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8575 CmpInst::Predicate Pred = ICI->getPredicate();
8576 if (match(TrueVal, m_ConstantInt(0)) &&
8577 match(FalseVal, m_ConstantInt(-1)))
8578 Pred = CmpInst::getInversePredicate(Pred);
8579 else if (!match(TrueVal, m_ConstantInt(-1)) ||
8580 !match(FalseVal, m_ConstantInt(0)))
8581 Pred = CmpInst::BAD_ICMP_PREDICATE;
8582 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8583 // If we are just checking for a icmp eq of a single bit and zext'ing it
8584 // to an integer, then shift the bit to the appropriate place and then
8585 // cast to integer to avoid the comparison.
8586 const APInt &Op1CV = CI->getValue();
8588 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8589 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8590 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8591 (Pred == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8592 Value *In = ICI->getOperand(0);
8593 Value *Sh = ConstantInt::get(In->getType(),
8594 In->getType()->getPrimitiveSizeInBits()-1);
8595 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8596 In->getName()+".lobit"),
8598 if (In->getType() != SI.getType())
8599 In = CastInst::CreateIntegerCast(In, SI.getType(),
8600 true/*SExt*/, "tmp", ICI);
8602 if (Pred == ICmpInst::ICMP_SGT)
8603 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8604 In->getName()+".not"), *ICI);
8606 return ReplaceInstUsesWith(SI, In);
8611 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8612 // Transform (X == Y) ? X : Y -> Y
8613 if (Pred == ICmpInst::ICMP_EQ)
8614 return ReplaceInstUsesWith(SI, FalseVal);
8615 // Transform (X != Y) ? X : Y -> X
8616 if (Pred == ICmpInst::ICMP_NE)
8617 return ReplaceInstUsesWith(SI, TrueVal);
8618 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8620 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8621 // Transform (X == Y) ? Y : X -> X
8622 if (Pred == ICmpInst::ICMP_EQ)
8623 return ReplaceInstUsesWith(SI, FalseVal);
8624 // Transform (X != Y) ? Y : X -> Y
8625 if (Pred == ICmpInst::ICMP_NE)
8626 return ReplaceInstUsesWith(SI, TrueVal);
8627 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8630 /// NOTE: if we wanted to, this is where to detect integer ABS
8632 return Changed ? &SI : 0;
8635 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8636 Value *CondVal = SI.getCondition();
8637 Value *TrueVal = SI.getTrueValue();
8638 Value *FalseVal = SI.getFalseValue();
8640 // select true, X, Y -> X
8641 // select false, X, Y -> Y
8642 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8643 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8645 // select C, X, X -> X
8646 if (TrueVal == FalseVal)
8647 return ReplaceInstUsesWith(SI, TrueVal);
8649 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8650 return ReplaceInstUsesWith(SI, FalseVal);
8651 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8652 return ReplaceInstUsesWith(SI, TrueVal);
8653 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8654 if (isa<Constant>(TrueVal))
8655 return ReplaceInstUsesWith(SI, TrueVal);
8657 return ReplaceInstUsesWith(SI, FalseVal);
8660 if (SI.getType() == Type::Int1Ty) {
8661 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8662 if (C->getZExtValue()) {
8663 // Change: A = select B, true, C --> A = or B, C
8664 return BinaryOperator::CreateOr(CondVal, FalseVal);
8666 // Change: A = select B, false, C --> A = and !B, C
8668 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8669 "not."+CondVal->getName()), SI);
8670 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8672 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8673 if (C->getZExtValue() == false) {
8674 // Change: A = select B, C, false --> A = and B, C
8675 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8677 // Change: A = select B, C, true --> A = or !B, C
8679 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8680 "not."+CondVal->getName()), SI);
8681 return BinaryOperator::CreateOr(NotCond, TrueVal);
8685 // select a, b, a -> a&b
8686 // select a, a, b -> a|b
8687 if (CondVal == TrueVal)
8688 return BinaryOperator::CreateOr(CondVal, FalseVal);
8689 else if (CondVal == FalseVal)
8690 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8693 // Selecting between two integer constants?
8694 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8695 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8696 // select C, 1, 0 -> zext C to int
8697 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8698 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8699 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8700 // select C, 0, 1 -> zext !C to int
8702 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8703 "not."+CondVal->getName()), SI);
8704 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8707 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8709 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8711 // (x <s 0) ? -1 : 0 -> ashr x, 31
8712 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8713 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8714 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8715 // The comparison constant and the result are not neccessarily the
8716 // same width. Make an all-ones value by inserting a AShr.
8717 Value *X = IC->getOperand(0);
8718 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8719 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8720 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8722 InsertNewInstBefore(SRA, SI);
8724 // Finally, convert to the type of the select RHS. We figure out
8725 // if this requires a SExt, Trunc or BitCast based on the sizes.
8726 Instruction::CastOps opc = Instruction::BitCast;
8727 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8728 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8729 if (SRASize < SISize)
8730 opc = Instruction::SExt;
8731 else if (SRASize > SISize)
8732 opc = Instruction::Trunc;
8733 return CastInst::Create(opc, SRA, SI.getType());
8738 // If one of the constants is zero (we know they can't both be) and we
8739 // have an icmp instruction with zero, and we have an 'and' with the
8740 // non-constant value, eliminate this whole mess. This corresponds to
8741 // cases like this: ((X & 27) ? 27 : 0)
8742 if (TrueValC->isZero() || FalseValC->isZero())
8743 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8744 cast<Constant>(IC->getOperand(1))->isNullValue())
8745 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8746 if (ICA->getOpcode() == Instruction::And &&
8747 isa<ConstantInt>(ICA->getOperand(1)) &&
8748 (ICA->getOperand(1) == TrueValC ||
8749 ICA->getOperand(1) == FalseValC) &&
8750 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8751 // Okay, now we know that everything is set up, we just don't
8752 // know whether we have a icmp_ne or icmp_eq and whether the
8753 // true or false val is the zero.
8754 bool ShouldNotVal = !TrueValC->isZero();
8755 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8758 V = InsertNewInstBefore(BinaryOperator::Create(
8759 Instruction::Xor, V, ICA->getOperand(1)), SI);
8760 return ReplaceInstUsesWith(SI, V);
8765 // See if we are selecting two values based on a comparison of the two values.
8766 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8767 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8768 // Transform (X == Y) ? X : Y -> Y
8769 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8770 // This is not safe in general for floating point:
8771 // consider X== -0, Y== +0.
8772 // It becomes safe if either operand is a nonzero constant.
8773 ConstantFP *CFPt, *CFPf;
8774 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8775 !CFPt->getValueAPF().isZero()) ||
8776 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8777 !CFPf->getValueAPF().isZero()))
8778 return ReplaceInstUsesWith(SI, FalseVal);
8780 // Transform (X != Y) ? X : Y -> X
8781 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8782 return ReplaceInstUsesWith(SI, TrueVal);
8783 // NOTE: if we wanted to, this is where to detect MIN/MAX
8785 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8786 // Transform (X == Y) ? Y : X -> X
8787 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8788 // This is not safe in general for floating point:
8789 // consider X== -0, Y== +0.
8790 // It becomes safe if either operand is a nonzero constant.
8791 ConstantFP *CFPt, *CFPf;
8792 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8793 !CFPt->getValueAPF().isZero()) ||
8794 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8795 !CFPf->getValueAPF().isZero()))
8796 return ReplaceInstUsesWith(SI, FalseVal);
8798 // Transform (X != Y) ? Y : X -> Y
8799 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8800 return ReplaceInstUsesWith(SI, TrueVal);
8801 // NOTE: if we wanted to, this is where to detect MIN/MAX
8803 // NOTE: if we wanted to, this is where to detect ABS
8806 // See if we are selecting two values based on a comparison of the two values.
8807 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8808 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8811 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8812 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8813 if (TI->hasOneUse() && FI->hasOneUse()) {
8814 Instruction *AddOp = 0, *SubOp = 0;
8816 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8817 if (TI->getOpcode() == FI->getOpcode())
8818 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8821 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8822 // even legal for FP.
8823 if (TI->getOpcode() == Instruction::Sub &&
8824 FI->getOpcode() == Instruction::Add) {
8825 AddOp = FI; SubOp = TI;
8826 } else if (FI->getOpcode() == Instruction::Sub &&
8827 TI->getOpcode() == Instruction::Add) {
8828 AddOp = TI; SubOp = FI;
8832 Value *OtherAddOp = 0;
8833 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8834 OtherAddOp = AddOp->getOperand(1);
8835 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8836 OtherAddOp = AddOp->getOperand(0);
8840 // So at this point we know we have (Y -> OtherAddOp):
8841 // select C, (add X, Y), (sub X, Z)
8842 Value *NegVal; // Compute -Z
8843 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8844 NegVal = ConstantExpr::getNeg(C);
8846 NegVal = InsertNewInstBefore(
8847 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8850 Value *NewTrueOp = OtherAddOp;
8851 Value *NewFalseOp = NegVal;
8853 std::swap(NewTrueOp, NewFalseOp);
8854 Instruction *NewSel =
8855 SelectInst::Create(CondVal, NewTrueOp,
8856 NewFalseOp, SI.getName() + ".p");
8858 NewSel = InsertNewInstBefore(NewSel, SI);
8859 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8864 // See if we can fold the select into one of our operands.
8865 if (SI.getType()->isInteger()) {
8866 // See the comment above GetSelectFoldableOperands for a description of the
8867 // transformation we are doing here.
8868 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8869 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8870 !isa<Constant>(FalseVal))
8871 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8872 unsigned OpToFold = 0;
8873 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8875 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8880 Constant *C = GetSelectFoldableConstant(TVI);
8881 Instruction *NewSel =
8882 SelectInst::Create(SI.getCondition(),
8883 TVI->getOperand(2-OpToFold), C);
8884 InsertNewInstBefore(NewSel, SI);
8885 NewSel->takeName(TVI);
8886 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8887 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8889 assert(0 && "Unknown instruction!!");
8894 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8895 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8896 !isa<Constant>(TrueVal))
8897 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8898 unsigned OpToFold = 0;
8899 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8901 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8906 Constant *C = GetSelectFoldableConstant(FVI);
8907 Instruction *NewSel =
8908 SelectInst::Create(SI.getCondition(), C,
8909 FVI->getOperand(2-OpToFold));
8910 InsertNewInstBefore(NewSel, SI);
8911 NewSel->takeName(FVI);
8912 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8913 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8915 assert(0 && "Unknown instruction!!");
8920 if (BinaryOperator::isNot(CondVal)) {
8921 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8922 SI.setOperand(1, FalseVal);
8923 SI.setOperand(2, TrueVal);
8930 /// EnforceKnownAlignment - If the specified pointer points to an object that
8931 /// we control, modify the object's alignment to PrefAlign. This isn't
8932 /// often possible though. If alignment is important, a more reliable approach
8933 /// is to simply align all global variables and allocation instructions to
8934 /// their preferred alignment from the beginning.
8936 static unsigned EnforceKnownAlignment(Value *V,
8937 unsigned Align, unsigned PrefAlign) {
8939 User *U = dyn_cast<User>(V);
8940 if (!U) return Align;
8942 switch (getOpcode(U)) {
8944 case Instruction::BitCast:
8945 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8946 case Instruction::GetElementPtr: {
8947 // If all indexes are zero, it is just the alignment of the base pointer.
8948 bool AllZeroOperands = true;
8949 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8950 if (!isa<Constant>(*i) ||
8951 !cast<Constant>(*i)->isNullValue()) {
8952 AllZeroOperands = false;
8956 if (AllZeroOperands) {
8957 // Treat this like a bitcast.
8958 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8964 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8965 // If there is a large requested alignment and we can, bump up the alignment
8967 if (!GV->isDeclaration()) {
8968 GV->setAlignment(PrefAlign);
8971 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8972 // If there is a requested alignment and if this is an alloca, round up. We
8973 // don't do this for malloc, because some systems can't respect the request.
