1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 std::vector<Instruction*> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass((intptr_t)&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 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 &CI);
223 Instruction *visitCallInst(CallInst &CI);
224 Instruction *visitInvokeInst(InvokeInst &II);
225 Instruction *visitPHINode(PHINode &PN);
226 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
227 Instruction *visitAllocationInst(AllocationInst &AI);
228 Instruction *visitFreeInst(FreeInst &FI);
229 Instruction *visitLoadInst(LoadInst &LI);
230 Instruction *visitStoreInst(StoreInst &SI);
231 Instruction *visitBranchInst(BranchInst &BI);
232 Instruction *visitSwitchInst(SwitchInst &SI);
233 Instruction *visitInsertElementInst(InsertElementInst &IE);
234 Instruction *visitExtractElementInst(ExtractElementInst &EI);
235 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
236 Instruction *visitExtractValueInst(ExtractValueInst &EV);
238 // visitInstruction - Specify what to return for unhandled instructions...
239 Instruction *visitInstruction(Instruction &I) { return 0; }
242 Instruction *visitCallSite(CallSite CS);
243 bool transformConstExprCastCall(CallSite CS);
244 Instruction *transformCallThroughTrampoline(CallSite CS);
245 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
246 bool DoXform = true);
247 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
250 // InsertNewInstBefore - insert an instruction New before instruction Old
251 // in the program. Add the new instruction to the worklist.
253 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
254 assert(New && New->getParent() == 0 &&
255 "New instruction already inserted into a basic block!");
256 BasicBlock *BB = Old.getParent();
257 BB->getInstList().insert(&Old, New); // Insert inst
262 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
263 /// This also adds the cast to the worklist. Finally, this returns the
265 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
267 if (V->getType() == Ty) return V;
269 if (Constant *CV = dyn_cast<Constant>(V))
270 return ConstantExpr::getCast(opc, CV, Ty);
272 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
277 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
278 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
282 // ReplaceInstUsesWith - This method is to be used when an instruction is
283 // found to be dead, replacable with another preexisting expression. Here
284 // we add all uses of I to the worklist, replace all uses of I with the new
285 // value, then return I, so that the inst combiner will know that I was
288 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
289 AddUsersToWorkList(I); // Add all modified instrs to worklist
291 I.replaceAllUsesWith(V);
294 // If we are replacing the instruction with itself, this must be in a
295 // segment of unreachable code, so just clobber the instruction.
296 I.replaceAllUsesWith(UndefValue::get(I.getType()));
301 // UpdateValueUsesWith - This method is to be used when an value is
302 // found to be replacable with another preexisting expression or was
303 // updated. Here we add all uses of I to the worklist, replace all uses of
304 // I with the new value (unless the instruction was just updated), then
305 // return true, so that the inst combiner will know that I was modified.
307 bool UpdateValueUsesWith(Value *Old, Value *New) {
308 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
310 Old->replaceAllUsesWith(New);
311 if (Instruction *I = dyn_cast<Instruction>(Old))
313 if (Instruction *I = dyn_cast<Instruction>(New))
318 // EraseInstFromFunction - When dealing with an instruction that has side
319 // effects or produces a void value, we can't rely on DCE to delete the
320 // instruction. Instead, visit methods should return the value returned by
322 Instruction *EraseInstFromFunction(Instruction &I) {
323 assert(I.use_empty() && "Cannot erase instruction that is used!");
324 AddUsesToWorkList(I);
325 RemoveFromWorkList(&I);
327 return 0; // Don't do anything with FI
330 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
331 APInt &KnownOne, unsigned Depth = 0) const {
332 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
335 bool MaskedValueIsZero(Value *V, const APInt &Mask,
336 unsigned Depth = 0) const {
337 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
339 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
340 return llvm::ComputeNumSignBits(Op, TD, Depth);
344 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
345 /// InsertBefore instruction. This is specialized a bit to avoid inserting
346 /// casts that are known to not do anything...
348 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
349 Value *V, const Type *DestTy,
350 Instruction *InsertBefore);
352 /// SimplifyCommutative - This performs a few simplifications for
353 /// commutative operators.
354 bool SimplifyCommutative(BinaryOperator &I);
356 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
357 /// most-complex to least-complex order.
358 bool SimplifyCompare(CmpInst &I);
360 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
361 /// on the demanded bits.
362 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
363 APInt& KnownZero, APInt& KnownOne,
366 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
367 uint64_t &UndefElts, unsigned Depth = 0);
369 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
370 // PHI node as operand #0, see if we can fold the instruction into the PHI
371 // (which is only possible if all operands to the PHI are constants).
372 Instruction *FoldOpIntoPhi(Instruction &I);
374 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
375 // operator and they all are only used by the PHI, PHI together their
376 // inputs, and do the operation once, to the result of the PHI.
377 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
378 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
381 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
382 ConstantInt *AndRHS, BinaryOperator &TheAnd);
384 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
385 bool isSub, Instruction &I);
386 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
387 bool isSigned, bool Inside, Instruction &IB);
388 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
389 Instruction *MatchBSwap(BinaryOperator &I);
390 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
391 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
392 Instruction *SimplifyMemSet(MemSetInst *MI);
395 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
397 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
399 int &NumCastsRemoved);
400 unsigned GetOrEnforceKnownAlignment(Value *V,
401 unsigned PrefAlign = 0);
406 char InstCombiner::ID = 0;
407 static RegisterPass<InstCombiner>
408 X("instcombine", "Combine redundant instructions");
410 // getComplexity: Assign a complexity or rank value to LLVM Values...
411 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
412 static unsigned getComplexity(Value *V) {
413 if (isa<Instruction>(V)) {
414 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
418 if (isa<Argument>(V)) return 3;
419 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
422 // isOnlyUse - Return true if this instruction will be deleted if we stop using
424 static bool isOnlyUse(Value *V) {
425 return V->hasOneUse() || isa<Constant>(V);
428 // getPromotedType - Return the specified type promoted as it would be to pass
429 // though a va_arg area...
430 static const Type *getPromotedType(const Type *Ty) {
431 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
432 if (ITy->getBitWidth() < 32)
433 return Type::Int32Ty;
438 /// getBitCastOperand - If the specified operand is a CastInst or a constant
439 /// expression bitcast, return the operand value, otherwise return null.
440 static Value *getBitCastOperand(Value *V) {
441 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
442 return I->getOperand(0);
443 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
444 if (CE->getOpcode() == Instruction::BitCast)
445 return CE->getOperand(0);
449 /// This function is a wrapper around CastInst::isEliminableCastPair. It
450 /// simply extracts arguments and returns what that function returns.
451 static Instruction::CastOps
452 isEliminableCastPair(
453 const CastInst *CI, ///< The first cast instruction
454 unsigned opcode, ///< The opcode of the second cast instruction
455 const Type *DstTy, ///< The target type for the second cast instruction
456 TargetData *TD ///< The target data for pointer size
459 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
460 const Type *MidTy = CI->getType(); // B from above
462 // Get the opcodes of the two Cast instructions
463 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
464 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
466 return Instruction::CastOps(
467 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
468 DstTy, TD->getIntPtrType()));
471 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
472 /// in any code being generated. It does not require codegen if V is simple
473 /// enough or if the cast can be folded into other casts.
474 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
475 const Type *Ty, TargetData *TD) {
476 if (V->getType() == Ty || isa<Constant>(V)) return false;
478 // If this is another cast that can be eliminated, it isn't codegen either.
479 if (const CastInst *CI = dyn_cast<CastInst>(V))
480 if (isEliminableCastPair(CI, opcode, Ty, TD))
485 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
486 /// InsertBefore instruction. This is specialized a bit to avoid inserting
487 /// casts that are known to not do anything...
489 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
490 Value *V, const Type *DestTy,
491 Instruction *InsertBefore) {
492 if (V->getType() == DestTy) return V;
493 if (Constant *C = dyn_cast<Constant>(V))
494 return ConstantExpr::getCast(opcode, C, DestTy);
496 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
499 // SimplifyCommutative - This performs a few simplifications for commutative
502 // 1. Order operands such that they are listed from right (least complex) to
503 // left (most complex). This puts constants before unary operators before
506 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
507 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
509 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
510 bool Changed = false;
511 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
512 Changed = !I.swapOperands();
514 if (!I.isAssociative()) return Changed;
515 Instruction::BinaryOps Opcode = I.getOpcode();
516 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
517 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
518 if (isa<Constant>(I.getOperand(1))) {
519 Constant *Folded = ConstantExpr::get(I.getOpcode(),
520 cast<Constant>(I.getOperand(1)),
521 cast<Constant>(Op->getOperand(1)));
522 I.setOperand(0, Op->getOperand(0));
523 I.setOperand(1, Folded);
525 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
526 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
527 isOnlyUse(Op) && isOnlyUse(Op1)) {
528 Constant *C1 = cast<Constant>(Op->getOperand(1));
529 Constant *C2 = cast<Constant>(Op1->getOperand(1));
531 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
532 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
533 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
537 I.setOperand(0, New);
538 I.setOperand(1, Folded);
545 /// SimplifyCompare - For a CmpInst this function just orders the operands
546 /// so that theyare listed from right (least complex) to left (most complex).
547 /// This puts constants before unary operators before binary operators.
548 bool InstCombiner::SimplifyCompare(CmpInst &I) {
549 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
552 // Compare instructions are not associative so there's nothing else we can do.
556 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
557 // if the LHS is a constant zero (which is the 'negate' form).
559 static inline Value *dyn_castNegVal(Value *V) {
560 if (BinaryOperator::isNeg(V))
561 return BinaryOperator::getNegArgument(V);
563 // Constants can be considered to be negated values if they can be folded.
564 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
565 return ConstantExpr::getNeg(C);
567 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
568 if (C->getType()->getElementType()->isInteger())
569 return ConstantExpr::getNeg(C);
574 static inline Value *dyn_castNotVal(Value *V) {
575 if (BinaryOperator::isNot(V))
576 return BinaryOperator::getNotArgument(V);
578 // Constants can be considered to be not'ed values...
579 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
580 return ConstantInt::get(~C->getValue());
584 // dyn_castFoldableMul - If this value is a multiply that can be folded into
585 // other computations (because it has a constant operand), return the
586 // non-constant operand of the multiply, and set CST to point to the multiplier.
587 // Otherwise, return null.
589 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
590 if (V->hasOneUse() && V->getType()->isInteger())
591 if (Instruction *I = dyn_cast<Instruction>(V)) {
592 if (I->getOpcode() == Instruction::Mul)
593 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
594 return I->getOperand(0);
595 if (I->getOpcode() == Instruction::Shl)
596 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
597 // The multiplier is really 1 << CST.
598 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
599 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
600 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
601 return I->getOperand(0);
607 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
608 /// expression, return it.
609 static User *dyn_castGetElementPtr(Value *V) {
610 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
611 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
612 if (CE->getOpcode() == Instruction::GetElementPtr)
613 return cast<User>(V);
617 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
618 /// opcode value. Otherwise return UserOp1.
619 static unsigned getOpcode(const Value *V) {
620 if (const Instruction *I = dyn_cast<Instruction>(V))
621 return I->getOpcode();
622 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
623 return CE->getOpcode();
624 // Use UserOp1 to mean there's no opcode.
625 return Instruction::UserOp1;
628 /// AddOne - Add one to a ConstantInt
629 static ConstantInt *AddOne(ConstantInt *C) {
630 APInt Val(C->getValue());
631 return ConstantInt::get(++Val);
633 /// SubOne - Subtract one from a ConstantInt
634 static ConstantInt *SubOne(ConstantInt *C) {
635 APInt Val(C->getValue());
636 return ConstantInt::get(--Val);
638 /// Add - Add two ConstantInts together
639 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
640 return ConstantInt::get(C1->getValue() + C2->getValue());
642 /// And - Bitwise AND two ConstantInts together
643 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
644 return ConstantInt::get(C1->getValue() & C2->getValue());
646 /// Subtract - Subtract one ConstantInt from another
647 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
648 return ConstantInt::get(C1->getValue() - C2->getValue());
650 /// Multiply - Multiply two ConstantInts together
651 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
652 return ConstantInt::get(C1->getValue() * C2->getValue());
654 /// MultiplyOverflows - True if the multiply can not be expressed in an int
656 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
657 uint32_t W = C1->getBitWidth();
658 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
667 APInt MulExt = LHSExt * RHSExt;
670 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
671 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
672 return MulExt.slt(Min) || MulExt.sgt(Max);
674 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
678 /// ShrinkDemandedConstant - Check to see if the specified operand of the
679 /// specified instruction is a constant integer. If so, check to see if there
680 /// are any bits set in the constant that are not demanded. If so, shrink the
681 /// constant and return true.
682 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
684 assert(I && "No instruction?");
685 assert(OpNo < I->getNumOperands() && "Operand index too large");
687 // If the operand is not a constant integer, nothing to do.
688 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
689 if (!OpC) return false;
691 // If there are no bits set that aren't demanded, nothing to do.
692 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
693 if ((~Demanded & OpC->getValue()) == 0)
696 // This instruction is producing bits that are not demanded. Shrink the RHS.
697 Demanded &= OpC->getValue();
698 I->setOperand(OpNo, ConstantInt::get(Demanded));
702 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
703 // set of known zero and one bits, compute the maximum and minimum values that
704 // could have the specified known zero and known one bits, returning them in
706 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
707 const APInt& KnownZero,
708 const APInt& KnownOne,
709 APInt& Min, APInt& Max) {
710 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
711 assert(KnownZero.getBitWidth() == BitWidth &&
712 KnownOne.getBitWidth() == BitWidth &&
713 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
714 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
715 APInt UnknownBits = ~(KnownZero|KnownOne);
717 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
718 // bit if it is unknown.
720 Max = KnownOne|UnknownBits;
722 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
724 Max.clear(BitWidth-1);
728 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
729 // a set of known zero and one bits, compute the maximum and minimum values that
730 // could have the specified known zero and known one bits, returning them in
732 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
733 const APInt &KnownZero,
734 const APInt &KnownOne,
735 APInt &Min, APInt &Max) {
736 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
737 assert(KnownZero.getBitWidth() == BitWidth &&
738 KnownOne.getBitWidth() == BitWidth &&
739 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
740 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
741 APInt UnknownBits = ~(KnownZero|KnownOne);
743 // The minimum value is when the unknown bits are all zeros.
745 // The maximum value is when the unknown bits are all ones.
746 Max = KnownOne|UnknownBits;
749 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
750 /// value based on the demanded bits. When this function is called, it is known
751 /// that only the bits set in DemandedMask of the result of V are ever used
752 /// downstream. Consequently, depending on the mask and V, it may be possible
753 /// to replace V with a constant or one of its operands. In such cases, this
754 /// function does the replacement and returns true. In all other cases, it
755 /// returns false after analyzing the expression and setting KnownOne and known
756 /// to be one in the expression. KnownZero contains all the bits that are known
757 /// to be zero in the expression. These are provided to potentially allow the
758 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
759 /// the expression. KnownOne and KnownZero always follow the invariant that
760 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
761 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
762 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
763 /// and KnownOne must all be the same.
764 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
765 APInt& KnownZero, APInt& KnownOne,
767 assert(V != 0 && "Null pointer of Value???");
768 assert(Depth <= 6 && "Limit Search Depth");
769 uint32_t BitWidth = DemandedMask.getBitWidth();
770 const IntegerType *VTy = cast<IntegerType>(V->getType());
771 assert(VTy->getBitWidth() == BitWidth &&
772 KnownZero.getBitWidth() == BitWidth &&
773 KnownOne.getBitWidth() == BitWidth &&
774 "Value *V, DemandedMask, KnownZero and KnownOne \
775 must have same BitWidth");
776 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
777 // We know all of the bits for a constant!
778 KnownOne = CI->getValue() & DemandedMask;
779 KnownZero = ~KnownOne & DemandedMask;
785 if (!V->hasOneUse()) { // Other users may use these bits.
786 if (Depth != 0) { // Not at the root.
787 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
788 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
791 // If this is the root being simplified, allow it to have multiple uses,
792 // just set the DemandedMask to all bits.
793 DemandedMask = APInt::getAllOnesValue(BitWidth);
794 } else if (DemandedMask == 0) { // Not demanding any bits from V.
795 if (V != UndefValue::get(VTy))
796 return UpdateValueUsesWith(V, UndefValue::get(VTy));
798 } else if (Depth == 6) { // Limit search depth.
802 Instruction *I = dyn_cast<Instruction>(V);
803 if (!I) return false; // Only analyze instructions.
805 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
806 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
807 switch (I->getOpcode()) {
809 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
811 case Instruction::And:
812 // If either the LHS or the RHS are Zero, the result is zero.
813 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
814 RHSKnownZero, RHSKnownOne, Depth+1))
816 assert((RHSKnownZero & RHSKnownOne) == 0 &&
817 "Bits known to be one AND zero?");
819 // If something is known zero on the RHS, the bits aren't demanded on the
821 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
822 LHSKnownZero, LHSKnownOne, Depth+1))
824 assert((LHSKnownZero & LHSKnownOne) == 0 &&
825 "Bits known to be one AND zero?");
827 // If all of the demanded bits are known 1 on one side, return the other.
828 // These bits cannot contribute to the result of the 'and'.
829 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
830 (DemandedMask & ~LHSKnownZero))
831 return UpdateValueUsesWith(I, I->getOperand(0));
832 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
833 (DemandedMask & ~RHSKnownZero))
834 return UpdateValueUsesWith(I, I->getOperand(1));
836 // If all of the demanded bits in the inputs are known zeros, return zero.
837 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
838 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
840 // If the RHS is a constant, see if we can simplify it.
841 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
842 return UpdateValueUsesWith(I, I);
844 // Output known-1 bits are only known if set in both the LHS & RHS.
845 RHSKnownOne &= LHSKnownOne;
846 // Output known-0 are known to be clear if zero in either the LHS | RHS.
847 RHSKnownZero |= LHSKnownZero;
849 case Instruction::Or:
850 // If either the LHS or the RHS are One, the result is One.
851 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
852 RHSKnownZero, RHSKnownOne, Depth+1))
854 assert((RHSKnownZero & RHSKnownOne) == 0 &&
855 "Bits known to be one AND zero?");
856 // If something is known one on the RHS, the bits aren't demanded on the
858 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
859 LHSKnownZero, LHSKnownOne, Depth+1))
861 assert((LHSKnownZero & LHSKnownOne) == 0 &&
862 "Bits known to be one AND zero?");
864 // If all of the demanded bits are known zero on one side, return the other.
865 // These bits cannot contribute to the result of the 'or'.
866 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
867 (DemandedMask & ~LHSKnownOne))
868 return UpdateValueUsesWith(I, I->getOperand(0));
869 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
870 (DemandedMask & ~RHSKnownOne))
871 return UpdateValueUsesWith(I, I->getOperand(1));
873 // If all of the potentially set bits on one side are known to be set on
874 // the other side, just use the 'other' side.
875 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
876 (DemandedMask & (~RHSKnownZero)))
877 return UpdateValueUsesWith(I, I->getOperand(0));
878 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
879 (DemandedMask & (~LHSKnownZero)))
880 return UpdateValueUsesWith(I, I->getOperand(1));
882 // If the RHS is a constant, see if we can simplify it.
883 if (ShrinkDemandedConstant(I, 1, DemandedMask))
884 return UpdateValueUsesWith(I, I);
886 // Output known-0 bits are only known if clear in both the LHS & RHS.
887 RHSKnownZero &= LHSKnownZero;
888 // Output known-1 are known to be set if set in either the LHS | RHS.
889 RHSKnownOne |= LHSKnownOne;
891 case Instruction::Xor: {
892 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
893 RHSKnownZero, RHSKnownOne, Depth+1))
895 assert((RHSKnownZero & RHSKnownOne) == 0 &&
896 "Bits known to be one AND zero?");
897 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
898 LHSKnownZero, LHSKnownOne, Depth+1))
900 assert((LHSKnownZero & LHSKnownOne) == 0 &&
901 "Bits known to be one AND zero?");
903 // If all of the demanded bits are known zero on one side, return the other.
904 // These bits cannot contribute to the result of the 'xor'.
905 if ((DemandedMask & RHSKnownZero) == DemandedMask)
906 return UpdateValueUsesWith(I, I->getOperand(0));
907 if ((DemandedMask & LHSKnownZero) == DemandedMask)
908 return UpdateValueUsesWith(I, I->getOperand(1));
910 // Output known-0 bits are known if clear or set in both the LHS & RHS.
911 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
912 (RHSKnownOne & LHSKnownOne);
913 // Output known-1 are known to be set if set in only one of the LHS, RHS.
914 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
915 (RHSKnownOne & LHSKnownZero);
917 // If all of the demanded bits are known to be zero on one side or the
918 // other, turn this into an *inclusive* or.
919 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
920 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
922 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
924 InsertNewInstBefore(Or, *I);
925 return UpdateValueUsesWith(I, Or);
928 // If all of the demanded bits on one side are known, and all of the set
929 // bits on that side are also known to be set on the other side, turn this
930 // into an AND, as we know the bits will be cleared.
931 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
932 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
934 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
935 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
937 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
938 InsertNewInstBefore(And, *I);
939 return UpdateValueUsesWith(I, And);
943 // If the RHS is a constant, see if we can simplify it.
944 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
945 if (ShrinkDemandedConstant(I, 1, DemandedMask))
946 return UpdateValueUsesWith(I, I);
948 RHSKnownZero = KnownZeroOut;
949 RHSKnownOne = KnownOneOut;
952 case Instruction::Select:
953 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
954 RHSKnownZero, RHSKnownOne, Depth+1))
956 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
957 LHSKnownZero, LHSKnownOne, Depth+1))
959 assert((RHSKnownZero & RHSKnownOne) == 0 &&
960 "Bits known to be one AND zero?");
961 assert((LHSKnownZero & LHSKnownOne) == 0 &&
962 "Bits known to be one AND zero?");
964 // If the operands are constants, see if we can simplify them.
965 if (ShrinkDemandedConstant(I, 1, DemandedMask))
966 return UpdateValueUsesWith(I, I);
967 if (ShrinkDemandedConstant(I, 2, DemandedMask))
968 return UpdateValueUsesWith(I, I);
970 // Only known if known in both the LHS and RHS.
971 RHSKnownOne &= LHSKnownOne;
972 RHSKnownZero &= LHSKnownZero;
974 case Instruction::Trunc: {
976 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
977 DemandedMask.zext(truncBf);
978 RHSKnownZero.zext(truncBf);
979 RHSKnownOne.zext(truncBf);
980 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
981 RHSKnownZero, RHSKnownOne, Depth+1))
983 DemandedMask.trunc(BitWidth);
984 RHSKnownZero.trunc(BitWidth);
985 RHSKnownOne.trunc(BitWidth);
986 assert((RHSKnownZero & RHSKnownOne) == 0 &&
987 "Bits known to be one AND zero?");
990 case Instruction::BitCast:
991 if (!I->getOperand(0)->getType()->isInteger())
994 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
995 RHSKnownZero, RHSKnownOne, Depth+1))
997 assert((RHSKnownZero & RHSKnownOne) == 0 &&
998 "Bits known to be one AND zero?");
1000 case Instruction::ZExt: {
1001 // Compute the bits in the result that are not present in the input.
1002 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1003 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1005 DemandedMask.trunc(SrcBitWidth);
1006 RHSKnownZero.trunc(SrcBitWidth);
1007 RHSKnownOne.trunc(SrcBitWidth);
1008 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1009 RHSKnownZero, RHSKnownOne, Depth+1))
1011 DemandedMask.zext(BitWidth);
1012 RHSKnownZero.zext(BitWidth);
1013 RHSKnownOne.zext(BitWidth);
1014 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1015 "Bits known to be one AND zero?");
1016 // The top bits are known to be zero.
1017 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1020 case Instruction::SExt: {
1021 // Compute the bits in the result that are not present in the input.
1022 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1023 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1025 APInt InputDemandedBits = DemandedMask &
1026 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1028 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1029 // If any of the sign extended bits are demanded, we know that the sign
1031 if ((NewBits & DemandedMask) != 0)
1032 InputDemandedBits.set(SrcBitWidth-1);
1034 InputDemandedBits.trunc(SrcBitWidth);
1035 RHSKnownZero.trunc(SrcBitWidth);
1036 RHSKnownOne.trunc(SrcBitWidth);
1037 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1038 RHSKnownZero, RHSKnownOne, Depth+1))
1040 InputDemandedBits.zext(BitWidth);
1041 RHSKnownZero.zext(BitWidth);
1042 RHSKnownOne.zext(BitWidth);
1043 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1044 "Bits known to be one AND zero?");
1046 // If the sign bit of the input is known set or clear, then we know the
1047 // top bits of the result.
1049 // If the input sign bit is known zero, or if the NewBits are not demanded
1050 // convert this into a zero extension.
1051 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1053 // Convert to ZExt cast
1054 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1055 return UpdateValueUsesWith(I, NewCast);
1056 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1057 RHSKnownOne |= NewBits;
1061 case Instruction::Add: {
1062 // Figure out what the input bits are. If the top bits of the and result
1063 // are not demanded, then the add doesn't demand them from its input
1065 uint32_t NLZ = DemandedMask.countLeadingZeros();
1067 // If there is a constant on the RHS, there are a variety of xformations
1069 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1070 // If null, this should be simplified elsewhere. Some of the xforms here
1071 // won't work if the RHS is zero.
1075 // If the top bit of the output is demanded, demand everything from the
1076 // input. Otherwise, we demand all the input bits except NLZ top bits.
1077 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1079 // Find information about known zero/one bits in the input.
1080 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1081 LHSKnownZero, LHSKnownOne, Depth+1))
1084 // If the RHS of the add has bits set that can't affect the input, reduce
1086 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1087 return UpdateValueUsesWith(I, I);
1089 // Avoid excess work.
1090 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1093 // Turn it into OR if input bits are zero.
1094 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1096 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1098 InsertNewInstBefore(Or, *I);
1099 return UpdateValueUsesWith(I, Or);
1102 // We can say something about the output known-zero and known-one bits,
1103 // depending on potential carries from the input constant and the
1104 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1105 // bits set and the RHS constant is 0x01001, then we know we have a known
1106 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1108 // To compute this, we first compute the potential carry bits. These are
1109 // the bits which may be modified. I'm not aware of a better way to do
1111 const APInt& RHSVal = RHS->getValue();
1112 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1114 // Now that we know which bits have carries, compute the known-1/0 sets.
1116 // Bits are known one if they are known zero in one operand and one in the
1117 // other, and there is no input carry.
1118 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1119 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1121 // Bits are known zero if they are known zero in both operands and there
1122 // is no input carry.
1123 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1125 // If the high-bits of this ADD are not demanded, then it does not demand
1126 // the high bits of its LHS or RHS.
1127 if (DemandedMask[BitWidth-1] == 0) {
1128 // Right fill the mask of bits for this ADD to demand the most
1129 // significant bit and all those below it.
1130 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1131 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1132 LHSKnownZero, LHSKnownOne, Depth+1))
1134 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1135 LHSKnownZero, LHSKnownOne, Depth+1))
1141 case Instruction::Sub:
1142 // If the high-bits of this SUB are not demanded, then it does not demand
1143 // the high bits of its LHS or RHS.
1144 if (DemandedMask[BitWidth-1] == 0) {
1145 // Right fill the mask of bits for this SUB to demand the most
1146 // significant bit and all those below it.
1147 uint32_t NLZ = DemandedMask.countLeadingZeros();
1148 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1149 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1150 LHSKnownZero, LHSKnownOne, Depth+1))
1152 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1153 LHSKnownZero, LHSKnownOne, Depth+1))
1156 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1157 // the known zeros and ones.
1158 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1160 case Instruction::Shl:
1161 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1162 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1163 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1164 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1165 RHSKnownZero, RHSKnownOne, Depth+1))
1167 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1168 "Bits known to be one AND zero?");
1169 RHSKnownZero <<= ShiftAmt;
1170 RHSKnownOne <<= ShiftAmt;
1171 // low bits known zero.
1173 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1176 case Instruction::LShr:
1177 // For a logical shift right
1178 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1179 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1181 // Unsigned shift right.
1182 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1183 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1184 RHSKnownZero, RHSKnownOne, Depth+1))
1186 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1187 "Bits known to be one AND zero?");
1188 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1189 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1191 // Compute the new bits that are at the top now.
1192 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1193 RHSKnownZero |= HighBits; // high bits known zero.
1197 case Instruction::AShr:
1198 // If this is an arithmetic shift right and only the low-bit is set, we can
1199 // always convert this into a logical shr, even if the shift amount is
1200 // variable. The low bit of the shift cannot be an input sign bit unless
1201 // the shift amount is >= the size of the datatype, which is undefined.
1202 if (DemandedMask == 1) {
1203 // Perform the logical shift right.
1204 Value *NewVal = BinaryOperator::CreateLShr(
1205 I->getOperand(0), I->getOperand(1), I->getName());
1206 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1207 return UpdateValueUsesWith(I, NewVal);
1210 // If the sign bit is the only bit demanded by this ashr, then there is no
1211 // need to do it, the shift doesn't change the high bit.
1212 if (DemandedMask.isSignBit())
1213 return UpdateValueUsesWith(I, I->getOperand(0));
1215 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1216 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1218 // Signed shift right.
1219 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1220 // If any of the "high bits" are demanded, we should set the sign bit as
1222 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1223 DemandedMaskIn.set(BitWidth-1);
1224 if (SimplifyDemandedBits(I->getOperand(0),
1226 RHSKnownZero, RHSKnownOne, Depth+1))
1228 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1229 "Bits known to be one AND zero?");
1230 // Compute the new bits that are at the top now.
1231 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1232 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1233 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1235 // Handle the sign bits.
1236 APInt SignBit(APInt::getSignBit(BitWidth));
1237 // Adjust to where it is now in the mask.
1238 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1240 // If the input sign bit is known to be zero, or if none of the top bits
1241 // are demanded, turn this into an unsigned shift right.
1242 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1243 (HighBits & ~DemandedMask) == HighBits) {
1244 // Perform the logical shift right.
1245 Value *NewVal = BinaryOperator::CreateLShr(
1246 I->getOperand(0), SA, I->getName());
1247 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1248 return UpdateValueUsesWith(I, NewVal);
1249 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1250 RHSKnownOne |= HighBits;
1254 case Instruction::SRem:
1255 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1256 APInt RA = Rem->getValue();
1257 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1258 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1259 return UpdateValueUsesWith(I, I->getOperand(0));
1261 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1262 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1263 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1264 LHSKnownZero, LHSKnownOne, Depth+1))
1267 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1268 LHSKnownZero |= ~LowBits;
1269 else if (LHSKnownOne[BitWidth-1])
1270 LHSKnownOne |= ~LowBits;
1272 KnownZero |= LHSKnownZero & DemandedMask;
1273 KnownOne |= LHSKnownOne & DemandedMask;
1275 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1279 case Instruction::URem: {
1280 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1281 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1282 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1283 KnownZero2, KnownOne2, Depth+1))
1286 uint32_t Leaders = KnownZero2.countLeadingOnes();
1287 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1288 KnownZero2, KnownOne2, Depth+1))
1291 Leaders = std::max(Leaders,
1292 KnownZero2.countLeadingOnes());
1293 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1296 case Instruction::Call:
1297 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1298 switch (II->getIntrinsicID()) {
1300 case Intrinsic::bswap: {
1301 // If the only bits demanded come from one byte of the bswap result,
1302 // just shift the input byte into position to eliminate the bswap.
1303 unsigned NLZ = DemandedMask.countLeadingZeros();
1304 unsigned NTZ = DemandedMask.countTrailingZeros();
1306 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1307 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1308 // have 14 leading zeros, round to 8.
1311 // If we need exactly one byte, we can do this transformation.
1312 if (BitWidth-NLZ-NTZ == 8) {
1313 unsigned ResultBit = NTZ;
1314 unsigned InputBit = BitWidth-NTZ-8;
1316 // Replace this with either a left or right shift to get the byte into
1318 Instruction *NewVal;
1319 if (InputBit > ResultBit)
1320 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1321 ConstantInt::get(I->getType(), InputBit-ResultBit));
1323 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1324 ConstantInt::get(I->getType(), ResultBit-InputBit));
1325 NewVal->takeName(I);
1326 InsertNewInstBefore(NewVal, *I);
1327 return UpdateValueUsesWith(I, NewVal);
1330 // TODO: Could compute known zero/one bits based on the input.
1335 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1339 // If the client is only demanding bits that we know, return the known
1341 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1342 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1347 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1348 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1349 /// actually used by the caller. This method analyzes which elements of the
1350 /// operand are undef and returns that information in UndefElts.
1352 /// If the information about demanded elements can be used to simplify the
1353 /// operation, the operation is simplified, then the resultant value is
1354 /// returned. This returns null if no change was made.
1355 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1356 uint64_t &UndefElts,
1358 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1359 assert(VWidth <= 64 && "Vector too wide to analyze!");
1360 uint64_t EltMask = ~0ULL >> (64-VWidth);
1361 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1362 "Invalid DemandedElts!");
1364 if (isa<UndefValue>(V)) {
1365 // If the entire vector is undefined, just return this info.
1366 UndefElts = EltMask;
1368 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1369 UndefElts = EltMask;
1370 return UndefValue::get(V->getType());
1374 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1375 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1376 Constant *Undef = UndefValue::get(EltTy);
1378 std::vector<Constant*> Elts;
1379 for (unsigned i = 0; i != VWidth; ++i)
1380 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1381 Elts.push_back(Undef);
1382 UndefElts |= (1ULL << i);
1383 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1384 Elts.push_back(Undef);
1385 UndefElts |= (1ULL << i);
1386 } else { // Otherwise, defined.
1387 Elts.push_back(CP->getOperand(i));
1390 // If we changed the constant, return it.
1391 Constant *NewCP = ConstantVector::get(Elts);
1392 return NewCP != CP ? NewCP : 0;
1393 } else if (isa<ConstantAggregateZero>(V)) {
1394 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1396 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1397 Constant *Zero = Constant::getNullValue(EltTy);
1398 Constant *Undef = UndefValue::get(EltTy);
1399 std::vector<Constant*> Elts;
1400 for (unsigned i = 0; i != VWidth; ++i)
1401 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1402 UndefElts = DemandedElts ^ EltMask;
1403 return ConstantVector::get(Elts);
1406 if (!V->hasOneUse()) { // Other users may use these bits.
1407 if (Depth != 0) { // Not at the root.
1408 // TODO: Just compute the UndefElts information recursively.
1412 } else if (Depth == 10) { // Limit search depth.
1416 Instruction *I = dyn_cast<Instruction>(V);
1417 if (!I) return false; // Only analyze instructions.