8974 if (isa<AllocaInst>(AI)) {
8975 AI->setAlignment(PrefAlign);
8983 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8984 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8985 /// and it is more than the alignment of the ultimate object, see if we can
8986 /// increase the alignment of the ultimate object, making this check succeed.
8987 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8988 unsigned PrefAlign) {
8989 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8990 sizeof(PrefAlign) * CHAR_BIT;
8991 APInt Mask = APInt::getAllOnesValue(BitWidth);
8992 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8993 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8994 unsigned TrailZ = KnownZero.countTrailingOnes();
8995 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8997 if (PrefAlign > Align)
8998 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9000 // We don't need to make any adjustment.
9004 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9005 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9006 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9007 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9008 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9010 if (CopyAlign < MinAlign) {
9011 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9015 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9017 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9018 if (MemOpLength == 0) return 0;
9020 // Source and destination pointer types are always "i8*" for intrinsic. See
9021 // if the size is something we can handle with a single primitive load/store.
9022 // A single load+store correctly handles overlapping memory in the memmove
9024 unsigned Size = MemOpLength->getZExtValue();
9025 if (Size == 0) return MI; // Delete this mem transfer.
9027 if (Size > 8 || (Size&(Size-1)))
9028 return 0; // If not 1/2/4/8 bytes, exit.
9030 // Use an integer load+store unless we can find something better.
9031 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9033 // Memcpy forces the use of i8* for the source and destination. That means
9034 // that if you're using memcpy to move one double around, you'll get a cast
9035 // from double* to i8*. We'd much rather use a double load+store rather than
9036 // an i64 load+store, here because this improves the odds that the source or
9037 // dest address will be promotable. See if we can find a better type than the
9038 // integer datatype.
9039 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9040 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9041 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9042 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9043 // down through these levels if so.
9044 while (!SrcETy->isSingleValueType()) {
9045 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9046 if (STy->getNumElements() == 1)
9047 SrcETy = STy->getElementType(0);
9050 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9051 if (ATy->getNumElements() == 1)
9052 SrcETy = ATy->getElementType();
9059 if (SrcETy->isSingleValueType())
9060 NewPtrTy = PointerType::getUnqual(SrcETy);
9065 // If the memcpy/memmove provides better alignment info than we can
9067 SrcAlign = std::max(SrcAlign, CopyAlign);
9068 DstAlign = std::max(DstAlign, CopyAlign);
9070 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9071 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9072 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9073 InsertNewInstBefore(L, *MI);
9074 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9076 // Set the size of the copy to 0, it will be deleted on the next iteration.
9077 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9081 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9082 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9083 if (MI->getAlignment()->getZExtValue() < Alignment) {
9084 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9088 // Extract the length and alignment and fill if they are constant.
9089 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9090 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9091 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9093 uint64_t Len = LenC->getZExtValue();
9094 Alignment = MI->getAlignment()->getZExtValue();
9096 // If the length is zero, this is a no-op
9097 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9099 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9100 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9101 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9103 Value *Dest = MI->getDest();
9104 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9106 // Alignment 0 is identity for alignment 1 for memset, but not store.
9107 if (Alignment == 0) Alignment = 1;
9109 // Extract the fill value and store.
9110 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9111 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9114 // Set the size of the copy to 0, it will be deleted on the next iteration.
9115 MI->setLength(Constant::getNullValue(LenC->getType()));
9123 /// visitCallInst - CallInst simplification. This mostly only handles folding
9124 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9125 /// the heavy lifting.
9127 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9128 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9129 if (!II) return visitCallSite(&CI);
9131 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9133 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9134 bool Changed = false;
9136 // memmove/cpy/set of zero bytes is a noop.
9137 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9138 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9140 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9141 if (CI->getZExtValue() == 1) {
9142 // Replace the instruction with just byte operations. We would
9143 // transform other cases to loads/stores, but we don't know if
9144 // alignment is sufficient.
9148 // If we have a memmove and the source operation is a constant global,
9149 // then the source and dest pointers can't alias, so we can change this
9150 // into a call to memcpy.
9151 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9152 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9153 if (GVSrc->isConstant()) {
9154 Module *M = CI.getParent()->getParent()->getParent();
9155 Intrinsic::ID MemCpyID;
9156 if (CI.getOperand(3)->getType() == Type::Int32Ty)
9157 MemCpyID = Intrinsic::memcpy_i32;
9159 MemCpyID = Intrinsic::memcpy_i64;
9160 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
9164 // memmove(x,x,size) -> noop.
9165 if (MMI->getSource() == MMI->getDest())
9166 return EraseInstFromFunction(CI);
9169 // If we can determine a pointer alignment that is bigger than currently
9170 // set, update the alignment.
9171 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9172 if (Instruction *I = SimplifyMemTransfer(MI))
9174 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9175 if (Instruction *I = SimplifyMemSet(MSI))
9179 if (Changed) return II;
9182 switch (II->getIntrinsicID()) {
9184 case Intrinsic::bswap:
9185 // bswap(bswap(x)) -> x
9186 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9187 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9188 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9190 case Intrinsic::ppc_altivec_lvx:
9191 case Intrinsic::ppc_altivec_lvxl:
9192 case Intrinsic::x86_sse_loadu_ps:
9193 case Intrinsic::x86_sse2_loadu_pd:
9194 case Intrinsic::x86_sse2_loadu_dq:
9195 // Turn PPC lvx -> load if the pointer is known aligned.
9196 // Turn X86 loadups -> load if the pointer is known aligned.
9197 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9198 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9199 PointerType::getUnqual(II->getType()),
9201 return new LoadInst(Ptr);
9204 case Intrinsic::ppc_altivec_stvx:
9205 case Intrinsic::ppc_altivec_stvxl:
9206 // Turn stvx -> store if the pointer is known aligned.
9207 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9208 const Type *OpPtrTy =
9209 PointerType::getUnqual(II->getOperand(1)->getType());
9210 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9211 return new StoreInst(II->getOperand(1), Ptr);
9214 case Intrinsic::x86_sse_storeu_ps:
9215 case Intrinsic::x86_sse2_storeu_pd:
9216 case Intrinsic::x86_sse2_storeu_dq:
9217 // Turn X86 storeu -> store if the pointer is known aligned.
9218 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9219 const Type *OpPtrTy =
9220 PointerType::getUnqual(II->getOperand(2)->getType());
9221 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9222 return new StoreInst(II->getOperand(2), Ptr);
9226 case Intrinsic::x86_sse_cvttss2si: {
9227 // These intrinsics only demands the 0th element of its input vector. If
9228 // we can simplify the input based on that, do so now.
9230 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9232 II->setOperand(1, V);
9238 case Intrinsic::ppc_altivec_vperm:
9239 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9240 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9241 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9243 // Check that all of the elements are integer constants or undefs.
9244 bool AllEltsOk = true;
9245 for (unsigned i = 0; i != 16; ++i) {
9246 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9247 !isa<UndefValue>(Mask->getOperand(i))) {
9254 // Cast the input vectors to byte vectors.
9255 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9256 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9257 Value *Result = UndefValue::get(Op0->getType());
9259 // Only extract each element once.
9260 Value *ExtractedElts[32];
9261 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9263 for (unsigned i = 0; i != 16; ++i) {
9264 if (isa<UndefValue>(Mask->getOperand(i)))
9266 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9267 Idx &= 31; // Match the hardware behavior.
9269 if (ExtractedElts[Idx] == 0) {
9271 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9272 InsertNewInstBefore(Elt, CI);
9273 ExtractedElts[Idx] = Elt;
9276 // Insert this value into the result vector.
9277 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9279 InsertNewInstBefore(cast<Instruction>(Result), CI);
9281 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9286 case Intrinsic::stackrestore: {
9287 // If the save is right next to the restore, remove the restore. This can
9288 // happen when variable allocas are DCE'd.
9289 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9290 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9291 BasicBlock::iterator BI = SS;
9293 return EraseInstFromFunction(CI);
9297 // Scan down this block to see if there is another stack restore in the
9298 // same block without an intervening call/alloca.
9299 BasicBlock::iterator BI = II;
9300 TerminatorInst *TI = II->getParent()->getTerminator();
9301 bool CannotRemove = false;
9302 for (++BI; &*BI != TI; ++BI) {
9303 if (isa<AllocaInst>(BI)) {
9304 CannotRemove = true;
9307 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9308 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9309 // If there is a stackrestore below this one, remove this one.
9310 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9311 return EraseInstFromFunction(CI);
9312 // Otherwise, ignore the intrinsic.
9314 // If we found a non-intrinsic call, we can't remove the stack
9316 CannotRemove = true;
9322 // If the stack restore is in a return/unwind block and if there are no
9323 // allocas or calls between the restore and the return, nuke the restore.
9324 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9325 return EraseInstFromFunction(CI);
9330 return visitCallSite(II);
9333 // InvokeInst simplification
9335 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9336 return visitCallSite(&II);
9339 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9340 /// passed through the varargs area, we can eliminate the use of the cast.
9341 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9342 const CastInst * const CI,
9343 const TargetData * const TD,
9345 if (!CI->isLosslessCast())
9348 // The size of ByVal arguments is derived from the type, so we
9349 // can't change to a type with a different size. If the size were
9350 // passed explicitly we could avoid this check.
9351 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9355 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9356 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9357 if (!SrcTy->isSized() || !DstTy->isSized())
9359 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9364 // visitCallSite - Improvements for call and invoke instructions.
9366 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9367 bool Changed = false;
9369 // If the callee is a constexpr cast of a function, attempt to move the cast
9370 // to the arguments of the call/invoke.
9371 if (transformConstExprCastCall(CS)) return 0;
9373 Value *Callee = CS.getCalledValue();
9375 if (Function *CalleeF = dyn_cast<Function>(Callee))
9376 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9377 Instruction *OldCall = CS.getInstruction();
9378 // If the call and callee calling conventions don't match, this call must
9379 // be unreachable, as the call is undefined.
9380 new StoreInst(ConstantInt::getTrue(),
9381 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9383 if (!OldCall->use_empty())
9384 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9385 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9386 return EraseInstFromFunction(*OldCall);
9390 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9391 // This instruction is not reachable, just remove it. We insert a store to
9392 // undef so that we know that this code is not reachable, despite the fact
9393 // that we can't modify the CFG here.
9394 new StoreInst(ConstantInt::getTrue(),
9395 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9396 CS.getInstruction());
9398 if (!CS.getInstruction()->use_empty())
9399 CS.getInstruction()->
9400 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9402 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9403 // Don't break the CFG, insert a dummy cond branch.
9404 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9405 ConstantInt::getTrue(), II);
9407 return EraseInstFromFunction(*CS.getInstruction());
9410 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9411 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9412 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9413 return transformCallThroughTrampoline(CS);
9415 const PointerType *PTy = cast<PointerType>(Callee->getType());
9416 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9417 if (FTy->isVarArg()) {
9418 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9419 // See if we can optimize any arguments passed through the varargs area of
9421 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9422 E = CS.arg_end(); I != E; ++I, ++ix) {
9423 CastInst *CI = dyn_cast<CastInst>(*I);
9424 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9425 *I = CI->getOperand(0);
9431 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9432 // Inline asm calls cannot throw - mark them 'nounwind'.
9433 CS.setDoesNotThrow();
9437 return Changed ? CS.getInstruction() : 0;
9440 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9441 // attempt to move the cast to the arguments of the call/invoke.
9443 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9444 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9445 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9446 if (CE->getOpcode() != Instruction::BitCast ||
9447 !isa<Function>(CE->getOperand(0)))
9449 Function *Callee = cast<Function>(CE->getOperand(0));
9450 Instruction *Caller = CS.getInstruction();
9451 const AttrListPtr &CallerPAL = CS.getAttributes();
9453 // Okay, this is a cast from a function to a different type. Unless doing so
9454 // would cause a type conversion of one of our arguments, change this call to
9455 // be a direct call with arguments casted to the appropriate types.
9457 const FunctionType *FT = Callee->getFunctionType();
9458 const Type *OldRetTy = Caller->getType();
9459 const Type *NewRetTy = FT->getReturnType();
9461 if (isa<StructType>(NewRetTy))
9462 return false; // TODO: Handle multiple return values.
9464 // Check to see if we are changing the return type...
9465 if (OldRetTy != NewRetTy) {
9466 if (Callee->isDeclaration() &&
9467 // Conversion is ok if changing from one pointer type to another or from
9468 // a pointer to an integer of the same size.
9469 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9470 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9471 return false; // Cannot transform this return value.
9473 if (!Caller->use_empty() &&
9474 // void -> non-void is handled specially
9475 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9476 return false; // Cannot transform this return value.
9478 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9479 Attributes RAttrs = CallerPAL.getRetAttributes();
9480 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9481 return false; // Attribute not compatible with transformed value.
9484 // If the callsite is an invoke instruction, and the return value is used by
9485 // a PHI node in a successor, we cannot change the return type of the call
9486 // because there is no place to put the cast instruction (without breaking
9487 // the critical edge). Bail out in this case.
9488 if (!Caller->use_empty())
9489 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9490 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9492 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9493 if (PN->getParent() == II->getNormalDest() ||
9494 PN->getParent() == II->getUnwindDest())
9498 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9499 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9501 CallSite::arg_iterator AI = CS.arg_begin();
9502 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9503 const Type *ParamTy = FT->getParamType(i);
9504 const Type *ActTy = (*AI)->getType();
9506 if (!CastInst::isCastable(ActTy, ParamTy))
9507 return false; // Cannot transform this parameter value.
9509 if (CallerPAL.getParamAttributes(i + 1)
9510 & Attribute::typeIncompatible(ParamTy))
9511 return false; // Attribute not compatible with transformed value.
9513 // Converting from one pointer type to another or between a pointer and an
9514 // integer of the same size is safe even if we do not have a body.
9515 bool isConvertible = ActTy == ParamTy ||
9516 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9517 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9518 if (Callee->isDeclaration() && !isConvertible) return false;
9521 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9522 Callee->isDeclaration())
9523 return false; // Do not delete arguments unless we have a function body.
9525 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9526 !CallerPAL.isEmpty())
9527 // In this case we have more arguments than the new function type, but we
9528 // won't be dropping them. Check that these extra arguments have attributes
9529 // that are compatible with being a vararg call argument.
9530 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9531 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9533 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9534 if (PAttrs & Attribute::VarArgsIncompatible)
9538 // Okay, we decided that this is a safe thing to do: go ahead and start
9539 // inserting cast instructions as necessary...
9540 std::vector<Value*> Args;
9541 Args.reserve(NumActualArgs);
9542 SmallVector<AttributeWithIndex, 8> attrVec;
9543 attrVec.reserve(NumCommonArgs);
9545 // Get any return attributes.
9546 Attributes RAttrs = CallerPAL.getRetAttributes();
9548 // If the return value is not being used, the type may not be compatible
9549 // with the existing attributes. Wipe out any problematic attributes.
9550 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9552 // Add the new return attributes.
9554 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9556 AI = CS.arg_begin();
9557 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9558 const Type *ParamTy = FT->getParamType(i);
9559 if ((*AI)->getType() == ParamTy) {
9560 Args.push_back(*AI);
9562 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9563 false, ParamTy, false);
9564 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9565 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9568 // Add any parameter attributes.
9569 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9570 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9573 // If the function takes more arguments than the call was taking, add them
9575 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9576 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9578 // If we are removing arguments to the function, emit an obnoxious warning...
9579 if (FT->getNumParams() < NumActualArgs) {
9580 if (!FT->isVarArg()) {
9581 cerr << "WARNING: While resolving call to function '"
9582 << Callee->getName() << "' arguments were dropped!\n";
9584 // Add all of the arguments in their promoted form to the arg list...
9585 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9586 const Type *PTy = getPromotedType((*AI)->getType());
9587 if (PTy != (*AI)->getType()) {
9588 // Must promote to pass through va_arg area!
9589 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9591 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9592 InsertNewInstBefore(Cast, *Caller);
9593 Args.push_back(Cast);
9595 Args.push_back(*AI);
9598 // Add any parameter attributes.
9599 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9600 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9605 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9606 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9608 if (NewRetTy == Type::VoidTy)
9609 Caller->setName(""); // Void type should not have a name.
9611 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9614 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9615 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9616 Args.begin(), Args.end(),
9617 Caller->getName(), Caller);
9618 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9619 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9621 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9622 Caller->getName(), Caller);
9623 CallInst *CI = cast<CallInst>(Caller);
9624 if (CI->isTailCall())
9625 cast<CallInst>(NC)->setTailCall();
9626 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9627 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9630 // Insert a cast of the return type as necessary.
9632 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9633 if (NV->getType() != Type::VoidTy) {
9634 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9636 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9638 // If this is an invoke instruction, we should insert it after the first
9639 // non-phi, instruction in the normal successor block.
9640 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9641 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9642 InsertNewInstBefore(NC, *I);
9644 // Otherwise, it's a call, just insert cast right after the call instr
9645 InsertNewInstBefore(NC, *Caller);
9647 AddUsersToWorkList(*Caller);
9649 NV = UndefValue::get(Caller->getType());
9653 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9654 Caller->replaceAllUsesWith(NV);
9655 Caller->eraseFromParent();
9656 RemoveFromWorkList(Caller);
9660 // transformCallThroughTrampoline - Turn a call to a function created by the
9661 // init_trampoline intrinsic into a direct call to the underlying function.
9663 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9664 Value *Callee = CS.getCalledValue();
9665 const PointerType *PTy = cast<PointerType>(Callee->getType());
9666 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9667 const AttrListPtr &Attrs = CS.getAttributes();
9669 // If the call already has the 'nest' attribute somewhere then give up -
9670 // otherwise 'nest' would occur twice after splicing in the chain.
9671 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9674 IntrinsicInst *Tramp =
9675 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9677 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9678 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9679 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9681 const AttrListPtr &NestAttrs = NestF->getAttributes();
9682 if (!NestAttrs.isEmpty()) {
9683 unsigned NestIdx = 1;
9684 const Type *NestTy = 0;
9685 Attributes NestAttr = Attribute::None;
9687 // Look for a parameter marked with the 'nest' attribute.
9688 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9689 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9690 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9691 // Record the parameter type and any other attributes.
9693 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9698 Instruction *Caller = CS.getInstruction();
9699 std::vector<Value*> NewArgs;
9700 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9702 SmallVector<AttributeWithIndex, 8> NewAttrs;
9703 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9705 // Insert the nest argument into the call argument list, which may
9706 // mean appending it. Likewise for attributes.
9708 // Add any result attributes.
9709 if (Attributes Attr = Attrs.getRetAttributes())
9710 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9714 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9716 if (Idx == NestIdx) {
9717 // Add the chain argument and attributes.
9718 Value *NestVal = Tramp->getOperand(3);
9719 if (NestVal->getType() != NestTy)
9720 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9721 NewArgs.push_back(NestVal);
9722 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9728 // Add the original argument and attributes.
9729 NewArgs.push_back(*I);
9730 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9732 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9738 // Add any function attributes.
9739 if (Attributes Attr = Attrs.getFnAttributes())
9740 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
9742 // The trampoline may have been bitcast to a bogus type (FTy).
9743 // Handle this by synthesizing a new function type, equal to FTy
9744 // with the chain parameter inserted.
9746 std::vector<const Type*> NewTypes;
9747 NewTypes.reserve(FTy->getNumParams()+1);
9749 // Insert the chain's type into the list of parameter types, which may
9750 // mean appending it.
9753 FunctionType::param_iterator I = FTy->param_begin(),
9754 E = FTy->param_end();
9758 // Add the chain's type.
9759 NewTypes.push_back(NestTy);
9764 // Add the original type.
9765 NewTypes.push_back(*I);
9771 // Replace the trampoline call with a direct call. Let the generic
9772 // code sort out any function type mismatches.
9773 FunctionType *NewFTy =
9774 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9775 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9776 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9777 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
9779 Instruction *NewCaller;
9780 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9781 NewCaller = InvokeInst::Create(NewCallee,
9782 II->getNormalDest(), II->getUnwindDest(),
9783 NewArgs.begin(), NewArgs.end(),
9784 Caller->getName(), Caller);
9785 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9786 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
9788 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9789 Caller->getName(), Caller);
9790 if (cast<CallInst>(Caller)->isTailCall())
9791 cast<CallInst>(NewCaller)->setTailCall();
9792 cast<CallInst>(NewCaller)->
9793 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9794 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
9796 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9797 Caller->replaceAllUsesWith(NewCaller);
9798 Caller->eraseFromParent();
9799 RemoveFromWorkList(Caller);
9804 // Replace the trampoline call with a direct call. Since there is no 'nest'
9805 // parameter, there is no need to adjust the argument list. Let the generic
9806 // code sort out any function type mismatches.
9807 Constant *NewCallee =
9808 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9809 CS.setCalledFunction(NewCallee);
9810 return CS.getInstruction();
9813 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9814 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9815 /// and a single binop.
9816 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9817 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9818 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9819 isa<CmpInst>(FirstInst));
9820 unsigned Opc = FirstInst->getOpcode();
9821 Value *LHSVal = FirstInst->getOperand(0);
9822 Value *RHSVal = FirstInst->getOperand(1);
9824 const Type *LHSType = LHSVal->getType();
9825 const Type *RHSType = RHSVal->getType();
9827 // Scan to see if all operands are the same opcode, all have one use, and all
9828 // kill their operands (i.e. the operands have one use).
9829 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9830 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9831 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9832 // Verify type of the LHS matches so we don't fold cmp's of different
9833 // types or GEP's with different index types.
9834 I->getOperand(0)->getType() != LHSType ||
9835 I->getOperand(1)->getType() != RHSType)
9838 // If they are CmpInst instructions, check their predicates
9839 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9840 if (cast<CmpInst>(I)->getPredicate() !=
9841 cast<CmpInst>(FirstInst)->getPredicate())
9844 // Keep track of which operand needs a phi node.
9845 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9846 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9849 // Otherwise, this is safe to transform, determine if it is profitable.
9851 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9852 // Indexes are often folded into load/store instructions, so we don't want to
9853 // hide them behind a phi.
9854 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9857 Value *InLHS = FirstInst->getOperand(0);
9858 Value *InRHS = FirstInst->getOperand(1);
9859 PHINode *NewLHS = 0, *NewRHS = 0;
9861 NewLHS = PHINode::Create(LHSType,
9862 FirstInst->getOperand(0)->getName() + ".pn");
9863 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9864 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9865 InsertNewInstBefore(NewLHS, PN);
9870 NewRHS = PHINode::Create(RHSType,
9871 FirstInst->getOperand(1)->getName() + ".pn");
9872 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9873 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9874 InsertNewInstBefore(NewRHS, PN);
9878 // Add all operands to the new PHIs.
9879 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9881 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9882 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9885 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9886 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9890 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9891 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9892 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9893 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9896 assert(isa<GetElementPtrInst>(FirstInst));
9897 return GetElementPtrInst::Create(LHSVal, RHSVal);
9901 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9902 /// of the block that defines it. This means that it must be obvious the value
9903 /// of the load is not changed from the point of the load to the end of the
9906 /// Finally, it is safe, but not profitable, to sink a load targetting a
9907 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9909 static bool isSafeToSinkLoad(LoadInst *L) {
9910 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9912 for (++BBI; BBI != E; ++BBI)
9913 if (BBI->mayWriteToMemory())
9916 // Check for non-address taken alloca. If not address-taken already, it isn't
9917 // profitable to do this xform.
9918 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9919 bool isAddressTaken = false;
9920 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9922 if (isa<LoadInst>(UI)) continue;
9923 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9924 // If storing TO the alloca, then the address isn't taken.
9925 if (SI->getOperand(1) == AI) continue;
9927 isAddressTaken = true;
9931 if (!isAddressTaken)
9939 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9940 // operator and they all are only used by the PHI, PHI together their
9941 // inputs, and do the operation once, to the result of the PHI.
9942 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9943 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9945 // Scan the instruction, looking for input operations that can be folded away.