1419 bool MadeChange = false;
1420 uint64_t UndefElts2;
1422 switch (I->getOpcode()) {
1425 case Instruction::InsertElement: {
1426 // If this is a variable index, we don't know which element it overwrites.
1427 // demand exactly the same input as we produce.
1428 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1430 // Note that we can't propagate undef elt info, because we don't know
1431 // which elt is getting updated.
1432 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1433 UndefElts2, Depth+1);
1434 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1438 // If this is inserting an element that isn't demanded, remove this
1440 unsigned IdxNo = Idx->getZExtValue();
1441 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1442 return AddSoonDeadInstToWorklist(*I, 0);
1444 // Otherwise, the element inserted overwrites whatever was there, so the
1445 // input demanded set is simpler than the output set.
1446 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1447 DemandedElts & ~(1ULL << IdxNo),
1448 UndefElts, Depth+1);
1449 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1451 // The inserted element is defined.
1452 UndefElts |= 1ULL << IdxNo;
1455 case Instruction::BitCast: {
1456 // Vector->vector casts only.
1457 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1459 unsigned InVWidth = VTy->getNumElements();
1460 uint64_t InputDemandedElts = 0;
1463 if (VWidth == InVWidth) {
1464 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1465 // elements as are demanded of us.
1467 InputDemandedElts = DemandedElts;
1468 } else if (VWidth > InVWidth) {
1472 // If there are more elements in the result than there are in the source,
1473 // then an input element is live if any of the corresponding output
1474 // elements are live.
1475 Ratio = VWidth/InVWidth;
1476 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1477 if (DemandedElts & (1ULL << OutIdx))
1478 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1484 // If there are more elements in the source than there are in the result,
1485 // then an input element is live if the corresponding output element is
1487 Ratio = InVWidth/VWidth;
1488 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1489 if (DemandedElts & (1ULL << InIdx/Ratio))
1490 InputDemandedElts |= 1ULL << InIdx;
1493 // div/rem demand all inputs, because they don't want divide by zero.
1494 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1495 UndefElts2, Depth+1);
1497 I->setOperand(0, TmpV);
1501 UndefElts = UndefElts2;
1502 if (VWidth > InVWidth) {
1503 assert(0 && "Unimp");
1504 // If there are more elements in the result than there are in the source,
1505 // then an output element is undef if the corresponding input element is
1507 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1508 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1509 UndefElts |= 1ULL << OutIdx;
1510 } else if (VWidth < InVWidth) {
1511 assert(0 && "Unimp");
1512 // If there are more elements in the source than there are in the result,
1513 // then a result element is undef if all of the corresponding input
1514 // elements are undef.
1515 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1516 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1517 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1518 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1522 case Instruction::And:
1523 case Instruction::Or:
1524 case Instruction::Xor:
1525 case Instruction::Add:
1526 case Instruction::Sub:
1527 case Instruction::Mul:
1528 // div/rem demand all inputs, because they don't want divide by zero.
1529 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1530 UndefElts, Depth+1);
1531 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1532 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1533 UndefElts2, Depth+1);
1534 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1536 // Output elements are undefined if both are undefined. Consider things
1537 // like undef&0. The result is known zero, not undef.
1538 UndefElts &= UndefElts2;
1541 case Instruction::Call: {
1542 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1544 switch (II->getIntrinsicID()) {
1547 // Binary vector operations that work column-wise. A dest element is a
1548 // function of the corresponding input elements from the two inputs.
1549 case Intrinsic::x86_sse_sub_ss:
1550 case Intrinsic::x86_sse_mul_ss:
1551 case Intrinsic::x86_sse_min_ss:
1552 case Intrinsic::x86_sse_max_ss:
1553 case Intrinsic::x86_sse2_sub_sd:
1554 case Intrinsic::x86_sse2_mul_sd:
1555 case Intrinsic::x86_sse2_min_sd:
1556 case Intrinsic::x86_sse2_max_sd:
1557 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1558 UndefElts, Depth+1);
1559 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1560 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1561 UndefElts2, Depth+1);
1562 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1564 // If only the low elt is demanded and this is a scalarizable intrinsic,
1565 // scalarize it now.
1566 if (DemandedElts == 1) {
1567 switch (II->getIntrinsicID()) {
1569 case Intrinsic::x86_sse_sub_ss:
1570 case Intrinsic::x86_sse_mul_ss:
1571 case Intrinsic::x86_sse2_sub_sd:
1572 case Intrinsic::x86_sse2_mul_sd:
1573 // TODO: Lower MIN/MAX/ABS/etc
1574 Value *LHS = II->getOperand(1);
1575 Value *RHS = II->getOperand(2);
1576 // Extract the element as scalars.
1577 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1578 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1580 switch (II->getIntrinsicID()) {
1581 default: assert(0 && "Case stmts out of sync!");
1582 case Intrinsic::x86_sse_sub_ss:
1583 case Intrinsic::x86_sse2_sub_sd:
1584 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1585 II->getName()), *II);
1587 case Intrinsic::x86_sse_mul_ss:
1588 case Intrinsic::x86_sse2_mul_sd:
1589 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1590 II->getName()), *II);
1595 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1597 InsertNewInstBefore(New, *II);
1598 AddSoonDeadInstToWorklist(*II, 0);
1603 // Output elements are undefined if both are undefined. Consider things
1604 // like undef&0. The result is known zero, not undef.
1605 UndefElts &= UndefElts2;
1611 return MadeChange ? I : 0;
1615 /// AssociativeOpt - Perform an optimization on an associative operator. This
1616 /// function is designed to check a chain of associative operators for a
1617 /// potential to apply a certain optimization. Since the optimization may be
1618 /// applicable if the expression was reassociated, this checks the chain, then
1619 /// reassociates the expression as necessary to expose the optimization
1620 /// opportunity. This makes use of a special Functor, which must define
1621 /// 'shouldApply' and 'apply' methods.
1623 template<typename Functor>
1624 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1625 unsigned Opcode = Root.getOpcode();
1626 Value *LHS = Root.getOperand(0);
1628 // Quick check, see if the immediate LHS matches...
1629 if (F.shouldApply(LHS))
1630 return F.apply(Root);
1632 // Otherwise, if the LHS is not of the same opcode as the root, return.
1633 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1634 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1635 // Should we apply this transform to the RHS?
1636 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1638 // If not to the RHS, check to see if we should apply to the LHS...
1639 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1640 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1644 // If the functor wants to apply the optimization to the RHS of LHSI,
1645 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1647 // Now all of the instructions are in the current basic block, go ahead
1648 // and perform the reassociation.
1649 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1651 // First move the selected RHS to the LHS of the root...
1652 Root.setOperand(0, LHSI->getOperand(1));
1654 // Make what used to be the LHS of the root be the user of the root...
1655 Value *ExtraOperand = TmpLHSI->getOperand(1);
1656 if (&Root == TmpLHSI) {
1657 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1660 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1661 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1662 BasicBlock::iterator ARI = &Root; ++ARI;
1663 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1666 // Now propagate the ExtraOperand down the chain of instructions until we
1668 while (TmpLHSI != LHSI) {
1669 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1670 // Move the instruction to immediately before the chain we are
1671 // constructing to avoid breaking dominance properties.
1672 NextLHSI->moveBefore(ARI);
1675 Value *NextOp = NextLHSI->getOperand(1);
1676 NextLHSI->setOperand(1, ExtraOperand);
1678 ExtraOperand = NextOp;
1681 // Now that the instructions are reassociated, have the functor perform
1682 // the transformation...
1683 return F.apply(Root);
1686 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1693 // AddRHS - Implements: X + X --> X << 1
1696 AddRHS(Value *rhs) : RHS(rhs) {}
1697 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1698 Instruction *apply(BinaryOperator &Add) const {
1699 return BinaryOperator::CreateShl(Add.getOperand(0),
1700 ConstantInt::get(Add.getType(), 1));
1704 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1706 struct AddMaskingAnd {
1708 AddMaskingAnd(Constant *c) : C2(c) {}
1709 bool shouldApply(Value *LHS) const {
1711 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1712 ConstantExpr::getAnd(C1, C2)->isNullValue();
1714 Instruction *apply(BinaryOperator &Add) const {
1715 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1721 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1723 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1724 if (Constant *SOC = dyn_cast<Constant>(SO))
1725 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1727 return IC->InsertNewInstBefore(CastInst::Create(
1728 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1731 // Figure out if the constant is the left or the right argument.
1732 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1733 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1735 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1737 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1738 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1741 Value *Op0 = SO, *Op1 = ConstOperand;
1743 std::swap(Op0, Op1);
1745 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1746 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1747 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1748 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1749 SO->getName()+".cmp");
1751 assert(0 && "Unknown binary instruction type!");
1754 return IC->InsertNewInstBefore(New, I);
1757 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1758 // constant as the other operand, try to fold the binary operator into the
1759 // select arguments. This also works for Cast instructions, which obviously do
1760 // not have a second operand.
1761 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1763 // Don't modify shared select instructions
1764 if (!SI->hasOneUse()) return 0;
1765 Value *TV = SI->getOperand(1);
1766 Value *FV = SI->getOperand(2);
1768 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1769 // Bool selects with constant operands can be folded to logical ops.
1770 if (SI->getType() == Type::Int1Ty) return 0;
1772 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1773 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1775 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1782 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1783 /// node as operand #0, see if we can fold the instruction into the PHI (which
1784 /// is only possible if all operands to the PHI are constants).
1785 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1786 PHINode *PN = cast<PHINode>(I.getOperand(0));
1787 unsigned NumPHIValues = PN->getNumIncomingValues();
1788 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1790 // Check to see if all of the operands of the PHI are constants. If there is
1791 // one non-constant value, remember the BB it is. If there is more than one
1792 // or if *it* is a PHI, bail out.
1793 BasicBlock *NonConstBB = 0;
1794 for (unsigned i = 0; i != NumPHIValues; ++i)
1795 if (!isa<Constant>(PN->getIncomingValue(i))) {
1796 if (NonConstBB) return 0; // More than one non-const value.
1797 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1798 NonConstBB = PN->getIncomingBlock(i);
1800 // If the incoming non-constant value is in I's block, we have an infinite
1802 if (NonConstBB == I.getParent())
1806 // If there is exactly one non-constant value, we can insert a copy of the
1807 // operation in that block. However, if this is a critical edge, we would be
1808 // inserting the computation one some other paths (e.g. inside a loop). Only
1809 // do this if the pred block is unconditionally branching into the phi block.
1811 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1812 if (!BI || !BI->isUnconditional()) return 0;
1815 // Okay, we can do the transformation: create the new PHI node.
1816 PHINode *NewPN = PHINode::Create(I.getType(), "");
1817 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1818 InsertNewInstBefore(NewPN, *PN);
1819 NewPN->takeName(PN);
1821 // Next, add all of the operands to the PHI.
1822 if (I.getNumOperands() == 2) {
1823 Constant *C = cast<Constant>(I.getOperand(1));
1824 for (unsigned i = 0; i != NumPHIValues; ++i) {
1826 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1827 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1828 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1830 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1832 assert(PN->getIncomingBlock(i) == NonConstBB);
1833 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1834 InV = BinaryOperator::Create(BO->getOpcode(),
1835 PN->getIncomingValue(i), C, "phitmp",
1836 NonConstBB->getTerminator());
1837 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1838 InV = CmpInst::Create(CI->getOpcode(),
1840 PN->getIncomingValue(i), C, "phitmp",
1841 NonConstBB->getTerminator());
1843 assert(0 && "Unknown binop!");
1845 AddToWorkList(cast<Instruction>(InV));
1847 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1850 CastInst *CI = cast<CastInst>(&I);
1851 const Type *RetTy = CI->getType();
1852 for (unsigned i = 0; i != NumPHIValues; ++i) {
1854 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1855 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1857 assert(PN->getIncomingBlock(i) == NonConstBB);
1858 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1859 I.getType(), "phitmp",
1860 NonConstBB->getTerminator());
1861 AddToWorkList(cast<Instruction>(InV));
1863 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1866 return ReplaceInstUsesWith(I, NewPN);
1870 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1871 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1872 /// This basically requires proving that the add in the original type would not
1873 /// overflow to change the sign bit or have a carry out.
1874 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1875 // There are different heuristics we can use for this. Here are some simple
1878 // Add has the property that adding any two 2's complement numbers can only
1879 // have one carry bit which can change a sign. As such, if LHS and RHS each
1880 // have at least two sign bits, we know that the addition of the two values will
1881 // sign extend fine.
1882 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1886 // If one of the operands only has one non-zero bit, and if the other operand
1887 // has a known-zero bit in a more significant place than it (not including the
1888 // sign bit) the ripple may go up to and fill the zero, but won't change the
1889 // sign. For example, (X & ~4) + 1.
1897 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1898 bool Changed = SimplifyCommutative(I);
1899 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1901 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1902 // X + undef -> undef
1903 if (isa<UndefValue>(RHS))
1904 return ReplaceInstUsesWith(I, RHS);
1907 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1908 if (RHSC->isNullValue())
1909 return ReplaceInstUsesWith(I, LHS);
1910 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1911 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1912 (I.getType())->getValueAPF()))
1913 return ReplaceInstUsesWith(I, LHS);
1916 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1917 // X + (signbit) --> X ^ signbit
1918 const APInt& Val = CI->getValue();
1919 uint32_t BitWidth = Val.getBitWidth();
1920 if (Val == APInt::getSignBit(BitWidth))
1921 return BinaryOperator::CreateXor(LHS, RHS);
1923 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1924 // (X & 254)+1 -> (X&254)|1
1925 if (!isa<VectorType>(I.getType())) {
1926 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1927 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1928 KnownZero, KnownOne))
1933 if (isa<PHINode>(LHS))
1934 if (Instruction *NV = FoldOpIntoPhi(I))
1937 ConstantInt *XorRHS = 0;
1939 if (isa<ConstantInt>(RHSC) &&
1940 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1941 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1942 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1944 uint32_t Size = TySizeBits / 2;
1945 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1946 APInt CFF80Val(-C0080Val);
1948 if (TySizeBits > Size) {
1949 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1950 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1951 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1952 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1953 // This is a sign extend if the top bits are known zero.
1954 if (!MaskedValueIsZero(XorLHS,
1955 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1956 Size = 0; // Not a sign ext, but can't be any others either.
1961 C0080Val = APIntOps::lshr(C0080Val, Size);
1962 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1963 } while (Size >= 1);
1965 // FIXME: This shouldn't be necessary. When the backends can handle types
1966 // with funny bit widths then this switch statement should be removed. It
1967 // is just here to get the size of the "middle" type back up to something
1968 // that the back ends can handle.
1969 const Type *MiddleType = 0;
1972 case 32: MiddleType = Type::Int32Ty; break;
1973 case 16: MiddleType = Type::Int16Ty; break;
1974 case 8: MiddleType = Type::Int8Ty; break;
1977 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1978 InsertNewInstBefore(NewTrunc, I);
1979 return new SExtInst(NewTrunc, I.getType(), I.getName());
1984 if (I.getType() == Type::Int1Ty)
1985 return BinaryOperator::CreateXor(LHS, RHS);
1988 if (I.getType()->isInteger()) {
1989 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1991 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1992 if (RHSI->getOpcode() == Instruction::Sub)
1993 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1994 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1996 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1997 if (LHSI->getOpcode() == Instruction::Sub)
1998 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1999 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2004 // -A + -B --> -(A + B)
2005 if (Value *LHSV = dyn_castNegVal(LHS)) {
2006 if (LHS->getType()->isIntOrIntVector()) {
2007 if (Value *RHSV = dyn_castNegVal(RHS)) {
2008 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2009 InsertNewInstBefore(NewAdd, I);
2010 return BinaryOperator::CreateNeg(NewAdd);
2014 return BinaryOperator::CreateSub(RHS, LHSV);
2018 if (!isa<Constant>(RHS))
2019 if (Value *V = dyn_castNegVal(RHS))
2020 return BinaryOperator::CreateSub(LHS, V);
2024 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2025 if (X == RHS) // X*C + X --> X * (C+1)
2026 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2028 // X*C1 + X*C2 --> X * (C1+C2)
2030 if (X == dyn_castFoldableMul(RHS, C1))
2031 return BinaryOperator::CreateMul(X, Add(C1, C2));
2034 // X + X*C --> X * (C+1)
2035 if (dyn_castFoldableMul(RHS, C2) == LHS)
2036 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2038 // X + ~X --> -1 since ~X = -X-1
2039 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2040 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2043 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2044 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2045 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2048 // A+B --> A|B iff A and B have no bits set in common.
2049 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2050 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2051 APInt LHSKnownOne(IT->getBitWidth(), 0);
2052 APInt LHSKnownZero(IT->getBitWidth(), 0);
2053 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2054 if (LHSKnownZero != 0) {
2055 APInt RHSKnownOne(IT->getBitWidth(), 0);
2056 APInt RHSKnownZero(IT->getBitWidth(), 0);
2057 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2059 // No bits in common -> bitwise or.
2060 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2061 return BinaryOperator::CreateOr(LHS, RHS);
2065 // W*X + Y*Z --> W * (X+Z) iff W == Y
2066 if (I.getType()->isIntOrIntVector()) {
2067 Value *W, *X, *Y, *Z;
2068 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2069 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2073 } else if (Y == X) {
2075 } else if (X == Z) {
2082 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2083 LHS->getName()), I);
2084 return BinaryOperator::CreateMul(W, NewAdd);
2089 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2091 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2092 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2094 // (X & FF00) + xx00 -> (X+xx00) & FF00
2095 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2096 Constant *Anded = And(CRHS, C2);
2097 if (Anded == CRHS) {
2098 // See if all bits from the first bit set in the Add RHS up are included
2099 // in the mask. First, get the rightmost bit.
2100 const APInt& AddRHSV = CRHS->getValue();
2102 // Form a mask of all bits from the lowest bit added through the top.
2103 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2105 // See if the and mask includes all of these bits.
2106 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2108 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2109 // Okay, the xform is safe. Insert the new add pronto.
2110 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2111 LHS->getName()), I);
2112 return BinaryOperator::CreateAnd(NewAdd, C2);
2117 // Try to fold constant add into select arguments.
2118 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2119 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2123 // add (cast *A to intptrtype) B ->
2124 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2126 CastInst *CI = dyn_cast<CastInst>(LHS);
2129 CI = dyn_cast<CastInst>(RHS);
2132 if (CI && CI->getType()->isSized() &&
2133 (CI->getType()->getPrimitiveSizeInBits() ==
2134 TD->getIntPtrType()->getPrimitiveSizeInBits())
2135 && isa<PointerType>(CI->getOperand(0)->getType())) {
2137 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2138 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2139 PointerType::get(Type::Int8Ty, AS), I);
2140 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2141 return new PtrToIntInst(I2, CI->getType());
2145 // add (select X 0 (sub n A)) A --> select X A n
2147 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2150 SI = dyn_cast<SelectInst>(RHS);
2153 if (SI && SI->hasOneUse()) {
2154 Value *TV = SI->getTrueValue();
2155 Value *FV = SI->getFalseValue();
2158 // Can we fold the add into the argument of the select?
2159 // We check both true and false select arguments for a matching subtract.
2160 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2161 A == Other) // Fold the add into the true select value.
2162 return SelectInst::Create(SI->getCondition(), N, A);
2163 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2164 A == Other) // Fold the add into the false select value.
2165 return SelectInst::Create(SI->getCondition(), A, N);
2169 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2170 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2171 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2172 return ReplaceInstUsesWith(I, LHS);
2174 // Check for (add (sext x), y), see if we can merge this into an
2175 // integer add followed by a sext.
2176 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2177 // (add (sext x), cst) --> (sext (add x, cst'))
2178 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2180 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2181 if (LHSConv->hasOneUse() &&
2182 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2183 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2184 // Insert the new, smaller add.
2185 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2187 InsertNewInstBefore(NewAdd, I);
2188 return new SExtInst(NewAdd, I.getType());
2192 // (add (sext x), (sext y)) --> (sext (add int x, y))
2193 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2194 // Only do this if x/y have the same type, if at last one of them has a
2195 // single use (so we don't increase the number of sexts), and if the
2196 // integer add will not overflow.
2197 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2198 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2199 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2200 RHSConv->getOperand(0))) {
2201 // Insert the new integer add.
2202 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2203 RHSConv->getOperand(0),
2205 InsertNewInstBefore(NewAdd, I);
2206 return new SExtInst(NewAdd, I.getType());
2211 // Check for (add double (sitofp x), y), see if we can merge this into an
2212 // integer add followed by a promotion.
2213 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2214 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2215 // ... if the constant fits in the integer value. This is useful for things
2216 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2217 // requires a constant pool load, and generally allows the add to be better
2219 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2221 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2222 if (LHSConv->hasOneUse() &&
2223 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2224 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2225 // Insert the new integer add.
2226 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2228 InsertNewInstBefore(NewAdd, I);
2229 return new SIToFPInst(NewAdd, I.getType());
2233 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2234 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2235 // Only do this if x/y have the same type, if at last one of them has a
2236 // single use (so we don't increase the number of int->fp conversions),
2237 // and if the integer add will not overflow.
2238 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2239 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2240 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2241 RHSConv->getOperand(0))) {
2242 // Insert the new integer add.
2243 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2244 RHSConv->getOperand(0),
2246 InsertNewInstBefore(NewAdd, I);
2247 return new SIToFPInst(NewAdd, I.getType());
2252 return Changed ? &I : 0;
2255 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2256 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2258 if (Op0 == Op1 && // sub X, X -> 0
2259 !I.getType()->isFPOrFPVector())
2260 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2262 // If this is a 'B = x-(-A)', change to B = x+A...
2263 if (Value *V = dyn_castNegVal(Op1))
2264 return BinaryOperator::CreateAdd(Op0, V);
2266 if (isa<UndefValue>(Op0))
2267 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2268 if (isa<UndefValue>(Op1))
2269 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2271 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2272 // Replace (-1 - A) with (~A)...
2273 if (C->isAllOnesValue())
2274 return BinaryOperator::CreateNot(Op1);
2276 // C - ~X == X + (1+C)
2278 if (match(Op1, m_Not(m_Value(X))))
2279 return BinaryOperator::CreateAdd(X, AddOne(C));
2281 // -(X >>u 31) -> (X >>s 31)
2282 // -(X >>s 31) -> (X >>u 31)
2284 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2285 if (SI->getOpcode() == Instruction::LShr) {
2286 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2287 // Check to see if we are shifting out everything but the sign bit.
2288 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2289 SI->getType()->getPrimitiveSizeInBits()-1) {
2290 // Ok, the transformation is safe. Insert AShr.
2291 return BinaryOperator::Create(Instruction::AShr,
2292 SI->getOperand(0), CU, SI->getName());
2296 else if (SI->getOpcode() == Instruction::AShr) {
2297 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2298 // Check to see if we are shifting out everything but the sign bit.
2299 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2300 SI->getType()->getPrimitiveSizeInBits()-1) {
2301 // Ok, the transformation is safe. Insert LShr.
2302 return BinaryOperator::CreateLShr(
2303 SI->getOperand(0), CU, SI->getName());
2310 // Try to fold constant sub into select arguments.
2311 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2312 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2315 if (isa<PHINode>(Op0))
2316 if (Instruction *NV = FoldOpIntoPhi(I))
2320 if (I.getType() == Type::Int1Ty)
2321 return BinaryOperator::CreateXor(Op0, Op1);
2323 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2324 if (Op1I->getOpcode() == Instruction::Add &&
2325 !Op0->getType()->isFPOrFPVector()) {
2326 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2327 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2328 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2329 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2330 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2331 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2332 // C1-(X+C2) --> (C1-C2)-X
2333 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2334 Op1I->getOperand(0));
2338 if (Op1I->hasOneUse()) {
2339 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2340 // is not used by anyone else...
2342 if (Op1I->getOpcode() == Instruction::Sub &&
2343 !Op1I->getType()->isFPOrFPVector()) {
2344 // Swap the two operands of the subexpr...
2345 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2346 Op1I->setOperand(0, IIOp1);
2347 Op1I->setOperand(1, IIOp0);
2349 // Create the new top level add instruction...
2350 return BinaryOperator::CreateAdd(Op0, Op1);
2353 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2355 if (Op1I->getOpcode() == Instruction::And &&
2356 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2357 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2360 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2361 return BinaryOperator::CreateAnd(Op0, NewNot);
2364 // 0 - (X sdiv C) -> (X sdiv -C)
2365 if (Op1I->getOpcode() == Instruction::SDiv)
2366 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2368 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2369 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2370 ConstantExpr::getNeg(DivRHS));
2372 // X - X*C --> X * (1-C)
2373 ConstantInt *C2 = 0;
2374 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2375 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2376 return BinaryOperator::CreateMul(Op0, CP1);
2379 // X - ((X / Y) * Y) --> X % Y
2380 if (Op1I->getOpcode() == Instruction::Mul)
2381 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2382 if (Op0 == I->getOperand(0) &&
2383 Op1I->getOperand(1) == I->getOperand(1)) {
2384 if (I->getOpcode() == Instruction::SDiv)
2385 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2386 if (I->getOpcode() == Instruction::UDiv)
2387 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2392 if (!Op0->getType()->isFPOrFPVector())
2393 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2394 if (Op0I->getOpcode() == Instruction::Add) {
2395 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2396 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2397 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2398 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2399 } else if (Op0I->getOpcode() == Instruction::Sub) {
2400 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2401 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2406 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2407 if (X == Op1) // X*C - X --> X * (C-1)
2408 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2410 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2411 if (X == dyn_castFoldableMul(Op1, C2))
2412 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2417 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2418 /// comparison only checks the sign bit. If it only checks the sign bit, set
2419 /// TrueIfSigned if the result of the comparison is true when the input value is
2421 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2422 bool &TrueIfSigned) {
2424 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2425 TrueIfSigned = true;
2426 return RHS->isZero();
2427 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2428 TrueIfSigned = true;
2429 return RHS->isAllOnesValue();
2430 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2431 TrueIfSigned = false;
2432 return RHS->isAllOnesValue();
2433 case ICmpInst::ICMP_UGT:
2434 // True if LHS u> RHS and RHS == high-bit-mask - 1
2435 TrueIfSigned = true;
2436 return RHS->getValue() ==
2437 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2438 case ICmpInst::ICMP_UGE:
2439 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2440 TrueIfSigned = true;
2441 return RHS->getValue().isSignBit();
2447 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2448 bool Changed = SimplifyCommutative(I);
2449 Value *Op0 = I.getOperand(0);
2451 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2452 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2454 // Simplify mul instructions with a constant RHS...
2455 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2456 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2458 // ((X << C1)*C2) == (X * (C2 << C1))
2459 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2460 if (SI->getOpcode() == Instruction::Shl)
2461 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2462 return BinaryOperator::CreateMul(SI->getOperand(0),
2463 ConstantExpr::getShl(CI, ShOp));
2466 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2467 if (CI->equalsInt(1)) // X * 1 == X
2468 return ReplaceInstUsesWith(I, Op0);
2469 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2470 return BinaryOperator::CreateNeg(Op0, I.getName());
2472 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2473 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2474 return BinaryOperator::CreateShl(Op0,
2475 ConstantInt::get(Op0->getType(), Val.logBase2()));
2477 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2478 if (Op1F->isNullValue())
2479 return ReplaceInstUsesWith(I, Op1);
2481 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2482 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2483 // We need a better interface for long double here.
2484 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2485 if (Op1F->isExactlyValue(1.0))
2486 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2489 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2490 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2491 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2492 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2493 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2495 InsertNewInstBefore(Add, I);
2496 Value *C1C2 = ConstantExpr::getMul(Op1,
2497 cast<Constant>(Op0I->getOperand(1)));
2498 return BinaryOperator::CreateAdd(Add, C1C2);
2502 // Try to fold constant mul into select arguments.
2503 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2504 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2507 if (isa<PHINode>(Op0))
2508 if (Instruction *NV = FoldOpIntoPhi(I))
2512 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2513 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2514 return BinaryOperator::CreateMul(Op0v, Op1v);
2516 if (I.getType() == Type::Int1Ty)
2517 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2519 // If one of the operands of the multiply is a cast from a boolean value, then
2520 // we know the bool is either zero or one, so this is a 'masking' multiply.
2521 // See if we can simplify things based on how the boolean was originally
2523 CastInst *BoolCast = 0;
2524 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2525 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2528 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2529 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2532 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2533 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2534 const Type *SCOpTy = SCIOp0->getType();
2537 // If the icmp is true iff the sign bit of X is set, then convert this
2538 // multiply into a shift/and combination.
2539 if (isa<ConstantInt>(SCIOp1) &&
2540 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2542 // Shift the X value right to turn it into "all signbits".
2543 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2544 SCOpTy->getPrimitiveSizeInBits()-1);
2546 InsertNewInstBefore(
2547 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2548 BoolCast->getOperand(0)->getName()+
2551 // If the multiply type is not the same as the source type, sign extend
2552 // or truncate to the multiply type.
2553 if (I.getType() != V->getType()) {
2554 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2555 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2556 Instruction::CastOps opcode =
2557 (SrcBits == DstBits ? Instruction::BitCast :
2558 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2559 V = InsertCastBefore(opcode, V, I.getType(), I);
2562 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2563 return BinaryOperator::CreateAnd(V, OtherOp);
2568 return Changed ? &I : 0;
2571 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2573 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2574 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2576 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2577 int NonNullOperand = -1;
2578 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2579 if (ST->isNullValue())
2581 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2582 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2583 if (ST->isNullValue())
2586 if (NonNullOperand == -1)
2589 Value *SelectCond = SI->getOperand(0);
2591 // Change the div/rem to use 'Y' instead of the select.
2592 I.setOperand(1, SI->getOperand(NonNullOperand));
2594 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2595 // problem. However, the select, or the condition of the select may have
2596 // multiple uses. Based on our knowledge that the operand must be non-zero,
2597 // propagate the known value for the select into other uses of it, and
2598 // propagate a known value of the condition into its other users.
2600 // If the select and condition only have a single use, don't bother with this,
2602 if (SI->use_empty() && SelectCond->hasOneUse())
2605 // Scan the current block backward, looking for other uses of SI.
2606 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2608 while (BBI != BBFront) {
2610 // If we found a call to a function, we can't assume it will return, so
2611 // information from below it cannot be propagated above it.
2612 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2615 // Replace uses of the select or its condition with the known values.
2616 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2619 *I = SI->getOperand(NonNullOperand);
2621 } else if (*I == SelectCond) {
2622 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2623 ConstantInt::getFalse();
2628 // If we past the instruction, quit looking for it.
2631 if (&*BBI == SelectCond)
2634 // If we ran out of things to eliminate, break out of the loop.
2635 if (SelectCond == 0 && SI == 0)
2643 /// This function implements the transforms on div instructions that work
2644 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2645 /// used by the visitors to those instructions.
2646 /// @brief Transforms common to all three div instructions
2647 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2648 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2650 // undef / X -> 0 for integer.
2651 // undef / X -> undef for FP (the undef could be a snan).
2652 if (isa<UndefValue>(Op0)) {
2653 if (Op0->getType()->isFPOrFPVector())
2654 return ReplaceInstUsesWith(I, Op0);
2655 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2658 // X / undef -> undef
2659 if (isa<UndefValue>(Op1))
2660 return ReplaceInstUsesWith(I, Op1);
2665 /// This function implements the transforms common to both integer division
2666 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2667 /// division instructions.
2668 /// @brief Common integer divide transforms
2669 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2670 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2672 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2674 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2675 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2676 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2677 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2680 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2681 return ReplaceInstUsesWith(I, CI);
2684 if (Instruction *Common = commonDivTransforms(I))
2687 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2688 // This does not apply for fdiv.
2689 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2692 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2694 if (RHS->equalsInt(1))
2695 return ReplaceInstUsesWith(I, Op0);
2697 // (X / C1) / C2 -> X / (C1*C2)
2698 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2699 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2700 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2701 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2702 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2704 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2705 Multiply(RHS, LHSRHS));
2708 if (!RHS->isZero()) { // avoid X udiv 0
2709 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2710 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2712 if (isa<PHINode>(Op0))
2713 if (Instruction *NV = FoldOpIntoPhi(I))
2718 // 0 / X == 0, we don't need to preserve faults!
2719 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2720 if (LHS->equalsInt(0))
2721 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2723 // It can't be division by zero, hence it must be division by one.
2724 if (I.getType() == Type::Int1Ty)
2725 return ReplaceInstUsesWith(I, Op0);
2730 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2731 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2733 // Handle the integer div common cases
2734 if (Instruction *Common = commonIDivTransforms(I))
2737 // X udiv C^2 -> X >> C
2738 // Check to see if this is an unsigned division with an exact power of 2,
2739 // if so, convert to a right shift.
2740 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2741 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2742 return BinaryOperator::CreateLShr(Op0,
2743 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2746 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2747 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2748 if (RHSI->getOpcode() == Instruction::Shl &&
2749 isa<ConstantInt>(RHSI->getOperand(0))) {
2750 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2751 if (C1.isPowerOf2()) {
2752 Value *N = RHSI->getOperand(1);
2753 const Type *NTy = N->getType();
2754 if (uint32_t C2 = C1.logBase2()) {
2755 Constant *C2V = ConstantInt::get(NTy, C2);
2756 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2758 return BinaryOperator::CreateLShr(Op0, N);
2763 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2764 // where C1&C2 are powers of two.
2765 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2766 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2767 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2768 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2769 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2770 // Compute the shift amounts
2771 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2772 // Construct the "on true" case of the select
2773 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2774 Instruction *TSI = BinaryOperator::CreateLShr(
2775 Op0, TC, SI->getName()+".t");
2776 TSI = InsertNewInstBefore(TSI, I);
2778 // Construct the "on false" case of the select
2779 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2780 Instruction *FSI = BinaryOperator::CreateLShr(
2781 Op0, FC, SI->getName()+".f");
2782 FSI = InsertNewInstBefore(FSI, I);
2784 // construct the select instruction and return it.
2785 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2791 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2792 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2794 // Handle the integer div common cases
2795 if (Instruction *Common = commonIDivTransforms(I))
2798 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2800 if (RHS->isAllOnesValue())
2801 return BinaryOperator::CreateNeg(Op0);
2804 if (Value *LHSNeg = dyn_castNegVal(Op0))
2805 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2808 // If the sign bits of both operands are zero (i.e. we can prove they are
2809 // unsigned inputs), turn this into a udiv.
2810 if (I.getType()->isInteger()) {
2811 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2812 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2813 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2814 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2821 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2822 return commonDivTransforms(I);
2825 /// This function implements the transforms on rem instructions that work
2826 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2827 /// is used by the visitors to those instructions.
2828 /// @brief Transforms common to all three rem instructions
2829 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2830 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2832 // 0 % X == 0 for integer, we don't need to preserve faults!
2833 if (Constant *LHS = dyn_cast<Constant>(Op0))
2834 if (LHS->isNullValue())
2835 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2837 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2838 if (I.getType()->isFPOrFPVector())
2839 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2840 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2842 if (isa<UndefValue>(Op1))
2843 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2845 // Handle cases involving: rem X, (select Cond, Y, Z)
2846 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2852 /// This function implements the transforms common to both integer remainder
2853 /// instructions (urem and srem). It is called by the visitors to those integer
2854 /// remainder instructions.
2855 /// @brief Common integer remainder transforms
2856 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2857 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2859 if (Instruction *common = commonRemTransforms(I))
2862 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2863 // X % 0 == undef, we don't need to preserve faults!
2864 if (RHS->equalsInt(0))
2865 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2867 if (RHS->equalsInt(1)) // X % 1 == 0
2868 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2870 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2871 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2872 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2874 } else if (isa<PHINode>(Op0I)) {
2875 if (Instruction *NV = FoldOpIntoPhi(I))
2879 // See if we can fold away this rem instruction.
2880 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2881 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2882 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2883 KnownZero, KnownOne))
2891 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2892 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2894 if (Instruction *common = commonIRemTransforms(I))
2897 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2898 // X urem C^2 -> X and C
2899 // Check to see if this is an unsigned remainder with an exact power of 2,
2900 // if so, convert to a bitwise and.
2901 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2902 if (C->getValue().isPowerOf2())
2903 return BinaryOperator::CreateAnd(Op0, SubOne(C));
2906 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2907 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2908 if (RHSI->getOpcode() == Instruction::Shl &&
2909 isa<ConstantInt>(RHSI->getOperand(0))) {
2910 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2911 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2912 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
2914 return BinaryOperator::CreateAnd(Op0, Add);
2919 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2920 // where C1&C2 are powers of two.
2921 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2922 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2923 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2924 // STO == 0 and SFO == 0 handled above.
2925 if ((STO->getValue().isPowerOf2()) &&
2926 (SFO->getValue().isPowerOf2())) {
2927 Value *TrueAnd = InsertNewInstBefore(
2928 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2929 Value *FalseAnd = InsertNewInstBefore(
2930 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2931 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
2939 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2940 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2942 // Handle the integer rem common cases
2943 if (Instruction *common = commonIRemTransforms(I))
2946 if (Value *RHSNeg = dyn_castNegVal(Op1))
2947 if (!isa<ConstantInt>(RHSNeg) ||
2948 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2950 AddUsesToWorkList(I);
2951 I.setOperand(1, RHSNeg);
2955 // If the sign bits of both operands are zero (i.e. we can prove they are
2956 // unsigned inputs), turn this into a urem.
2957 if (I.getType()->isInteger()) {
2958 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2959 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2960 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2961 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
2968 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2969 return commonRemTransforms(I);
2972 // isOneBitSet - Return true if there is exactly one bit set in the specified
2974 static bool isOneBitSet(const ConstantInt *CI) {
2975 return CI->getValue().isPowerOf2();
2978 // isHighOnes - Return true if the constant is of the form 1+0+.
2979 // This is the same as lowones(~X).
2980 static bool isHighOnes(const ConstantInt *CI) {
2981 return (~CI->getValue() + 1).isPowerOf2();
2984 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2985 /// are carefully arranged to allow folding of expressions such as:
2987 /// (A < B) | (A > B) --> (A != B)
2989 /// Note that this is only valid if the first and second predicates have the
2990 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2992 /// Three bits are used to represent the condition, as follows:
2997 /// <=> Value Definition
2998 /// 000 0 Always false
3005 /// 111 7 Always true
3007 static unsigned getICmpCode(const ICmpInst *ICI) {
3008 switch (ICI->getPredicate()) {
3010 case ICmpInst::ICMP_UGT: return 1; // 001
3011 case ICmpInst::ICMP_SGT: return 1; // 001
3012 case ICmpInst::ICMP_EQ: return 2; // 010
3013 case ICmpInst::ICMP_UGE: return 3; // 011
3014 case ICmpInst::ICMP_SGE: return 3; // 011
3015 case ICmpInst::ICMP_ULT: return 4; // 100
3016 case ICmpInst::ICMP_SLT: return 4; // 100
3017 case ICmpInst::ICMP_NE: return 5; // 101
3018 case ICmpInst::ICMP_ULE: return 6; // 110
3019 case ICmpInst::ICMP_SLE: return 6; // 110
3022 assert(0 && "Invalid ICmp predicate!");
3027 /// getICmpValue - This is the complement of getICmpCode, which turns an
3028 /// opcode and two operands into either a constant true or false, or a brand
3029 /// new ICmp instruction. The sign is passed in to determine which kind
3030 /// of predicate to use in new icmp instructions.
3031 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3033 default: assert(0 && "Illegal ICmp code!");
3034 case 0: return ConstantInt::getFalse();
3037 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3039 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3040 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3043 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3045 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3048 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3050 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3051 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3054 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3056 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3057 case 7: return ConstantInt::getTrue();
3061 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3062 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3063 (ICmpInst::isSignedPredicate(p1) &&
3064 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3065 (ICmpInst::isSignedPredicate(p2) &&
3066 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3070 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3071 struct FoldICmpLogical {
3074 ICmpInst::Predicate pred;
3075 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3076 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3077 pred(ICI->getPredicate()) {}
3078 bool shouldApply(Value *V) const {
3079 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3080 if (PredicatesFoldable(pred, ICI->getPredicate()))
3081 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3082 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3085 Instruction *apply(Instruction &Log) const {
3086 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3087 if (ICI->getOperand(0) != LHS) {
3088 assert(ICI->getOperand(1) == LHS);
3089 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3092 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3093 unsigned LHSCode = getICmpCode(ICI);
3094 unsigned RHSCode = getICmpCode(RHSICI);
3096 switch (Log.getOpcode()) {
3097 case Instruction::And: Code = LHSCode & RHSCode; break;
3098 case Instruction::Or: Code = LHSCode | RHSCode; break;
3099 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3100 default: assert(0 && "Illegal logical opcode!"); return 0;
3103 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3104 ICmpInst::isSignedPredicate(ICI->getPredicate());
3106 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3107 if (Instruction *I = dyn_cast<Instruction>(RV))
3109 // Otherwise, it's a constant boolean value...
3110 return IC.ReplaceInstUsesWith(Log, RV);
3113 } // end anonymous namespace
3115 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3116 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3117 // guaranteed to be a binary operator.
3118 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3120 ConstantInt *AndRHS,
3121 BinaryOperator &TheAnd) {
3122 Value *X = Op->getOperand(0);
3123 Constant *Together = 0;
3125 Together = And(AndRHS, OpRHS);
3127 switch (Op->getOpcode()) {
3128 case Instruction::Xor:
3129 if (Op->hasOneUse()) {
3130 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3131 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3132 InsertNewInstBefore(And, TheAnd);
3134 return BinaryOperator::CreateXor(And, Together);
3137 case Instruction::Or:
3138 if (Together == AndRHS) // (X | C) & C --> C
3139 return ReplaceInstUsesWith(TheAnd, AndRHS);
3141 if (Op->hasOneUse() && Together != OpRHS) {
3142 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3143 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3144 InsertNewInstBefore(Or, TheAnd);
3146 return BinaryOperator::CreateAnd(Or, AndRHS);
3149 case Instruction::Add:
3150 if (Op->hasOneUse()) {
3151 // Adding a one to a single bit bit-field should be turned into an XOR
3152 // of the bit. First thing to check is to see if this AND is with a
3153 // single bit constant.
3154 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3156 // If there is only one bit set...
3157 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3158 // Ok, at this point, we know that we are masking the result of the
3159 // ADD down to exactly one bit. If the constant we are adding has
3160 // no bits set below this bit, then we can eliminate the ADD.
3161 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3163 // Check to see if any bits below the one bit set in AndRHSV are set.
3164 if ((AddRHS & (AndRHSV-1)) == 0) {
3165 // If not, the only thing that can effect the output of the AND is
3166 // the bit specified by AndRHSV. If that bit is set, the effect of
3167 // the XOR is to toggle the bit. If it is clear, then the ADD has
3169 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3170 TheAnd.setOperand(0, X);
3173 // Pull the XOR out of the AND.
3174 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3175 InsertNewInstBefore(NewAnd, TheAnd);
3176 NewAnd->takeName(Op);
3177 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3184 case Instruction::Shl: {
3185 // We know that the AND will not produce any of the bits shifted in, so if
3186 // the anded constant includes them, clear them now!
3188 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3189 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3190 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3191 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3193 if (CI->getValue() == ShlMask) {
3194 // Masking out bits that the shift already masks
3195 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3196 } else if (CI != AndRHS) { // Reducing bits set in and.
3197 TheAnd.setOperand(1, CI);
3202 case Instruction::LShr:
3204 // We know that the AND will not produce any of the bits shifted in, so if
3205 // the anded constant includes them, clear them now! This only applies to
3206 // unsigned shifts, because a signed shr may bring in set bits!
3208 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3209 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3210 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3211 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3213 if (CI->getValue() == ShrMask) {
3214 // Masking out bits that the shift already masks.
3215 return ReplaceInstUsesWith(TheAnd, Op);
3216 } else if (CI != AndRHS) {
3217 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3222 case Instruction::AShr:
3224 // See if this is shifting in some sign extension, then masking it out
3226 if (Op->hasOneUse()) {
3227 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3228 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3229 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3230 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3231 if (C == AndRHS) { // Masking out bits shifted in.
3232 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3233 // Make the argument unsigned.
3234 Value *ShVal = Op->getOperand(0);
3235 ShVal = InsertNewInstBefore(
3236 BinaryOperator::CreateLShr(ShVal, OpRHS,
3237 Op->getName()), TheAnd);
3238 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3247 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3248 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3249 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3250 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3251 /// insert new instructions.
3252 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3253 bool isSigned, bool Inside,
3255 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3256 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3257 "Lo is not <= Hi in range emission code!");
3260 if (Lo == Hi) // Trivially false.
3261 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3263 // V >= Min && V < Hi --> V < Hi
3264 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3265 ICmpInst::Predicate pred = (isSigned ?
3266 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3267 return new ICmpInst(pred, V, Hi);
3270 // Emit V-Lo <u Hi-Lo
3271 Constant *NegLo = ConstantExpr::getNeg(Lo);
3272 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3273 InsertNewInstBefore(Add, IB);
3274 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3275 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3278 if (Lo == Hi) // Trivially true.
3279 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3281 // V < Min || V >= Hi -> V > Hi-1
3282 Hi = SubOne(cast<ConstantInt>(Hi));
3283 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3284 ICmpInst::Predicate pred = (isSigned ?
3285 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3286 return new ICmpInst(pred, V, Hi);
3289 // Emit V-Lo >u Hi-1-Lo
3290 // Note that Hi has already had one subtracted from it, above.
3291 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3292 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3293 InsertNewInstBefore(Add, IB);
3294 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3295 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3298 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3299 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3300 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3301 // not, since all 1s are not contiguous.
3302 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3303 const APInt& V = Val->getValue();
3304 uint32_t BitWidth = Val->getType()->getBitWidth();
3305 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3307 // look for the first zero bit after the run of ones
3308 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3309 // look for the first non-zero bit
3310 ME = V.getActiveBits();
3314 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3315 /// where isSub determines whether the operator is a sub. If we can fold one of
3316 /// the following xforms:
3318 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3319 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3320 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3322 /// return (A +/- B).
3324 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3325 ConstantInt *Mask, bool isSub,
3327 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3328 if (!LHSI || LHSI->getNumOperands() != 2 ||
3329 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3331 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3333 switch (LHSI->getOpcode()) {
3335 case Instruction::And:
3336 if (And(N, Mask) == Mask) {
3337 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3338 if ((Mask->getValue().countLeadingZeros() +
3339 Mask->getValue().countPopulation()) ==
3340 Mask->getValue().getBitWidth())
3343 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3344 // part, we don't need any explicit masks to take them out of A. If that
3345 // is all N is, ignore it.
3346 uint32_t MB = 0, ME = 0;
3347 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3348 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3349 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3350 if (MaskedValueIsZero(RHS, Mask))
3355 case Instruction::Or:
3356 case Instruction::Xor:
3357 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3358 if ((Mask->getValue().countLeadingZeros() +
3359 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3360 && And(N, Mask)->isZero())
3367 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3369 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3370 return InsertNewInstBefore(New, I);
3373 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3374 bool Changed = SimplifyCommutative(I);
3375 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3377 if (isa<UndefValue>(Op1)) // X & undef -> 0
3378 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3382 return ReplaceInstUsesWith(I, Op1);
3384 // See if we can simplify any instructions used by the instruction whose sole
3385 // purpose is to compute bits we don't care about.
3386 if (!isa<VectorType>(I.getType())) {
3387 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3388 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3389 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3390 KnownZero, KnownOne))
3393 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3394 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3395 return ReplaceInstUsesWith(I, I.getOperand(0));
3396 } else if (isa<ConstantAggregateZero>(Op1)) {
3397 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3401 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3402 const APInt& AndRHSMask = AndRHS->getValue();
3403 APInt NotAndRHS(~AndRHSMask);
3405 // Optimize a variety of ((val OP C1) & C2) combinations...
3406 if (isa<BinaryOperator>(Op0)) {
3407 Instruction *Op0I = cast<Instruction>(Op0);
3408 Value *Op0LHS = Op0I->getOperand(0);
3409 Value *Op0RHS = Op0I->getOperand(1);
3410 switch (Op0I->getOpcode()) {
3411 case Instruction::Xor:
3412 case Instruction::Or:
3413 // If the mask is only needed on one incoming arm, push it up.
3414 if (Op0I->hasOneUse()) {
3415 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3416 // Not masking anything out for the LHS, move to RHS.
3417 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3418 Op0RHS->getName()+".masked");
3419 InsertNewInstBefore(NewRHS, I);
3420 return BinaryOperator::Create(
3421 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3423 if (!isa<Constant>(Op0RHS) &&
3424 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3425 // Not masking anything out for the RHS, move to LHS.
3426 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3427 Op0LHS->getName()+".masked");
3428 InsertNewInstBefore(NewLHS, I);
3429 return BinaryOperator::Create(
3430 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3435 case Instruction::Add:
3436 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3437 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3438 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3439 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3440 return BinaryOperator::CreateAnd(V, AndRHS);
3441 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3442 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3445 case Instruction::Sub:
3446 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3447 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3448 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3449 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3450 return BinaryOperator::CreateAnd(V, AndRHS);
3452 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3453 // has 1's for all bits that the subtraction with A might affect.
3454 if (Op0I->hasOneUse()) {
3455 uint32_t BitWidth = AndRHSMask.getBitWidth();
3456 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3457 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3459 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3460 if (!(A && A->isZero()) && // avoid infinite recursion.
3461 MaskedValueIsZero(Op0LHS, Mask)) {
3462 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3463 InsertNewInstBefore(NewNeg, I);
3464 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3469 case Instruction::Shl:
3470 case Instruction::LShr:
3471 // (1 << x) & 1 --> zext(x == 0)
3472 // (1 >> x) & 1 --> zext(x == 0)
3473 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3474 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3475 Constant::getNullValue(I.getType()));
3476 InsertNewInstBefore(NewICmp, I);
3477 return new ZExtInst(NewICmp, I.getType());
3482 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3483 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3485 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3486 // If this is an integer truncation or change from signed-to-unsigned, and
3487 // if the source is an and/or with immediate, transform it. This
3488 // frequently occurs for bitfield accesses.
3489 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3490 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3491 CastOp->getNumOperands() == 2)
3492 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3493 if (CastOp->getOpcode() == Instruction::And) {
3494 // Change: and (cast (and X, C1) to T), C2
3495 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3496 // This will fold the two constants together, which may allow
3497 // other simplifications.
3498 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3499 CastOp->getOperand(0), I.getType(),
3500 CastOp->getName()+".shrunk");
3501 NewCast = InsertNewInstBefore(NewCast, I);
3502 // trunc_or_bitcast(C1)&C2
3503 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3504 C3 = ConstantExpr::getAnd(C3, AndRHS);
3505 return BinaryOperator::CreateAnd(NewCast, C3);
3506 } else if (CastOp->getOpcode() == Instruction::Or) {
3507 // Change: and (cast (or X, C1) to T), C2
3508 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3509 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3510 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3511 return ReplaceInstUsesWith(I, AndRHS);
3517 // Try to fold constant and into select arguments.
3518 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3519 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3521 if (isa<PHINode>(Op0))
3522 if (Instruction *NV = FoldOpIntoPhi(I))
3526 Value *Op0NotVal = dyn_castNotVal(Op0);
3527 Value *Op1NotVal = dyn_castNotVal(Op1);
3529 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3530 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3532 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3533 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3534 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3535 I.getName()+".demorgan");
3536 InsertNewInstBefore(Or, I);
3537 return BinaryOperator::CreateNot(Or);
3541 Value *A = 0, *B = 0, *C = 0, *D = 0;
3542 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3543 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3544 return ReplaceInstUsesWith(I, Op1);
3546 // (A|B) & ~(A&B) -> A^B
3547 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3548 if ((A == C && B == D) || (A == D && B == C))
3549 return BinaryOperator::CreateXor(A, B);
3553 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3554 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3555 return ReplaceInstUsesWith(I, Op0);
3557 // ~(A&B) & (A|B) -> A^B
3558 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3559 if ((A == C && B == D) || (A == D && B == C))
3560 return BinaryOperator::CreateXor(A, B);
3564 if (Op0->hasOneUse() &&
3565 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3566 if (A == Op1) { // (A^B)&A -> A&(A^B)
3567 I.swapOperands(); // Simplify below
3568 std::swap(Op0, Op1);
3569 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3570 cast<BinaryOperator>(Op0)->swapOperands();
3571 I.swapOperands(); // Simplify below
3572 std::swap(Op0, Op1);
3575 if (Op1->hasOneUse() &&
3576 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3577 if (B == Op0) { // B&(A^B) -> B&(B^A)
3578 cast<BinaryOperator>(Op1)->swapOperands();
3581 if (A == Op0) { // A&(A^B) -> A & ~B
3582 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3583 InsertNewInstBefore(NotB, I);
3584 return BinaryOperator::CreateAnd(A, NotB);
3589 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3590 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3591 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3594 Value *LHSVal, *RHSVal;
3595 ConstantInt *LHSCst, *RHSCst;
3596 ICmpInst::Predicate LHSCC, RHSCC;
3597 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3598 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3599 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3600 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3601 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3602 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3603 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3604 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3606 // Don't try to fold ICMP_SLT + ICMP_ULT.
3607 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3608 ICmpInst::isSignedPredicate(LHSCC) ==
3609 ICmpInst::isSignedPredicate(RHSCC))) {
3610 // Ensure that the larger constant is on the RHS.
3611 ICmpInst::Predicate GT;
3612 if (ICmpInst::isSignedPredicate(LHSCC) ||
3613 (ICmpInst::isEquality(LHSCC) &&
3614 ICmpInst::isSignedPredicate(RHSCC)))
3615 GT = ICmpInst::ICMP_SGT;
3617 GT = ICmpInst::ICMP_UGT;
3619 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3620 ICmpInst *LHS = cast<ICmpInst>(Op0);
3621 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3622 std::swap(LHS, RHS);
3623 std::swap(LHSCst, RHSCst);
3624 std::swap(LHSCC, RHSCC);
3627 // At this point, we know we have have two icmp instructions
3628 // comparing a value against two constants and and'ing the result
3629 // together. Because of the above check, we know that we only have
3630 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3631 // (from the FoldICmpLogical check above), that the two constants
3632 // are not equal and that the larger constant is on the RHS
3633 assert(LHSCst != RHSCst && "Compares not folded above?");
3636 default: assert(0 && "Unknown integer condition code!");
3637 case ICmpInst::ICMP_EQ:
3639 default: assert(0 && "Unknown integer condition code!");
3640 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3641 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3642 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3643 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3644 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3645 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3646 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3647 return ReplaceInstUsesWith(I, LHS);
3649 case ICmpInst::ICMP_NE:
3651 default: assert(0 && "Unknown integer condition code!");
3652 case ICmpInst::ICMP_ULT:
3653 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3654 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3655 break; // (X != 13 & X u< 15) -> no change
3656 case ICmpInst::ICMP_SLT:
3657 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3658 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3659 break; // (X != 13 & X s< 15) -> no change
3660 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3661 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3662 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3663 return ReplaceInstUsesWith(I, RHS);
3664 case ICmpInst::ICMP_NE:
3665 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3666 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3667 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
3668 LHSVal->getName()+".off");
3669 InsertNewInstBefore(Add, I);
3670 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3671 ConstantInt::get(Add->getType(), 1));
3673 break; // (X != 13 & X != 15) -> no change
3676 case ICmpInst::ICMP_ULT:
3678 default: assert(0 && "Unknown integer condition code!");
3679 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3680 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3681 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3682 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3684 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3685 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3686 return ReplaceInstUsesWith(I, LHS);
3687 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3691 case ICmpInst::ICMP_SLT:
3693 default: assert(0 && "Unknown integer condition code!");
3694 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3695 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3696 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3697 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3699 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3700 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3701 return ReplaceInstUsesWith(I, LHS);
3702 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3706 case ICmpInst::ICMP_UGT:
3708 default: assert(0 && "Unknown integer condition code!");
3709 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3710 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3711 return ReplaceInstUsesWith(I, RHS);
3712 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3714 case ICmpInst::ICMP_NE:
3715 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3716 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3717 break; // (X u> 13 & X != 15) -> no change
3718 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3719 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3721 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3725 case ICmpInst::ICMP_SGT:
3727 default: assert(0 && "Unknown integer condition code!");
3728 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3729 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3730 return ReplaceInstUsesWith(I, RHS);
3731 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3733 case ICmpInst::ICMP_NE:
3734 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3735 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3736 break; // (X s> 13 & X != 15) -> no change
3737 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3738 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3740 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3748 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3749 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3750 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3751 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3752 const Type *SrcTy = Op0C->getOperand(0)->getType();
3753 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3754 // Only do this if the casts both really cause code to be generated.
3755 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3757 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3759 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3760 Op1C->getOperand(0),
3762 InsertNewInstBefore(NewOp, I);
3763 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3767 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3768 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3769 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3770 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3771 SI0->getOperand(1) == SI1->getOperand(1) &&
3772 (SI0->hasOneUse() || SI1->hasOneUse())) {
3773 Instruction *NewOp =
3774 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3776 SI0->getName()), I);
3777 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3778 SI1->getOperand(1));
3782 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3783 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3784 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3785 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3786 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3787 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3788 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3789 // If either of the constants are nans, then the whole thing returns
3791 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3792 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3793 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3794 RHS->getOperand(0));
3799 return Changed ? &I : 0;
3802 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3803 /// in the result. If it does, and if the specified byte hasn't been filled in
3804 /// yet, fill it in and return false.
3805 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3806 Instruction *I = dyn_cast<Instruction>(V);
3807 if (I == 0) return true;
3809 // If this is an or instruction, it is an inner node of the bswap.
3810 if (I->getOpcode() == Instruction::Or)
3811 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3812 CollectBSwapParts(I->getOperand(1), ByteValues);
3814 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3815 // If this is a shift by a constant int, and it is "24", then its operand
3816 // defines a byte. We only handle unsigned types here.
3817 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3818 // Not shifting the entire input by N-1 bytes?
3819 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3820 8*(ByteValues.size()-1))
3824 if (I->getOpcode() == Instruction::Shl) {
3825 // X << 24 defines the top byte with the lowest of the input bytes.
3826 DestNo = ByteValues.size()-1;
3828 // X >>u 24 defines the low byte with the highest of the input bytes.
3832 // If the destination byte value is already defined, the values are or'd
3833 // together, which isn't a bswap (unless it's an or of the same bits).
3834 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3836 ByteValues[DestNo] = I->getOperand(0);
3840 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3842 Value *Shift = 0, *ShiftLHS = 0;
3843 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3844 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3845 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3847 Instruction *SI = cast<Instruction>(Shift);
3849 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3850 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3851 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3854 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3856 if (AndAmt->getValue().getActiveBits() > 64)
3858 uint64_t AndAmtVal = AndAmt->getZExtValue();
3859 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3860 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3862 // Unknown mask for bswap.
3863 if (DestByte == ByteValues.size()) return true;
3865 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3867 if (SI->getOpcode() == Instruction::Shl)
3868 SrcByte = DestByte - ShiftBytes;
3870 SrcByte = DestByte + ShiftBytes;
3872 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3873 if (SrcByte != ByteValues.size()-DestByte-1)
3876 // If the destination byte value is already defined, the values are or'd
3877 // together, which isn't a bswap (unless it's an or of the same bits).
3878 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3880 ByteValues[DestByte] = SI->getOperand(0);
3884 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3885 /// If so, insert the new bswap intrinsic and return it.
3886 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3887 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3888 if (!ITy || ITy->getBitWidth() % 16)
3889 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3891 /// ByteValues - For each byte of the result, we keep track of which value
3892 /// defines each byte.
3893 SmallVector<Value*, 8> ByteValues;
3894 ByteValues.resize(ITy->getBitWidth()/8);
3896 // Try to find all the pieces corresponding to the bswap.
3897 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3898 CollectBSwapParts(I.getOperand(1), ByteValues))
3901 // Check to see if all of the bytes come from the same value.
3902 Value *V = ByteValues[0];
3903 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3905 // Check to make sure that all of the bytes come from the same value.
3906 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3907 if (ByteValues[i] != V)
3909 const Type *Tys[] = { ITy };
3910 Module *M = I.getParent()->getParent()->getParent();
3911 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3912 return CallInst::Create(F, V);
3916 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3917 bool Changed = SimplifyCommutative(I);
3918 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3920 if (isa<UndefValue>(Op1)) // X | undef -> -1
3921 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3925 return ReplaceInstUsesWith(I, Op0);
3927 // See if we can simplify any instructions used by the instruction whose sole
3928 // purpose is to compute bits we don't care about.
3929 if (!isa<VectorType>(I.getType())) {
3930 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3931 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3932 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3933 KnownZero, KnownOne))
3935 } else if (isa<ConstantAggregateZero>(Op1)) {
3936 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3937 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3938 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3939 return ReplaceInstUsesWith(I, I.getOperand(1));
3945 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3946 ConstantInt *C1 = 0; Value *X = 0;
3947 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3948 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3949 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3950 InsertNewInstBefore(Or, I);
3952 return BinaryOperator::CreateAnd(Or,
3953 ConstantInt::get(RHS->getValue() | C1->getValue()));
3956 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3957 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3958 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3959 InsertNewInstBefore(Or, I);
3961 return BinaryOperator::CreateXor(Or,
3962 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3965 // Try to fold constant and into select arguments.
3966 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3967 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3969 if (isa<PHINode>(Op0))
3970 if (Instruction *NV = FoldOpIntoPhi(I))
3974 Value *A = 0, *B = 0;
3975 ConstantInt *C1 = 0, *C2 = 0;
3977 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3978 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3979 return ReplaceInstUsesWith(I, Op1);
3980 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3981 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3982 return ReplaceInstUsesWith(I, Op0);
3984 // (A | B) | C and A | (B | C) -> bswap if possible.
3985 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3986 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3987 match(Op1, m_Or(m_Value(), m_Value())) ||
3988 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3989 match(Op1, m_Shift(m_Value(), m_Value())))) {
3990 if (Instruction *BSwap = MatchBSwap(I))
3994 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3995 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3996 MaskedValueIsZero(Op1, C1->getValue())) {
3997 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
3998 InsertNewInstBefore(NOr, I);
4000 return BinaryOperator::CreateXor(NOr, C1);
4003 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4004 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4005 MaskedValueIsZero(Op0, C1->getValue())) {
4006 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4007 InsertNewInstBefore(NOr, I);
4009 return BinaryOperator::CreateXor(NOr, C1);
4013 Value *C = 0, *D = 0;
4014 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4015 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4016 Value *V1 = 0, *V2 = 0, *V3 = 0;
4017 C1 = dyn_cast<ConstantInt>(C);
4018 C2 = dyn_cast<ConstantInt>(D);
4019 if (C1 && C2) { // (A & C1)|(B & C2)
4020 // If we have: ((V + N) & C1) | (V & C2)
4021 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4022 // replace with V+N.
4023 if (C1->getValue() == ~C2->getValue()) {
4024 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4025 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4026 // Add commutes, try both ways.
4027 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4028 return ReplaceInstUsesWith(I, A);
4029 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4030 return ReplaceInstUsesWith(I, A);
4032 // Or commutes, try both ways.
4033 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4034 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4035 // Add commutes, try both ways.
4036 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4037 return ReplaceInstUsesWith(I, B);
4038 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4039 return ReplaceInstUsesWith(I, B);
4042 V1 = 0; V2 = 0; V3 = 0;
4045 // Check to see if we have any common things being and'ed. If so, find the
4046 // terms for V1 & (V2|V3).
4047 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4048 if (A == B) // (A & C)|(A & D) == A & (C|D)
4049 V1 = A, V2 = C, V3 = D;
4050 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4051 V1 = A, V2 = B, V3 = C;
4052 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4053 V1 = C, V2 = A, V3 = D;
4054 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4055 V1 = C, V2 = A, V3 = B;
4059 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4060 return BinaryOperator::CreateAnd(V1, Or);
4065 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4066 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4067 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4068 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4069 SI0->getOperand(1) == SI1->getOperand(1) &&
4070 (SI0->hasOneUse() || SI1->hasOneUse())) {
4071 Instruction *NewOp =
4072 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4074 SI0->getName()), I);
4075 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4076 SI1->getOperand(1));
4080 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4081 if (A == Op1) // ~A | A == -1
4082 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4086 // Note, A is still live here!
4087 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4089 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4091 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4092 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4093 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4094 I.getName()+".demorgan"), I);
4095 return BinaryOperator::CreateNot(And);
4099 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4100 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4101 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4104 Value *LHSVal, *RHSVal;
4105 ConstantInt *LHSCst, *RHSCst;
4106 ICmpInst::Predicate LHSCC, RHSCC;
4107 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4108 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4109 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4110 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4111 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4112 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4113 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4114 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4115 // We can't fold (ugt x, C) | (sgt x, C2).
4116 PredicatesFoldable(LHSCC, RHSCC)) {
4117 // Ensure that the larger constant is on the RHS.
4118 ICmpInst *LHS = cast<ICmpInst>(Op0);
4120 if (ICmpInst::isSignedPredicate(LHSCC))
4121 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4123 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4126 std::swap(LHS, RHS);
4127 std::swap(LHSCst, RHSCst);
4128 std::swap(LHSCC, RHSCC);
4131 // At this point, we know we have have two icmp instructions
4132 // comparing a value against two constants and or'ing the result
4133 // together. Because of the above check, we know that we only have
4134 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4135 // FoldICmpLogical check above), that the two constants are not
4137 assert(LHSCst != RHSCst && "Compares not folded above?");
4140 default: assert(0 && "Unknown integer condition code!");
4141 case ICmpInst::ICMP_EQ:
4143 default: assert(0 && "Unknown integer condition code!");
4144 case ICmpInst::ICMP_EQ:
4145 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4146 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4147 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4148 LHSVal->getName()+".off");
4149 InsertNewInstBefore(Add, I);
4150 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4151 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4153 break; // (X == 13 | X == 15) -> no change
4154 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4155 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4157 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4158 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4159 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4160 return ReplaceInstUsesWith(I, RHS);
4163 case ICmpInst::ICMP_NE:
4165 default: assert(0 && "Unknown integer condition code!");
4166 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4167 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4168 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4169 return ReplaceInstUsesWith(I, LHS);
4170 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4171 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4172 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4173 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4176 case ICmpInst::ICMP_ULT:
4178 default: assert(0 && "Unknown integer condition code!");
4179 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4181 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4182 // If RHSCst is [us]MAXINT, it is always false. Not handling
4183 // this can cause overflow.
4184 if (RHSCst->isMaxValue(false))
4185 return ReplaceInstUsesWith(I, LHS);
4186 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4188 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4190 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4191 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4192 return ReplaceInstUsesWith(I, RHS);
4193 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4197 case ICmpInst::ICMP_SLT:
4199 default: assert(0 && "Unknown integer condition code!");
4200 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4202 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4203 // If RHSCst is [us]MAXINT, it is always false. Not handling
4204 // this can cause overflow.
4205 if (RHSCst->isMaxValue(true))
4206 return ReplaceInstUsesWith(I, LHS);
4207 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4209 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4211 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4212 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4213 return ReplaceInstUsesWith(I, RHS);
4214 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4218 case ICmpInst::ICMP_UGT:
4220 default: assert(0 && "Unknown integer condition code!");
4221 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4222 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4223 return ReplaceInstUsesWith(I, LHS);
4224 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4226 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4227 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4228 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4229 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4233 case ICmpInst::ICMP_SGT:
4235 default: assert(0 && "Unknown integer condition code!");
4236 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4237 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4238 return ReplaceInstUsesWith(I, LHS);
4239 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4241 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4242 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4243 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4244 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4252 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4253 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4254 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4255 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4256 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4257 !isa<ICmpInst>(Op1C->getOperand(0))) {
4258 const Type *SrcTy = Op0C->getOperand(0)->getType();
4259 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4260 // Only do this if the casts both really cause code to be
4262 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4264 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4266 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4267 Op1C->getOperand(0),
4269 InsertNewInstBefore(NewOp, I);
4270 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4277 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4278 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4279 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4280 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4281 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4282 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4283 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4284 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4285 // If either of the constants are nans, then the whole thing returns
4287 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4288 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4290 // Otherwise, no need to compare the two constants, compare the
4292 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4293 RHS->getOperand(0));
4298 return Changed ? &I : 0;
4303 // XorSelf - Implements: X ^ X --> 0
4306 XorSelf(Value *rhs) : RHS(rhs) {}
4307 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4308 Instruction *apply(BinaryOperator &Xor) const {
4315 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4316 bool Changed = SimplifyCommutative(I);
4317 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4319 if (isa<UndefValue>(Op1)) {
4320 if (isa<UndefValue>(Op0))
4321 // Handle undef ^ undef -> 0 special case. This is a common
4323 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4324 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4327 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4328 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4329 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4330 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4333 // See if we can simplify any instructions used by the instruction whose sole
4334 // purpose is to compute bits we don't care about.