9946 // If all input operands to the phi are the same instruction (e.g. a cast from
9947 // the same type or "+42") we can pull the operation through the PHI, reducing
9948 // code size and simplifying code.
9949 Constant *ConstantOp = 0;
9950 const Type *CastSrcTy = 0;
9951 bool isVolatile = false;
9952 if (isa<CastInst>(FirstInst)) {
9953 CastSrcTy = FirstInst->getOperand(0)->getType();
9954 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9955 // Can fold binop, compare or shift here if the RHS is a constant,
9956 // otherwise call FoldPHIArgBinOpIntoPHI.
9957 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9958 if (ConstantOp == 0)
9959 return FoldPHIArgBinOpIntoPHI(PN);
9960 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9961 isVolatile = LI->isVolatile();
9962 // We can't sink the load if the loaded value could be modified between the
9963 // load and the PHI.
9964 if (LI->getParent() != PN.getIncomingBlock(0) ||
9965 !isSafeToSinkLoad(LI))
9968 // If the PHI is of volatile loads and the load block has multiple
9969 // successors, sinking it would remove a load of the volatile value from
9970 // the path through the other successor.
9972 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9975 } else if (isa<GetElementPtrInst>(FirstInst)) {
9976 if (FirstInst->getNumOperands() == 2)
9977 return FoldPHIArgBinOpIntoPHI(PN);
9978 // Can't handle general GEPs yet.
9981 return 0; // Cannot fold this operation.
9984 // Check to see if all arguments are the same operation.
9985 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9986 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9987 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9988 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9991 if (I->getOperand(0)->getType() != CastSrcTy)
9992 return 0; // Cast operation must match.
9993 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9994 // We can't sink the load if the loaded value could be modified between
9995 // the load and the PHI.
9996 if (LI->isVolatile() != isVolatile ||
9997 LI->getParent() != PN.getIncomingBlock(i) ||
9998 !isSafeToSinkLoad(LI))
10001 // If the PHI is of volatile loads and the load block has multiple
10002 // successors, sinking it would remove a load of the volatile value from
10003 // the path through the other successor.
10005 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10009 } else if (I->getOperand(1) != ConstantOp) {
10014 // Okay, they are all the same operation. Create a new PHI node of the
10015 // correct type, and PHI together all of the LHS's of the instructions.
10016 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10017 PN.getName()+".in");
10018 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10020 Value *InVal = FirstInst->getOperand(0);
10021 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10023 // Add all operands to the new PHI.
10024 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10025 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10026 if (NewInVal != InVal)
10028 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10033 // The new PHI unions all of the same values together. This is really
10034 // common, so we handle it intelligently here for compile-time speed.
10038 InsertNewInstBefore(NewPN, PN);
10042 // Insert and return the new operation.
10043 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10044 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10045 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10046 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10047 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10048 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10049 PhiVal, ConstantOp);
10050 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10052 // If this was a volatile load that we are merging, make sure to loop through
10053 // and mark all the input loads as non-volatile. If we don't do this, we will
10054 // insert a new volatile load and the old ones will not be deletable.
10056 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10057 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10059 return new LoadInst(PhiVal, "", isVolatile);
10062 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10064 static bool DeadPHICycle(PHINode *PN,
10065 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10066 if (PN->use_empty()) return true;
10067 if (!PN->hasOneUse()) return false;
10069 // Remember this node, and if we find the cycle, return.
10070 if (!PotentiallyDeadPHIs.insert(PN))
10073 // Don't scan crazily complex things.
10074 if (PotentiallyDeadPHIs.size() == 16)
10077 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10078 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10083 /// PHIsEqualValue - Return true if this phi node is always equal to
10084 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10085 /// z = some value; x = phi (y, z); y = phi (x, z)
10086 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10087 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10088 // See if we already saw this PHI node.
10089 if (!ValueEqualPHIs.insert(PN))
10092 // Don't scan crazily complex things.
10093 if (ValueEqualPHIs.size() == 16)
10096 // Scan the operands to see if they are either phi nodes or are equal to
10098 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10099 Value *Op = PN->getIncomingValue(i);
10100 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10101 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10103 } else if (Op != NonPhiInVal)
10111 // PHINode simplification
10113 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10114 // If LCSSA is around, don't mess with Phi nodes
10115 if (MustPreserveLCSSA) return 0;
10117 if (Value *V = PN.hasConstantValue())
10118 return ReplaceInstUsesWith(PN, V);
10120 // If all PHI operands are the same operation, pull them through the PHI,
10121 // reducing code size.
10122 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10123 PN.getIncomingValue(0)->hasOneUse())
10124 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10127 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10128 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10129 // PHI)... break the cycle.
10130 if (PN.hasOneUse()) {
10131 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10132 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10133 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10134 PotentiallyDeadPHIs.insert(&PN);
10135 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10136 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10139 // If this phi has a single use, and if that use just computes a value for
10140 // the next iteration of a loop, delete the phi. This occurs with unused
10141 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10142 // common case here is good because the only other things that catch this
10143 // are induction variable analysis (sometimes) and ADCE, which is only run
10145 if (PHIUser->hasOneUse() &&
10146 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10147 PHIUser->use_back() == &PN) {
10148 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10152 // We sometimes end up with phi cycles that non-obviously end up being the
10153 // same value, for example:
10154 // z = some value; x = phi (y, z); y = phi (x, z)
10155 // where the phi nodes don't necessarily need to be in the same block. Do a
10156 // quick check to see if the PHI node only contains a single non-phi value, if
10157 // so, scan to see if the phi cycle is actually equal to that value.
10159 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10160 // Scan for the first non-phi operand.
10161 while (InValNo != NumOperandVals &&
10162 isa<PHINode>(PN.getIncomingValue(InValNo)))
10165 if (InValNo != NumOperandVals) {
10166 Value *NonPhiInVal = PN.getOperand(InValNo);
10168 // Scan the rest of the operands to see if there are any conflicts, if so
10169 // there is no need to recursively scan other phis.
10170 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10171 Value *OpVal = PN.getIncomingValue(InValNo);
10172 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10176 // If we scanned over all operands, then we have one unique value plus
10177 // phi values. Scan PHI nodes to see if they all merge in each other or
10179 if (InValNo == NumOperandVals) {
10180 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10181 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10182 return ReplaceInstUsesWith(PN, NonPhiInVal);
10189 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10190 Instruction *InsertPoint,
10191 InstCombiner *IC) {
10192 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10193 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10194 // We must cast correctly to the pointer type. Ensure that we
10195 // sign extend the integer value if it is smaller as this is
10196 // used for address computation.
10197 Instruction::CastOps opcode =
10198 (VTySize < PtrSize ? Instruction::SExt :
10199 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10200 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10204 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10205 Value *PtrOp = GEP.getOperand(0);
10206 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10207 // If so, eliminate the noop.
10208 if (GEP.getNumOperands() == 1)
10209 return ReplaceInstUsesWith(GEP, PtrOp);
10211 if (isa<UndefValue>(GEP.getOperand(0)))
10212 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10214 bool HasZeroPointerIndex = false;
10215 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10216 HasZeroPointerIndex = C->isNullValue();
10218 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10219 return ReplaceInstUsesWith(GEP, PtrOp);
10221 // Eliminate unneeded casts for indices.
10222 bool MadeChange = false;
10224 gep_type_iterator GTI = gep_type_begin(GEP);
10225 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10226 i != e; ++i, ++GTI) {
10227 if (isa<SequentialType>(*GTI)) {
10228 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10229 if (CI->getOpcode() == Instruction::ZExt ||
10230 CI->getOpcode() == Instruction::SExt) {
10231 const Type *SrcTy = CI->getOperand(0)->getType();
10232 // We can eliminate a cast from i32 to i64 iff the target
10233 // is a 32-bit pointer target.
10234 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10236 *i = CI->getOperand(0);
10240 // If we are using a wider index than needed for this platform, shrink it
10241 // to what we need. If narrower, sign-extend it to what we need.
10242 // If the incoming value needs a cast instruction,
10243 // insert it. This explicit cast can make subsequent optimizations more
10246 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10247 if (Constant *C = dyn_cast<Constant>(Op)) {
10248 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10251 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10256 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10257 if (Constant *C = dyn_cast<Constant>(Op)) {
10258 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10261 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10269 if (MadeChange) return &GEP;
10271 // If this GEP instruction doesn't move the pointer, and if the input operand
10272 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10273 // real input to the dest type.
10274 if (GEP.hasAllZeroIndices()) {
10275 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10276 // If the bitcast is of an allocation, and the allocation will be
10277 // converted to match the type of the cast, don't touch this.
10278 if (isa<AllocationInst>(BCI->getOperand(0))) {
10279 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10280 if (Instruction *I = visitBitCast(*BCI)) {
10283 BCI->getParent()->getInstList().insert(BCI, I);
10284 ReplaceInstUsesWith(*BCI, I);
10289 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10293 // Combine Indices - If the source pointer to this getelementptr instruction
10294 // is a getelementptr instruction, combine the indices of the two
10295 // getelementptr instructions into a single instruction.
10297 SmallVector<Value*, 8> SrcGEPOperands;
10298 if (User *Src = dyn_castGetElementPtr(PtrOp))
10299 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10301 if (!SrcGEPOperands.empty()) {
10302 // Note that if our source is a gep chain itself that we wait for that
10303 // chain to be resolved before we perform this transformation. This
10304 // avoids us creating a TON of code in some cases.
10306 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10307 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10308 return 0; // Wait until our source is folded to completion.
10310 SmallVector<Value*, 8> Indices;
10312 // Find out whether the last index in the source GEP is a sequential idx.
10313 bool EndsWithSequential = false;
10314 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10315 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10316 EndsWithSequential = !isa<StructType>(*I);
10318 // Can we combine the two pointer arithmetics offsets?
10319 if (EndsWithSequential) {
10320 // Replace: gep (gep %P, long B), long A, ...
10321 // With: T = long A+B; gep %P, T, ...
10323 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10324 if (SO1 == Constant::getNullValue(SO1->getType())) {
10326 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10329 // If they aren't the same type, convert both to an integer of the
10330 // target's pointer size.
10331 if (SO1->getType() != GO1->getType()) {
10332 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10333 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10334 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10335 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10337 unsigned PS = TD->getPointerSizeInBits();
10338 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10339 // Convert GO1 to SO1's type.
10340 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10342 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10343 // Convert SO1 to GO1's type.
10344 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10346 const Type *PT = TD->getIntPtrType();
10347 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10348 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10352 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10353 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10355 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10356 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10360 // Recycle the GEP we already have if possible.
10361 if (SrcGEPOperands.size() == 2) {
10362 GEP.setOperand(0, SrcGEPOperands[0]);
10363 GEP.setOperand(1, Sum);
10366 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10367 SrcGEPOperands.end()-1);
10368 Indices.push_back(Sum);
10369 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10371 } else if (isa<Constant>(*GEP.idx_begin()) &&
10372 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10373 SrcGEPOperands.size() != 1) {
10374 // Otherwise we can do the fold if the first index of the GEP is a zero
10375 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10376 SrcGEPOperands.end());
10377 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10380 if (!Indices.empty())
10381 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10382 Indices.end(), GEP.getName());
10384 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10385 // GEP of global variable. If all of the indices for this GEP are
10386 // constants, we can promote this to a constexpr instead of an instruction.
10388 // Scan for nonconstants...
10389 SmallVector<Constant*, 8> Indices;
10390 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10391 for (; I != E && isa<Constant>(*I); ++I)
10392 Indices.push_back(cast<Constant>(*I));
10394 if (I == E) { // If they are all constants...
10395 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10396 &Indices[0],Indices.size());
10398 // Replace all uses of the GEP with the new constexpr...
10399 return ReplaceInstUsesWith(GEP, CE);
10401 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10402 if (!isa<PointerType>(X->getType())) {
10403 // Not interesting. Source pointer must be a cast from pointer.
10404 } else if (HasZeroPointerIndex) {
10405 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10406 // into : GEP [10 x i8]* X, i32 0, ...