4335 if (!isa<VectorType>(I.getType())) {
4336 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4337 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4338 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4339 KnownZero, KnownOne))
4341 } else if (isa<ConstantAggregateZero>(Op1)) {
4342 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4345 // Is this a ~ operation?
4346 if (Value *NotOp = dyn_castNotVal(&I)) {
4347 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4348 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4349 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4350 if (Op0I->getOpcode() == Instruction::And ||
4351 Op0I->getOpcode() == Instruction::Or) {
4352 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4353 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4355 BinaryOperator::CreateNot(Op0I->getOperand(1),
4356 Op0I->getOperand(1)->getName()+".not");
4357 InsertNewInstBefore(NotY, I);
4358 if (Op0I->getOpcode() == Instruction::And)
4359 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4361 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4368 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4369 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4370 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4371 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4372 return new ICmpInst(ICI->getInversePredicate(),
4373 ICI->getOperand(0), ICI->getOperand(1));
4375 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4376 return new FCmpInst(FCI->getInversePredicate(),
4377 FCI->getOperand(0), FCI->getOperand(1));
4380 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4381 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4382 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4383 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4384 Instruction::CastOps Opcode = Op0C->getOpcode();
4385 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4386 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4387 Op0C->getDestTy())) {
4388 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4389 CI->getOpcode(), CI->getInversePredicate(),
4390 CI->getOperand(0), CI->getOperand(1)), I);
4391 NewCI->takeName(CI);
4392 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4399 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4400 // ~(c-X) == X-c-1 == X+(-c-1)
4401 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4402 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4403 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4404 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4405 ConstantInt::get(I.getType(), 1));
4406 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4409 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4410 if (Op0I->getOpcode() == Instruction::Add) {
4411 // ~(X-c) --> (-c-1)-X
4412 if (RHS->isAllOnesValue()) {
4413 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4414 return BinaryOperator::CreateSub(
4415 ConstantExpr::getSub(NegOp0CI,
4416 ConstantInt::get(I.getType(), 1)),
4417 Op0I->getOperand(0));
4418 } else if (RHS->getValue().isSignBit()) {
4419 // (X + C) ^ signbit -> (X + C + signbit)
4420 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4421 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4424 } else if (Op0I->getOpcode() == Instruction::Or) {
4425 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4426 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4427 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4428 // Anything in both C1 and C2 is known to be zero, remove it from
4430 Constant *CommonBits = And(Op0CI, RHS);
4431 NewRHS = ConstantExpr::getAnd(NewRHS,
4432 ConstantExpr::getNot(CommonBits));
4433 AddToWorkList(Op0I);
4434 I.setOperand(0, Op0I->getOperand(0));
4435 I.setOperand(1, NewRHS);
4442 // Try to fold constant and into select arguments.
4443 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4444 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4446 if (isa<PHINode>(Op0))
4447 if (Instruction *NV = FoldOpIntoPhi(I))
4451 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4453 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4455 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4457 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4460 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4463 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4464 if (A == Op0) { // B^(B|A) == (A|B)^B
4465 Op1I->swapOperands();
4467 std::swap(Op0, Op1);
4468 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4469 I.swapOperands(); // Simplified below.
4470 std::swap(Op0, Op1);
4472 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4473 if (Op0 == A) // A^(A^B) == B
4474 return ReplaceInstUsesWith(I, B);
4475 else if (Op0 == B) // A^(B^A) == B
4476 return ReplaceInstUsesWith(I, A);
4477 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4478 if (A == Op0) { // A^(A&B) -> A^(B&A)
4479 Op1I->swapOperands();
4482 if (B == Op0) { // A^(B&A) -> (B&A)^A
4483 I.swapOperands(); // Simplified below.
4484 std::swap(Op0, Op1);
4489 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4492 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4493 if (A == Op1) // (B|A)^B == (A|B)^B
4495 if (B == Op1) { // (A|B)^B == A & ~B
4497 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4498 return BinaryOperator::CreateAnd(A, NotB);
4500 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4501 if (Op1 == A) // (A^B)^A == B
4502 return ReplaceInstUsesWith(I, B);
4503 else if (Op1 == B) // (B^A)^A == B
4504 return ReplaceInstUsesWith(I, A);
4505 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4506 if (A == Op1) // (A&B)^A -> (B&A)^A
4508 if (B == Op1 && // (B&A)^A == ~B & A
4509 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4511 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4512 return BinaryOperator::CreateAnd(N, Op1);
4517 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4518 if (Op0I && Op1I && Op0I->isShift() &&
4519 Op0I->getOpcode() == Op1I->getOpcode() &&
4520 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4521 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4522 Instruction *NewOp =
4523 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4524 Op1I->getOperand(0),
4525 Op0I->getName()), I);
4526 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4527 Op1I->getOperand(1));
4531 Value *A, *B, *C, *D;
4532 // (A & B)^(A | B) -> A ^ B
4533 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4534 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4535 if ((A == C && B == D) || (A == D && B == C))
4536 return BinaryOperator::CreateXor(A, B);
4538 // (A | B)^(A & B) -> A ^ B
4539 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4540 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4541 if ((A == C && B == D) || (A == D && B == C))
4542 return BinaryOperator::CreateXor(A, B);
4546 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4547 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4548 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4549 // (X & Y)^(X & Y) -> (Y^Z) & X
4550 Value *X = 0, *Y = 0, *Z = 0;
4552 X = A, Y = B, Z = D;
4554 X = A, Y = B, Z = C;
4556 X = B, Y = A, Z = D;
4558 X = B, Y = A, Z = C;
4561 Instruction *NewOp =
4562 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4563 return BinaryOperator::CreateAnd(NewOp, X);
4568 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4569 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4570 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4573 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4574 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4575 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4576 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4577 const Type *SrcTy = Op0C->getOperand(0)->getType();
4578 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4579 // Only do this if the casts both really cause code to be generated.
4580 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4582 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4584 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4585 Op1C->getOperand(0),
4587 InsertNewInstBefore(NewOp, I);
4588 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4593 return Changed ? &I : 0;
4596 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4597 /// overflowed for this type.
4598 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4599 ConstantInt *In2, bool IsSigned = false) {
4600 Result = cast<ConstantInt>(Add(In1, In2));
4603 if (In2->getValue().isNegative())
4604 return Result->getValue().sgt(In1->getValue());
4606 return Result->getValue().slt(In1->getValue());
4608 return Result->getValue().ult(In1->getValue());
4611 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4612 /// code necessary to compute the offset from the base pointer (without adding
4613 /// in the base pointer). Return the result as a signed integer of intptr size.
4614 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4615 TargetData &TD = IC.getTargetData();
4616 gep_type_iterator GTI = gep_type_begin(GEP);
4617 const Type *IntPtrTy = TD.getIntPtrType();
4618 Value *Result = Constant::getNullValue(IntPtrTy);
4620 // Build a mask for high order bits.
4621 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4622 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4624 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
4627 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4628 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4629 if (OpC->isZero()) continue;
4631 // Handle a struct index, which adds its field offset to the pointer.
4632 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4633 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4635 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4636 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4638 Result = IC.InsertNewInstBefore(
4639 BinaryOperator::CreateAdd(Result,
4640 ConstantInt::get(IntPtrTy, Size),
4641 GEP->getName()+".offs"), I);
4645 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4646 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4647 Scale = ConstantExpr::getMul(OC, Scale);
4648 if (Constant *RC = dyn_cast<Constant>(Result))
4649 Result = ConstantExpr::getAdd(RC, Scale);
4651 // Emit an add instruction.
4652 Result = IC.InsertNewInstBefore(
4653 BinaryOperator::CreateAdd(Result, Scale,
4654 GEP->getName()+".offs"), I);
4658 // Convert to correct type.
4659 if (Op->getType() != IntPtrTy) {
4660 if (Constant *OpC = dyn_cast<Constant>(Op))
4661 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4663 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4664 Op->getName()+".c"), I);
4667 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4668 if (Constant *OpC = dyn_cast<Constant>(Op))
4669 Op = ConstantExpr::getMul(OpC, Scale);
4670 else // We'll let instcombine(mul) convert this to a shl if possible.
4671 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
4672 GEP->getName()+".idx"), I);
4675 // Emit an add instruction.
4676 if (isa<Constant>(Op) && isa<Constant>(Result))
4677 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4678 cast<Constant>(Result));
4680 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
4681 GEP->getName()+".offs"), I);
4687 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4688 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4689 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4690 /// complex, and scales are involved. The above expression would also be legal
4691 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4692 /// later form is less amenable to optimization though, and we are allowed to
4693 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4695 /// If we can't emit an optimized form for this expression, this returns null.
4697 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4699 TargetData &TD = IC.getTargetData();
4700 gep_type_iterator GTI = gep_type_begin(GEP);
4702 // Check to see if this gep only has a single variable index. If so, and if
4703 // any constant indices are a multiple of its scale, then we can compute this
4704 // in terms of the scale of the variable index. For example, if the GEP
4705 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4706 // because the expression will cross zero at the same point.
4707 unsigned i, e = GEP->getNumOperands();
4709 for (i = 1; i != e; ++i, ++GTI) {
4710 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4711 // Compute the aggregate offset of constant indices.
4712 if (CI->isZero()) continue;
4714 // Handle a struct index, which adds its field offset to the pointer.
4715 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4716 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4718 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4719 Offset += Size*CI->getSExtValue();
4722 // Found our variable index.
4727 // If there are no variable indices, we must have a constant offset, just
4728 // evaluate it the general way.
4729 if (i == e) return 0;
4731 Value *VariableIdx = GEP->getOperand(i);
4732 // Determine the scale factor of the variable element. For example, this is
4733 // 4 if the variable index is into an array of i32.
4734 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4736 // Verify that there are no other variable indices. If so, emit the hard way.
4737 for (++i, ++GTI; i != e; ++i, ++GTI) {
4738 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4741 // Compute the aggregate offset of constant indices.
4742 if (CI->isZero()) continue;
4744 // Handle a struct index, which adds its field offset to the pointer.
4745 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4746 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4748 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4749 Offset += Size*CI->getSExtValue();
4753 // Okay, we know we have a single variable index, which must be a
4754 // pointer/array/vector index. If there is no offset, life is simple, return
4756 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4758 // Cast to intptrty in case a truncation occurs. If an extension is needed,
4759 // we don't need to bother extending: the extension won't affect where the
4760 // computation crosses zero.
4761 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
4762 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
4763 VariableIdx->getNameStart(), &I);
4767 // Otherwise, there is an index. The computation we will do will be modulo
4768 // the pointer size, so get it.
4769 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4771 Offset &= PtrSizeMask;
4772 VariableScale &= PtrSizeMask;
4774 // To do this transformation, any constant index must be a multiple of the
4775 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
4776 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
4777 // multiple of the variable scale.
4778 int64_t NewOffs = Offset / (int64_t)VariableScale;
4779 if (Offset != NewOffs*(int64_t)VariableScale)
4782 // Okay, we can do this evaluation. Start by converting the index to intptr.
4783 const Type *IntPtrTy = TD.getIntPtrType();
4784 if (VariableIdx->getType() != IntPtrTy)
4785 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
4787 VariableIdx->getNameStart(), &I);
4788 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
4789 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
4793 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4794 /// else. At this point we know that the GEP is on the LHS of the comparison.
4795 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4796 ICmpInst::Predicate Cond,
4798 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4800 // Look through bitcasts.
4801 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
4802 RHS = BCI->getOperand(0);
4804 Value *PtrBase = GEPLHS->getOperand(0);
4805 if (PtrBase == RHS) {
4806 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4807 // This transformation (ignoring the base and scales) is valid because we
4808 // know pointers can't overflow. See if we can output an optimized form.
4809 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
4811 // If not, synthesize the offset the hard way.
4813 Offset = EmitGEPOffset(GEPLHS, I, *this);
4814 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4815 Constant::getNullValue(Offset->getType()));
4816 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4817 // If the base pointers are different, but the indices are the same, just
4818 // compare the base pointer.
4819 if (PtrBase != GEPRHS->getOperand(0)) {
4820 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4821 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4822 GEPRHS->getOperand(0)->getType();
4824 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4825 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4826 IndicesTheSame = false;
4830 // If all indices are the same, just compare the base pointers.
4832 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4833 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4835 // Otherwise, the base pointers are different and the indices are
4836 // different, bail out.
4840 // If one of the GEPs has all zero indices, recurse.
4841 bool AllZeros = true;
4842 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4843 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4844 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4849 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4850 ICmpInst::getSwappedPredicate(Cond), I);
4852 // If the other GEP has all zero indices, recurse.
4854 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4855 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4856 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4861 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4863 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4864 // If the GEPs only differ by one index, compare it.
4865 unsigned NumDifferences = 0; // Keep track of # differences.
4866 unsigned DiffOperand = 0; // The operand that differs.
4867 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4868 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4869 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4870 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4871 // Irreconcilable differences.
4875 if (NumDifferences++) break;
4880 if (NumDifferences == 0) // SAME GEP?
4881 return ReplaceInstUsesWith(I, // No comparison is needed here.
4882 ConstantInt::get(Type::Int1Ty,
4883 ICmpInst::isTrueWhenEqual(Cond)));
4885 else if (NumDifferences == 1) {
4886 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4887 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4888 // Make sure we do a signed comparison here.
4889 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4893 // Only lower this if the icmp is the only user of the GEP or if we expect
4894 // the result to fold to a constant!
4895 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4896 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4897 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4898 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4899 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4900 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4906 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
4908 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
4911 if (!isa<ConstantFP>(RHSC)) return 0;
4912 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4914 // Get the width of the mantissa. We don't want to hack on conversions that
4915 // might lose information from the integer, e.g. "i64 -> float"
4916 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4917 if (MantissaWidth == -1) return 0; // Unknown.
4919 // Check to see that the input is converted from an integer type that is small
4920 // enough that preserves all bits. TODO: check here for "known" sign bits.
4921 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4922 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
4924 // If this is a uitofp instruction, we need an extra bit to hold the sign.
4925 if (isa<UIToFPInst>(LHSI))
4928 // If the conversion would lose info, don't hack on this.
4929 if ((int)InputSize > MantissaWidth)
4932 // Otherwise, we can potentially simplify the comparison. We know that it
4933 // will always come through as an integer value and we know the constant is
4934 // not a NAN (it would have been previously simplified).
4935 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4937 ICmpInst::Predicate Pred;
4938 switch (I.getPredicate()) {
4939 default: assert(0 && "Unexpected predicate!");
4940 case FCmpInst::FCMP_UEQ:
4941 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
4942 case FCmpInst::FCMP_UGT:
4943 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
4944 case FCmpInst::FCMP_UGE:
4945 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
4946 case FCmpInst::FCMP_ULT:
4947 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
4948 case FCmpInst::FCMP_ULE:
4949 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
4950 case FCmpInst::FCMP_UNE:
4951 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
4952 case FCmpInst::FCMP_ORD:
4953 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4954 case FCmpInst::FCMP_UNO:
4955 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4958 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4960 // Now we know that the APFloat is a normal number, zero or inf.
4962 // See if the FP constant is too large for the integer. For example,
4963 // comparing an i8 to 300.0.
4964 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
4966 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4967 // and large values.
4968 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
4969 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4970 APFloat::rmNearestTiesToEven);
4971 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4972 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4973 Pred == ICmpInst::ICMP_SLE)
4974 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4975 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4978 // See if the RHS value is < SignedMin.
4979 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
4980 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4981 APFloat::rmNearestTiesToEven);
4982 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4983 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4984 Pred == ICmpInst::ICMP_SGE)
4985 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4986 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4989 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
4990 // it may still be fractional. See if it is fractional by casting the FP
4991 // value to the integer value and back, checking for equality. Don't do this
4992 // for zero, because -0.0 is not fractional.
4993 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
4994 if (!RHS.isZero() &&
4995 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
4996 // If we had a comparison against a fractional value, we have to adjust
4997 // the compare predicate and sometimes the value. RHSC is rounded towards
4998 // zero at this point.
5000 default: assert(0 && "Unexpected integer comparison!");
5001 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5002 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5003 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5004 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5005 case ICmpInst::ICMP_SLE:
5006 // (float)int <= 4.4 --> int <= 4
5007 // (float)int <= -4.4 --> int < -4
5008 if (RHS.isNegative())
5009 Pred = ICmpInst::ICMP_SLT;
5011 case ICmpInst::ICMP_SLT:
5012 // (float)int < -4.4 --> int < -4
5013 // (float)int < 4.4 --> int <= 4
5014 if (!RHS.isNegative())
5015 Pred = ICmpInst::ICMP_SLE;
5017 case ICmpInst::ICMP_SGT:
5018 // (float)int > 4.4 --> int > 4
5019 // (float)int > -4.4 --> int >= -4
5020 if (RHS.isNegative())
5021 Pred = ICmpInst::ICMP_SGE;
5023 case ICmpInst::ICMP_SGE:
5024 // (float)int >= -4.4 --> int >= -4
5025 // (float)int >= 4.4 --> int > 4
5026 if (!RHS.isNegative())
5027 Pred = ICmpInst::ICMP_SGT;
5032 // Lower this FP comparison into an appropriate integer version of the
5034 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5037 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5038 bool Changed = SimplifyCompare(I);
5039 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5041 // Fold trivial predicates.
5042 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5043 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5044 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5045 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5047 // Simplify 'fcmp pred X, X'
5049 switch (I.getPredicate()) {
5050 default: assert(0 && "Unknown predicate!");
5051 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5052 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5053 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5054 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5055 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5056 case FCmpInst::FCMP_OLT: // True if ordered and less than
5057 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5058 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5060 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5061 case FCmpInst::FCMP_ULT: // True if unordered or less than
5062 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5063 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5064 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5065 I.setPredicate(FCmpInst::FCMP_UNO);
5066 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5069 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5070 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5071 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5072 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5073 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5074 I.setPredicate(FCmpInst::FCMP_ORD);
5075 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5080 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5081 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5083 // Handle fcmp with constant RHS
5084 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5085 // If the constant is a nan, see if we can fold the comparison based on it.
5086 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5087 if (CFP->getValueAPF().isNaN()) {
5088 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5089 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5090 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5091 "Comparison must be either ordered or unordered!");
5092 // True if unordered.
5093 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5097 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5098 switch (LHSI->getOpcode()) {
5099 case Instruction::PHI:
5100 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5101 // block. If in the same block, we're encouraging jump threading. If
5102 // not, we are just pessimizing the code by making an i1 phi.
5103 if (LHSI->getParent() == I.getParent())
5104 if (Instruction *NV = FoldOpIntoPhi(I))
5107 case Instruction::SIToFP:
5108 case Instruction::UIToFP:
5109 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5112 case Instruction::Select:
5113 // If either operand of the select is a constant, we can fold the
5114 // comparison into the select arms, which will cause one to be
5115 // constant folded and the select turned into a bitwise or.
5116 Value *Op1 = 0, *Op2 = 0;
5117 if (LHSI->hasOneUse()) {
5118 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5119 // Fold the known value into the constant operand.
5120 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5121 // Insert a new FCmp of the other select operand.
5122 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5123 LHSI->getOperand(2), RHSC,
5125 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5126 // Fold the known value into the constant operand.
5127 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5128 // Insert a new FCmp of the other select operand.
5129 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5130 LHSI->getOperand(1), RHSC,
5136 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5141 return Changed ? &I : 0;
5144 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5145 bool Changed = SimplifyCompare(I);
5146 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5147 const Type *Ty = Op0->getType();
5151 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5152 I.isTrueWhenEqual()));
5154 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5155 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5157 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5158 // addresses never equal each other! We already know that Op0 != Op1.
5159 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5160 isa<ConstantPointerNull>(Op0)) &&
5161 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5162 isa<ConstantPointerNull>(Op1)))
5163 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5164 !I.isTrueWhenEqual()));
5166 // icmp's with boolean values can always be turned into bitwise operations
5167 if (Ty == Type::Int1Ty) {
5168 switch (I.getPredicate()) {
5169 default: assert(0 && "Invalid icmp instruction!");
5170 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5171 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5172 InsertNewInstBefore(Xor, I);
5173 return BinaryOperator::CreateNot(Xor);
5175 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5176 return BinaryOperator::CreateXor(Op0, Op1);
5178 case ICmpInst::ICMP_UGT:
5179 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5181 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5182 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5183 InsertNewInstBefore(Not, I);
5184 return BinaryOperator::CreateAnd(Not, Op1);
5186 case ICmpInst::ICMP_SGT:
5187 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5189 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5190 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5191 InsertNewInstBefore(Not, I);
5192 return BinaryOperator::CreateAnd(Not, Op0);
5194 case ICmpInst::ICMP_UGE:
5195 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5197 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5198 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5199 InsertNewInstBefore(Not, I);
5200 return BinaryOperator::CreateOr(Not, Op1);
5202 case ICmpInst::ICMP_SGE:
5203 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5205 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5206 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5207 InsertNewInstBefore(Not, I);
5208 return BinaryOperator::CreateOr(Not, Op0);
5213 // See if we are doing a comparison between a constant and an instruction that
5214 // can be folded into the comparison.
5215 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5218 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5219 if (I.isEquality() && CI->isNullValue() &&
5220 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5221 // (icmp cond A B) if cond is equality
5222 return new ICmpInst(I.getPredicate(), A, B);
5225 // If we have a icmp le or icmp ge instruction, turn it into the appropriate
5226 // icmp lt or icmp gt instruction. This allows us to rely on them being
5227 // folded in the code below.
5228 switch (I.getPredicate()) {
5230 case ICmpInst::ICMP_ULE:
5231 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5232 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5233 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5234 case ICmpInst::ICMP_SLE:
5235 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5236 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5237 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5238 case ICmpInst::ICMP_UGE:
5239 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5240 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5241 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5242 case ICmpInst::ICMP_SGE:
5243 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5244 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5245 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5248 // See if we can fold the comparison based on range information we can get
5249 // by checking whether bits are known to be zero or one in the input.
5250 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5251 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5253 // If this comparison is a normal comparison, it demands all
5254 // bits, if it is a sign bit comparison, it only demands the sign bit.
5256 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5258 if (SimplifyDemandedBits(Op0,
5259 isSignBit ? APInt::getSignBit(BitWidth)
5260 : APInt::getAllOnesValue(BitWidth),
5261 KnownZero, KnownOne, 0))
5264 // Given the known and unknown bits, compute a range that the LHS could be
5265 // in. Compute the Min, Max and RHS values based on the known bits. For the
5266 // EQ and NE we use unsigned values.
5267 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5268 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5269 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5271 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5273 // If Min and Max are known to be the same, then SimplifyDemandedBits
5274 // figured out that the LHS is a constant. Just constant fold this now so
5275 // that code below can assume that Min != Max.
5277 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5278 ConstantInt::get(Min),
5281 // Based on the range information we know about the LHS, see if we can
5282 // simplify this comparison. For example, (x&4) < 8 is always true.
5283 const APInt &RHSVal = CI->getValue();
5284 switch (I.getPredicate()) { // LE/GE have been folded already.
5285 default: assert(0 && "Unknown icmp opcode!");
5286 case ICmpInst::ICMP_EQ:
5287 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5288 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5290 case ICmpInst::ICMP_NE:
5291 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5292 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5294 case ICmpInst::ICMP_ULT:
5295 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5296 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5297 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5298 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5299 if (RHSVal == Max) // A <u MAX -> A != MAX
5300 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5301 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5302 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5304 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5305 if (CI->isMinValue(true))
5306 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5307 ConstantInt::getAllOnesValue(Op0->getType()));
5309 case ICmpInst::ICMP_UGT:
5310 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5311 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5312 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5313 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5315 if (RHSVal == Min) // A >u MIN -> A != MIN
5316 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5317 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5318 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5320 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5321 if (CI->isMaxValue(true))
5322 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5323 ConstantInt::getNullValue(Op0->getType()));
5325 case ICmpInst::ICMP_SLT:
5326 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5327 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5328 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5329 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5330 if (RHSVal == Max) // A <s MAX -> A != MAX
5331 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5332 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5333 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5335 case ICmpInst::ICMP_SGT:
5336 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5337 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5338 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5339 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5341 if (RHSVal == Min) // A >s MIN -> A != MIN
5342 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5343 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5344 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5348 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5349 // instruction, see if that instruction also has constants so that the
5350 // instruction can be folded into the icmp
5351 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5352 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5356 // Handle icmp with constant (but not simple integer constant) RHS
5357 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5358 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5359 switch (LHSI->getOpcode()) {
5360 case Instruction::GetElementPtr:
5361 if (RHSC->isNullValue()) {
5362 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5363 bool isAllZeros = true;
5364 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5365 if (!isa<Constant>(LHSI->getOperand(i)) ||
5366 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5371 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5372 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5376 case Instruction::PHI:
5377 // Only fold icmp into the PHI if the phi and fcmp are in the same
5378 // block. If in the same block, we're encouraging jump threading. If
5379 // not, we are just pessimizing the code by making an i1 phi.
5380 if (LHSI->getParent() == I.getParent())
5381 if (Instruction *NV = FoldOpIntoPhi(I))
5384 case Instruction::Select: {
5385 // If either operand of the select is a constant, we can fold the
5386 // comparison into the select arms, which will cause one to be
5387 // constant folded and the select turned into a bitwise or.
5388 Value *Op1 = 0, *Op2 = 0;
5389 if (LHSI->hasOneUse()) {
5390 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5391 // Fold the known value into the constant operand.
5392 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5393 // Insert a new ICmp of the other select operand.
5394 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5395 LHSI->getOperand(2), RHSC,
5397 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5398 // Fold the known value into the constant operand.
5399 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5400 // Insert a new ICmp of the other select operand.
5401 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5402 LHSI->getOperand(1), RHSC,
5408 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5411 case Instruction::Malloc:
5412 // If we have (malloc != null), and if the malloc has a single use, we
5413 // can assume it is successful and remove the malloc.
5414 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5415 AddToWorkList(LHSI);
5416 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5417 !I.isTrueWhenEqual()));
5423 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5424 if (User *GEP = dyn_castGetElementPtr(Op0))
5425 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5427 if (User *GEP = dyn_castGetElementPtr(Op1))
5428 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5429 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5432 // Test to see if the operands of the icmp are casted versions of other
5433 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5435 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5436 if (isa<PointerType>(Op0->getType()) &&
5437 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5438 // We keep moving the cast from the left operand over to the right
5439 // operand, where it can often be eliminated completely.
5440 Op0 = CI->getOperand(0);
5442 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5443 // so eliminate it as well.
5444 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5445 Op1 = CI2->getOperand(0);
5447 // If Op1 is a constant, we can fold the cast into the constant.
5448 if (Op0->getType() != Op1->getType()) {
5449 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5450 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5452 // Otherwise, cast the RHS right before the icmp
5453 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5456 return new ICmpInst(I.getPredicate(), Op0, Op1);
5460 if (isa<CastInst>(Op0)) {
5461 // Handle the special case of: icmp (cast bool to X), <cst>
5462 // This comes up when you have code like
5465 // For generality, we handle any zero-extension of any operand comparison
5466 // with a constant or another cast from the same type.
5467 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5468 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5472 // See if it's the same type of instruction on the left and right.
5473 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5474 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
5475 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
5476 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
5478 switch (Op0I->getOpcode()) {
5480 case Instruction::Add:
5481 case Instruction::Sub:
5482 case Instruction::Xor:
5483 // a+x icmp eq/ne b+x --> a icmp b
5484 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
5485 Op1I->getOperand(0));
5487 case Instruction::Mul:
5488 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5489 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
5490 // Mask = -1 >> count-trailing-zeros(Cst).
5491 if (!CI->isZero() && !CI->isOne()) {
5492 const APInt &AP = CI->getValue();
5493 ConstantInt *Mask = ConstantInt::get(
5494 APInt::getLowBitsSet(AP.getBitWidth(),
5496 AP.countTrailingZeros()));
5497 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
5499 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
5501 InsertNewInstBefore(And1, I);
5502 InsertNewInstBefore(And2, I);
5503 return new ICmpInst(I.getPredicate(), And1, And2);
5512 // ~x < ~y --> y < x
5514 if (match(Op0, m_Not(m_Value(A))) &&
5515 match(Op1, m_Not(m_Value(B))))
5516 return new ICmpInst(I.getPredicate(), B, A);
5519 if (I.isEquality()) {
5520 Value *A, *B, *C, *D;
5522 // -x == -y --> x == y
5523 if (match(Op0, m_Neg(m_Value(A))) &&
5524 match(Op1, m_Neg(m_Value(B))))
5525 return new ICmpInst(I.getPredicate(), A, B);
5527 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5528 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5529 Value *OtherVal = A == Op1 ? B : A;
5530 return new ICmpInst(I.getPredicate(), OtherVal,
5531 Constant::getNullValue(A->getType()));
5534 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5535 // A^c1 == C^c2 --> A == C^(c1^c2)
5536 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5537 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5538 if (Op1->hasOneUse()) {
5539 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5540 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5541 return new ICmpInst(I.getPredicate(), A,
5542 InsertNewInstBefore(Xor, I));
5545 // A^B == A^D -> B == D
5546 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5547 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5548 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5549 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5553 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5554 (A == Op0 || B == Op0)) {
5555 // A == (A^B) -> B == 0
5556 Value *OtherVal = A == Op0 ? B : A;
5557 return new ICmpInst(I.getPredicate(), OtherVal,
5558 Constant::getNullValue(A->getType()));
5560 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5561 // (A-B) == A -> B == 0
5562 return new ICmpInst(I.getPredicate(), B,
5563 Constant::getNullValue(B->getType()));
5565 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5566 // A == (A-B) -> B == 0
5567 return new ICmpInst(I.getPredicate(), B,
5568 Constant::getNullValue(B->getType()));
5571 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5572 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5573 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5574 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5575 Value *X = 0, *Y = 0, *Z = 0;
5578 X = B; Y = D; Z = A;
5579 } else if (A == D) {
5580 X = B; Y = C; Z = A;
5581 } else if (B == C) {
5582 X = A; Y = D; Z = B;
5583 } else if (B == D) {
5584 X = A; Y = C; Z = B;
5587 if (X) { // Build (X^Y) & Z
5588 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
5589 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
5590 I.setOperand(0, Op1);
5591 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5596 return Changed ? &I : 0;
5600 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5601 /// and CmpRHS are both known to be integer constants.
5602 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5603 ConstantInt *DivRHS) {
5604 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5605 const APInt &CmpRHSV = CmpRHS->getValue();
5607 // FIXME: If the operand types don't match the type of the divide
5608 // then don't attempt this transform. The code below doesn't have the
5609 // logic to deal with a signed divide and an unsigned compare (and
5610 // vice versa). This is because (x /s C1) <s C2 produces different
5611 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5612 // (x /u C1) <u C2. Simply casting the operands and result won't
5613 // work. :( The if statement below tests that condition and bails
5615 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5616 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5618 if (DivRHS->isZero())
5619 return 0; // The ProdOV computation fails on divide by zero.
5621 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5622 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5623 // C2 (CI). By solving for X we can turn this into a range check
5624 // instead of computing a divide.
5625 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5627 // Determine if the product overflows by seeing if the product is
5628 // not equal to the divide. Make sure we do the same kind of divide
5629 // as in the LHS instruction that we're folding.
5630 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5631 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5633 // Get the ICmp opcode
5634 ICmpInst::Predicate Pred = ICI.getPredicate();
5636 // Figure out the interval that is being checked. For example, a comparison
5637 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5638 // Compute this interval based on the constants involved and the signedness of
5639 // the compare/divide. This computes a half-open interval, keeping track of
5640 // whether either value in the interval overflows. After analysis each
5641 // overflow variable is set to 0 if it's corresponding bound variable is valid
5642 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5643 int LoOverflow = 0, HiOverflow = 0;
5644 ConstantInt *LoBound = 0, *HiBound = 0;
5647 if (!DivIsSigned) { // udiv
5648 // e.g. X/5 op 3 --> [15, 20)
5650 HiOverflow = LoOverflow = ProdOV;
5652 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5653 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5654 if (CmpRHSV == 0) { // (X / pos) op 0
5655 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5656 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5658 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5659 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5660 HiOverflow = LoOverflow = ProdOV;
5662 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5663 } else { // (X / pos) op neg
5664 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5665 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5666 LoOverflow = AddWithOverflow(LoBound, Prod,
5667 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5668 HiBound = AddOne(Prod);
5669 HiOverflow = ProdOV ? -1 : 0;
5671 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5672 if (CmpRHSV == 0) { // (X / neg) op 0
5673 // e.g. X/-5 op 0 --> [-4, 5)
5674 LoBound = AddOne(DivRHS);
5675 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5676 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5677 HiOverflow = 1; // [INTMIN+1, overflow)
5678 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5680 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5681 // e.g. X/-5 op 3 --> [-19, -14)
5682 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5684 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5685 HiBound = AddOne(Prod);
5686 } else { // (X / neg) op neg
5687 // e.g. X/-5 op -3 --> [15, 20)
5689 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5690 HiBound = Subtract(Prod, DivRHS);
5693 // Dividing by a negative swaps the condition. LT <-> GT
5694 Pred = ICmpInst::getSwappedPredicate(Pred);
5697 Value *X = DivI->getOperand(0);
5699 default: assert(0 && "Unhandled icmp opcode!");
5700 case ICmpInst::ICMP_EQ:
5701 if (LoOverflow && HiOverflow)
5702 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5703 else if (HiOverflow)
5704 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5705 ICmpInst::ICMP_UGE, X, LoBound);
5706 else if (LoOverflow)
5707 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5708 ICmpInst::ICMP_ULT, X, HiBound);
5710 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5711 case ICmpInst::ICMP_NE:
5712 if (LoOverflow && HiOverflow)
5713 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5714 else if (HiOverflow)
5715 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5716 ICmpInst::ICMP_ULT, X, LoBound);
5717 else if (LoOverflow)
5718 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5719 ICmpInst::ICMP_UGE, X, HiBound);
5721 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5722 case ICmpInst::ICMP_ULT:
5723 case ICmpInst::ICMP_SLT:
5724 if (LoOverflow == +1) // Low bound is greater than input range.
5725 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5726 if (LoOverflow == -1) // Low bound is less than input range.
5727 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5728 return new ICmpInst(Pred, X, LoBound);
5729 case ICmpInst::ICMP_UGT:
5730 case ICmpInst::ICMP_SGT:
5731 if (HiOverflow == +1) // High bound greater than input range.
5732 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5733 else if (HiOverflow == -1) // High bound less than input range.
5734 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5735 if (Pred == ICmpInst::ICMP_UGT)
5736 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5738 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5743 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5745 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5748 const APInt &RHSV = RHS->getValue();
5750 switch (LHSI->getOpcode()) {
5751 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5752 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5753 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5755 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5756 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5757 Value *CompareVal = LHSI->getOperand(0);
5759 // If the sign bit of the XorCST is not set, there is no change to
5760 // the operation, just stop using the Xor.
5761 if (!XorCST->getValue().isNegative()) {
5762 ICI.setOperand(0, CompareVal);
5763 AddToWorkList(LHSI);
5767 // Was the old condition true if the operand is positive?
5768 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5770 // If so, the new one isn't.
5771 isTrueIfPositive ^= true;
5773 if (isTrueIfPositive)
5774 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5776 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5780 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5781 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5782 LHSI->getOperand(0)->hasOneUse()) {
5783 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5785 // If the LHS is an AND of a truncating cast, we can widen the
5786 // and/compare to be the input width without changing the value
5787 // produced, eliminating a cast.
5788 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5789 // We can do this transformation if either the AND constant does not
5790 // have its sign bit set or if it is an equality comparison.
5791 // Extending a relational comparison when we're checking the sign
5792 // bit would not work.
5793 if (Cast->hasOneUse() &&
5794 (ICI.isEquality() ||
5795 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5797 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5798 APInt NewCST = AndCST->getValue();
5799 NewCST.zext(BitWidth);
5801 NewCI.zext(BitWidth);
5802 Instruction *NewAnd =
5803 BinaryOperator::CreateAnd(Cast->getOperand(0),
5804 ConstantInt::get(NewCST),LHSI->getName());
5805 InsertNewInstBefore(NewAnd, ICI);
5806 return new ICmpInst(ICI.getPredicate(), NewAnd,
5807 ConstantInt::get(NewCI));
5811 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5812 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5813 // happens a LOT in code produced by the C front-end, for bitfield
5815 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5816 if (Shift && !Shift->isShift())
5820 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5821 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5822 const Type *AndTy = AndCST->getType(); // Type of the and.
5824 // We can fold this as long as we can't shift unknown bits
5825 // into the mask. This can only happen with signed shift
5826 // rights, as they sign-extend.
5828 bool CanFold = Shift->isLogicalShift();
5830 // To test for the bad case of the signed shr, see if any
5831 // of the bits shifted in could be tested after the mask.
5832 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5833 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5835 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5836 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5837 AndCST->getValue()) == 0)
5843 if (Shift->getOpcode() == Instruction::Shl)
5844 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5846 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5848 // Check to see if we are shifting out any of the bits being
5850 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5851 // If we shifted bits out, the fold is not going to work out.
5852 // As a special case, check to see if this means that the
5853 // result is always true or false now.
5854 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5855 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5856 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5857 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5859 ICI.setOperand(1, NewCst);
5860 Constant *NewAndCST;
5861 if (Shift->getOpcode() == Instruction::Shl)
5862 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5864 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5865 LHSI->setOperand(1, NewAndCST);
5866 LHSI->setOperand(0, Shift->getOperand(0));
5867 AddToWorkList(Shift); // Shift is dead.
5868 AddUsesToWorkList(ICI);
5874 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5875 // preferable because it allows the C<<Y expression to be hoisted out
5876 // of a loop if Y is invariant and X is not.
5877 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5878 ICI.isEquality() && !Shift->isArithmeticShift() &&
5879 isa<Instruction>(Shift->getOperand(0))) {
5882 if (Shift->getOpcode() == Instruction::LShr) {
5883 NS = BinaryOperator::CreateShl(AndCST,
5884 Shift->getOperand(1), "tmp");
5886 // Insert a logical shift.
5887 NS = BinaryOperator::CreateLShr(AndCST,
5888 Shift->getOperand(1), "tmp");
5890 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5892 // Compute X & (C << Y).
5893 Instruction *NewAnd =
5894 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
5895 InsertNewInstBefore(NewAnd, ICI);
5897 ICI.setOperand(0, NewAnd);
5903 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5904 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5907 uint32_t TypeBits = RHSV.getBitWidth();
5909 // Check that the shift amount is in range. If not, don't perform
5910 // undefined shifts. When the shift is visited it will be
5912 if (ShAmt->uge(TypeBits))
5915 if (ICI.isEquality()) {
5916 // If we are comparing against bits always shifted out, the
5917 // comparison cannot succeed.
5919 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5920 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5921 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5922 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5923 return ReplaceInstUsesWith(ICI, Cst);
5926 if (LHSI->hasOneUse()) {
5927 // Otherwise strength reduce the shift into an and.
5928 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5930 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5933 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5934 Mask, LHSI->getName()+".mask");
5935 Value *And = InsertNewInstBefore(AndI, ICI);
5936 return new ICmpInst(ICI.getPredicate(), And,
5937 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5941 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5942 bool TrueIfSigned = false;
5943 if (LHSI->hasOneUse() &&
5944 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5945 // (X << 31) <s 0 --> (X&1) != 0
5946 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5947 (TypeBits-ShAmt->getZExtValue()-1));
5949 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5950 Mask, LHSI->getName()+".mask");
5951 Value *And = InsertNewInstBefore(AndI, ICI);
5953 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5954 And, Constant::getNullValue(And->getType()));
5959 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5960 case Instruction::AShr: {
5961 // Only handle equality comparisons of shift-by-constant.
5962 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5963 if (!ShAmt || !ICI.isEquality()) break;
5965 // Check that the shift amount is in range. If not, don't perform
5966 // undefined shifts. When the shift is visited it will be
5968 uint32_t TypeBits = RHSV.getBitWidth();
5969 if (ShAmt->uge(TypeBits))
5972 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5974 // If we are comparing against bits always shifted out, the
5975 // comparison cannot succeed.
5976 APInt Comp = RHSV << ShAmtVal;
5977 if (LHSI->getOpcode() == Instruction::LShr)
5978 Comp = Comp.lshr(ShAmtVal);
5980 Comp = Comp.ashr(ShAmtVal);
5982 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5983 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5984 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5985 return ReplaceInstUsesWith(ICI, Cst);
5988 // Otherwise, check to see if the bits shifted out are known to be zero.
5989 // If so, we can compare against the unshifted value:
5990 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
5991 if (LHSI->hasOneUse() &&
5992 MaskedValueIsZero(LHSI->getOperand(0),
5993 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
5994 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
5995 ConstantExpr::getShl(RHS, ShAmt));
5998 if (LHSI->hasOneUse()) {
5999 // Otherwise strength reduce the shift into an and.
6000 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6001 Constant *Mask = ConstantInt::get(Val);
6004 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6005 Mask, LHSI->getName()+".mask");
6006 Value *And = InsertNewInstBefore(AndI, ICI);
6007 return new ICmpInst(ICI.getPredicate(), And,
6008 ConstantExpr::getShl(RHS, ShAmt));
6013 case Instruction::SDiv:
6014 case Instruction::UDiv:
6015 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6016 // Fold this div into the comparison, producing a range check.
6017 // Determine, based on the divide type, what the range is being
6018 // checked. If there is an overflow on the low or high side, remember
6019 // it, otherwise compute the range [low, hi) bounding the new value.
6020 // See: InsertRangeTest above for the kinds of replacements possible.
6021 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6022 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6027 case Instruction::Add:
6028 // Fold: icmp pred (add, X, C1), C2
6030 if (!ICI.isEquality()) {
6031 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6033 const APInt &LHSV = LHSC->getValue();
6035 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6038 if (ICI.isSignedPredicate()) {
6039 if (CR.getLower().isSignBit()) {
6040 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6041 ConstantInt::get(CR.getUpper()));
6042 } else if (CR.getUpper().isSignBit()) {
6043 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6044 ConstantInt::get(CR.getLower()));
6047 if (CR.getLower().isMinValue()) {
6048 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6049 ConstantInt::get(CR.getUpper()));
6050 } else if (CR.getUpper().isMinValue()) {
6051 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6052 ConstantInt::get(CR.getLower()));
6059 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6060 if (ICI.isEquality()) {
6061 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6063 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6064 // the second operand is a constant, simplify a bit.
6065 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6066 switch (BO->getOpcode()) {
6067 case Instruction::SRem:
6068 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6069 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6070 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6071 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6072 Instruction *NewRem =
6073 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6075 InsertNewInstBefore(NewRem, ICI);
6076 return new ICmpInst(ICI.getPredicate(), NewRem,
6077 Constant::getNullValue(BO->getType()));
6081 case Instruction::Add:
6082 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6083 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6084 if (BO->hasOneUse())
6085 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6086 Subtract(RHS, BOp1C));
6087 } else if (RHSV == 0) {
6088 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6089 // efficiently invertible, or if the add has just this one use.
6090 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6092 if (Value *NegVal = dyn_castNegVal(BOp1))
6093 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6094 else if (Value *NegVal = dyn_castNegVal(BOp0))
6095 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6096 else if (BO->hasOneUse()) {
6097 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6098 InsertNewInstBefore(Neg, ICI);
6100 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6104 case Instruction::Xor:
6105 // For the xor case, we can xor two constants together, eliminating
6106 // the explicit xor.
6107 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6108 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6109 ConstantExpr::getXor(RHS, BOC));
6112 case Instruction::Sub:
6113 // Replace (([sub|xor] A, B) != 0) with (A != B)
6115 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6119 case Instruction::Or:
6120 // If bits are being or'd in that are not present in the constant we
6121 // are comparing against, then the comparison could never succeed!
6122 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6123 Constant *NotCI = ConstantExpr::getNot(RHS);
6124 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6125 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6130 case Instruction::And:
6131 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6132 // If bits are being compared against that are and'd out, then the
6133 // comparison can never succeed!
6134 if ((RHSV & ~BOC->getValue()) != 0)
6135 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6138 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6139 if (RHS == BOC && RHSV.isPowerOf2())
6140 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6141 ICmpInst::ICMP_NE, LHSI,
6142 Constant::getNullValue(RHS->getType()));
6144 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6145 if (BOC->getValue().isSignBit()) {
6146 Value *X = BO->getOperand(0);
6147 Constant *Zero = Constant::getNullValue(X->getType());
6148 ICmpInst::Predicate pred = isICMP_NE ?
6149 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6150 return new ICmpInst(pred, X, Zero);
6153 // ((X & ~7) == 0) --> X < 8
6154 if (RHSV == 0 && isHighOnes(BOC)) {
6155 Value *X = BO->getOperand(0);
6156 Constant *NegX = ConstantExpr::getNeg(BOC);
6157 ICmpInst::Predicate pred = isICMP_NE ?
6158 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6159 return new ICmpInst(pred, X, NegX);
6164 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6165 // Handle icmp {eq|ne} <intrinsic>, intcst.
6166 if (II->getIntrinsicID() == Intrinsic::bswap) {
6168 ICI.setOperand(0, II->getOperand(1));
6169 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6173 } else { // Not a ICMP_EQ/ICMP_NE
6174 // If the LHS is a cast from an integral value of the same size,
6175 // then since we know the RHS is a constant, try to simlify.
6176 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6177 Value *CastOp = Cast->getOperand(0);
6178 const Type *SrcTy = CastOp->getType();
6179 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6180 if (SrcTy->isInteger() &&
6181 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6182 // If this is an unsigned comparison, try to make the comparison use
6183 // smaller constant values.
6184 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6185 // X u< 128 => X s> -1
6186 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6187 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6188 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6189 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6190 // X u> 127 => X s< 0
6191 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6192 Constant::getNullValue(SrcTy));
6200 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6201 /// We only handle extending casts so far.
6203 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6204 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6205 Value *LHSCIOp = LHSCI->getOperand(0);
6206 const Type *SrcTy = LHSCIOp->getType();
6207 const Type *DestTy = LHSCI->getType();
6210 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6211 // integer type is the same size as the pointer type.
6212 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6213 getTargetData().getPointerSizeInBits() ==
6214 cast<IntegerType>(DestTy)->getBitWidth()) {
6216 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6217 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6218 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6219 RHSOp = RHSC->getOperand(0);
6220 // If the pointer types don't match, insert a bitcast.
6221 if (LHSCIOp->getType() != RHSOp->getType())
6222 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6226 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6229 // The code below only handles extension cast instructions, so far.
6231 if (LHSCI->getOpcode() != Instruction::ZExt &&
6232 LHSCI->getOpcode() != Instruction::SExt)
6235 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6236 bool isSignedCmp = ICI.isSignedPredicate();
6238 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6239 // Not an extension from the same type?
6240 RHSCIOp = CI->getOperand(0);
6241 if (RHSCIOp->getType() != LHSCIOp->getType())
6244 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6245 // and the other is a zext), then we can't handle this.
6246 if (CI->getOpcode() != LHSCI->getOpcode())
6249 // Deal with equality cases early.
6250 if (ICI.isEquality())
6251 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6253 // A signed comparison of sign extended values simplifies into a
6254 // signed comparison.
6255 if (isSignedCmp && isSignedExt)
6256 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6258 // The other three cases all fold into an unsigned comparison.
6259 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6262 // If we aren't dealing with a constant on the RHS, exit early
6263 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6267 // Compute the constant that would happen if we truncated to SrcTy then
6268 // reextended to DestTy.
6269 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6270 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6272 // If the re-extended constant didn't change...
6274 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6275 // For example, we might have:
6276 // %A = sext short %X to uint
6277 // %B = icmp ugt uint %A, 1330
6278 // It is incorrect to transform this into
6279 // %B = icmp ugt short %X, 1330
6280 // because %A may have negative value.
6282 // However, we allow this when the compare is EQ/NE, because they are
6284 if (isSignedExt == isSignedCmp || ICI.isEquality())
6285 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6289 // The re-extended constant changed so the constant cannot be represented
6290 // in the shorter type. Consequently, we cannot emit a simple comparison.
6292 // First, handle some easy cases. We know the result cannot be equal at this
6293 // point so handle the ICI.isEquality() cases
6294 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6295 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6296 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6297 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6299 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6300 // should have been folded away previously and not enter in here.
6303 // We're performing a signed comparison.
6304 if (cast<ConstantInt>(CI)->getValue().isNegative())
6305 Result = ConstantInt::getFalse(); // X < (small) --> false
6307 Result = ConstantInt::getTrue(); // X < (large) --> true
6309 // We're performing an unsigned comparison.
6311 // We're performing an unsigned comp with a sign extended value.
6312 // This is true if the input is >= 0. [aka >s -1]
6313 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6314 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6315 NegOne, ICI.getName()), ICI);
6317 // Unsigned extend & unsigned compare -> always true.
6318 Result = ConstantInt::getTrue();
6322 // Finally, return the value computed.
6323 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6324 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6325 return ReplaceInstUsesWith(ICI, Result);
6327 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6328 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6329 "ICmp should be folded!");
6330 if (Constant *CI = dyn_cast<Constant>(Result))
6331 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6332 return BinaryOperator::CreateNot(Result);
6335 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6336 return commonShiftTransforms(I);
6339 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6340 return commonShiftTransforms(I);
6343 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6344 if (Instruction *R = commonShiftTransforms(I))
6347 Value *Op0 = I.getOperand(0);
6349 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6350 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6351 if (CSI->isAllOnesValue())
6352 return ReplaceInstUsesWith(I, CSI);
6354 // See if we can turn a signed shr into an unsigned shr.
6355 if (!isa<VectorType>(I.getType()) &&
6356 MaskedValueIsZero(Op0,
6357 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6358 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6363 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6364 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6365 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6367 // shl X, 0 == X and shr X, 0 == X
6368 // shl 0, X == 0 and shr 0, X == 0
6369 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6370 Op0 == Constant::getNullValue(Op0->getType()))
6371 return ReplaceInstUsesWith(I, Op0);
6373 if (isa<UndefValue>(Op0)) {
6374 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6375 return ReplaceInstUsesWith(I, Op0);
6376 else // undef << X -> 0, undef >>u X -> 0
6377 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6379 if (isa<UndefValue>(Op1)) {
6380 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6381 return ReplaceInstUsesWith(I, Op0);
6382 else // X << undef, X >>u undef -> 0
6383 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6386 // Try to fold constant and into select arguments.
6387 if (isa<Constant>(Op0))
6388 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6389 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6392 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6393 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6398 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6399 BinaryOperator &I) {
6400 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6402 // See if we can simplify any instructions used by the instruction whose sole
6403 // purpose is to compute bits we don't care about.
6404 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6405 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6406 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6407 KnownZero, KnownOne))
6410 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6411 // of a signed value.
6413 if (Op1->uge(TypeBits)) {
6414 if (I.getOpcode() != Instruction::AShr)
6415 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6417 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6422 // ((X*C1) << C2) == (X * (C1 << C2))
6423 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6424 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6425 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6426 return BinaryOperator::CreateMul(BO->getOperand(0),
6427 ConstantExpr::getShl(BOOp, Op1));
6429 // Try to fold constant and into select arguments.
6430 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6431 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6433 if (isa<PHINode>(Op0))
6434 if (Instruction *NV = FoldOpIntoPhi(I))
6437 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6438 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6439 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6440 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6441 // place. Don't try to do this transformation in this case. Also, we
6442 // require that the input operand is a shift-by-constant so that we have
6443 // confidence that the shifts will get folded together. We could do this
6444 // xform in more cases, but it is unlikely to be profitable.
6445 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6446 isa<ConstantInt>(TrOp->getOperand(1))) {
6447 // Okay, we'll do this xform. Make the shift of shift.
6448 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6449 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6451 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6453 // For logical shifts, the truncation has the effect of making the high
6454 // part of the register be zeros. Emulate this by inserting an AND to
6455 // clear the top bits as needed. This 'and' will usually be zapped by
6456 // other xforms later if dead.
6457 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6458 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6459 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6461 // The mask we constructed says what the trunc would do if occurring
6462 // between the shifts. We want to know the effect *after* the second
6463 // shift. We know that it is a logical shift by a constant, so adjust the
6464 // mask as appropriate.
6465 if (I.getOpcode() == Instruction::Shl)
6466 MaskV <<= Op1->getZExtValue();
6468 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6469 MaskV = MaskV.lshr(Op1->getZExtValue());
6472 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6474 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6476 // Return the value truncated to the interesting size.
6477 return new TruncInst(And, I.getType());
6481 if (Op0->hasOneUse()) {
6482 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6483 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6486 switch (Op0BO->getOpcode()) {
6488 case Instruction::Add:
6489 case Instruction::And:
6490 case Instruction::Or:
6491 case Instruction::Xor: {
6492 // These operators commute.
6493 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6494 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6495 match(Op0BO->getOperand(1),
6496 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6497 Instruction *YS = BinaryOperator::CreateShl(
6498 Op0BO->getOperand(0), Op1,
6500 InsertNewInstBefore(YS, I); // (Y << C)
6502 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6503 Op0BO->getOperand(1)->getName());
6504 InsertNewInstBefore(X, I); // (X + (Y << C))
6505 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6506 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6507 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6510 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6511 Value *Op0BOOp1 = Op0BO->getOperand(1);
6512 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6514 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6515 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6517 Instruction *YS = BinaryOperator::CreateShl(
6518 Op0BO->getOperand(0), Op1,
6520 InsertNewInstBefore(YS, I); // (Y << C)
6522 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6523 V1->getName()+".mask");
6524 InsertNewInstBefore(XM, I); // X & (CC << C)
6526 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6531 case Instruction::Sub: {
6532 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6533 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6534 match(Op0BO->getOperand(0),
6535 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6536 Instruction *YS = BinaryOperator::CreateShl(
6537 Op0BO->getOperand(1), Op1,
6539 InsertNewInstBefore(YS, I); // (Y << C)
6541 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6542 Op0BO->getOperand(0)->getName());
6543 InsertNewInstBefore(X, I); // (X + (Y << C))
6544 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6545 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6546 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6549 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6550 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6551 match(Op0BO->getOperand(0),
6552 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6553 m_ConstantInt(CC))) && V2 == Op1 &&
6554 cast<BinaryOperator>(Op0BO->getOperand(0))
6555 ->getOperand(0)->hasOneUse()) {
6556 Instruction *YS = BinaryOperator::CreateShl(
6557 Op0BO->getOperand(1), Op1,
6559 InsertNewInstBefore(YS, I); // (Y << C)
6561 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6562 V1->getName()+".mask");
6563 InsertNewInstBefore(XM, I); // X & (CC << C)
6565 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
6573 // If the operand is an bitwise operator with a constant RHS, and the
6574 // shift is the only use, we can pull it out of the shift.
6575 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6576 bool isValid = true; // Valid only for And, Or, Xor
6577 bool highBitSet = false; // Transform if high bit of constant set?
6579 switch (Op0BO->getOpcode()) {
6580 default: isValid = false; break; // Do not perform transform!
6581 case Instruction::Add:
6582 isValid = isLeftShift;
6584 case Instruction::Or:
6585 case Instruction::Xor:
6588 case Instruction::And:
6593 // If this is a signed shift right, and the high bit is modified
6594 // by the logical operation, do not perform the transformation.
6595 // The highBitSet boolean indicates the value of the high bit of
6596 // the constant which would cause it to be modified for this
6599 if (isValid && I.getOpcode() == Instruction::AShr)
6600 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6603 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6605 Instruction *NewShift =
6606 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6607 InsertNewInstBefore(NewShift, I);
6608 NewShift->takeName(Op0BO);
6610 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
6617 // Find out if this is a shift of a shift by a constant.
6618 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6619 if (ShiftOp && !ShiftOp->isShift())
6622 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6623 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6624 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6625 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6626 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6627 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6628 Value *X = ShiftOp->getOperand(0);
6630 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6631 if (AmtSum > TypeBits)
6634 const IntegerType *Ty = cast<IntegerType>(I.getType());
6636 // Check for (X << c1) << c2 and (X >> c1) >> c2
6637 if (I.getOpcode() == ShiftOp->getOpcode()) {
6638 return BinaryOperator::Create(I.getOpcode(), X,
6639 ConstantInt::get(Ty, AmtSum));
6640 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6641 I.getOpcode() == Instruction::AShr) {
6642 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6643 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
6644 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6645 I.getOpcode() == Instruction::LShr) {
6646 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6647 Instruction *Shift =
6648 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
6649 InsertNewInstBefore(Shift, I);
6651 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6652 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6655 // Okay, if we get here, one shift must be left, and the other shift must be
6656 // right. See if the amounts are equal.
6657 if (ShiftAmt1 == ShiftAmt2) {
6658 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6659 if (I.getOpcode() == Instruction::Shl) {
6660 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6661 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6663 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6664 if (I.getOpcode() == Instruction::LShr) {
6665 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6666 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6668 // We can simplify ((X << C) >>s C) into a trunc + sext.
6669 // NOTE: we could do this for any C, but that would make 'unusual' integer
6670 // types. For now, just stick to ones well-supported by the code
6672 const Type *SExtType = 0;
6673 switch (Ty->getBitWidth() - ShiftAmt1) {
6680 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6685 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6686 InsertNewInstBefore(NewTrunc, I);
6687 return new SExtInst(NewTrunc, Ty);
6689 // Otherwise, we can't handle it yet.
6690 } else if (ShiftAmt1 < ShiftAmt2) {
6691 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6693 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6694 if (I.getOpcode() == Instruction::Shl) {
6695 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6696 ShiftOp->getOpcode() == Instruction::AShr);
6697 Instruction *Shift =
6698 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6699 InsertNewInstBefore(Shift, I);
6701 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6702 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6705 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6706 if (I.getOpcode() == Instruction::LShr) {
6707 assert(ShiftOp->getOpcode() == Instruction::Shl);
6708 Instruction *Shift =
6709 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
6710 InsertNewInstBefore(Shift, I);
6712 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6713 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6716 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6718 assert(ShiftAmt2 < ShiftAmt1);
6719 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6721 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6722 if (I.getOpcode() == Instruction::Shl) {
6723 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6724 ShiftOp->getOpcode() == Instruction::AShr);
6725 Instruction *Shift =
6726 BinaryOperator::Create(ShiftOp->getOpcode(), X,
6727 ConstantInt::get(Ty, ShiftDiff));
6728 InsertNewInstBefore(Shift, I);
6730 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6731 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6734 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6735 if (I.getOpcode() == Instruction::LShr) {
6736 assert(ShiftOp->getOpcode() == Instruction::Shl);
6737 Instruction *Shift =
6738 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6739 InsertNewInstBefore(Shift, I);
6741 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6742 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6745 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6752 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6753 /// expression. If so, decompose it, returning some value X, such that Val is
6756 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6758 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6759 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6760 Offset = CI->getZExtValue();
6762 return ConstantInt::get(Type::Int32Ty, 0);
6763 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6764 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6765 if (I->getOpcode() == Instruction::Shl) {
6766 // This is a value scaled by '1 << the shift amt'.
6767 Scale = 1U << RHS->getZExtValue();
6769 return I->getOperand(0);
6770 } else if (I->getOpcode() == Instruction::Mul) {
6771 // This value is scaled by 'RHS'.
6772 Scale = RHS->getZExtValue();
6774 return I->getOperand(0);
6775 } else if (I->getOpcode() == Instruction::Add) {
6776 // We have X+C. Check to see if we really have (X*C2)+C1,
6777 // where C1 is divisible by C2.
6780 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6781 Offset += RHS->getZExtValue();
6788 // Otherwise, we can't look past this.
6795 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6796 /// try to eliminate the cast by moving the type information into the alloc.
6797 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6798 AllocationInst &AI) {
6799 const PointerType *PTy = cast<PointerType>(CI.getType());
6801 // Remove any uses of AI that are dead.
6802 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6804 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6805 Instruction *User = cast<Instruction>(*UI++);
6806 if (isInstructionTriviallyDead(User)) {
6807 while (UI != E && *UI == User)
6808 ++UI; // If this instruction uses AI more than once, don't break UI.
6811 DOUT << "IC: DCE: " << *User;
6812 EraseInstFromFunction(*User);
6816 // Get the type really allocated and the type casted to.
6817 const Type *AllocElTy = AI.getAllocatedType();
6818 const Type *CastElTy = PTy->getElementType();
6819 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6821 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6822 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6823 if (CastElTyAlign < AllocElTyAlign) return 0;
6825 // If the allocation has multiple uses, only promote it if we are strictly
6826 // increasing the alignment of the resultant allocation. If we keep it the
6827 // same, we open the door to infinite loops of various kinds.
6828 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6830 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6831 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6832 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6834 // See if we can satisfy the modulus by pulling a scale out of the array
6836 unsigned ArraySizeScale;
6838 Value *NumElements = // See if the array size is a decomposable linear expr.
6839 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6841 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6843 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6844 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6846 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6851 // If the allocation size is constant, form a constant mul expression
6852 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6853 if (isa<ConstantInt>(NumElements))
6854 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6855 // otherwise multiply the amount and the number of elements
6856 else if (Scale != 1) {
6857 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
6858 Amt = InsertNewInstBefore(Tmp, AI);
6862 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6863 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6864 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
6865 Amt = InsertNewInstBefore(Tmp, AI);
6868 AllocationInst *New;
6869 if (isa<MallocInst>(AI))
6870 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6872 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6873 InsertNewInstBefore(New, AI);
6876 // If the allocation has multiple uses, insert a cast and change all things
6877 // that used it to use the new cast. This will also hack on CI, but it will
6879 if (!AI.hasOneUse()) {
6880 AddUsesToWorkList(AI);
6881 // New is the allocation instruction, pointer typed. AI is the original
6882 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6883 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6884 InsertNewInstBefore(NewCast, AI);
6885 AI.replaceAllUsesWith(NewCast);
6887 return ReplaceInstUsesWith(CI, New);
6890 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6891 /// and return it as type Ty without inserting any new casts and without
6892 /// changing the computed value. This is used by code that tries to decide
6893 /// whether promoting or shrinking integer operations to wider or smaller types
6894 /// will allow us to eliminate a truncate or extend.
6896 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6897 /// extension operation if Ty is larger.
6899 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
6900 /// should return true if trunc(V) can be computed by computing V in the smaller
6901 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
6902 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
6903 /// efficiently truncated.
6905 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
6906 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
6907 /// the final result.
6908 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6910 int &NumCastsRemoved) {
6911 // We can always evaluate constants in another type.
6912 if (isa<ConstantInt>(V))
6915 Instruction *I = dyn_cast<Instruction>(V);
6916 if (!I) return false;
6918 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6920 // If this is an extension or truncate, we can often eliminate it.
6921 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6922 // If this is a cast from the destination type, we can trivially eliminate
6923 // it, and this will remove a cast overall.
6924 if (I->getOperand(0)->getType() == Ty) {
6925 // If the first operand is itself a cast, and is eliminable, do not count
6926 // this as an eliminable cast. We would prefer to eliminate those two
6928 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
6934 // We can't extend or shrink something that has multiple uses: doing so would
6935 // require duplicating the instruction in general, which isn't profitable.
6936 if (!I->hasOneUse()) return false;
6938 switch (I->getOpcode()) {
6939 case Instruction::Add:
6940 case Instruction::Sub:
6941 case Instruction::Mul:
6942 case Instruction::And:
6943 case Instruction::Or:
6944 case Instruction::Xor:
6945 // These operators can all arbitrarily be extended or truncated.
6946 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6948 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6951 case Instruction::Shl:
6952 // If we are truncating the result of this SHL, and if it's a shift of a
6953 // constant amount, we can always perform a SHL in a smaller type.
6954 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6955 uint32_t BitWidth = Ty->getBitWidth();
6956 if (BitWidth < OrigTy->getBitWidth() &&
6957 CI->getLimitedValue(BitWidth) < BitWidth)
6958 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6962 case Instruction::LShr:
6963 // If this is a truncate of a logical shr, we can truncate it to a smaller
6964 // lshr iff we know that the bits we would otherwise be shifting in are
6966 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6967 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6968 uint32_t BitWidth = Ty->getBitWidth();
6969 if (BitWidth < OrigBitWidth &&
6970 MaskedValueIsZero(I->getOperand(0),
6971 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6972 CI->getLimitedValue(BitWidth) < BitWidth) {
6973 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6978 case Instruction::ZExt:
6979 case Instruction::SExt:
6980 case Instruction::Trunc:
6981 // If this is the same kind of case as our original (e.g. zext+zext), we
6982 // can safely replace it. Note that replacing it does not reduce the number
6983 // of casts in the input.
6984 if (I->getOpcode() == CastOpc)
6987 case Instruction::Select: {
6988 SelectInst *SI = cast<SelectInst>(I);
6989 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
6991 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
6994 case Instruction::PHI: {
6995 // We can change a phi if we can change all operands.
6996 PHINode *PN = cast<PHINode>(I);
6997 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
6998 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7004 // TODO: Can handle more cases here.
7011 /// EvaluateInDifferentType - Given an expression that
7012 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7013 /// evaluate the expression.
7014 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7016 if (Constant *C = dyn_cast<Constant>(V))
7017 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7019 // Otherwise, it must be an instruction.
7020 Instruction *I = cast<Instruction>(V);
7021 Instruction *Res = 0;
7022 switch (I->getOpcode()) {
7023 case Instruction::Add:
7024 case Instruction::Sub:
7025 case Instruction::Mul:
7026 case Instruction::And:
7027 case Instruction::Or:
7028 case Instruction::Xor:
7029 case Instruction::AShr:
7030 case Instruction::LShr:
7031 case Instruction::Shl: {
7032 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7033 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7034 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7038 case Instruction::Trunc:
7039 case Instruction::ZExt:
7040 case Instruction::SExt:
7041 // If the source type of the cast is the type we're trying for then we can
7042 // just return the source. There's no need to insert it because it is not
7044 if (I->getOperand(0)->getType() == Ty)
7045 return I->getOperand(0);
7047 // Otherwise, must be the same type of cast, so just reinsert a new one.
7048 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7051 case Instruction::Select: {
7052 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7053 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7054 Res = SelectInst::Create(I->getOperand(0), True, False);
7057 case Instruction::PHI: {
7058 PHINode *OPN = cast<PHINode>(I);
7059 PHINode *NPN = PHINode::Create(Ty);
7060 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7061 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7062 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7068 // TODO: Can handle more cases here.
7069 assert(0 && "Unreachable!");
7074 return InsertNewInstBefore(Res, *I);
7077 /// @brief Implement the transforms common to all CastInst visitors.
7078 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7079 Value *Src = CI.getOperand(0);
7081 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7082 // eliminate it now.
7083 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7084 if (Instruction::CastOps opc =
7085 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7086 // The first cast (CSrc) is eliminable so we need to fix up or replace
7087 // the second cast (CI). CSrc will then have a good chance of being dead.
7088 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7092 // If we are casting a select then fold the cast into the select
7093 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7094 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7097 // If we are casting a PHI then fold the cast into the PHI
7098 if (isa<PHINode>(Src))
7099 if (Instruction *NV = FoldOpIntoPhi(CI))
7105 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7106 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7107 Value *Src = CI.getOperand(0);
7109 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7110 // If casting the result of a getelementptr instruction with no offset, turn
7111 // this into a cast of the original pointer!
7112 if (GEP->hasAllZeroIndices()) {
7113 // Changing the cast operand is usually not a good idea but it is safe
7114 // here because the pointer operand is being replaced with another
7115 // pointer operand so the opcode doesn't need to change.
7117 CI.setOperand(0, GEP->getOperand(0));
7121 // If the GEP has a single use, and the base pointer is a bitcast, and the
7122 // GEP computes a constant offset, see if we can convert these three
7123 // instructions into fewer. This typically happens with unions and other
7124 // non-type-safe code.
7125 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7126 if (GEP->hasAllConstantIndices()) {
7127 // We are guaranteed to get a constant from EmitGEPOffset.
7128 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7129 int64_t Offset = OffsetV->getSExtValue();
7131 // Get the base pointer input of the bitcast, and the type it points to.
7132 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7133 const Type *GEPIdxTy =
7134 cast<PointerType>(OrigBase->getType())->getElementType();
7135 if (GEPIdxTy->isSized()) {
7136 SmallVector<Value*, 8> NewIndices;
7138 // Start with the index over the outer type. Note that the type size
7139 // might be zero (even if the offset isn't zero) if the indexed type
7140 // is something like [0 x {int, int}]
7141 const Type *IntPtrTy = TD->getIntPtrType();
7142 int64_t FirstIdx = 0;
7143 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7144 FirstIdx = Offset/TySize;
7147 // Handle silly modulus not returning values values [0..TySize).
7151 assert(Offset >= 0);
7153 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7156 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7158 // Index into the types. If we fail, set OrigBase to null.
7160 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7161 const StructLayout *SL = TD->getStructLayout(STy);
7162 if (Offset < (int64_t)SL->getSizeInBytes()) {
7163 unsigned Elt = SL->getElementContainingOffset(Offset);
7164 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7166 Offset -= SL->getElementOffset(Elt);
7167 GEPIdxTy = STy->getElementType(Elt);
7169 // Otherwise, we can't index into this, bail out.