10408 // This occurs when the program declares an array extern like "int X[];"
10410 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10411 const PointerType *XTy = cast<PointerType>(X->getType());
10412 if (const ArrayType *XATy =
10413 dyn_cast<ArrayType>(XTy->getElementType()))
10414 if (const ArrayType *CATy =
10415 dyn_cast<ArrayType>(CPTy->getElementType()))
10416 if (CATy->getElementType() == XATy->getElementType()) {
10417 // At this point, we know that the cast source type is a pointer
10418 // to an array of the same type as the destination pointer
10419 // array. Because the array type is never stepped over (there
10420 // is a leading zero) we can fold the cast into this GEP.
10421 GEP.setOperand(0, X);
10424 } else if (GEP.getNumOperands() == 2) {
10425 // Transform things like:
10426 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10427 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10428 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10429 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10430 if (isa<ArrayType>(SrcElTy) &&
10431 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10432 TD->getABITypeSize(ResElTy)) {
10434 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10435 Idx[1] = GEP.getOperand(1);
10436 Value *V = InsertNewInstBefore(
10437 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10438 // V and GEP are both pointer types --> BitCast
10439 return new BitCastInst(V, GEP.getType());
10442 // Transform things like:
10443 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10444 // (where tmp = 8*tmp2) into:
10445 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10447 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10448 uint64_t ArrayEltSize =
10449 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10451 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10452 // allow either a mul, shift, or constant here.
10454 ConstantInt *Scale = 0;
10455 if (ArrayEltSize == 1) {
10456 NewIdx = GEP.getOperand(1);
10457 Scale = ConstantInt::get(NewIdx->getType(), 1);
10458 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10459 NewIdx = ConstantInt::get(CI->getType(), 1);
10461 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10462 if (Inst->getOpcode() == Instruction::Shl &&
10463 isa<ConstantInt>(Inst->getOperand(1))) {
10464 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10465 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10466 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10467 NewIdx = Inst->getOperand(0);
10468 } else if (Inst->getOpcode() == Instruction::Mul &&
10469 isa<ConstantInt>(Inst->getOperand(1))) {
10470 Scale = cast<ConstantInt>(Inst->getOperand(1));
10471 NewIdx = Inst->getOperand(0);
10475 // If the index will be to exactly the right offset with the scale taken
10476 // out, perform the transformation. Note, we don't know whether Scale is
10477 // signed or not. We'll use unsigned version of division/modulo
10478 // operation after making sure Scale doesn't have the sign bit set.
10479 if (Scale && Scale->getSExtValue() >= 0LL &&
10480 Scale->getZExtValue() % ArrayEltSize == 0) {
10481 Scale = ConstantInt::get(Scale->getType(),
10482 Scale->getZExtValue() / ArrayEltSize);
10483 if (Scale->getZExtValue() != 1) {
10484 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10486 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10487 NewIdx = InsertNewInstBefore(Sc, GEP);
10490 // Insert the new GEP instruction.
10492 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10494 Instruction *NewGEP =
10495 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10496 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10497 // The NewGEP must be pointer typed, so must the old one -> BitCast
10498 return new BitCastInst(NewGEP, GEP.getType());
10507 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10508 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10509 if (AI.isArrayAllocation()) { // Check C != 1
10510 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10511 const Type *NewTy =
10512 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10513 AllocationInst *New = 0;
10515 // Create and insert the replacement instruction...
10516 if (isa<MallocInst>(AI))
10517 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10519 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10520 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10523 InsertNewInstBefore(New, AI);
10525 // Scan to the end of the allocation instructions, to skip over a block of
10526 // allocas if possible...
10528 BasicBlock::iterator It = New;
10529 while (isa<AllocationInst>(*It)) ++It;
10531 // Now that I is pointing to the first non-allocation-inst in the block,
10532 // insert our getelementptr instruction...
10534 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10538 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10539 New->getName()+".sub", It);
10541 // Now make everything use the getelementptr instead of the original
10543 return ReplaceInstUsesWith(AI, V);
10544 } else if (isa<UndefValue>(AI.getArraySize())) {
10545 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10549 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10550 // Note that we only do this for alloca's, because malloc should allocate and
10551 // return a unique pointer, even for a zero byte allocation.
10552 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10553 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10554 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10559 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10560 Value *Op = FI.getOperand(0);
10562 // free undef -> unreachable.
10563 if (isa<UndefValue>(Op)) {
10564 // Insert a new store to null because we cannot modify the CFG here.
10565 new StoreInst(ConstantInt::getTrue(),
10566 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10567 return EraseInstFromFunction(FI);
10570 // If we have 'free null' delete the instruction. This can happen in stl code
10571 // when lots of inlining happens.
10572 if (isa<ConstantPointerNull>(Op))
10573 return EraseInstFromFunction(FI);
10575 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10576 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10577 FI.setOperand(0, CI->getOperand(0));
10581 // Change free (gep X, 0,0,0,0) into free(X)
10582 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10583 if (GEPI->hasAllZeroIndices()) {
10584 AddToWorkList(GEPI);
10585 FI.setOperand(0, GEPI->getOperand(0));
10590 // Change free(malloc) into nothing, if the malloc has a single use.
10591 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10592 if (MI->hasOneUse()) {
10593 EraseInstFromFunction(FI);
10594 return EraseInstFromFunction(*MI);
10601 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10602 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10603 const TargetData *TD) {
10604 User *CI = cast<User>(LI.getOperand(0));
10605 Value *CastOp = CI->getOperand(0);
10607 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10608 // Instead of loading constant c string, use corresponding integer value
10609 // directly if string length is small enough.
10611 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10612 unsigned len = Str.length();
10613 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10614 unsigned numBits = Ty->getPrimitiveSizeInBits();
10615 // Replace LI with immediate integer store.
10616 if ((numBits >> 3) == len + 1) {
10617 APInt StrVal(numBits, 0);
10618 APInt SingleChar(numBits, 0);
10619 if (TD->isLittleEndian()) {
10620 for (signed i = len-1; i >= 0; i--) {
10621 SingleChar = (uint64_t) Str[i];
10622 StrVal = (StrVal << 8) | SingleChar;
10625 for (unsigned i = 0; i < len; i++) {
10626 SingleChar = (uint64_t) Str[i];
10627 StrVal = (StrVal << 8) | SingleChar;
10629 // Append NULL at the end.
10631 StrVal = (StrVal << 8) | SingleChar;
10633 Value *NL = ConstantInt::get(StrVal);
10634 return IC.ReplaceInstUsesWith(LI, NL);
10639 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10640 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10641 const Type *SrcPTy = SrcTy->getElementType();
10643 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10644 isa<VectorType>(DestPTy)) {
10645 // If the source is an array, the code below will not succeed. Check to
10646 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10648 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10649 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10650 if (ASrcTy->getNumElements() != 0) {
10652 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10653 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10654 SrcTy = cast<PointerType>(CastOp->getType());
10655 SrcPTy = SrcTy->getElementType();
10658 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10659 isa<VectorType>(SrcPTy)) &&
10660 // Do not allow turning this into a load of an integer, which is then
10661 // casted to a pointer, this pessimizes pointer analysis a lot.
10662 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10663 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10664 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10666 // Okay, we are casting from one integer or pointer type to another of
10667 // the same size. Instead of casting the pointer before the load, cast
10668 // the result of the loaded value.
10669 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10671 LI.isVolatile()),LI);
10672 // Now cast the result of the load.
10673 return new BitCastInst(NewLoad, LI.getType());
10680 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10681 /// from this value cannot trap. If it is not obviously safe to load from the
10682 /// specified pointer, we do a quick local scan of the basic block containing
10683 /// ScanFrom, to determine if the address is already accessed.
10684 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10685 // If it is an alloca it is always safe to load from.
10686 if (isa<AllocaInst>(V)) return true;
10688 // If it is a global variable it is mostly safe to load from.
10689 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10690 // Don't try to evaluate aliases. External weak GV can be null.
10691 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10693 // Otherwise, be a little bit agressive by scanning the local block where we
10694 // want to check to see if the pointer is already being loaded or stored
10695 // from/to. If so, the previous load or store would have already trapped,
10696 // so there is no harm doing an extra load (also, CSE will later eliminate
10697 // the load entirely).
10698 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10703 // If we see a free or a call (which might do a free) the pointer could be
10705 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10708 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10709 if (LI->getOperand(0) == V) return true;
10710 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10711 if (SI->getOperand(1) == V) return true;
10718 /// equivalentAddressValues - Test if A and B will obviously have the same
10719 /// value. This includes recognizing that %t0 and %t1 will have the same
10720 /// value in code like this:
10721 /// %t0 = getelementptr @a, 0, 3
10722 /// store i32 0, i32* %t0
10723 /// %t1 = getelementptr @a, 0, 3
10724 /// %t2 = load i32* %t1
10726 static bool equivalentAddressValues(Value *A, Value *B) {
10727 // Test if the values are trivially equivalent.
10728 if (A == B) return true;
10730 // Test if the values come form identical arithmetic instructions.
10731 if (isa<BinaryOperator>(A) ||
10732 isa<CastInst>(A) ||
10734 isa<GetElementPtrInst>(A))
10735 if (Instruction *BI = dyn_cast<Instruction>(B))
10736 if (cast<Instruction>(A)->isIdenticalTo(BI))
10739 // Otherwise they may not be equivalent.
10743 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10744 Value *Op = LI.getOperand(0);
10746 // Attempt to improve the alignment.
10747 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10749 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10750 LI.getAlignment()))
10751 LI.setAlignment(KnownAlign);
10753 // load (cast X) --> cast (load X) iff safe
10754 if (isa<CastInst>(Op))
10755 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10758 // None of the following transforms are legal for volatile loads.
10759 if (LI.isVolatile()) return 0;
10761 // Do really simple store-to-load forwarding and load CSE, to catch cases
10762 // where there are several consequtive memory accesses to the same location,
10763 // separated by a few arithmetic operations.
10764 BasicBlock::iterator BBI = &LI;
10765 for (unsigned ScanInsts = 6; BBI != LI.getParent()->begin() && ScanInsts;
10769 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10770 if (equivalentAddressValues(SI->getOperand(1), LI.getOperand(0)))
10771 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10772 } else if (LoadInst *LIB = dyn_cast<LoadInst>(BBI)) {
10773 if (equivalentAddressValues(LIB->getOperand(0), LI.getOperand(0)))
10774 return ReplaceInstUsesWith(LI, LIB);
10777 // Don't skip over things that can modify memory.
10778 if (BBI->mayWriteToMemory())
10782 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10783 const Value *GEPI0 = GEPI->getOperand(0);
10784 // TODO: Consider a target hook for valid address spaces for this xform.
10785 if (isa<ConstantPointerNull>(GEPI0) &&
10786 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10787 // Insert a new store to null instruction before the load to indicate
10788 // that this code is not reachable. We do this instead of inserting
10789 // an unreachable instruction directly because we cannot modify the
10791 new StoreInst(UndefValue::get(LI.getType()),
10792 Constant::getNullValue(Op->getType()), &LI);
10793 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10797 if (Constant *C = dyn_cast<Constant>(Op)) {
10798 // load null/undef -> undef
10799 // TODO: Consider a target hook for valid address spaces for this xform.
10800 if (isa<UndefValue>(C) || (C->isNullValue() &&
10801 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10802 // Insert a new store to null instruction before the load to indicate that
10803 // this code is not reachable. We do this instead of inserting an
10804 // unreachable instruction directly because we cannot modify the CFG.
10805 new StoreInst(UndefValue::get(LI.getType()),
10806 Constant::getNullValue(Op->getType()), &LI);
10807 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10810 // Instcombine load (constant global) into the value loaded.