7173 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7174 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7175 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7176 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7179 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7181 GEPIdxTy = STy->getElementType();
7183 // Otherwise, we can't index into this, bail out.
7189 // If we were able to index down into an element, create the GEP
7190 // and bitcast the result. This eliminates one bitcast, potentially
7192 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7194 NewIndices.end(), "");
7195 InsertNewInstBefore(NGEP, CI);
7196 NGEP->takeName(GEP);
7198 if (isa<BitCastInst>(CI))
7199 return new BitCastInst(NGEP, CI.getType());
7200 assert(isa<PtrToIntInst>(CI));
7201 return new PtrToIntInst(NGEP, CI.getType());
7208 return commonCastTransforms(CI);
7213 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7214 /// integer types. This function implements the common transforms for all those
7216 /// @brief Implement the transforms common to CastInst with integer operands
7217 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7218 if (Instruction *Result = commonCastTransforms(CI))
7221 Value *Src = CI.getOperand(0);
7222 const Type *SrcTy = Src->getType();
7223 const Type *DestTy = CI.getType();
7224 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7225 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7227 // See if we can simplify any instructions used by the LHS whose sole
7228 // purpose is to compute bits we don't care about.
7229 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7230 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7231 KnownZero, KnownOne))
7234 // If the source isn't an instruction or has more than one use then we
7235 // can't do anything more.
7236 Instruction *SrcI = dyn_cast<Instruction>(Src);
7237 if (!SrcI || !Src->hasOneUse())
7240 // Attempt to propagate the cast into the instruction for int->int casts.
7241 int NumCastsRemoved = 0;
7242 if (!isa<BitCastInst>(CI) &&
7243 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7244 CI.getOpcode(), NumCastsRemoved)) {
7245 // If this cast is a truncate, evaluting in a different type always
7246 // eliminates the cast, so it is always a win. If this is a zero-extension,
7247 // we need to do an AND to maintain the clear top-part of the computation,
7248 // so we require that the input have eliminated at least one cast. If this
7249 // is a sign extension, we insert two new casts (to do the extension) so we
7250 // require that two casts have been eliminated.
7252 switch (CI.getOpcode()) {
7254 // All the others use floating point so we shouldn't actually
7255 // get here because of the check above.
7256 assert(0 && "Unknown cast type");
7257 case Instruction::Trunc:
7260 case Instruction::ZExt:
7261 DoXForm = NumCastsRemoved >= 1;
7263 case Instruction::SExt:
7264 DoXForm = NumCastsRemoved >= 2;
7269 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7270 CI.getOpcode() == Instruction::SExt);
7271 assert(Res->getType() == DestTy);
7272 switch (CI.getOpcode()) {
7273 default: assert(0 && "Unknown cast type!");
7274 case Instruction::Trunc:
7275 case Instruction::BitCast:
7276 // Just replace this cast with the result.
7277 return ReplaceInstUsesWith(CI, Res);
7278 case Instruction::ZExt: {
7279 // We need to emit an AND to clear the high bits.
7280 assert(SrcBitSize < DestBitSize && "Not a zext?");
7281 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7283 return BinaryOperator::CreateAnd(Res, C);
7285 case Instruction::SExt:
7286 // We need to emit a cast to truncate, then a cast to sext.
7287 return CastInst::Create(Instruction::SExt,
7288 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7294 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7295 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7297 switch (SrcI->getOpcode()) {
7298 case Instruction::Add:
7299 case Instruction::Mul:
7300 case Instruction::And:
7301 case Instruction::Or:
7302 case Instruction::Xor:
7303 // If we are discarding information, rewrite.
7304 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7305 // Don't insert two casts if they cannot be eliminated. We allow
7306 // two casts to be inserted if the sizes are the same. This could
7307 // only be converting signedness, which is a noop.
7308 if (DestBitSize == SrcBitSize ||
7309 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7310 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7311 Instruction::CastOps opcode = CI.getOpcode();
7312 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7313 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7314 return BinaryOperator::Create(
7315 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7319 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7320 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7321 SrcI->getOpcode() == Instruction::Xor &&
7322 Op1 == ConstantInt::getTrue() &&
7323 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7324 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7325 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7328 case Instruction::SDiv:
7329 case Instruction::UDiv:
7330 case Instruction::SRem:
7331 case Instruction::URem:
7332 // If we are just changing the sign, rewrite.
7333 if (DestBitSize == SrcBitSize) {
7334 // Don't insert two casts if they cannot be eliminated. We allow
7335 // two casts to be inserted if the sizes are the same. This could
7336 // only be converting signedness, which is a noop.
7337 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7338 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7339 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7341 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7343 return BinaryOperator::Create(
7344 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7349 case Instruction::Shl:
7350 // Allow changing the sign of the source operand. Do not allow
7351 // changing the size of the shift, UNLESS the shift amount is a
7352 // constant. We must not change variable sized shifts to a smaller
7353 // size, because it is undefined to shift more bits out than exist
7355 if (DestBitSize == SrcBitSize ||
7356 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7357 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7358 Instruction::BitCast : Instruction::Trunc);
7359 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7360 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7361 return BinaryOperator::CreateShl(Op0c, Op1c);
7364 case Instruction::AShr:
7365 // If this is a signed shr, and if all bits shifted in are about to be
7366 // truncated off, turn it into an unsigned shr to allow greater
7368 if (DestBitSize < SrcBitSize &&
7369 isa<ConstantInt>(Op1)) {
7370 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7371 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7372 // Insert the new logical shift right.
7373 return BinaryOperator::CreateLShr(Op0, Op1);
7381 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7382 if (Instruction *Result = commonIntCastTransforms(CI))
7385 Value *Src = CI.getOperand(0);
7386 const Type *Ty = CI.getType();
7387 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7388 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7390 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7391 switch (SrcI->getOpcode()) {
7393 case Instruction::LShr:
7394 // We can shrink lshr to something smaller if we know the bits shifted in
7395 // are already zeros.
7396 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7397 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7399 // Get a mask for the bits shifting in.
7400 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7401 Value* SrcIOp0 = SrcI->getOperand(0);
7402 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7403 if (ShAmt >= DestBitWidth) // All zeros.
7404 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7406 // Okay, we can shrink this. Truncate the input, then return a new
7408 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7409 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7411 return BinaryOperator::CreateLShr(V1, V2);
7413 } else { // This is a variable shr.
7415 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7416 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7417 // loop-invariant and CSE'd.
7418 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7419 Value *One = ConstantInt::get(SrcI->getType(), 1);
7421 Value *V = InsertNewInstBefore(
7422 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7424 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7425 SrcI->getOperand(0),
7427 Value *Zero = Constant::getNullValue(V->getType());
7428 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7438 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7439 /// in order to eliminate the icmp.
7440 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7442 // If we are just checking for a icmp eq of a single bit and zext'ing it
7443 // to an integer, then shift the bit to the appropriate place and then
7444 // cast to integer to avoid the comparison.
7445 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7446 const APInt &Op1CV = Op1C->getValue();
7448 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7449 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7450 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7451 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7452 if (!DoXform) return ICI;
7454 Value *In = ICI->getOperand(0);
7455 Value *Sh = ConstantInt::get(In->getType(),
7456 In->getType()->getPrimitiveSizeInBits()-1);
7457 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7458 In->getName()+".lobit"),
7460 if (In->getType() != CI.getType())
7461 In = CastInst::CreateIntegerCast(In, CI.getType(),
7462 false/*ZExt*/, "tmp", &CI);
7464 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7465 Constant *One = ConstantInt::get(In->getType(), 1);
7466 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7467 In->getName()+".not"),
7471 return ReplaceInstUsesWith(CI, In);
7476 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7477 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7478 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7479 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7480 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7481 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7482 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7483 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7484 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7485 // This only works for EQ and NE
7486 ICI->isEquality()) {
7487 // If Op1C some other power of two, convert:
7488 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7489 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7490 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7491 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7493 APInt KnownZeroMask(~KnownZero);
7494 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7495 if (!DoXform) return ICI;
7497 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7498 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7499 // (X&4) == 2 --> false
7500 // (X&4) != 2 --> true
7501 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7502 Res = ConstantExpr::getZExt(Res, CI.getType());
7503 return ReplaceInstUsesWith(CI, Res);
7506 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7507 Value *In = ICI->getOperand(0);
7509 // Perform a logical shr by shiftamt.
7510 // Insert the shift to put the result in the low bit.
7511 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7512 ConstantInt::get(In->getType(), ShiftAmt),
7513 In->getName()+".lobit"), CI);
7516 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7517 Constant *One = ConstantInt::get(In->getType(), 1);
7518 In = BinaryOperator::CreateXor(In, One, "tmp");
7519 InsertNewInstBefore(cast<Instruction>(In), CI);
7522 if (CI.getType() == In->getType())
7523 return ReplaceInstUsesWith(CI, In);
7525 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7533 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7534 // If one of the common conversion will work ..
7535 if (Instruction *Result = commonIntCastTransforms(CI))
7538 Value *Src = CI.getOperand(0);
7540 // If this is a cast of a cast
7541 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7542 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7543 // types and if the sizes are just right we can convert this into a logical
7544 // 'and' which will be much cheaper than the pair of casts.
7545 if (isa<TruncInst>(CSrc)) {
7546 // Get the sizes of the types involved
7547 Value *A = CSrc->getOperand(0);
7548 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7549 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7550 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7551 // If we're actually extending zero bits and the trunc is a no-op
7552 if (MidSize < DstSize && SrcSize == DstSize) {
7553 // Replace both of the casts with an And of the type mask.
7554 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7555 Constant *AndConst = ConstantInt::get(AndValue);
7557 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
7558 // Unfortunately, if the type changed, we need to cast it back.
7559 if (And->getType() != CI.getType()) {
7560 And->setName(CSrc->getName()+".mask");
7561 InsertNewInstBefore(And, CI);
7562 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
7569 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7570 return transformZExtICmp(ICI, CI);
7572 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7573 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7574 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7575 // of the (zext icmp) will be transformed.
7576 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7577 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7578 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7579 (transformZExtICmp(LHS, CI, false) ||
7580 transformZExtICmp(RHS, CI, false))) {
7581 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7582 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7583 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
7590 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7591 if (Instruction *I = commonIntCastTransforms(CI))
7594 Value *Src = CI.getOperand(0);
7596 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7597 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7598 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7599 // If we are just checking for a icmp eq of a single bit and zext'ing it
7600 // to an integer, then shift the bit to the appropriate place and then
7601 // cast to integer to avoid the comparison.
7602 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7603 const APInt &Op1CV = Op1C->getValue();
7605 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7606 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7607 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7608 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7609 Value *In = ICI->getOperand(0);
7610 Value *Sh = ConstantInt::get(In->getType(),
7611 In->getType()->getPrimitiveSizeInBits()-1);
7612 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
7613 In->getName()+".lobit"),
7615 if (In->getType() != CI.getType())
7616 In = CastInst::CreateIntegerCast(In, CI.getType(),
7617 true/*SExt*/, "tmp", &CI);
7619 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7620 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
7621 In->getName()+".not"), CI);
7623 return ReplaceInstUsesWith(CI, In);
7628 // See if the value being truncated is already sign extended. If so, just
7629 // eliminate the trunc/sext pair.
7630 if (getOpcode(Src) == Instruction::Trunc) {
7631 Value *Op = cast<User>(Src)->getOperand(0);
7632 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
7633 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
7634 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
7635 unsigned NumSignBits = ComputeNumSignBits(Op);
7637 if (OpBits == DestBits) {
7638 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7639 // bits, it is already ready.
7640 if (NumSignBits > DestBits-MidBits)
7641 return ReplaceInstUsesWith(CI, Op);
7642 } else if (OpBits < DestBits) {
7643 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7644 // bits, just sext from i32.
7645 if (NumSignBits > OpBits-MidBits)
7646 return new SExtInst(Op, CI.getType(), "tmp");
7648 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7649 // bits, just truncate to i32.
7650 if (NumSignBits > OpBits-MidBits)
7651 return new TruncInst(Op, CI.getType(), "tmp");
7658 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7659 /// in the specified FP type without changing its value.
7660 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7661 APFloat F = CFP->getValueAPF();
7662 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7663 return ConstantFP::get(F);
7667 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7668 /// through it until we get the source value.
7669 static Value *LookThroughFPExtensions(Value *V) {
7670 if (Instruction *I = dyn_cast<Instruction>(V))
7671 if (I->getOpcode() == Instruction::FPExt)
7672 return LookThroughFPExtensions(I->getOperand(0));
7674 // If this value is a constant, return the constant in the smallest FP type
7675 // that can accurately represent it. This allows us to turn
7676 // (float)((double)X+2.0) into x+2.0f.
7677 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7678 if (CFP->getType() == Type::PPC_FP128Ty)
7679 return V; // No constant folding of this.
7680 // See if the value can be truncated to float and then reextended.
7681 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7683 if (CFP->getType() == Type::DoubleTy)
7684 return V; // Won't shrink.
7685 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7687 // Don't try to shrink to various long double types.
7693 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7694 if (Instruction *I = commonCastTransforms(CI))
7697 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7698 // smaller than the destination type, we can eliminate the truncate by doing
7699 // the add as the smaller type. This applies to add/sub/mul/div as well as
7700 // many builtins (sqrt, etc).
7701 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7702 if (OpI && OpI->hasOneUse()) {
7703 switch (OpI->getOpcode()) {
7705 case Instruction::Add:
7706 case Instruction::Sub:
7707 case Instruction::Mul:
7708 case Instruction::FDiv:
7709 case Instruction::FRem:
7710 const Type *SrcTy = OpI->getType();
7711 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7712 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7713 if (LHSTrunc->getType() != SrcTy &&
7714 RHSTrunc->getType() != SrcTy) {
7715 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7716 // If the source types were both smaller than the destination type of
7717 // the cast, do this xform.
7718 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7719 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7720 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7722 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7724 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7733 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7734 return commonCastTransforms(CI);
7737 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
7738 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
7739 // mantissa to accurately represent all values of X. For example, do not
7740 // do this with i64->float->i64.
7741 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
7742 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7743 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
7744 SrcI->getType()->getFPMantissaWidth())
7745 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7747 return commonCastTransforms(FI);
7750 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
7751 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
7752 // mantissa to accurately represent all values of X. For example, do not
7753 // do this with i64->float->i64.
7754 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
7755 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7756 (int)FI.getType()->getPrimitiveSizeInBits() <=
7757 SrcI->getType()->getFPMantissaWidth())
7758 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7760 return commonCastTransforms(FI);
7763 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7764 return commonCastTransforms(CI);
7767 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7768 return commonCastTransforms(CI);
7771 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7772 return commonPointerCastTransforms(CI);
7775 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7776 if (Instruction *I = commonCastTransforms(CI))
7779 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7780 if (!DestPointee->isSized()) return 0;
7782 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7785 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7786 m_ConstantInt(Cst)))) {
7787 // If the source and destination operands have the same type, see if this
7788 // is a single-index GEP.
7789 if (X->getType() == CI.getType()) {
7790 // Get the size of the pointee type.
7791 uint64_t Size = TD->getABITypeSize(DestPointee);
7793 // Convert the constant to intptr type.
7794 APInt Offset = Cst->getValue();
7795 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7797 // If Offset is evenly divisible by Size, we can do this xform.
7798 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7799 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7800 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7803 // TODO: Could handle other cases, e.g. where add is indexing into field of
7805 } else if (CI.getOperand(0)->hasOneUse() &&
7806 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7807 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7808 // "inttoptr+GEP" instead of "add+intptr".
7810 // Get the size of the pointee type.
7811 uint64_t Size = TD->getABITypeSize(DestPointee);
7813 // Convert the constant to intptr type.
7814 APInt Offset = Cst->getValue();
7815 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7817 // If Offset is evenly divisible by Size, we can do this xform.
7818 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7819 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7821 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7823 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7829 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7830 // If the operands are integer typed then apply the integer transforms,
7831 // otherwise just apply the common ones.
7832 Value *Src = CI.getOperand(0);
7833 const Type *SrcTy = Src->getType();
7834 const Type *DestTy = CI.getType();
7836 if (SrcTy->isInteger() && DestTy->isInteger()) {
7837 if (Instruction *Result = commonIntCastTransforms(CI))
7839 } else if (isa<PointerType>(SrcTy)) {
7840 if (Instruction *I = commonPointerCastTransforms(CI))
7843 if (Instruction *Result = commonCastTransforms(CI))
7848 // Get rid of casts from one type to the same type. These are useless and can
7849 // be replaced by the operand.
7850 if (DestTy == Src->getType())
7851 return ReplaceInstUsesWith(CI, Src);
7853 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7854 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7855 const Type *DstElTy = DstPTy->getElementType();
7856 const Type *SrcElTy = SrcPTy->getElementType();
7858 // If the address spaces don't match, don't eliminate the bitcast, which is
7859 // required for changing types.
7860 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7863 // If we are casting a malloc or alloca to a pointer to a type of the same
7864 // size, rewrite the allocation instruction to allocate the "right" type.
7865 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7866 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7869 // If the source and destination are pointers, and this cast is equivalent
7870 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7871 // This can enhance SROA and other transforms that want type-safe pointers.
7872 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7873 unsigned NumZeros = 0;
7874 while (SrcElTy != DstElTy &&
7875 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7876 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7877 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7881 // If we found a path from the src to dest, create the getelementptr now.
7882 if (SrcElTy == DstElTy) {
7883 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7884 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7885 ((Instruction*) NULL));
7889 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7890 if (SVI->hasOneUse()) {
7891 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7892 // a bitconvert to a vector with the same # elts.
7893 if (isa<VectorType>(DestTy) &&
7894 cast<VectorType>(DestTy)->getNumElements() ==
7895 SVI->getType()->getNumElements()) {
7897 // If either of the operands is a cast from CI.getType(), then
7898 // evaluating the shuffle in the casted destination's type will allow
7899 // us to eliminate at least one cast.
7900 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7901 Tmp->getOperand(0)->getType() == DestTy) ||
7902 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7903 Tmp->getOperand(0)->getType() == DestTy)) {
7904 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7905 SVI->getOperand(0), DestTy, &CI);
7906 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7907 SVI->getOperand(1), DestTy, &CI);
7908 // Return a new shuffle vector. Use the same element ID's, as we
7909 // know the vector types match #elts.
7910 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7918 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7920 /// %D = select %cond, %C, %A
7922 /// %C = select %cond, %B, 0
7925 /// Assuming that the specified instruction is an operand to the select, return
7926 /// a bitmask indicating which operands of this instruction are foldable if they
7927 /// equal the other incoming value of the select.
7929 static unsigned GetSelectFoldableOperands(Instruction *I) {
7930 switch (I->getOpcode()) {
7931 case Instruction::Add:
7932 case Instruction::Mul:
7933 case Instruction::And:
7934 case Instruction::Or:
7935 case Instruction::Xor:
7936 return 3; // Can fold through either operand.
7937 case Instruction::Sub: // Can only fold on the amount subtracted.
7938 case Instruction::Shl: // Can only fold on the shift amount.
7939 case Instruction::LShr:
7940 case Instruction::AShr:
7943 return 0; // Cannot fold
7947 /// GetSelectFoldableConstant - For the same transformation as the previous
7948 /// function, return the identity constant that goes into the select.
7949 static Constant *GetSelectFoldableConstant(Instruction *I) {
7950 switch (I->getOpcode()) {
7951 default: assert(0 && "This cannot happen!"); abort();
7952 case Instruction::Add:
7953 case Instruction::Sub:
7954 case Instruction::Or:
7955 case Instruction::Xor:
7956 case Instruction::Shl:
7957 case Instruction::LShr:
7958 case Instruction::AShr:
7959 return Constant::getNullValue(I->getType());
7960 case Instruction::And:
7961 return Constant::getAllOnesValue(I->getType());
7962 case Instruction::Mul:
7963 return ConstantInt::get(I->getType(), 1);
7967 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7968 /// have the same opcode and only one use each. Try to simplify this.
7969 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7971 if (TI->getNumOperands() == 1) {
7972 // If this is a non-volatile load or a cast from the same type,
7975 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7978 return 0; // unknown unary op.
7981 // Fold this by inserting a select from the input values.
7982 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
7983 FI->getOperand(0), SI.getName()+".v");
7984 InsertNewInstBefore(NewSI, SI);
7985 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
7989 // Only handle binary operators here.
7990 if (!isa<BinaryOperator>(TI))
7993 // Figure out if the operations have any operands in common.
7994 Value *MatchOp, *OtherOpT, *OtherOpF;
7996 if (TI->getOperand(0) == FI->getOperand(0)) {
7997 MatchOp = TI->getOperand(0);
7998 OtherOpT = TI->getOperand(1);
7999 OtherOpF = FI->getOperand(1);
8000 MatchIsOpZero = true;
8001 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8002 MatchOp = TI->getOperand(1);
8003 OtherOpT = TI->getOperand(0);
8004 OtherOpF = FI->getOperand(0);
8005 MatchIsOpZero = false;
8006 } else if (!TI->isCommutative()) {
8008 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8009 MatchOp = TI->getOperand(0);
8010 OtherOpT = TI->getOperand(1);
8011 OtherOpF = FI->getOperand(0);
8012 MatchIsOpZero = true;
8013 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8014 MatchOp = TI->getOperand(1);
8015 OtherOpT = TI->getOperand(0);
8016 OtherOpF = FI->getOperand(1);
8017 MatchIsOpZero = true;
8022 // If we reach here, they do have operations in common.
8023 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8024 OtherOpF, SI.getName()+".v");
8025 InsertNewInstBefore(NewSI, SI);
8027 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8029 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8031 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8033 assert(0 && "Shouldn't get here");
8037 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8038 Value *CondVal = SI.getCondition();
8039 Value *TrueVal = SI.getTrueValue();
8040 Value *FalseVal = SI.getFalseValue();
8042 // select true, X, Y -> X
8043 // select false, X, Y -> Y
8044 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8045 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8047 // select C, X, X -> X
8048 if (TrueVal == FalseVal)
8049 return ReplaceInstUsesWith(SI, TrueVal);
8051 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8052 return ReplaceInstUsesWith(SI, FalseVal);
8053 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8054 return ReplaceInstUsesWith(SI, TrueVal);
8055 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8056 if (isa<Constant>(TrueVal))
8057 return ReplaceInstUsesWith(SI, TrueVal);
8059 return ReplaceInstUsesWith(SI, FalseVal);
8062 if (SI.getType() == Type::Int1Ty) {
8063 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8064 if (C->getZExtValue()) {
8065 // Change: A = select B, true, C --> A = or B, C
8066 return BinaryOperator::CreateOr(CondVal, FalseVal);
8068 // Change: A = select B, false, C --> A = and !B, C
8070 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8071 "not."+CondVal->getName()), SI);
8072 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8074 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8075 if (C->getZExtValue() == false) {
8076 // Change: A = select B, C, false --> A = and B, C
8077 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8079 // Change: A = select B, C, true --> A = or !B, C
8081 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8082 "not."+CondVal->getName()), SI);
8083 return BinaryOperator::CreateOr(NotCond, TrueVal);
8087 // select a, b, a -> a&b
8088 // select a, a, b -> a|b
8089 if (CondVal == TrueVal)
8090 return BinaryOperator::CreateOr(CondVal, FalseVal);
8091 else if (CondVal == FalseVal)
8092 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8095 // Selecting between two integer constants?
8096 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8097 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8098 // select C, 1, 0 -> zext C to int
8099 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8100 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8101 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8102 // select C, 0, 1 -> zext !C to int
8104 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8105 "not."+CondVal->getName()), SI);
8106 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8109 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8111 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8113 // (x <s 0) ? -1 : 0 -> ashr x, 31
8114 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8115 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8116 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8117 // The comparison constant and the result are not neccessarily the
8118 // same width. Make an all-ones value by inserting a AShr.
8119 Value *X = IC->getOperand(0);
8120 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8121 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8122 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8124 InsertNewInstBefore(SRA, SI);
8126 // Finally, convert to the type of the select RHS. We figure out
8127 // if this requires a SExt, Trunc or BitCast based on the sizes.
8128 Instruction::CastOps opc = Instruction::BitCast;
8129 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8130 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8131 if (SRASize < SISize)
8132 opc = Instruction::SExt;
8133 else if (SRASize > SISize)
8134 opc = Instruction::Trunc;
8135 return CastInst::Create(opc, SRA, SI.getType());
8140 // If one of the constants is zero (we know they can't both be) and we
8141 // have an icmp instruction with zero, and we have an 'and' with the
8142 // non-constant value, eliminate this whole mess. This corresponds to
8143 // cases like this: ((X & 27) ? 27 : 0)
8144 if (TrueValC->isZero() || FalseValC->isZero())
8145 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8146 cast<Constant>(IC->getOperand(1))->isNullValue())
8147 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8148 if (ICA->getOpcode() == Instruction::And &&
8149 isa<ConstantInt>(ICA->getOperand(1)) &&
8150 (ICA->getOperand(1) == TrueValC ||
8151 ICA->getOperand(1) == FalseValC) &&
8152 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8153 // Okay, now we know that everything is set up, we just don't
8154 // know whether we have a icmp_ne or icmp_eq and whether the
8155 // true or false val is the zero.
8156 bool ShouldNotVal = !TrueValC->isZero();
8157 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8160 V = InsertNewInstBefore(BinaryOperator::Create(
8161 Instruction::Xor, V, ICA->getOperand(1)), SI);
8162 return ReplaceInstUsesWith(SI, V);
8167 // See if we are selecting two values based on a comparison of the two values.
8168 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8169 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8170 // Transform (X == Y) ? X : Y -> Y
8171 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8172 // This is not safe in general for floating point:
8173 // consider X== -0, Y== +0.
8174 // It becomes safe if either operand is a nonzero constant.
8175 ConstantFP *CFPt, *CFPf;
8176 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8177 !CFPt->getValueAPF().isZero()) ||
8178 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8179 !CFPf->getValueAPF().isZero()))
8180 return ReplaceInstUsesWith(SI, FalseVal);
8182 // Transform (X != Y) ? X : Y -> X
8183 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8184 return ReplaceInstUsesWith(SI, TrueVal);
8185 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8187 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8188 // Transform (X == Y) ? Y : X -> X
8189 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8190 // This is not safe in general for floating point:
8191 // consider X== -0, Y== +0.
8192 // It becomes safe if either operand is a nonzero constant.
8193 ConstantFP *CFPt, *CFPf;
8194 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8195 !CFPt->getValueAPF().isZero()) ||
8196 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8197 !CFPf->getValueAPF().isZero()))
8198 return ReplaceInstUsesWith(SI, FalseVal);
8200 // Transform (X != Y) ? Y : X -> Y
8201 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8202 return ReplaceInstUsesWith(SI, TrueVal);
8203 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8207 // See if we are selecting two values based on a comparison of the two values.
8208 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8209 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8210 // Transform (X == Y) ? X : Y -> Y
8211 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8212 return ReplaceInstUsesWith(SI, FalseVal);
8213 // Transform (X != Y) ? X : Y -> X
8214 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8215 return ReplaceInstUsesWith(SI, TrueVal);
8216 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8218 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8219 // Transform (X == Y) ? Y : X -> X
8220 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8221 return ReplaceInstUsesWith(SI, FalseVal);
8222 // Transform (X != Y) ? Y : X -> Y
8223 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8224 return ReplaceInstUsesWith(SI, TrueVal);
8225 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8229 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8230 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8231 if (TI->hasOneUse() && FI->hasOneUse()) {
8232 Instruction *AddOp = 0, *SubOp = 0;
8234 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8235 if (TI->getOpcode() == FI->getOpcode())
8236 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8239 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8240 // even legal for FP.
8241 if (TI->getOpcode() == Instruction::Sub &&
8242 FI->getOpcode() == Instruction::Add) {
8243 AddOp = FI; SubOp = TI;
8244 } else if (FI->getOpcode() == Instruction::Sub &&
8245 TI->getOpcode() == Instruction::Add) {
8246 AddOp = TI; SubOp = FI;
8250 Value *OtherAddOp = 0;
8251 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8252 OtherAddOp = AddOp->getOperand(1);
8253 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8254 OtherAddOp = AddOp->getOperand(0);
8258 // So at this point we know we have (Y -> OtherAddOp):
8259 // select C, (add X, Y), (sub X, Z)
8260 Value *NegVal; // Compute -Z
8261 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8262 NegVal = ConstantExpr::getNeg(C);
8264 NegVal = InsertNewInstBefore(
8265 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8268 Value *NewTrueOp = OtherAddOp;
8269 Value *NewFalseOp = NegVal;
8271 std::swap(NewTrueOp, NewFalseOp);
8272 Instruction *NewSel =
8273 SelectInst::Create(CondVal, NewTrueOp,
8274 NewFalseOp, SI.getName() + ".p");
8276 NewSel = InsertNewInstBefore(NewSel, SI);
8277 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8282 // See if we can fold the select into one of our operands.
8283 if (SI.getType()->isInteger()) {
8284 // See the comment above GetSelectFoldableOperands for a description of the
8285 // transformation we are doing here.
8286 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8287 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8288 !isa<Constant>(FalseVal))
8289 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8290 unsigned OpToFold = 0;
8291 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8293 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8298 Constant *C = GetSelectFoldableConstant(TVI);
8299 Instruction *NewSel =
8300 SelectInst::Create(SI.getCondition(),
8301 TVI->getOperand(2-OpToFold), C);
8302 InsertNewInstBefore(NewSel, SI);
8303 NewSel->takeName(TVI);
8304 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8305 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8307 assert(0 && "Unknown instruction!!");
8312 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8313 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8314 !isa<Constant>(TrueVal))
8315 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8316 unsigned OpToFold = 0;
8317 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8319 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8324 Constant *C = GetSelectFoldableConstant(FVI);
8325 Instruction *NewSel =
8326 SelectInst::Create(SI.getCondition(), C,
8327 FVI->getOperand(2-OpToFold));
8328 InsertNewInstBefore(NewSel, SI);
8329 NewSel->takeName(FVI);
8330 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8331 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8333 assert(0 && "Unknown instruction!!");
8338 if (BinaryOperator::isNot(CondVal)) {
8339 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8340 SI.setOperand(1, FalseVal);
8341 SI.setOperand(2, TrueVal);
8348 /// EnforceKnownAlignment - If the specified pointer points to an object that
8349 /// we control, modify the object's alignment to PrefAlign. This isn't
8350 /// often possible though. If alignment is important, a more reliable approach
8351 /// is to simply align all global variables and allocation instructions to
8352 /// their preferred alignment from the beginning.
8354 static unsigned EnforceKnownAlignment(Value *V,
8355 unsigned Align, unsigned PrefAlign) {
8357 User *U = dyn_cast<User>(V);
8358 if (!U) return Align;
8360 switch (getOpcode(U)) {
8362 case Instruction::BitCast:
8363 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8364 case Instruction::GetElementPtr: {
8365 // If all indexes are zero, it is just the alignment of the base pointer.
8366 bool AllZeroOperands = true;
8367 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8368 if (!isa<Constant>(*i) ||
8369 !cast<Constant>(*i)->isNullValue()) {
8370 AllZeroOperands = false;
8374 if (AllZeroOperands) {
8375 // Treat this like a bitcast.
8376 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8382 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8383 // If there is a large requested alignment and we can, bump up the alignment
8385 if (!GV->isDeclaration()) {
8386 GV->setAlignment(PrefAlign);
8389 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8390 // If there is a requested alignment and if this is an alloca, round up. We
8391 // don't do this for malloc, because some systems can't respect the request.
8392 if (isa<AllocaInst>(AI)) {
8393 AI->setAlignment(PrefAlign);
8401 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8402 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8403 /// and it is more than the alignment of the ultimate object, see if we can
8404 /// increase the alignment of the ultimate object, making this check succeed.
8405 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8406 unsigned PrefAlign) {
8407 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8408 sizeof(PrefAlign) * CHAR_BIT;
8409 APInt Mask = APInt::getAllOnesValue(BitWidth);
8410 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8411 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8412 unsigned TrailZ = KnownZero.countTrailingOnes();
8413 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8415 if (PrefAlign > Align)
8416 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8418 // We don't need to make any adjustment.
8422 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8423 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8424 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8425 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8426 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8428 if (CopyAlign < MinAlign) {
8429 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8433 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8435 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8436 if (MemOpLength == 0) return 0;
8438 // Source and destination pointer types are always "i8*" for intrinsic. See
8439 // if the size is something we can handle with a single primitive load/store.
8440 // A single load+store correctly handles overlapping memory in the memmove
8442 unsigned Size = MemOpLength->getZExtValue();
8443 if (Size == 0) return MI; // Delete this mem transfer.
8445 if (Size > 8 || (Size&(Size-1)))
8446 return 0; // If not 1/2/4/8 bytes, exit.
8448 // Use an integer load+store unless we can find something better.
8449 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8451 // Memcpy forces the use of i8* for the source and destination. That means
8452 // that if you're using memcpy to move one double around, you'll get a cast
8453 // from double* to i8*. We'd much rather use a double load+store rather than
8454 // an i64 load+store, here because this improves the odds that the source or
8455 // dest address will be promotable. See if we can find a better type than the
8456 // integer datatype.
8457 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8458 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8459 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8460 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8461 // down through these levels if so.
8462 while (!SrcETy->isSingleValueType()) {
8463 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8464 if (STy->getNumElements() == 1)
8465 SrcETy = STy->getElementType(0);
8468 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8469 if (ATy->getNumElements() == 1)
8470 SrcETy = ATy->getElementType();
8477 if (SrcETy->isSingleValueType())
8478 NewPtrTy = PointerType::getUnqual(SrcETy);
8483 // If the memcpy/memmove provides better alignment info than we can
8485 SrcAlign = std::max(SrcAlign, CopyAlign);
8486 DstAlign = std::max(DstAlign, CopyAlign);
8488 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8489 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8490 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8491 InsertNewInstBefore(L, *MI);
8492 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8494 // Set the size of the copy to 0, it will be deleted on the next iteration.
8495 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8499 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8500 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8501 if (MI->getAlignment()->getZExtValue() < Alignment) {
8502 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8506 // Extract the length and alignment and fill if they are constant.
8507 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8508 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8509 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8511 uint64_t Len = LenC->getZExtValue();
8512 Alignment = MI->getAlignment()->getZExtValue();
8514 // If the length is zero, this is a no-op
8515 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8517 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8518 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8519 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8521 Value *Dest = MI->getDest();
8522 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8524 // Alignment 0 is identity for alignment 1 for memset, but not store.
8525 if (Alignment == 0) Alignment = 1;
8527 // Extract the fill value and store.
8528 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8529 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8532 // Set the size of the copy to 0, it will be deleted on the next iteration.