10811 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10812 if (GV->isConstant() && !GV->isDeclaration())
10813 return ReplaceInstUsesWith(LI, GV->getInitializer());
10815 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10816 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10817 if (CE->getOpcode() == Instruction::GetElementPtr) {
10818 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10819 if (GV->isConstant() && !GV->isDeclaration())
10821 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10822 return ReplaceInstUsesWith(LI, V);
10823 if (CE->getOperand(0)->isNullValue()) {
10824 // Insert a new store to null instruction before the load to indicate
10825 // that this code is not reachable. We do this instead of inserting
10826 // an unreachable instruction directly because we cannot modify the
10828 new StoreInst(UndefValue::get(LI.getType()),
10829 Constant::getNullValue(Op->getType()), &LI);
10830 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10833 } else if (CE->isCast()) {
10834 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10840 // If this load comes from anywhere in a constant global, and if the global
10841 // is all undef or zero, we know what it loads.
10842 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
10843 if (GV->isConstant() && GV->hasInitializer()) {
10844 if (GV->getInitializer()->isNullValue())
10845 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10846 else if (isa<UndefValue>(GV->getInitializer()))
10847 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10851 if (Op->hasOneUse()) {
10852 // Change select and PHI nodes to select values instead of addresses: this
10853 // helps alias analysis out a lot, allows many others simplifications, and
10854 // exposes redundancy in the code.
10856 // Note that we cannot do the transformation unless we know that the
10857 // introduced loads cannot trap! Something like this is valid as long as
10858 // the condition is always false: load (select bool %C, int* null, int* %G),
10859 // but it would not be valid if we transformed it to load from null
10860 // unconditionally.
10862 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10863 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10864 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10865 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10866 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10867 SI->getOperand(1)->getName()+".val"), LI);
10868 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10869 SI->getOperand(2)->getName()+".val"), LI);
10870 return SelectInst::Create(SI->getCondition(), V1, V2);
10873 // load (select (cond, null, P)) -> load P
10874 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10875 if (C->isNullValue()) {
10876 LI.setOperand(0, SI->getOperand(2));
10880 // load (select (cond, P, null)) -> load P
10881 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10882 if (C->isNullValue()) {
10883 LI.setOperand(0, SI->getOperand(1));
10891 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10893 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10894 User *CI = cast<User>(SI.getOperand(1));
10895 Value *CastOp = CI->getOperand(0);
10897 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10898 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10899 const Type *SrcPTy = SrcTy->getElementType();
10901 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10902 // If the source is an array, the code below will not succeed. Check to
10903 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10905 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10906 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10907 if (ASrcTy->getNumElements() != 0) {
10909 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10910 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10911 SrcTy = cast<PointerType>(CastOp->getType());
10912 SrcPTy = SrcTy->getElementType();
10915 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10916 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10917 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10919 // Okay, we are casting from one integer or pointer type to another of
10920 // the same size. Instead of casting the pointer before
10921 // the store, cast the value to be stored.
10923 Value *SIOp0 = SI.getOperand(0);
10924 Instruction::CastOps opcode = Instruction::BitCast;
10925 const Type* CastSrcTy = SIOp0->getType();
10926 const Type* CastDstTy = SrcPTy;
10927 if (isa<PointerType>(CastDstTy)) {
10928 if (CastSrcTy->isInteger())
10929 opcode = Instruction::IntToPtr;
10930 } else if (isa<IntegerType>(CastDstTy)) {
10931 if (isa<PointerType>(SIOp0->getType()))
10932 opcode = Instruction::PtrToInt;
10934 if (Constant *C = dyn_cast<Constant>(SIOp0))
10935 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10937 NewCast = IC.InsertNewInstBefore(
10938 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10940 return new StoreInst(NewCast, CastOp);
10947 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10948 Value *Val = SI.getOperand(0);
10949 Value *Ptr = SI.getOperand(1);
10951 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10952 EraseInstFromFunction(SI);
10957 // If the RHS is an alloca with a single use, zapify the store, making the
10959 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10960 if (isa<AllocaInst>(Ptr)) {
10961 EraseInstFromFunction(SI);
10966 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10967 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10968 GEP->getOperand(0)->hasOneUse()) {
10969 EraseInstFromFunction(SI);
10975 // Attempt to improve the alignment.
10976 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10978 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10979 SI.getAlignment()))
10980 SI.setAlignment(KnownAlign);
10982 // Do really simple DSE, to catch cases where there are several consequtive
10983 // stores to the same location, separated by a few arithmetic operations. This
10984 // situation often occurs with bitfield accesses.
10985 BasicBlock::iterator BBI = &SI;
10986 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10990 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10991 // Prev store isn't volatile, and stores to the same location?
10992 if (!PrevSI->isVolatile() && equivalentAddressValues(PrevSI->getOperand(1),
10993 SI.getOperand(1))) {
10996 EraseInstFromFunction(*PrevSI);
11002 // If this is a load, we have to stop. However, if the loaded value is from
11003 // the pointer we're loading and is producing the pointer we're storing,
11004 // then *this* store is dead (X = load P; store X -> P).
11005 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11006 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11007 !SI.isVolatile()) {
11008 EraseInstFromFunction(SI);
11012 // Otherwise, this is a load from some other location. Stores before it
11013 // may not be dead.
11017 // Don't skip over loads or things that can modify memory.
11018 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11023 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11025 // store X, null -> turns into 'unreachable' in SimplifyCFG
11026 if (isa<ConstantPointerNull>(Ptr)) {
11027 if (!isa<UndefValue>(Val)) {
11028 SI.setOperand(0, UndefValue::get(Val->getType()));
11029 if (Instruction *U = dyn_cast<Instruction>(Val))
11030 AddToWorkList(U); // Dropped a use.
11033 return 0; // Do not modify these!
11036 // store undef, Ptr -> noop
11037 if (isa<UndefValue>(Val)) {
11038 EraseInstFromFunction(SI);
11043 // If the pointer destination is a cast, see if we can fold the cast into the
11045 if (isa<CastInst>(Ptr))
11046 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11048 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11050 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11054 // If this store is the last instruction in the basic block, and if the block
11055 // ends with an unconditional branch, try to move it to the successor block.
11057 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11058 if (BI->isUnconditional())
11059 if (SimplifyStoreAtEndOfBlock(SI))
11060 return 0; // xform done!
11065 /// SimplifyStoreAtEndOfBlock - Turn things like:
11066 /// if () { *P = v1; } else { *P = v2 }
11067 /// into a phi node with a store in the successor.
11069 /// Simplify things like:
11070 /// *P = v1; if () { *P = v2; }
11071 /// into a phi node with a store in the successor.
11073 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11074 BasicBlock *StoreBB = SI.getParent();
11076 // Check to see if the successor block has exactly two incoming edges. If
11077 // so, see if the other predecessor contains a store to the same location.
11078 // if so, insert a PHI node (if needed) and move the stores down.
11079 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11081 // Determine whether Dest has exactly two predecessors and, if so, compute
11082 // the other predecessor.
11083 pred_iterator PI = pred_begin(DestBB);
11084 BasicBlock *OtherBB = 0;
11085 if (*PI != StoreBB)
11088 if (PI == pred_end(DestBB))
11091 if (*PI != StoreBB) {
11096 if (++PI != pred_end(DestBB))
11099 // Bail out if all the relevant blocks aren't distinct (this can happen,
11100 // for example, if SI is in an infinite loop)
11101 if (StoreBB == DestBB || OtherBB == DestBB)
11104 // Verify that the other block ends in a branch and is not otherwise empty.
11105 BasicBlock::iterator BBI = OtherBB->getTerminator();
11106 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11107 if (!OtherBr || BBI == OtherBB->begin())
11110 // If the other block ends in an unconditional branch, check for the 'if then
11111 // else' case. there is an instruction before the branch.
11112 StoreInst *OtherStore = 0;
11113 if (OtherBr->isUnconditional()) {
11114 // If this isn't a store, or isn't a store to the same location, bail out.
11116 OtherStore = dyn_cast<StoreInst>(BBI);
11117 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11120 // Otherwise, the other block ended with a conditional branch. If one of the
11121 // destinations is StoreBB, then we have the if/then case.
11122 if (OtherBr->getSuccessor(0) != StoreBB &&
11123 OtherBr->getSuccessor(1) != StoreBB)
11126 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11127 // if/then triangle. See if there is a store to the same ptr as SI that
11128 // lives in OtherBB.
11130 // Check to see if we find the matching store.
11131 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11132 if (OtherStore->getOperand(1) != SI.getOperand(1))
11136 // If we find something that may be using or overwriting the stored
11137 // value, or if we run out of instructions, we can't do the xform.
11138 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11139 BBI == OtherBB->begin())
11143 // In order to eliminate the store in OtherBr, we have to
11144 // make sure nothing reads or overwrites the stored value in
11146 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11147 // FIXME: This should really be AA driven.
11148 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11153 // Insert a PHI node now if we need it.
11154 Value *MergedVal = OtherStore->getOperand(0);
11155 if (MergedVal != SI.getOperand(0)) {
11156 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11157 PN->reserveOperandSpace(2);
11158 PN->addIncoming(SI.getOperand(0), SI.getParent());
11159 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11160 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11163 // Advance to a place where it is safe to insert the new store and
11165 BBI = DestBB->getFirstNonPHI();
11166 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11167 OtherStore->isVolatile()), *BBI);
11169 // Nuke the old stores.
11170 EraseInstFromFunction(SI);
11171 EraseInstFromFunction(*OtherStore);
11177 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11178 // Change br (not X), label True, label False to: br X, label False, True
11180 BasicBlock *TrueDest;
11181 BasicBlock *FalseDest;
11182 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11183 !isa<Constant>(X)) {
11184 // Swap Destinations and condition...
11185 BI.setCondition(X);
11186 BI.setSuccessor(0, FalseDest);
11187 BI.setSuccessor(1, TrueDest);
11191 // Cannonicalize fcmp_one -> fcmp_oeq
11192 FCmpInst::Predicate FPred; Value *Y;
11193 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11194 TrueDest, FalseDest)))
11195 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11196 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11197 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11198 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11199 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11200 NewSCC->takeName(I);
11201 // Swap Destinations and condition...
11202 BI.setCondition(NewSCC);
11203 BI.setSuccessor(0, FalseDest);
11204 BI.setSuccessor(1, TrueDest);
11205 RemoveFromWorkList(I);
11206 I->eraseFromParent();
11207 AddToWorkList(NewSCC);
11211 // Cannonicalize icmp_ne -> icmp_eq
11212 ICmpInst::Predicate IPred;
11213 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11214 TrueDest, FalseDest)))
11215 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11216 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11217 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11218 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11219 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11220 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11221 NewSCC->takeName(I);
11222 // Swap Destinations and condition...
11223 BI.setCondition(NewSCC);
11224 BI.setSuccessor(0, FalseDest);
11225 BI.setSuccessor(1, TrueDest);
11226 RemoveFromWorkList(I);
11227 I->eraseFromParent();;
11228 AddToWorkList(NewSCC);
11235 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11236 Value *Cond = SI.getCondition();
11237 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11238 if (I->getOpcode() == Instruction::Add)
11239 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11240 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11241 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11242 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11244 SI.setOperand(0, I->getOperand(0));
11252 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11253 Value *Agg = EV.getAggregateOperand();
11255 if (!EV.hasIndices())
11256 return ReplaceInstUsesWith(EV, Agg);
11258 if (Constant *C = dyn_cast<Constant>(Agg)) {
11259 if (isa<UndefValue>(C))
11260 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11262 if (isa<ConstantAggregateZero>(C))
11263 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11265 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11266 // Extract the element indexed by the first index out of the constant
11267 Value *V = C->getOperand(*EV.idx_begin());
11268 if (EV.getNumIndices() > 1)
11269 // Extract the remaining indices out of the constant indexed by the
11271 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11273 return ReplaceInstUsesWith(EV, V);
11275 return 0; // Can't handle other constants
11277 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11278 // We're extracting from an insertvalue instruction, compare the indices
11279 const unsigned *exti, *exte, *insi, *inse;
11280 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11281 exte = EV.idx_end(), inse = IV->idx_end();
11282 exti != exte && insi != inse;
11284 if (*insi != *exti)
11285 // The insert and extract both reference distinctly different elements.
11286 // This means the extract is not influenced by the insert, and we can
11287 // replace the aggregate operand of the extract with the aggregate
11288 // operand of the insert. i.e., replace
11289 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11290 // %E = extractvalue { i32, { i32 } } %I, 0
11292 // %E = extractvalue { i32, { i32 } } %A, 0
11293 return ExtractValueInst::Create(IV->getAggregateOperand(),
11294 EV.idx_begin(), EV.idx_end());
11296 if (exti == exte && insi == inse)
11297 // Both iterators are at the end: Index lists are identical. Replace
11298 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11299 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11301 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11302 if (exti == exte) {
11303 // The extract list is a prefix of the insert list. i.e. replace
11304 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11305 // %E = extractvalue { i32, { i32 } } %I, 1
11307 // %X = extractvalue { i32, { i32 } } %A, 1
11308 // %E = insertvalue { i32 } %X, i32 42, 0
11309 // by switching the order of the insert and extract (though the
11310 // insertvalue should be left in, since it may have other uses).
11311 Value *NewEV = InsertNewInstBefore(
11312 ExtractValueInst::Create(IV->getAggregateOperand(),
11313 EV.idx_begin(), EV.idx_end()),
11315 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11319 // The insert list is a prefix of the extract list
11320 // We can simply remove the common indices from the extract and make it
11321 // operate on the inserted value instead of the insertvalue result.
11323 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11324 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11326 // %E extractvalue { i32 } { i32 42 }, 0
11327 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11330 // Can't simplify extracts from other values. Note that nested extracts are
11331 // already simplified implicitely by the above (extract ( extract (insert) )
11332 // will be translated into extract ( insert ( extract ) ) first and then just
11333 // the value inserted, if appropriate).
11337 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11338 /// is to leave as a vector operation.
11339 static bool CheapToScalarize(Value *V, bool isConstant) {
11340 if (isa<ConstantAggregateZero>(V))
11342 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11343 if (isConstant) return true;
11344 // If all elts are the same, we can extract.
11345 Constant *Op0 = C->getOperand(0);
11346 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11347 if (C->getOperand(i) != Op0)
11351 Instruction *I = dyn_cast<Instruction>(V);
11352 if (!I) return false;
11354 // Insert element gets simplified to the inserted element or is deleted if
11355 // this is constant idx extract element and its a constant idx insertelt.
11356 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11357 isa<ConstantInt>(I->getOperand(2)))
11359 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11361 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11362 if (BO->hasOneUse() &&
11363 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11364 CheapToScalarize(BO->getOperand(1), isConstant)))
11366 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11367 if (CI->hasOneUse() &&
11368 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11369 CheapToScalarize(CI->getOperand(1), isConstant)))
11375 /// Read and decode a shufflevector mask.
11377 /// It turns undef elements into values that are larger than the number of
11378 /// elements in the input.
11379 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11380 unsigned NElts = SVI->getType()->getNumElements();
11381 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11382 return std::vector<unsigned>(NElts, 0);
11383 if (isa<UndefValue>(SVI->getOperand(2)))
11384 return std::vector<unsigned>(NElts, 2*NElts);
11386 std::vector<unsigned> Result;
11387 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11388 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11389 if (isa<UndefValue>(*i))
11390 Result.push_back(NElts*2); // undef -> 8
11392 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11396 /// FindScalarElement - Given a vector and an element number, see if the scalar
11397 /// value is already around as a register, for example if it were inserted then
11398 /// extracted from the vector.
11399 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11400 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11401 const VectorType *PTy = cast<VectorType>(V->getType());
11402 unsigned Width = PTy->getNumElements();
11403 if (EltNo >= Width) // Out of range access.
11404 return UndefValue::get(PTy->getElementType());
11406 if (isa<UndefValue>(V))
11407 return UndefValue::get(PTy->getElementType());
11408 else if (isa<ConstantAggregateZero>(V))
11409 return Constant::getNullValue(PTy->getElementType());
11410 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11411 return CP->getOperand(EltNo);
11412 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11413 // If this is an insert to a variable element, we don't know what it is.
11414 if (!isa<ConstantInt>(III->getOperand(2)))
11416 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11418 // If this is an insert to the element we are looking for, return the
11420 if (EltNo == IIElt)
11421 return III->getOperand(1);
11423 // Otherwise, the insertelement doesn't modify the value, recurse on its
11425 return FindScalarElement(III->getOperand(0), EltNo);
11426 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11427 unsigned LHSWidth =
11428 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11429 unsigned InEl = getShuffleMask(SVI)[EltNo];
11430 if (InEl < LHSWidth)
11431 return FindScalarElement(SVI->getOperand(0), InEl);
11432 else if (InEl < LHSWidth*2)
11433 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11435 return UndefValue::get(PTy->getElementType());
11438 // Otherwise, we don't know.
11442 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11443 // If vector val is undef, replace extract with scalar undef.
11444 if (isa<UndefValue>(EI.getOperand(0)))
11445 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11447 // If vector val is constant 0, replace extract with scalar 0.
11448 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11449 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11451 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11452 // If vector val is constant with all elements the same, replace EI with
11453 // that element. When the elements are not identical, we cannot replace yet
11454 // (we do that below, but only when the index is constant).
11455 Constant *op0 = C->getOperand(0);
11456 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11457 if (C->getOperand(i) != op0) {
11462 return ReplaceInstUsesWith(EI, op0);
11465 // If extracting a specified index from the vector, see if we can recursively
11466 // find a previously computed scalar that was inserted into the vector.
11467 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11468 unsigned IndexVal = IdxC->getZExtValue();
11469 unsigned VectorWidth =
11470 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11472 // If this is extracting an invalid index, turn this into undef, to avoid
11473 // crashing the code below.
11474 if (IndexVal >= VectorWidth)
11475 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11477 // This instruction only demands the single element from the input vector.
11478 // If the input vector has a single use, simplify it based on this use
11480 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11481 uint64_t UndefElts;
11482 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11485 EI.setOperand(0, V);
11490 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11491 return ReplaceInstUsesWith(EI, Elt);
11493 // If the this extractelement is directly using a bitcast from a vector of
11494 // the same number of elements, see if we can find the source element from
11495 // it. In this case, we will end up needing to bitcast the scalars.
11496 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11497 if (const VectorType *VT =
11498 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11499 if (VT->getNumElements() == VectorWidth)
11500 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11501 return new BitCastInst(Elt, EI.getType());
11505 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11506 if (I->hasOneUse()) {
11507 // Push extractelement into predecessor operation if legal and
11508 // profitable to do so
11509 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11510 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11511 if (CheapToScalarize(BO, isConstantElt)) {
11512 ExtractElementInst *newEI0 =
11513 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11514 EI.getName()+".lhs");
11515 ExtractElementInst *newEI1 =
11516 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11517 EI.getName()+".rhs");
11518 InsertNewInstBefore(newEI0, EI);
11519 InsertNewInstBefore(newEI1, EI);
11520 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11522 } else if (isa<LoadInst>(I)) {
11524 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11525 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11526 PointerType::get(EI.getType(), AS),EI);
11527 GetElementPtrInst *GEP =
11528 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11529 InsertNewInstBefore(GEP, EI);
11530 return new LoadInst(GEP);
11533 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11534 // Extracting the inserted element?
11535 if (IE->getOperand(2) == EI.getOperand(1))
11536 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11537 // If the inserted and extracted elements are constants, they must not
11538 // be the same value, extract from the pre-inserted value instead.
11539 if (isa<Constant>(IE->getOperand(2)) &&
11540 isa<Constant>(EI.getOperand(1))) {
11541 AddUsesToWorkList(EI);
11542 EI.setOperand(0, IE->getOperand(0));
11545 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11546 // If this is extracting an element from a shufflevector, figure out where
11547 // it came from and extract from the appropriate input element instead.
11548 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11549 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11551 unsigned LHSWidth =
11552 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11554 if (SrcIdx < LHSWidth)
11555 Src = SVI->getOperand(0);
11556 else if (SrcIdx < LHSWidth*2) {
11557 SrcIdx -= LHSWidth;
11558 Src = SVI->getOperand(1);
11560 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11562 return new ExtractElementInst(Src, SrcIdx);
11569 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11570 /// elements from either LHS or RHS, return the shuffle mask and true.
11571 /// Otherwise, return false.
11572 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11573 std::vector<Constant*> &Mask) {
11574 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11575 "Invalid CollectSingleShuffleElements");
11576 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11578 if (isa<UndefValue>(V)) {
11579 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11581 } else if (V == LHS) {
11582 for (unsigned i = 0; i != NumElts; ++i)
11583 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11585 } else if (V == RHS) {
11586 for (unsigned i = 0; i != NumElts; ++i)
11587 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11589 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11590 // If this is an insert of an extract from some other vector, include it.
11591 Value *VecOp = IEI->getOperand(0);
11592 Value *ScalarOp = IEI->getOperand(1);
11593 Value *IdxOp = IEI->getOperand(2);
11595 if (!isa<ConstantInt>(IdxOp))
11597 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11599 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11600 // Okay, we can handle this if the vector we are insertinting into is
11601 // transitively ok.
11602 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11603 // If so, update the mask to reflect the inserted undef.
11604 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11607 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11608 if (isa<ConstantInt>(EI->getOperand(1)) &&
11609 EI->getOperand(0)->getType() == V->getType()) {
11610 unsigned ExtractedIdx =
11611 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11613 // This must be extracting from either LHS or RHS.
11614 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11615 // Okay, we can handle this if the vector we are insertinting into is
11616 // transitively ok.
11617 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11618 // If so, update the mask to reflect the inserted value.
11619 if (EI->getOperand(0) == LHS) {
11620 Mask[InsertedIdx % NumElts] =
11621 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11623 assert(EI->getOperand(0) == RHS);
11624 Mask[InsertedIdx % NumElts] =
11625 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11634 // TODO: Handle shufflevector here!
11639 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11640 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11641 /// that computes V and the LHS value of the shuffle.
11642 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11644 assert(isa<VectorType>(V->getType()) &&
11645 (RHS == 0 || V->getType() == RHS->getType()) &&
11646 "Invalid shuffle!");
11647 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11649 if (isa<UndefValue>(V)) {
11650 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11652 } else if (isa<ConstantAggregateZero>(V)) {
11653 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11655 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11656 // If this is an insert of an extract from some other vector, include it.
11657 Value *VecOp = IEI->getOperand(0);
11658 Value *ScalarOp = IEI->getOperand(1);
11659 Value *IdxOp = IEI->getOperand(2);
11661 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11662 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11663 EI->getOperand(0)->getType() == V->getType()) {
11664 unsigned ExtractedIdx =
11665 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11666 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11668 // Either the extracted from or inserted into vector must be RHSVec,
11669 // otherwise we'd end up with a shuffle of three inputs.
11670 if (EI->getOperand(0) == RHS || RHS == 0) {
11671 RHS = EI->getOperand(0);
11672 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11673 Mask[InsertedIdx % NumElts] =
11674 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11678 if (VecOp == RHS) {
11679 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11680 // Everything but the extracted element is replaced with the RHS.
11681 for (unsigned i = 0; i != NumElts; ++i) {
11682 if (i != InsertedIdx)
11683 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11688 // If this insertelement is a chain that comes from exactly these two
11689 // vectors, return the vector and the effective shuffle.
11690 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11691 return EI->getOperand(0);
11696 // TODO: Handle shufflevector here!
11698 // Otherwise, can't do anything fancy. Return an identity vector.
11699 for (unsigned i = 0; i != NumElts; ++i)
11700 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11704 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11705 Value *VecOp = IE.getOperand(0);
11706 Value *ScalarOp = IE.getOperand(1);
11707 Value *IdxOp = IE.getOperand(2);
11709 // Inserting an undef or into an undefined place, remove this.
11710 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11711 ReplaceInstUsesWith(IE, VecOp);
11713 // If the inserted element was extracted from some other vector, and if the
11714 // indexes are constant, try to turn this into a shufflevector operation.