8533 MI->setLength(Constant::getNullValue(LenC->getType()));
8541 /// visitCallInst - CallInst simplification. This mostly only handles folding
8542 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8543 /// the heavy lifting.
8545 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8546 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8547 if (!II) return visitCallSite(&CI);
8549 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8551 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8552 bool Changed = false;
8554 // memmove/cpy/set of zero bytes is a noop.
8555 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8556 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8558 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8559 if (CI->getZExtValue() == 1) {
8560 // Replace the instruction with just byte operations. We would
8561 // transform other cases to loads/stores, but we don't know if
8562 // alignment is sufficient.
8566 // If we have a memmove and the source operation is a constant global,
8567 // then the source and dest pointers can't alias, so we can change this
8568 // into a call to memcpy.
8569 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8570 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8571 if (GVSrc->isConstant()) {
8572 Module *M = CI.getParent()->getParent()->getParent();
8573 Intrinsic::ID MemCpyID;
8574 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8575 MemCpyID = Intrinsic::memcpy_i32;
8577 MemCpyID = Intrinsic::memcpy_i64;
8578 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8582 // memmove(x,x,size) -> noop.
8583 if (MMI->getSource() == MMI->getDest())
8584 return EraseInstFromFunction(CI);
8587 // If we can determine a pointer alignment that is bigger than currently
8588 // set, update the alignment.
8589 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8590 if (Instruction *I = SimplifyMemTransfer(MI))
8592 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8593 if (Instruction *I = SimplifyMemSet(MSI))
8597 if (Changed) return II;
8600 switch (II->getIntrinsicID()) {
8602 case Intrinsic::bswap:
8603 // bswap(bswap(x)) -> x
8604 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
8605 if (Operand->getIntrinsicID() == Intrinsic::bswap)
8606 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
8608 case Intrinsic::ppc_altivec_lvx:
8609 case Intrinsic::ppc_altivec_lvxl:
8610 case Intrinsic::x86_sse_loadu_ps:
8611 case Intrinsic::x86_sse2_loadu_pd:
8612 case Intrinsic::x86_sse2_loadu_dq:
8613 // Turn PPC lvx -> load if the pointer is known aligned.
8614 // Turn X86 loadups -> load if the pointer is known aligned.
8615 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8616 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8617 PointerType::getUnqual(II->getType()),
8619 return new LoadInst(Ptr);
8622 case Intrinsic::ppc_altivec_stvx:
8623 case Intrinsic::ppc_altivec_stvxl:
8624 // Turn stvx -> store if the pointer is known aligned.
8625 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8626 const Type *OpPtrTy =
8627 PointerType::getUnqual(II->getOperand(1)->getType());
8628 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8629 return new StoreInst(II->getOperand(1), Ptr);
8632 case Intrinsic::x86_sse_storeu_ps:
8633 case Intrinsic::x86_sse2_storeu_pd:
8634 case Intrinsic::x86_sse2_storeu_dq:
8635 // Turn X86 storeu -> store if the pointer is known aligned.
8636 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8637 const Type *OpPtrTy =
8638 PointerType::getUnqual(II->getOperand(2)->getType());
8639 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8640 return new StoreInst(II->getOperand(2), Ptr);
8644 case Intrinsic::x86_sse_cvttss2si: {
8645 // These intrinsics only demands the 0th element of its input vector. If
8646 // we can simplify the input based on that, do so now.
8648 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8650 II->setOperand(1, V);
8656 case Intrinsic::ppc_altivec_vperm:
8657 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8658 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8659 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8661 // Check that all of the elements are integer constants or undefs.
8662 bool AllEltsOk = true;
8663 for (unsigned i = 0; i != 16; ++i) {
8664 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8665 !isa<UndefValue>(Mask->getOperand(i))) {
8672 // Cast the input vectors to byte vectors.
8673 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8674 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8675 Value *Result = UndefValue::get(Op0->getType());
8677 // Only extract each element once.
8678 Value *ExtractedElts[32];
8679 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8681 for (unsigned i = 0; i != 16; ++i) {
8682 if (isa<UndefValue>(Mask->getOperand(i)))
8684 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8685 Idx &= 31; // Match the hardware behavior.
8687 if (ExtractedElts[Idx] == 0) {
8689 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8690 InsertNewInstBefore(Elt, CI);
8691 ExtractedElts[Idx] = Elt;
8694 // Insert this value into the result vector.
8695 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
8697 InsertNewInstBefore(cast<Instruction>(Result), CI);
8699 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
8704 case Intrinsic::stackrestore: {
8705 // If the save is right next to the restore, remove the restore. This can
8706 // happen when variable allocas are DCE'd.
8707 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8708 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8709 BasicBlock::iterator BI = SS;
8711 return EraseInstFromFunction(CI);
8715 // Scan down this block to see if there is another stack restore in the
8716 // same block without an intervening call/alloca.
8717 BasicBlock::iterator BI = II;
8718 TerminatorInst *TI = II->getParent()->getTerminator();
8719 bool CannotRemove = false;
8720 for (++BI; &*BI != TI; ++BI) {
8721 if (isa<AllocaInst>(BI)) {
8722 CannotRemove = true;
8725 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
8726 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
8727 // If there is a stackrestore below this one, remove this one.
8728 if (II->getIntrinsicID() == Intrinsic::stackrestore)
8729 return EraseInstFromFunction(CI);
8730 // Otherwise, ignore the intrinsic.
8732 // If we found a non-intrinsic call, we can't remove the stack
8734 CannotRemove = true;
8740 // If the stack restore is in a return/unwind block and if there are no
8741 // allocas or calls between the restore and the return, nuke the restore.
8742 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8743 return EraseInstFromFunction(CI);
8748 return visitCallSite(II);
8751 // InvokeInst simplification
8753 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8754 return visitCallSite(&II);
8757 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8758 /// passed through the varargs area, we can eliminate the use of the cast.
8759 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8760 const CastInst * const CI,
8761 const TargetData * const TD,
8763 if (!CI->isLosslessCast())
8766 // The size of ByVal arguments is derived from the type, so we
8767 // can't change to a type with a different size. If the size were
8768 // passed explicitly we could avoid this check.
8769 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
8773 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
8774 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
8775 if (!SrcTy->isSized() || !DstTy->isSized())
8777 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
8782 // visitCallSite - Improvements for call and invoke instructions.
8784 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8785 bool Changed = false;
8787 // If the callee is a constexpr cast of a function, attempt to move the cast
8788 // to the arguments of the call/invoke.
8789 if (transformConstExprCastCall(CS)) return 0;
8791 Value *Callee = CS.getCalledValue();
8793 if (Function *CalleeF = dyn_cast<Function>(Callee))
8794 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8795 Instruction *OldCall = CS.getInstruction();
8796 // If the call and callee calling conventions don't match, this call must
8797 // be unreachable, as the call is undefined.
8798 new StoreInst(ConstantInt::getTrue(),
8799 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8801 if (!OldCall->use_empty())
8802 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8803 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8804 return EraseInstFromFunction(*OldCall);
8808 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8809 // This instruction is not reachable, just remove it. We insert a store to
8810 // undef so that we know that this code is not reachable, despite the fact
8811 // that we can't modify the CFG here.
8812 new StoreInst(ConstantInt::getTrue(),
8813 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8814 CS.getInstruction());
8816 if (!CS.getInstruction()->use_empty())
8817 CS.getInstruction()->
8818 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8820 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8821 // Don't break the CFG, insert a dummy cond branch.
8822 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8823 ConstantInt::getTrue(), II);
8825 return EraseInstFromFunction(*CS.getInstruction());
8828 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8829 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8830 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8831 return transformCallThroughTrampoline(CS);
8833 const PointerType *PTy = cast<PointerType>(Callee->getType());
8834 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8835 if (FTy->isVarArg()) {
8836 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
8837 // See if we can optimize any arguments passed through the varargs area of
8839 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8840 E = CS.arg_end(); I != E; ++I, ++ix) {
8841 CastInst *CI = dyn_cast<CastInst>(*I);
8842 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
8843 *I = CI->getOperand(0);
8849 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8850 // Inline asm calls cannot throw - mark them 'nounwind'.
8851 CS.setDoesNotThrow();
8855 return Changed ? CS.getInstruction() : 0;
8858 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8859 // attempt to move the cast to the arguments of the call/invoke.
8861 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8862 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8863 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8864 if (CE->getOpcode() != Instruction::BitCast ||
8865 !isa<Function>(CE->getOperand(0)))
8867 Function *Callee = cast<Function>(CE->getOperand(0));
8868 Instruction *Caller = CS.getInstruction();
8869 const PAListPtr &CallerPAL = CS.getParamAttrs();
8871 // Okay, this is a cast from a function to a different type. Unless doing so
8872 // would cause a type conversion of one of our arguments, change this call to
8873 // be a direct call with arguments casted to the appropriate types.
8875 const FunctionType *FT = Callee->getFunctionType();
8876 const Type *OldRetTy = Caller->getType();
8877 const Type *NewRetTy = FT->getReturnType();
8879 if (isa<StructType>(NewRetTy))
8880 return false; // TODO: Handle multiple return values.
8882 // Check to see if we are changing the return type...
8883 if (OldRetTy != NewRetTy) {
8884 if (Callee->isDeclaration() &&
8885 // Conversion is ok if changing from one pointer type to another or from
8886 // a pointer to an integer of the same size.
8887 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
8888 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
8889 return false; // Cannot transform this return value.
8891 if (!Caller->use_empty() &&
8892 // void -> non-void is handled specially
8893 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
8894 return false; // Cannot transform this return value.
8896 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8897 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8898 if (RAttrs & ParamAttr::typeIncompatible(NewRetTy))
8899 return false; // Attribute not compatible with transformed value.
8902 // If the callsite is an invoke instruction, and the return value is used by
8903 // a PHI node in a successor, we cannot change the return type of the call
8904 // because there is no place to put the cast instruction (without breaking
8905 // the critical edge). Bail out in this case.
8906 if (!Caller->use_empty())
8907 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8908 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8910 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8911 if (PN->getParent() == II->getNormalDest() ||
8912 PN->getParent() == II->getUnwindDest())
8916 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8917 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8919 CallSite::arg_iterator AI = CS.arg_begin();
8920 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8921 const Type *ParamTy = FT->getParamType(i);
8922 const Type *ActTy = (*AI)->getType();
8924 if (!CastInst::isCastable(ActTy, ParamTy))
8925 return false; // Cannot transform this parameter value.
8927 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8928 return false; // Attribute not compatible with transformed value.
8930 // Converting from one pointer type to another or between a pointer and an
8931 // integer of the same size is safe even if we do not have a body.
8932 bool isConvertible = ActTy == ParamTy ||
8933 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
8934 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
8935 if (Callee->isDeclaration() && !isConvertible) return false;
8938 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8939 Callee->isDeclaration())
8940 return false; // Do not delete arguments unless we have a function body.
8942 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
8943 !CallerPAL.isEmpty())
8944 // In this case we have more arguments than the new function type, but we
8945 // won't be dropping them. Check that these extra arguments have attributes
8946 // that are compatible with being a vararg call argument.
8947 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
8948 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
8950 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
8951 if (PAttrs & ParamAttr::VarArgsIncompatible)
8955 // Okay, we decided that this is a safe thing to do: go ahead and start
8956 // inserting cast instructions as necessary...
8957 std::vector<Value*> Args;
8958 Args.reserve(NumActualArgs);
8959 SmallVector<ParamAttrsWithIndex, 8> attrVec;
8960 attrVec.reserve(NumCommonArgs);
8962 // Get any return attributes.
8963 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8965 // If the return value is not being used, the type may not be compatible
8966 // with the existing attributes. Wipe out any problematic attributes.
8967 RAttrs &= ~ParamAttr::typeIncompatible(NewRetTy);
8969 // Add the new return attributes.
8971 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8973 AI = CS.arg_begin();
8974 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8975 const Type *ParamTy = FT->getParamType(i);
8976 if ((*AI)->getType() == ParamTy) {
8977 Args.push_back(*AI);
8979 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8980 false, ParamTy, false);
8981 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
8982 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8985 // Add any parameter attributes.
8986 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8987 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8990 // If the function takes more arguments than the call was taking, add them
8992 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8993 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8995 // If we are removing arguments to the function, emit an obnoxious warning...
8996 if (FT->getNumParams() < NumActualArgs) {
8997 if (!FT->isVarArg()) {
8998 cerr << "WARNING: While resolving call to function '"
8999 << Callee->getName() << "' arguments were dropped!\n";
9001 // Add all of the arguments in their promoted form to the arg list...
9002 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9003 const Type *PTy = getPromotedType((*AI)->getType());
9004 if (PTy != (*AI)->getType()) {
9005 // Must promote to pass through va_arg area!
9006 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9008 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9009 InsertNewInstBefore(Cast, *Caller);
9010 Args.push_back(Cast);
9012 Args.push_back(*AI);
9015 // Add any parameter attributes.
9016 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9017 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9022 if (NewRetTy == Type::VoidTy)
9023 Caller->setName(""); // Void type should not have a name.
9025 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9028 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9029 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9030 Args.begin(), Args.end(),
9031 Caller->getName(), Caller);
9032 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9033 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9035 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9036 Caller->getName(), Caller);
9037 CallInst *CI = cast<CallInst>(Caller);
9038 if (CI->isTailCall())
9039 cast<CallInst>(NC)->setTailCall();
9040 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9041 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9044 // Insert a cast of the return type as necessary.
9046 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9047 if (NV->getType() != Type::VoidTy) {
9048 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9050 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9052 // If this is an invoke instruction, we should insert it after the first
9053 // non-phi, instruction in the normal successor block.
9054 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9055 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9056 InsertNewInstBefore(NC, *I);
9058 // Otherwise, it's a call, just insert cast right after the call instr
9059 InsertNewInstBefore(NC, *Caller);
9061 AddUsersToWorkList(*Caller);
9063 NV = UndefValue::get(Caller->getType());
9067 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9068 Caller->replaceAllUsesWith(NV);
9069 Caller->eraseFromParent();
9070 RemoveFromWorkList(Caller);
9074 // transformCallThroughTrampoline - Turn a call to a function created by the
9075 // init_trampoline intrinsic into a direct call to the underlying function.
9077 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9078 Value *Callee = CS.getCalledValue();
9079 const PointerType *PTy = cast<PointerType>(Callee->getType());
9080 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9081 const PAListPtr &Attrs = CS.getParamAttrs();
9083 // If the call already has the 'nest' attribute somewhere then give up -
9084 // otherwise 'nest' would occur twice after splicing in the chain.
9085 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9088 IntrinsicInst *Tramp =
9089 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9091 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9092 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9093 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9095 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9096 if (!NestAttrs.isEmpty()) {
9097 unsigned NestIdx = 1;
9098 const Type *NestTy = 0;
9099 ParameterAttributes NestAttr = ParamAttr::None;
9101 // Look for a parameter marked with the 'nest' attribute.
9102 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9103 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9104 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9105 // Record the parameter type and any other attributes.
9107 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9112 Instruction *Caller = CS.getInstruction();
9113 std::vector<Value*> NewArgs;
9114 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9116 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9117 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9119 // Insert the nest argument into the call argument list, which may
9120 // mean appending it. Likewise for attributes.
9122 // Add any function result attributes.
9123 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9124 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9128 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9130 if (Idx == NestIdx) {
9131 // Add the chain argument and attributes.
9132 Value *NestVal = Tramp->getOperand(3);
9133 if (NestVal->getType() != NestTy)
9134 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9135 NewArgs.push_back(NestVal);
9136 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9142 // Add the original argument and attributes.
9143 NewArgs.push_back(*I);
9144 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9146 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9152 // The trampoline may have been bitcast to a bogus type (FTy).
9153 // Handle this by synthesizing a new function type, equal to FTy
9154 // with the chain parameter inserted.
9156 std::vector<const Type*> NewTypes;
9157 NewTypes.reserve(FTy->getNumParams()+1);
9159 // Insert the chain's type into the list of parameter types, which may
9160 // mean appending it.
9163 FunctionType::param_iterator I = FTy->param_begin(),
9164 E = FTy->param_end();
9168 // Add the chain's type.
9169 NewTypes.push_back(NestTy);
9174 // Add the original type.
9175 NewTypes.push_back(*I);
9181 // Replace the trampoline call with a direct call. Let the generic
9182 // code sort out any function type mismatches.
9183 FunctionType *NewFTy =
9184 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9185 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9186 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9187 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9189 Instruction *NewCaller;
9190 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9191 NewCaller = InvokeInst::Create(NewCallee,
9192 II->getNormalDest(), II->getUnwindDest(),
9193 NewArgs.begin(), NewArgs.end(),
9194 Caller->getName(), Caller);
9195 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9196 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9198 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9199 Caller->getName(), Caller);
9200 if (cast<CallInst>(Caller)->isTailCall())
9201 cast<CallInst>(NewCaller)->setTailCall();
9202 cast<CallInst>(NewCaller)->
9203 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9204 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9206 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9207 Caller->replaceAllUsesWith(NewCaller);
9208 Caller->eraseFromParent();
9209 RemoveFromWorkList(Caller);
9214 // Replace the trampoline call with a direct call. Since there is no 'nest'
9215 // parameter, there is no need to adjust the argument list. Let the generic
9216 // code sort out any function type mismatches.
9217 Constant *NewCallee =
9218 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9219 CS.setCalledFunction(NewCallee);
9220 return CS.getInstruction();
9223 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9224 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9225 /// and a single binop.
9226 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9227 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9228 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9229 isa<CmpInst>(FirstInst));
9230 unsigned Opc = FirstInst->getOpcode();
9231 Value *LHSVal = FirstInst->getOperand(0);
9232 Value *RHSVal = FirstInst->getOperand(1);
9234 const Type *LHSType = LHSVal->getType();
9235 const Type *RHSType = RHSVal->getType();
9237 // Scan to see if all operands are the same opcode, all have one use, and all
9238 // kill their operands (i.e. the operands have one use).
9239 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9240 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9241 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9242 // Verify type of the LHS matches so we don't fold cmp's of different
9243 // types or GEP's with different index types.
9244 I->getOperand(0)->getType() != LHSType ||
9245 I->getOperand(1)->getType() != RHSType)
9248 // If they are CmpInst instructions, check their predicates
9249 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9250 if (cast<CmpInst>(I)->getPredicate() !=
9251 cast<CmpInst>(FirstInst)->getPredicate())
9254 // Keep track of which operand needs a phi node.
9255 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9256 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9259 // Otherwise, this is safe to transform, determine if it is profitable.
9261 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9262 // Indexes are often folded into load/store instructions, so we don't want to
9263 // hide them behind a phi.
9264 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9267 Value *InLHS = FirstInst->getOperand(0);
9268 Value *InRHS = FirstInst->getOperand(1);
9269 PHINode *NewLHS = 0, *NewRHS = 0;
9271 NewLHS = PHINode::Create(LHSType,
9272 FirstInst->getOperand(0)->getName() + ".pn");
9273 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9274 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9275 InsertNewInstBefore(NewLHS, PN);
9280 NewRHS = PHINode::Create(RHSType,
9281 FirstInst->getOperand(1)->getName() + ".pn");
9282 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9283 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9284 InsertNewInstBefore(NewRHS, PN);
9288 // Add all operands to the new PHIs.
9289 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9291 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9292 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9295 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9296 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9300 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9301 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9302 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9303 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9306 assert(isa<GetElementPtrInst>(FirstInst));
9307 return GetElementPtrInst::Create(LHSVal, RHSVal);
9311 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9312 /// of the block that defines it. This means that it must be obvious the value
9313 /// of the load is not changed from the point of the load to the end of the
9316 /// Finally, it is safe, but not profitable, to sink a load targetting a
9317 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9319 static bool isSafeToSinkLoad(LoadInst *L) {
9320 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9322 for (++BBI; BBI != E; ++BBI)
9323 if (BBI->mayWriteToMemory())
9326 // Check for non-address taken alloca. If not address-taken already, it isn't
9327 // profitable to do this xform.
9328 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9329 bool isAddressTaken = false;
9330 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9332 if (isa<LoadInst>(UI)) continue;
9333 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9334 // If storing TO the alloca, then the address isn't taken.
9335 if (SI->getOperand(1) == AI) continue;
9337 isAddressTaken = true;
9341 if (!isAddressTaken)
9349 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9350 // operator and they all are only used by the PHI, PHI together their
9351 // inputs, and do the operation once, to the result of the PHI.
9352 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9353 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9355 // Scan the instruction, looking for input operations that can be folded away.
9356 // If all input operands to the phi are the same instruction (e.g. a cast from
9357 // the same type or "+42") we can pull the operation through the PHI, reducing
9358 // code size and simplifying code.
9359 Constant *ConstantOp = 0;
9360 const Type *CastSrcTy = 0;
9361 bool isVolatile = false;
9362 if (isa<CastInst>(FirstInst)) {
9363 CastSrcTy = FirstInst->getOperand(0)->getType();
9364 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9365 // Can fold binop, compare or shift here if the RHS is a constant,
9366 // otherwise call FoldPHIArgBinOpIntoPHI.
9367 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9368 if (ConstantOp == 0)
9369 return FoldPHIArgBinOpIntoPHI(PN);
9370 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9371 isVolatile = LI->isVolatile();
9372 // We can't sink the load if the loaded value could be modified between the
9373 // load and the PHI.
9374 if (LI->getParent() != PN.getIncomingBlock(0) ||
9375 !isSafeToSinkLoad(LI))
9378 // If the PHI is of volatile loads and the load block has multiple
9379 // successors, sinking it would remove a load of the volatile value from
9380 // the path through the other successor.
9382 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9385 } else if (isa<GetElementPtrInst>(FirstInst)) {
9386 if (FirstInst->getNumOperands() == 2)
9387 return FoldPHIArgBinOpIntoPHI(PN);
9388 // Can't handle general GEPs yet.
9391 return 0; // Cannot fold this operation.
9394 // Check to see if all arguments are the same operation.
9395 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9396 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9397 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9398 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9401 if (I->getOperand(0)->getType() != CastSrcTy)
9402 return 0; // Cast operation must match.
9403 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9404 // We can't sink the load if the loaded value could be modified between
9405 // the load and the PHI.
9406 if (LI->isVolatile() != isVolatile ||
9407 LI->getParent() != PN.getIncomingBlock(i) ||
9408 !isSafeToSinkLoad(LI))
9411 // If the PHI is of volatile loads and the load block has multiple
9412 // successors, sinking it would remove a load of the volatile value from
9413 // the path through the other successor.
9415 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9419 } else if (I->getOperand(1) != ConstantOp) {
9424 // Okay, they are all the same operation. Create a new PHI node of the
9425 // correct type, and PHI together all of the LHS's of the instructions.
9426 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9427 PN.getName()+".in");
9428 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9430 Value *InVal = FirstInst->getOperand(0);
9431 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9433 // Add all operands to the new PHI.
9434 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9435 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9436 if (NewInVal != InVal)
9438 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9443 // The new PHI unions all of the same values together. This is really
9444 // common, so we handle it intelligently here for compile-time speed.
9448 InsertNewInstBefore(NewPN, PN);
9452 // Insert and return the new operation.
9453 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9454 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9455 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9456 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9457 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9458 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9459 PhiVal, ConstantOp);
9460 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9462 // If this was a volatile load that we are merging, make sure to loop through
9463 // and mark all the input loads as non-volatile. If we don't do this, we will
9464 // insert a new volatile load and the old ones will not be deletable.
9466 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9467 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9469 return new LoadInst(PhiVal, "", isVolatile);
9472 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9474 static bool DeadPHICycle(PHINode *PN,
9475 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9476 if (PN->use_empty()) return true;
9477 if (!PN->hasOneUse()) return false;
9479 // Remember this node, and if we find the cycle, return.
9480 if (!PotentiallyDeadPHIs.insert(PN))
9483 // Don't scan crazily complex things.
9484 if (PotentiallyDeadPHIs.size() == 16)
9487 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9488 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9493 /// PHIsEqualValue - Return true if this phi node is always equal to
9494 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9495 /// z = some value; x = phi (y, z); y = phi (x, z)
9496 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9497 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9498 // See if we already saw this PHI node.
9499 if (!ValueEqualPHIs.insert(PN))
9502 // Don't scan crazily complex things.
9503 if (ValueEqualPHIs.size() == 16)
9506 // Scan the operands to see if they are either phi nodes or are equal to
9508 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9509 Value *Op = PN->getIncomingValue(i);
9510 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9511 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9513 } else if (Op != NonPhiInVal)
9521 // PHINode simplification
9523 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9524 // If LCSSA is around, don't mess with Phi nodes
9525 if (MustPreserveLCSSA) return 0;
9527 if (Value *V = PN.hasConstantValue())
9528 return ReplaceInstUsesWith(PN, V);
9530 // If all PHI operands are the same operation, pull them through the PHI,
9531 // reducing code size.
9532 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9533 PN.getIncomingValue(0)->hasOneUse())
9534 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9537 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9538 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9539 // PHI)... break the cycle.
9540 if (PN.hasOneUse()) {
9541 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9542 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9543 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9544 PotentiallyDeadPHIs.insert(&PN);
9545 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9546 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9549 // If this phi has a single use, and if that use just computes a value for
9550 // the next iteration of a loop, delete the phi. This occurs with unused
9551 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9552 // common case here is good because the only other things that catch this
9553 // are induction variable analysis (sometimes) and ADCE, which is only run
9555 if (PHIUser->hasOneUse() &&
9556 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9557 PHIUser->use_back() == &PN) {
9558 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9562 // We sometimes end up with phi cycles that non-obviously end up being the
9563 // same value, for example:
9564 // z = some value; x = phi (y, z); y = phi (x, z)
9565 // where the phi nodes don't necessarily need to be in the same block. Do a
9566 // quick check to see if the PHI node only contains a single non-phi value, if
9567 // so, scan to see if the phi cycle is actually equal to that value.
9569 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9570 // Scan for the first non-phi operand.
9571 while (InValNo != NumOperandVals &&
9572 isa<PHINode>(PN.getIncomingValue(InValNo)))
9575 if (InValNo != NumOperandVals) {
9576 Value *NonPhiInVal = PN.getOperand(InValNo);
9578 // Scan the rest of the operands to see if there are any conflicts, if so
9579 // there is no need to recursively scan other phis.
9580 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9581 Value *OpVal = PN.getIncomingValue(InValNo);
9582 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9586 // If we scanned over all operands, then we have one unique value plus
9587 // phi values. Scan PHI nodes to see if they all merge in each other or
9589 if (InValNo == NumOperandVals) {
9590 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9591 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9592 return ReplaceInstUsesWith(PN, NonPhiInVal);
9599 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9600 Instruction *InsertPoint,
9602 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9603 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9604 // We must cast correctly to the pointer type. Ensure that we
9605 // sign extend the integer value if it is smaller as this is
9606 // used for address computation.
9607 Instruction::CastOps opcode =
9608 (VTySize < PtrSize ? Instruction::SExt :
9609 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9610 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9614 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9615 Value *PtrOp = GEP.getOperand(0);
9616 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9617 // If so, eliminate the noop.
9618 if (GEP.getNumOperands() == 1)
9619 return ReplaceInstUsesWith(GEP, PtrOp);
9621 if (isa<UndefValue>(GEP.getOperand(0)))
9622 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9624 bool HasZeroPointerIndex = false;
9625 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9626 HasZeroPointerIndex = C->isNullValue();
9628 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9629 return ReplaceInstUsesWith(GEP, PtrOp);
9631 // Eliminate unneeded casts for indices.
9632 bool MadeChange = false;
9634 gep_type_iterator GTI = gep_type_begin(GEP);
9635 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
9636 i != e; ++i, ++GTI) {
9637 if (isa<SequentialType>(*GTI)) {
9638 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
9639 if (CI->getOpcode() == Instruction::ZExt ||
9640 CI->getOpcode() == Instruction::SExt) {
9641 const Type *SrcTy = CI->getOperand(0)->getType();
9642 // We can eliminate a cast from i32 to i64 iff the target
9643 // is a 32-bit pointer target.
9644 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9646 *i = CI->getOperand(0);
9650 // If we are using a wider index than needed for this platform, shrink it
9651 // to what we need. If the incoming value needs a cast instruction,
9652 // insert it. This explicit cast can make subsequent optimizations more
9655 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9656 if (Constant *C = dyn_cast<Constant>(Op)) {
9657 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
9660 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9668 if (MadeChange) return &GEP;
9670 // If this GEP instruction doesn't move the pointer, and if the input operand
9671 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9672 // real input to the dest type.
9673 if (GEP.hasAllZeroIndices()) {
9674 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9675 // If the bitcast is of an allocation, and the allocation will be
9676 // converted to match the type of the cast, don't touch this.
9677 if (isa<AllocationInst>(BCI->getOperand(0))) {
9678 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9679 if (Instruction *I = visitBitCast(*BCI)) {
9682 BCI->getParent()->getInstList().insert(BCI, I);
9683 ReplaceInstUsesWith(*BCI, I);
9688 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9692 // Combine Indices - If the source pointer to this getelementptr instruction
9693 // is a getelementptr instruction, combine the indices of the two
9694 // getelementptr instructions into a single instruction.
9696 SmallVector<Value*, 8> SrcGEPOperands;
9697 if (User *Src = dyn_castGetElementPtr(PtrOp))
9698 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9700 if (!SrcGEPOperands.empty()) {
9701 // Note that if our source is a gep chain itself that we wait for that
9702 // chain to be resolved before we perform this transformation. This
9703 // avoids us creating a TON of code in some cases.
9705 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9706 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9707 return 0; // Wait until our source is folded to completion.
9709 SmallVector<Value*, 8> Indices;
9711 // Find out whether the last index in the source GEP is a sequential idx.
9712 bool EndsWithSequential = false;
9713 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9714 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9715 EndsWithSequential = !isa<StructType>(*I);
9717 // Can we combine the two pointer arithmetics offsets?
9718 if (EndsWithSequential) {
9719 // Replace: gep (gep %P, long B), long A, ...
9720 // With: T = long A+B; gep %P, T, ...
9722 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9723 if (SO1 == Constant::getNullValue(SO1->getType())) {
9725 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9728 // If they aren't the same type, convert both to an integer of the
9729 // target's pointer size.
9730 if (SO1->getType() != GO1->getType()) {
9731 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9732 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9733 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9734 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9736 unsigned PS = TD->getPointerSizeInBits();
9737 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9738 // Convert GO1 to SO1's type.
9739 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9741 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9742 // Convert SO1 to GO1's type.
9743 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9745 const Type *PT = TD->getIntPtrType();
9746 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9747 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9751 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9752 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9754 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
9755 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9759 // Recycle the GEP we already have if possible.
9760 if (SrcGEPOperands.size() == 2) {
9761 GEP.setOperand(0, SrcGEPOperands[0]);
9762 GEP.setOperand(1, Sum);
9765 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9766 SrcGEPOperands.end()-1);
9767 Indices.push_back(Sum);
9768 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9770 } else if (isa<Constant>(*GEP.idx_begin()) &&
9771 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9772 SrcGEPOperands.size() != 1) {
9773 // Otherwise we can do the fold if the first index of the GEP is a zero
9774 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9775 SrcGEPOperands.end());
9776 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9779 if (!Indices.empty())
9780 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9781 Indices.end(), GEP.getName());
9783 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9784 // GEP of global variable. If all of the indices for this GEP are
9785 // constants, we can promote this to a constexpr instead of an instruction.
9787 // Scan for nonconstants...
9788 SmallVector<Constant*, 8> Indices;
9789 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9790 for (; I != E && isa<Constant>(*I); ++I)
9791 Indices.push_back(cast<Constant>(*I));
9793 if (I == E) { // If they are all constants...
9794 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9795 &Indices[0],Indices.size());
9797 // Replace all uses of the GEP with the new constexpr...
9798 return ReplaceInstUsesWith(GEP, CE);
9800 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9801 if (!isa<PointerType>(X->getType())) {
9802 // Not interesting. Source pointer must be a cast from pointer.
9803 } else if (HasZeroPointerIndex) {
9804 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9805 // into : GEP [10 x i8]* X, i32 0, ...
9807 // This occurs when the program declares an array extern like "int X[];"
9809 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9810 const PointerType *XTy = cast<PointerType>(X->getType());
9811 if (const ArrayType *XATy =
9812 dyn_cast<ArrayType>(XTy->getElementType()))
9813 if (const ArrayType *CATy =
9814 dyn_cast<ArrayType>(CPTy->getElementType()))
9815 if (CATy->getElementType() == XATy->getElementType()) {
9816 // At this point, we know that the cast source type is a pointer
9817 // to an array of the same type as the destination pointer
9818 // array. Because the array type is never stepped over (there
9819 // is a leading zero) we can fold the cast into this GEP.
9820 GEP.setOperand(0, X);
9823 } else if (GEP.getNumOperands() == 2) {
9824 // Transform things like:
9825 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9826 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9827 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9828 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9829 if (isa<ArrayType>(SrcElTy) &&
9830 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9831 TD->getABITypeSize(ResElTy)) {
9833 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9834 Idx[1] = GEP.getOperand(1);
9835 Value *V = InsertNewInstBefore(
9836 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9837 // V and GEP are both pointer types --> BitCast
9838 return new BitCastInst(V, GEP.getType());
9841 // Transform things like:
9842 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9843 // (where tmp = 8*tmp2) into:
9844 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9846 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9847 uint64_t ArrayEltSize =
9848 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9850 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9851 // allow either a mul, shift, or constant here.
9853 ConstantInt *Scale = 0;
9854 if (ArrayEltSize == 1) {
9855 NewIdx = GEP.getOperand(1);
9856 Scale = ConstantInt::get(NewIdx->getType(), 1);
9857 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9858 NewIdx = ConstantInt::get(CI->getType(), 1);
9860 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9861 if (Inst->getOpcode() == Instruction::Shl &&
9862 isa<ConstantInt>(Inst->getOperand(1))) {
9863 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9864 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9865 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9866 NewIdx = Inst->getOperand(0);
9867 } else if (Inst->getOpcode() == Instruction::Mul &&
9868 isa<ConstantInt>(Inst->getOperand(1))) {
9869 Scale = cast<ConstantInt>(Inst->getOperand(1));
9870 NewIdx = Inst->getOperand(0);
9874 // If the index will be to exactly the right offset with the scale taken
9875 // out, perform the transformation. Note, we don't know whether Scale is
9876 // signed or not. We'll use unsigned version of division/modulo
9877 // operation after making sure Scale doesn't have the sign bit set.