11715 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11716 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11717 EI->getOperand(0)->getType() == IE.getType()) {
11718 unsigned NumVectorElts = IE.getType()->getNumElements();
11719 unsigned ExtractedIdx =
11720 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11721 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11723 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11724 return ReplaceInstUsesWith(IE, VecOp);
11726 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11727 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11729 // If we are extracting a value from a vector, then inserting it right
11730 // back into the same place, just use the input vector.
11731 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11732 return ReplaceInstUsesWith(IE, VecOp);
11734 // We could theoretically do this for ANY input. However, doing so could
11735 // turn chains of insertelement instructions into a chain of shufflevector
11736 // instructions, and right now we do not merge shufflevectors. As such,
11737 // only do this in a situation where it is clear that there is benefit.
11738 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11739 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11740 // the values of VecOp, except then one read from EIOp0.
11741 // Build a new shuffle mask.
11742 std::vector<Constant*> Mask;
11743 if (isa<UndefValue>(VecOp))
11744 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11746 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11747 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11750 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11751 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11752 ConstantVector::get(Mask));
11755 // If this insertelement isn't used by some other insertelement, turn it
11756 // (and any insertelements it points to), into one big shuffle.
11757 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11758 std::vector<Constant*> Mask;
11760 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11761 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11762 // We now have a shuffle of LHS, RHS, Mask.
11763 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11772 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11773 Value *LHS = SVI.getOperand(0);
11774 Value *RHS = SVI.getOperand(1);
11775 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11777 bool MadeChange = false;
11779 // Undefined shuffle mask -> undefined value.
11780 if (isa<UndefValue>(SVI.getOperand(2)))
11781 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11783 uint64_t UndefElts;
11784 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11786 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
11789 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
11790 if (VWidth <= 64 &&
11791 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
11792 LHS = SVI.getOperand(0);
11793 RHS = SVI.getOperand(1);
11797 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11798 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11799 if (LHS == RHS || isa<UndefValue>(LHS)) {
11800 if (isa<UndefValue>(LHS) && LHS == RHS) {
11801 // shuffle(undef,undef,mask) -> undef.
11802 return ReplaceInstUsesWith(SVI, LHS);
11805 // Remap any references to RHS to use LHS.
11806 std::vector<Constant*> Elts;
11807 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11808 if (Mask[i] >= 2*e)
11809 Elts.push_back(UndefValue::get(Type::Int32Ty));
11811 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11812 (Mask[i] < e && isa<UndefValue>(LHS))) {
11813 Mask[i] = 2*e; // Turn into undef.
11814 Elts.push_back(UndefValue::get(Type::Int32Ty));
11816 Mask[i] = Mask[i] % e; // Force to LHS.
11817 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11821 SVI.setOperand(0, SVI.getOperand(1));
11822 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11823 SVI.setOperand(2, ConstantVector::get(Elts));
11824 LHS = SVI.getOperand(0);
11825 RHS = SVI.getOperand(1);
11829 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11830 bool isLHSID = true, isRHSID = true;
11832 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11833 if (Mask[i] >= e*2) continue; // Ignore undef values.
11834 // Is this an identity shuffle of the LHS value?
11835 isLHSID &= (Mask[i] == i);
11837 // Is this an identity shuffle of the RHS value?
11838 isRHSID &= (Mask[i]-e == i);
11841 // Eliminate identity shuffles.
11842 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11843 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11845 // If the LHS is a shufflevector itself, see if we can combine it with this
11846 // one without producing an unusual shuffle. Here we are really conservative:
11847 // we are absolutely afraid of producing a shuffle mask not in the input
11848 // program, because the code gen may not be smart enough to turn a merged
11849 // shuffle into two specific shuffles: it may produce worse code. As such,
11850 // we only merge two shuffles if the result is one of the two input shuffle
11851 // masks. In this case, merging the shuffles just removes one instruction,
11852 // which we know is safe. This is good for things like turning:
11853 // (splat(splat)) -> splat.
11854 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11855 if (isa<UndefValue>(RHS)) {
11856 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11858 std::vector<unsigned> NewMask;
11859 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11860 if (Mask[i] >= 2*e)
11861 NewMask.push_back(2*e);
11863 NewMask.push_back(LHSMask[Mask[i]]);
11865 // If the result mask is equal to the src shuffle or this shuffle mask, do
11866 // the replacement.
11867 if (NewMask == LHSMask || NewMask == Mask) {
11868 std::vector<Constant*> Elts;
11869 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11870 if (NewMask[i] >= e*2) {
11871 Elts.push_back(UndefValue::get(Type::Int32Ty));
11873 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11876 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11877 LHSSVI->getOperand(1),
11878 ConstantVector::get(Elts));
11883 return MadeChange ? &SVI : 0;
11889 /// TryToSinkInstruction - Try to move the specified instruction from its
11890 /// current block into the beginning of DestBlock, which can only happen if it's
11891 /// safe to move the instruction past all of the instructions between it and the
11892 /// end of its block.
11893 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11894 assert(I->hasOneUse() && "Invariants didn't hold!");
11896 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11897 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11900 // Do not sink alloca instructions out of the entry block.
11901 if (isa<AllocaInst>(I) && I->getParent() ==
11902 &DestBlock->getParent()->getEntryBlock())
11905 // We can only sink load instructions if there is nothing between the load and
11906 // the end of block that could change the value.
11907 if (I->mayReadFromMemory()) {
11908 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11910 if (Scan->mayWriteToMemory())
11914 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11916 I->moveBefore(InsertPos);
11922 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11923 /// all reachable code to the worklist.
11925 /// This has a couple of tricks to make the code faster and more powerful. In
11926 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11927 /// them to the worklist (this significantly speeds up instcombine on code where
11928 /// many instructions are dead or constant). Additionally, if we find a branch
11929 /// whose condition is a known constant, we only visit the reachable successors.
11931 static void AddReachableCodeToWorklist(BasicBlock *BB,
11932 SmallPtrSet<BasicBlock*, 64> &Visited,
11934 const TargetData *TD) {
11935 SmallVector<BasicBlock*, 256> Worklist;
11936 Worklist.push_back(BB);
11938 while (!Worklist.empty()) {
11939 BB = Worklist.back();
11940 Worklist.pop_back();
11942 // We have now visited this block! If we've already been here, ignore it.
11943 if (!Visited.insert(BB)) continue;
11945 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11946 Instruction *Inst = BBI++;
11948 // DCE instruction if trivially dead.
11949 if (isInstructionTriviallyDead(Inst)) {
11951 DOUT << "IC: DCE: " << *Inst;
11952 Inst->eraseFromParent();
11956 // ConstantProp instruction if trivially constant.
11957 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11958 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11959 Inst->replaceAllUsesWith(C);
11961 Inst->eraseFromParent();
11965 IC.AddToWorkList(Inst);
11968 // Recursively visit successors. If this is a branch or switch on a
11969 // constant, only visit the reachable successor.
11970 TerminatorInst *TI = BB->getTerminator();
11971 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11972 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11973 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11974 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11975 Worklist.push_back(ReachableBB);
11978 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11979 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11980 // See if this is an explicit destination.
11981 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11982 if (SI->getCaseValue(i) == Cond) {
11983 BasicBlock *ReachableBB = SI->getSuccessor(i);
11984 Worklist.push_back(ReachableBB);
11988 // Otherwise it is the default destination.
11989 Worklist.push_back(SI->getSuccessor(0));
11994 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11995 Worklist.push_back(TI->getSuccessor(i));
11999 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12000 bool Changed = false;
12001 TD = &getAnalysis<TargetData>();
12003 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12004 << F.getNameStr() << "\n");
12007 // Do a depth-first traversal of the function, populate the worklist with
12008 // the reachable instructions. Ignore blocks that are not reachable. Keep
12009 // track of which blocks we visit.
12010 SmallPtrSet<BasicBlock*, 64> Visited;
12011 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12013 // Do a quick scan over the function. If we find any blocks that are
12014 // unreachable, remove any instructions inside of them. This prevents
12015 // the instcombine code from having to deal with some bad special cases.
12016 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12017 if (!Visited.count(BB)) {
12018 Instruction *Term = BB->getTerminator();
12019 while (Term != BB->begin()) { // Remove instrs bottom-up
12020 BasicBlock::iterator I = Term; --I;
12022 DOUT << "IC: DCE: " << *I;
12025 if (!I->use_empty())
12026 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12027 I->eraseFromParent();
12032 while (!Worklist.empty()) {
12033 Instruction *I = RemoveOneFromWorkList();
12034 if (I == 0) continue; // skip null values.
12036 // Check to see if we can DCE the instruction.
12037 if (isInstructionTriviallyDead(I)) {
12038 // Add operands to the worklist.
12039 if (I->getNumOperands() < 4)
12040 AddUsesToWorkList(*I);
12043 DOUT << "IC: DCE: " << *I;
12045 I->eraseFromParent();
12046 RemoveFromWorkList(I);
12050 // Instruction isn't dead, see if we can constant propagate it.
12051 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12052 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12054 // Add operands to the worklist.
12055 AddUsesToWorkList(*I);
12056 ReplaceInstUsesWith(*I, C);
12059 I->eraseFromParent();
12060 RemoveFromWorkList(I);
12064 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12065 // See if we can constant fold its operands.
12066 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
12067 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
12068 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12074 // See if we can trivially sink this instruction to a successor basic block.
12075 if (I->hasOneUse()) {
12076 BasicBlock *BB = I->getParent();
12077 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12078 if (UserParent != BB) {
12079 bool UserIsSuccessor = false;
12080 // See if the user is one of our successors.
12081 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12082 if (*SI == UserParent) {
12083 UserIsSuccessor = true;
12087 // If the user is one of our immediate successors, and if that successor
12088 // only has us as a predecessors (we'd have to split the critical edge
12089 // otherwise), we can keep going.
12090 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12091 next(pred_begin(UserParent)) == pred_end(UserParent))
12092 // Okay, the CFG is simple enough, try to sink this instruction.
12093 Changed |= TryToSinkInstruction(I, UserParent);
12097 // Now that we have an instruction, try combining it to simplify it...
12101 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12102 if (Instruction *Result = visit(*I)) {
12104 // Should we replace the old instruction with a new one?
12106 DOUT << "IC: Old = " << *I
12107 << " New = " << *Result;
12109 // Everything uses the new instruction now.
12110 I->replaceAllUsesWith(Result);
12112 // Push the new instruction and any users onto the worklist.
12113 AddToWorkList(Result);
12114 AddUsersToWorkList(*Result);
12116 // Move the name to the new instruction first.
12117 Result->takeName(I);
12119 // Insert the new instruction into the basic block...
12120 BasicBlock *InstParent = I->getParent();
12121 BasicBlock::iterator InsertPos = I;
12123 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12124 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12127 InstParent->getInstList().insert(InsertPos, Result);
12129 // Make sure that we reprocess all operands now that we reduced their
12131 AddUsesToWorkList(*I);
12133 // Instructions can end up on the worklist more than once. Make sure
12134 // we do not process an instruction that has been deleted.
12135 RemoveFromWorkList(I);
12137 // Erase the old instruction.
12138 InstParent->getInstList().erase(I);
12141 DOUT << "IC: Mod = " << OrigI
12142 << " New = " << *I;
12145 // If the instruction was modified, it's possible that it is now dead.
12146 // if so, remove it.
12147 if (isInstructionTriviallyDead(I)) {
12148 // Make sure we process all operands now that we are reducing their
12150 AddUsesToWorkList(*I);
12152 // Instructions may end up in the worklist more than once. Erase all
12153 // occurrences of this instruction.
12154 RemoveFromWorkList(I);
12155 I->eraseFromParent();
12158 AddUsersToWorkList(*I);
12165 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12167 // Do an explicit clear, this shrinks the map if needed.
12168 WorklistMap.clear();
12173 bool InstCombiner::runOnFunction(Function &F) {
12174 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12176 bool EverMadeChange = false;
12178 // Iterate while there is work to do.
12179 unsigned Iteration = 0;
12180 while (DoOneIteration(F, Iteration++))
12181 EverMadeChange = true;
12182 return EverMadeChange;
12185 FunctionPass *llvm::createInstructionCombiningPass() {
12186 return new InstCombiner();