9878 if (Scale && Scale->getSExtValue() >= 0LL &&
9879 Scale->getZExtValue() % ArrayEltSize == 0) {
9880 Scale = ConstantInt::get(Scale->getType(),
9881 Scale->getZExtValue() / ArrayEltSize);
9882 if (Scale->getZExtValue() != 1) {
9883 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9885 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
9886 NewIdx = InsertNewInstBefore(Sc, GEP);
9889 // Insert the new GEP instruction.
9891 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9893 Instruction *NewGEP =
9894 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9895 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9896 // The NewGEP must be pointer typed, so must the old one -> BitCast
9897 return new BitCastInst(NewGEP, GEP.getType());
9906 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9907 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9908 if (AI.isArrayAllocation()) { // Check C != 1
9909 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9911 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9912 AllocationInst *New = 0;
9914 // Create and insert the replacement instruction...
9915 if (isa<MallocInst>(AI))
9916 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9918 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9919 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9922 InsertNewInstBefore(New, AI);
9924 // Scan to the end of the allocation instructions, to skip over a block of
9925 // allocas if possible...
9927 BasicBlock::iterator It = New;
9928 while (isa<AllocationInst>(*It)) ++It;
9930 // Now that I is pointing to the first non-allocation-inst in the block,
9931 // insert our getelementptr instruction...
9933 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9937 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9938 New->getName()+".sub", It);
9940 // Now make everything use the getelementptr instead of the original
9942 return ReplaceInstUsesWith(AI, V);
9943 } else if (isa<UndefValue>(AI.getArraySize())) {
9944 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9948 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9949 // Note that we only do this for alloca's, because malloc should allocate and
9950 // return a unique pointer, even for a zero byte allocation.
9951 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9952 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9953 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9958 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9959 Value *Op = FI.getOperand(0);
9961 // free undef -> unreachable.
9962 if (isa<UndefValue>(Op)) {
9963 // Insert a new store to null because we cannot modify the CFG here.
9964 new StoreInst(ConstantInt::getTrue(),
9965 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9966 return EraseInstFromFunction(FI);
9969 // If we have 'free null' delete the instruction. This can happen in stl code
9970 // when lots of inlining happens.
9971 if (isa<ConstantPointerNull>(Op))
9972 return EraseInstFromFunction(FI);
9974 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9975 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9976 FI.setOperand(0, CI->getOperand(0));
9980 // Change free (gep X, 0,0,0,0) into free(X)
9981 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9982 if (GEPI->hasAllZeroIndices()) {
9983 AddToWorkList(GEPI);
9984 FI.setOperand(0, GEPI->getOperand(0));
9989 // Change free(malloc) into nothing, if the malloc has a single use.
9990 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9991 if (MI->hasOneUse()) {
9992 EraseInstFromFunction(FI);
9993 return EraseInstFromFunction(*MI);
10000 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10001 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10002 const TargetData *TD) {
10003 User *CI = cast<User>(LI.getOperand(0));
10004 Value *CastOp = CI->getOperand(0);
10006 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10007 // Instead of loading constant c string, use corresponding integer value
10008 // directly if string length is small enough.
10010 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10011 unsigned len = Str.length();
10012 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10013 unsigned numBits = Ty->getPrimitiveSizeInBits();
10014 // Replace LI with immediate integer store.
10015 if ((numBits >> 3) == len + 1) {
10016 APInt StrVal(numBits, 0);
10017 APInt SingleChar(numBits, 0);
10018 if (TD->isLittleEndian()) {
10019 for (signed i = len-1; i >= 0; i--) {
10020 SingleChar = (uint64_t) Str[i];
10021 StrVal = (StrVal << 8) | SingleChar;
10024 for (unsigned i = 0; i < len; i++) {
10025 SingleChar = (uint64_t) Str[i];
10026 StrVal = (StrVal << 8) | SingleChar;
10028 // Append NULL at the end.
10030 StrVal = (StrVal << 8) | SingleChar;
10032 Value *NL = ConstantInt::get(StrVal);
10033 return IC.ReplaceInstUsesWith(LI, NL);
10038 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10039 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10040 const Type *SrcPTy = SrcTy->getElementType();
10042 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10043 isa<VectorType>(DestPTy)) {
10044 // If the source is an array, the code below will not succeed. Check to
10045 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10047 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10048 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10049 if (ASrcTy->getNumElements() != 0) {
10051 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10052 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10053 SrcTy = cast<PointerType>(CastOp->getType());
10054 SrcPTy = SrcTy->getElementType();
10057 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10058 isa<VectorType>(SrcPTy)) &&
10059 // Do not allow turning this into a load of an integer, which is then
10060 // casted to a pointer, this pessimizes pointer analysis a lot.
10061 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10062 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10063 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10065 // Okay, we are casting from one integer or pointer type to another of
10066 // the same size. Instead of casting the pointer before the load, cast
10067 // the result of the loaded value.
10068 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10070 LI.isVolatile()),LI);
10071 // Now cast the result of the load.
10072 return new BitCastInst(NewLoad, LI.getType());
10079 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10080 /// from this value cannot trap. If it is not obviously safe to load from the
10081 /// specified pointer, we do a quick local scan of the basic block containing
10082 /// ScanFrom, to determine if the address is already accessed.
10083 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10084 // If it is an alloca it is always safe to load from.
10085 if (isa<AllocaInst>(V)) return true;
10087 // If it is a global variable it is mostly safe to load from.
10088 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10089 // Don't try to evaluate aliases. External weak GV can be null.
10090 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10092 // Otherwise, be a little bit agressive by scanning the local block where we
10093 // want to check to see if the pointer is already being loaded or stored
10094 // from/to. If so, the previous load or store would have already trapped,
10095 // so there is no harm doing an extra load (also, CSE will later eliminate
10096 // the load entirely).
10097 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10102 // If we see a free or a call (which might do a free) the pointer could be
10104 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10107 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10108 if (LI->getOperand(0) == V) return true;
10109 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10110 if (SI->getOperand(1) == V) return true;
10117 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10118 /// until we find the underlying object a pointer is referring to or something
10119 /// we don't understand. Note that the returned pointer may be offset from the
10120 /// input, because we ignore GEP indices.
10121 static Value *GetUnderlyingObject(Value *Ptr) {
10123 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10124 if (CE->getOpcode() == Instruction::BitCast ||
10125 CE->getOpcode() == Instruction::GetElementPtr)
10126 Ptr = CE->getOperand(0);
10129 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10130 Ptr = BCI->getOperand(0);
10131 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10132 Ptr = GEP->getOperand(0);
10139 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10140 Value *Op = LI.getOperand(0);
10142 // Attempt to improve the alignment.
10143 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10145 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10146 LI.getAlignment()))
10147 LI.setAlignment(KnownAlign);
10149 // load (cast X) --> cast (load X) iff safe
10150 if (isa<CastInst>(Op))
10151 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10154 // None of the following transforms are legal for volatile loads.
10155 if (LI.isVolatile()) return 0;
10157 if (&LI.getParent()->front() != &LI) {
10158 BasicBlock::iterator BBI = &LI; --BBI;
10159 // If the instruction immediately before this is a store to the same
10160 // address, do a simple form of store->load forwarding.
10161 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10162 if (SI->getOperand(1) == LI.getOperand(0))
10163 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10164 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10165 if (LIB->getOperand(0) == LI.getOperand(0))
10166 return ReplaceInstUsesWith(LI, LIB);
10169 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10170 const Value *GEPI0 = GEPI->getOperand(0);
10171 // TODO: Consider a target hook for valid address spaces for this xform.
10172 if (isa<ConstantPointerNull>(GEPI0) &&
10173 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10174 // Insert a new store to null instruction before the load to indicate
10175 // that this code is not reachable. We do this instead of inserting
10176 // an unreachable instruction directly because we cannot modify the
10178 new StoreInst(UndefValue::get(LI.getType()),
10179 Constant::getNullValue(Op->getType()), &LI);
10180 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10184 if (Constant *C = dyn_cast<Constant>(Op)) {
10185 // load null/undef -> undef
10186 // TODO: Consider a target hook for valid address spaces for this xform.
10187 if (isa<UndefValue>(C) || (C->isNullValue() &&
10188 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10189 // Insert a new store to null instruction before the load to indicate that
10190 // this code is not reachable. We do this instead of inserting an
10191 // unreachable instruction directly because we cannot modify the CFG.
10192 new StoreInst(UndefValue::get(LI.getType()),
10193 Constant::getNullValue(Op->getType()), &LI);
10194 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10197 // Instcombine load (constant global) into the value loaded.
10198 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10199 if (GV->isConstant() && !GV->isDeclaration())
10200 return ReplaceInstUsesWith(LI, GV->getInitializer());
10202 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10203 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10204 if (CE->getOpcode() == Instruction::GetElementPtr) {
10205 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10206 if (GV->isConstant() && !GV->isDeclaration())
10208 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10209 return ReplaceInstUsesWith(LI, V);
10210 if (CE->getOperand(0)->isNullValue()) {
10211 // Insert a new store to null instruction before the load to indicate
10212 // that this code is not reachable. We do this instead of inserting
10213 // an unreachable instruction directly because we cannot modify the
10215 new StoreInst(UndefValue::get(LI.getType()),
10216 Constant::getNullValue(Op->getType()), &LI);
10217 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10220 } else if (CE->isCast()) {
10221 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10227 // If this load comes from anywhere in a constant global, and if the global
10228 // is all undef or zero, we know what it loads.
10229 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10230 if (GV->isConstant() && GV->hasInitializer()) {
10231 if (GV->getInitializer()->isNullValue())
10232 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10233 else if (isa<UndefValue>(GV->getInitializer()))
10234 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10238 if (Op->hasOneUse()) {
10239 // Change select and PHI nodes to select values instead of addresses: this
10240 // helps alias analysis out a lot, allows many others simplifications, and
10241 // exposes redundancy in the code.
10243 // Note that we cannot do the transformation unless we know that the
10244 // introduced loads cannot trap! Something like this is valid as long as
10245 // the condition is always false: load (select bool %C, int* null, int* %G),
10246 // but it would not be valid if we transformed it to load from null
10247 // unconditionally.
10249 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10250 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10251 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10252 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10253 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10254 SI->getOperand(1)->getName()+".val"), LI);
10255 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10256 SI->getOperand(2)->getName()+".val"), LI);
10257 return SelectInst::Create(SI->getCondition(), V1, V2);
10260 // load (select (cond, null, P)) -> load P
10261 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10262 if (C->isNullValue()) {
10263 LI.setOperand(0, SI->getOperand(2));
10267 // load (select (cond, P, null)) -> load P
10268 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10269 if (C->isNullValue()) {
10270 LI.setOperand(0, SI->getOperand(1));
10278 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10280 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10281 User *CI = cast<User>(SI.getOperand(1));
10282 Value *CastOp = CI->getOperand(0);
10284 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10285 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10286 const Type *SrcPTy = SrcTy->getElementType();
10288 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10289 // If the source is an array, the code below will not succeed. Check to
10290 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10292 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10293 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10294 if (ASrcTy->getNumElements() != 0) {
10296 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10297 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10298 SrcTy = cast<PointerType>(CastOp->getType());
10299 SrcPTy = SrcTy->getElementType();
10302 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10303 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10304 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10306 // Okay, we are casting from one integer or pointer type to another of
10307 // the same size. Instead of casting the pointer before
10308 // the store, cast the value to be stored.
10310 Value *SIOp0 = SI.getOperand(0);
10311 Instruction::CastOps opcode = Instruction::BitCast;
10312 const Type* CastSrcTy = SIOp0->getType();
10313 const Type* CastDstTy = SrcPTy;
10314 if (isa<PointerType>(CastDstTy)) {
10315 if (CastSrcTy->isInteger())
10316 opcode = Instruction::IntToPtr;
10317 } else if (isa<IntegerType>(CastDstTy)) {
10318 if (isa<PointerType>(SIOp0->getType()))
10319 opcode = Instruction::PtrToInt;
10321 if (Constant *C = dyn_cast<Constant>(SIOp0))
10322 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10324 NewCast = IC.InsertNewInstBefore(
10325 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10327 return new StoreInst(NewCast, CastOp);
10334 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10335 Value *Val = SI.getOperand(0);
10336 Value *Ptr = SI.getOperand(1);
10338 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10339 EraseInstFromFunction(SI);
10344 // If the RHS is an alloca with a single use, zapify the store, making the
10346 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10347 if (isa<AllocaInst>(Ptr)) {
10348 EraseInstFromFunction(SI);
10353 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10354 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10355 GEP->getOperand(0)->hasOneUse()) {
10356 EraseInstFromFunction(SI);
10362 // Attempt to improve the alignment.
10363 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10365 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10366 SI.getAlignment()))
10367 SI.setAlignment(KnownAlign);
10369 // Do really simple DSE, to catch cases where there are several consequtive
10370 // stores to the same location, separated by a few arithmetic operations. This
10371 // situation often occurs with bitfield accesses.
10372 BasicBlock::iterator BBI = &SI;
10373 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10377 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10378 // Prev store isn't volatile, and stores to the same location?
10379 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10382 EraseInstFromFunction(*PrevSI);
10388 // If this is a load, we have to stop. However, if the loaded value is from
10389 // the pointer we're loading and is producing the pointer we're storing,
10390 // then *this* store is dead (X = load P; store X -> P).
10391 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10392 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10393 EraseInstFromFunction(SI);
10397 // Otherwise, this is a load from some other location. Stores before it
10398 // may not be dead.
10402 // Don't skip over loads or things that can modify memory.
10403 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10408 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10410 // store X, null -> turns into 'unreachable' in SimplifyCFG
10411 if (isa<ConstantPointerNull>(Ptr)) {
10412 if (!isa<UndefValue>(Val)) {
10413 SI.setOperand(0, UndefValue::get(Val->getType()));
10414 if (Instruction *U = dyn_cast<Instruction>(Val))
10415 AddToWorkList(U); // Dropped a use.
10418 return 0; // Do not modify these!
10421 // store undef, Ptr -> noop
10422 if (isa<UndefValue>(Val)) {
10423 EraseInstFromFunction(SI);
10428 // If the pointer destination is a cast, see if we can fold the cast into the
10430 if (isa<CastInst>(Ptr))
10431 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10433 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10435 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10439 // If this store is the last instruction in the basic block, and if the block
10440 // ends with an unconditional branch, try to move it to the successor block.
10442 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10443 if (BI->isUnconditional())
10444 if (SimplifyStoreAtEndOfBlock(SI))
10445 return 0; // xform done!
10450 /// SimplifyStoreAtEndOfBlock - Turn things like:
10451 /// if () { *P = v1; } else { *P = v2 }
10452 /// into a phi node with a store in the successor.
10454 /// Simplify things like:
10455 /// *P = v1; if () { *P = v2; }
10456 /// into a phi node with a store in the successor.
10458 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10459 BasicBlock *StoreBB = SI.getParent();
10461 // Check to see if the successor block has exactly two incoming edges. If
10462 // so, see if the other predecessor contains a store to the same location.
10463 // if so, insert a PHI node (if needed) and move the stores down.
10464 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10466 // Determine whether Dest has exactly two predecessors and, if so, compute
10467 // the other predecessor.
10468 pred_iterator PI = pred_begin(DestBB);
10469 BasicBlock *OtherBB = 0;
10470 if (*PI != StoreBB)
10473 if (PI == pred_end(DestBB))
10476 if (*PI != StoreBB) {
10481 if (++PI != pred_end(DestBB))
10484 // Bail out if all the relevant blocks aren't distinct (this can happen,
10485 // for example, if SI is in an infinite loop)
10486 if (StoreBB == DestBB || OtherBB == DestBB)
10489 // Verify that the other block ends in a branch and is not otherwise empty.
10490 BasicBlock::iterator BBI = OtherBB->getTerminator();
10491 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10492 if (!OtherBr || BBI == OtherBB->begin())
10495 // If the other block ends in an unconditional branch, check for the 'if then
10496 // else' case. there is an instruction before the branch.
10497 StoreInst *OtherStore = 0;
10498 if (OtherBr->isUnconditional()) {
10499 // If this isn't a store, or isn't a store to the same location, bail out.
10501 OtherStore = dyn_cast<StoreInst>(BBI);
10502 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10505 // Otherwise, the other block ended with a conditional branch. If one of the
10506 // destinations is StoreBB, then we have the if/then case.
10507 if (OtherBr->getSuccessor(0) != StoreBB &&
10508 OtherBr->getSuccessor(1) != StoreBB)
10511 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10512 // if/then triangle. See if there is a store to the same ptr as SI that
10513 // lives in OtherBB.
10515 // Check to see if we find the matching store.
10516 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10517 if (OtherStore->getOperand(1) != SI.getOperand(1))
10521 // If we find something that may be using or overwriting the stored
10522 // value, or if we run out of instructions, we can't do the xform.
10523 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
10524 BBI == OtherBB->begin())
10528 // In order to eliminate the store in OtherBr, we have to
10529 // make sure nothing reads or overwrites the stored value in
10531 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10532 // FIXME: This should really be AA driven.
10533 if (I->mayReadFromMemory() || I->mayWriteToMemory())
10538 // Insert a PHI node now if we need it.
10539 Value *MergedVal = OtherStore->getOperand(0);
10540 if (MergedVal != SI.getOperand(0)) {
10541 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10542 PN->reserveOperandSpace(2);
10543 PN->addIncoming(SI.getOperand(0), SI.getParent());
10544 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10545 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10548 // Advance to a place where it is safe to insert the new store and
10550 BBI = DestBB->getFirstNonPHI();
10551 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10552 OtherStore->isVolatile()), *BBI);
10554 // Nuke the old stores.
10555 EraseInstFromFunction(SI);
10556 EraseInstFromFunction(*OtherStore);
10562 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10563 // Change br (not X), label True, label False to: br X, label False, True
10565 BasicBlock *TrueDest;
10566 BasicBlock *FalseDest;
10567 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10568 !isa<Constant>(X)) {
10569 // Swap Destinations and condition...
10570 BI.setCondition(X);
10571 BI.setSuccessor(0, FalseDest);
10572 BI.setSuccessor(1, TrueDest);
10576 // Cannonicalize fcmp_one -> fcmp_oeq
10577 FCmpInst::Predicate FPred; Value *Y;
10578 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10579 TrueDest, FalseDest)))
10580 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10581 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10582 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10583 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10584 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10585 NewSCC->takeName(I);
10586 // Swap Destinations and condition...
10587 BI.setCondition(NewSCC);
10588 BI.setSuccessor(0, FalseDest);
10589 BI.setSuccessor(1, TrueDest);
10590 RemoveFromWorkList(I);
10591 I->eraseFromParent();
10592 AddToWorkList(NewSCC);
10596 // Cannonicalize icmp_ne -> icmp_eq
10597 ICmpInst::Predicate IPred;
10598 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10599 TrueDest, FalseDest)))
10600 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10601 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10602 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10603 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10604 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10605 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10606 NewSCC->takeName(I);
10607 // Swap Destinations and condition...
10608 BI.setCondition(NewSCC);
10609 BI.setSuccessor(0, FalseDest);
10610 BI.setSuccessor(1, TrueDest);
10611 RemoveFromWorkList(I);
10612 I->eraseFromParent();;
10613 AddToWorkList(NewSCC);
10620 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10621 Value *Cond = SI.getCondition();
10622 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10623 if (I->getOpcode() == Instruction::Add)
10624 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10625 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10626 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10627 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10629 SI.setOperand(0, I->getOperand(0));
10637 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
10638 Value *Agg = EV.getAggregateOperand();
10640 if (!EV.hasIndices())
10641 return ReplaceInstUsesWith(EV, Agg);
10643 if (Constant *C = dyn_cast<Constant>(Agg)) {
10644 if (isa<UndefValue>(C))
10645 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
10647 if (isa<ConstantAggregateZero>(C))
10648 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
10650 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
10651 // Extract the element indexed by the first index out of the constant
10652 Value *V = C->getOperand(*EV.idx_begin());
10653 if (EV.getNumIndices() > 1)
10654 // Extract the remaining indices out of the constant indexed by the
10656 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
10658 return ReplaceInstUsesWith(EV, V);
10660 return 0; // Can't handle other constants
10662 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
10663 // We're extracting from an insertvalue instruction, compare the indices
10664 const unsigned *exti, *exte, *insi, *inse;
10665 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
10666 exte = EV.idx_end(), inse = IV->idx_end();
10667 exti != exte && insi != inse;
10669 if (*insi != *exti)
10670 // The insert and extract both reference distinctly different elements.
10671 // This means the extract is not influenced by the insert, and we can
10672 // replace the aggregate operand of the extract with the aggregate
10673 // operand of the insert. i.e., replace
10674 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
10675 // %E = extractvalue { i32, { i32 } } %I, 0
10677 // %E = extractvalue { i32, { i32 } } %A, 0
10678 return ExtractValueInst::Create(IV->getAggregateOperand(),
10679 EV.idx_begin(), EV.idx_end());
10681 if (exti == exte && insi == inse)
10682 // Both iterators are at the end: Index lists are identical. Replace
10683 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
10684 // %C = extractvalue { i32, { i32 } } %B, 1, 0
10686 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
10687 if (exti == exte) {
10688 // The extract list is a prefix of the insert list. i.e. replace
10689 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
10690 // %E = extractvalue { i32, { i32 } } %I, 1
10692 // %X = extractvalue { i32, { i32 } } %A, 1
10693 // %E = insertvalue { i32 } %X, i32 42, 0
10694 // by switching the order of the insert and extract (though the
10695 // insertvalue should be left in, since it may have other uses).
10696 Value *NewEV = InsertNewInstBefore(
10697 ExtractValueInst::Create(IV->getAggregateOperand(),
10698 EV.idx_begin(), EV.idx_end()),
10700 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
10704 // The insert list is a prefix of the extract list
10705 // We can simply remove the common indices from the extract and make it
10706 // operate on the inserted value instead of the insertvalue result.
10708 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
10709 // %E = extractvalue { i32, { i32 } } %I, 1, 0
10711 // %E extractvalue { i32 } { i32 42 }, 0
10712 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
10715 // Can't simplify extracts from other values. Note that nested extracts are
10716 // already simplified implicitely by the above (extract ( extract (insert) )
10717 // will be translated into extract ( insert ( extract ) ) first and then just
10718 // the value inserted, if appropriate).
10722 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10723 /// is to leave as a vector operation.
10724 static bool CheapToScalarize(Value *V, bool isConstant) {
10725 if (isa<ConstantAggregateZero>(V))
10727 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10728 if (isConstant) return true;
10729 // If all elts are the same, we can extract.
10730 Constant *Op0 = C->getOperand(0);
10731 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10732 if (C->getOperand(i) != Op0)
10736 Instruction *I = dyn_cast<Instruction>(V);
10737 if (!I) return false;
10739 // Insert element gets simplified to the inserted element or is deleted if
10740 // this is constant idx extract element and its a constant idx insertelt.
10741 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10742 isa<ConstantInt>(I->getOperand(2)))
10744 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10746 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10747 if (BO->hasOneUse() &&
10748 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10749 CheapToScalarize(BO->getOperand(1), isConstant)))
10751 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10752 if (CI->hasOneUse() &&
10753 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10754 CheapToScalarize(CI->getOperand(1), isConstant)))
10760 /// Read and decode a shufflevector mask.
10762 /// It turns undef elements into values that are larger than the number of
10763 /// elements in the input.
10764 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10765 unsigned NElts = SVI->getType()->getNumElements();
10766 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10767 return std::vector<unsigned>(NElts, 0);
10768 if (isa<UndefValue>(SVI->getOperand(2)))
10769 return std::vector<unsigned>(NElts, 2*NElts);
10771 std::vector<unsigned> Result;
10772 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10773 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
10774 if (isa<UndefValue>(*i))
10775 Result.push_back(NElts*2); // undef -> 8
10777 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
10781 /// FindScalarElement - Given a vector and an element number, see if the scalar
10782 /// value is already around as a register, for example if it were inserted then
10783 /// extracted from the vector.
10784 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10785 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10786 const VectorType *PTy = cast<VectorType>(V->getType());
10787 unsigned Width = PTy->getNumElements();
10788 if (EltNo >= Width) // Out of range access.
10789 return UndefValue::get(PTy->getElementType());
10791 if (isa<UndefValue>(V))
10792 return UndefValue::get(PTy->getElementType());
10793 else if (isa<ConstantAggregateZero>(V))
10794 return Constant::getNullValue(PTy->getElementType());
10795 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10796 return CP->getOperand(EltNo);
10797 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10798 // If this is an insert to a variable element, we don't know what it is.
10799 if (!isa<ConstantInt>(III->getOperand(2)))
10801 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10803 // If this is an insert to the element we are looking for, return the
10805 if (EltNo == IIElt)
10806 return III->getOperand(1);
10808 // Otherwise, the insertelement doesn't modify the value, recurse on its
10810 return FindScalarElement(III->getOperand(0), EltNo);
10811 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10812 unsigned InEl = getShuffleMask(SVI)[EltNo];
10814 return FindScalarElement(SVI->getOperand(0), InEl);
10815 else if (InEl < Width*2)
10816 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10818 return UndefValue::get(PTy->getElementType());
10821 // Otherwise, we don't know.
10825 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10826 // If vector val is undef, replace extract with scalar undef.
10827 if (isa<UndefValue>(EI.getOperand(0)))
10828 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10830 // If vector val is constant 0, replace extract with scalar 0.
10831 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10832 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10834 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10835 // If vector val is constant with all elements the same, replace EI with
10836 // that element. When the elements are not identical, we cannot replace yet
10837 // (we do that below, but only when the index is constant).
10838 Constant *op0 = C->getOperand(0);
10839 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10840 if (C->getOperand(i) != op0) {
10845 return ReplaceInstUsesWith(EI, op0);
10848 // If extracting a specified index from the vector, see if we can recursively
10849 // find a previously computed scalar that was inserted into the vector.
10850 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10851 unsigned IndexVal = IdxC->getZExtValue();
10852 unsigned VectorWidth =
10853 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10855 // If this is extracting an invalid index, turn this into undef, to avoid
10856 // crashing the code below.
10857 if (IndexVal >= VectorWidth)
10858 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10860 // This instruction only demands the single element from the input vector.
10861 // If the input vector has a single use, simplify it based on this use
10863 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10864 uint64_t UndefElts;
10865 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10868 EI.setOperand(0, V);
10873 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10874 return ReplaceInstUsesWith(EI, Elt);
10876 // If the this extractelement is directly using a bitcast from a vector of
10877 // the same number of elements, see if we can find the source element from
10878 // it. In this case, we will end up needing to bitcast the scalars.
10879 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10880 if (const VectorType *VT =
10881 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10882 if (VT->getNumElements() == VectorWidth)
10883 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10884 return new BitCastInst(Elt, EI.getType());
10888 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10889 if (I->hasOneUse()) {
10890 // Push extractelement into predecessor operation if legal and
10891 // profitable to do so
10892 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10893 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10894 if (CheapToScalarize(BO, isConstantElt)) {
10895 ExtractElementInst *newEI0 =
10896 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10897 EI.getName()+".lhs");
10898 ExtractElementInst *newEI1 =
10899 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10900 EI.getName()+".rhs");
10901 InsertNewInstBefore(newEI0, EI);
10902 InsertNewInstBefore(newEI1, EI);
10903 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
10905 } else if (isa<LoadInst>(I)) {
10907 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10908 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10909 PointerType::get(EI.getType(), AS),EI);
10910 GetElementPtrInst *GEP =
10911 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
10912 InsertNewInstBefore(GEP, EI);
10913 return new LoadInst(GEP);
10916 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10917 // Extracting the inserted element?
10918 if (IE->getOperand(2) == EI.getOperand(1))
10919 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10920 // If the inserted and extracted elements are constants, they must not
10921 // be the same value, extract from the pre-inserted value instead.
10922 if (isa<Constant>(IE->getOperand(2)) &&
10923 isa<Constant>(EI.getOperand(1))) {
10924 AddUsesToWorkList(EI);
10925 EI.setOperand(0, IE->getOperand(0));
10928 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10929 // If this is extracting an element from a shufflevector, figure out where
10930 // it came from and extract from the appropriate input element instead.
10931 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10932 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10934 if (SrcIdx < SVI->getType()->getNumElements())
10935 Src = SVI->getOperand(0);
10936 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10937 SrcIdx -= SVI->getType()->getNumElements();
10938 Src = SVI->getOperand(1);
10940 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10942 return new ExtractElementInst(Src, SrcIdx);
10949 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10950 /// elements from either LHS or RHS, return the shuffle mask and true.
10951 /// Otherwise, return false.
10952 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10953 std::vector<Constant*> &Mask) {
10954 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10955 "Invalid CollectSingleShuffleElements");
10956 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10958 if (isa<UndefValue>(V)) {
10959 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10961 } else if (V == LHS) {
10962 for (unsigned i = 0; i != NumElts; ++i)
10963 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10965 } else if (V == RHS) {
10966 for (unsigned i = 0; i != NumElts; ++i)
10967 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10969 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10970 // If this is an insert of an extract from some other vector, include it.
10971 Value *VecOp = IEI->getOperand(0);
10972 Value *ScalarOp = IEI->getOperand(1);
10973 Value *IdxOp = IEI->getOperand(2);
10975 if (!isa<ConstantInt>(IdxOp))
10977 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10979 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10980 // Okay, we can handle this if the vector we are insertinting into is
10981 // transitively ok.
10982 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10983 // If so, update the mask to reflect the inserted undef.
10984 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10987 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10988 if (isa<ConstantInt>(EI->getOperand(1)) &&
10989 EI->getOperand(0)->getType() == V->getType()) {
10990 unsigned ExtractedIdx =
10991 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10993 // This must be extracting from either LHS or RHS.
10994 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10995 // Okay, we can handle this if the vector we are insertinting into is
10996 // transitively ok.
10997 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10998 // If so, update the mask to reflect the inserted value.
10999 if (EI->getOperand(0) == LHS) {
11000 Mask[InsertedIdx & (NumElts-1)] =
11001 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11003 assert(EI->getOperand(0) == RHS);
11004 Mask[InsertedIdx & (NumElts-1)] =
11005 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11014 // TODO: Handle shufflevector here!
11019 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11020 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11021 /// that computes V and the LHS value of the shuffle.
11022 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11024 assert(isa<VectorType>(V->getType()) &&
11025 (RHS == 0 || V->getType() == RHS->getType()) &&
11026 "Invalid shuffle!");
11027 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11029 if (isa<UndefValue>(V)) {
11030 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11032 } else if (isa<ConstantAggregateZero>(V)) {
11033 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11035 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11036 // If this is an insert of an extract from some other vector, include it.
11037 Value *VecOp = IEI->getOperand(0);
11038 Value *ScalarOp = IEI->getOperand(1);
11039 Value *IdxOp = IEI->getOperand(2);
11041 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11042 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11043 EI->getOperand(0)->getType() == V->getType()) {
11044 unsigned ExtractedIdx =
11045 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11046 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11048 // Either the extracted from or inserted into vector must be RHSVec,
11049 // otherwise we'd end up with a shuffle of three inputs.
11050 if (EI->getOperand(0) == RHS || RHS == 0) {
11051 RHS = EI->getOperand(0);
11052 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11053 Mask[InsertedIdx & (NumElts-1)] =
11054 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11058 if (VecOp == RHS) {
11059 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11060 // Everything but the extracted element is replaced with the RHS.
11061 for (unsigned i = 0; i != NumElts; ++i) {
11062 if (i != InsertedIdx)
11063 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11068 // If this insertelement is a chain that comes from exactly these two
11069 // vectors, return the vector and the effective shuffle.
11070 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11071 return EI->getOperand(0);
11076 // TODO: Handle shufflevector here!
11078 // Otherwise, can't do anything fancy. Return an identity vector.
11079 for (unsigned i = 0; i != NumElts; ++i)
11080 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11084 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11085 Value *VecOp = IE.getOperand(0);
11086 Value *ScalarOp = IE.getOperand(1);
11087 Value *IdxOp = IE.getOperand(2);
11089 // Inserting an undef or into an undefined place, remove this.
11090 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11091 ReplaceInstUsesWith(IE, VecOp);
11093 // If the inserted element was extracted from some other vector, and if the
11094 // indexes are constant, try to turn this into a shufflevector operation.
11095 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11096 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11097 EI->getOperand(0)->getType() == IE.getType()) {
11098 unsigned NumVectorElts = IE.getType()->getNumElements();
11099 unsigned ExtractedIdx =
11100 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11101 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11103 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11104 return ReplaceInstUsesWith(IE, VecOp);
11106 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11107 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11109 // If we are extracting a value from a vector, then inserting it right
11110 // back into the same place, just use the input vector.
11111 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11112 return ReplaceInstUsesWith(IE, VecOp);
11114 // We could theoretically do this for ANY input. However, doing so could
11115 // turn chains of insertelement instructions into a chain of shufflevector
11116 // instructions, and right now we do not merge shufflevectors. As such,
11117 // only do this in a situation where it is clear that there is benefit.
11118 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11119 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11120 // the values of VecOp, except then one read from EIOp0.
11121 // Build a new shuffle mask.
11122 std::vector<Constant*> Mask;
11123 if (isa<UndefValue>(VecOp))
11124 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11126 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11127 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11130 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11131 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11132 ConstantVector::get(Mask));
11135 // If this insertelement isn't used by some other insertelement, turn it
11136 // (and any insertelements it points to), into one big shuffle.
11137 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11138 std::vector<Constant*> Mask;
11140 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11141 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11142 // We now have a shuffle of LHS, RHS, Mask.
11143 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11152 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11153 Value *LHS = SVI.getOperand(0);
11154 Value *RHS = SVI.getOperand(1);
11155 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11157 bool MadeChange = false;
11159 // Undefined shuffle mask -> undefined value.
11160 if (isa<UndefValue>(SVI.getOperand(2)))
11161 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11163 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11164 // the undef, change them to undefs.
11165 if (isa<UndefValue>(SVI.getOperand(1))) {
11166 // Scan to see if there are any references to the RHS. If so, replace them
11167 // with undef element refs and set MadeChange to true.
11168 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11169 if (Mask[i] >= e && Mask[i] != 2*e) {
11176 // Remap any references to RHS to use LHS.
11177 std::vector<Constant*> Elts;
11178 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11179 if (Mask[i] == 2*e)
11180 Elts.push_back(UndefValue::get(Type::Int32Ty));
11182 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11184 SVI.setOperand(2, ConstantVector::get(Elts));
11188 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11189 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11190 if (LHS == RHS || isa<UndefValue>(LHS)) {
11191 if (isa<UndefValue>(LHS) && LHS == RHS) {
11192 // shuffle(undef,undef,mask) -> undef.
11193 return ReplaceInstUsesWith(SVI, LHS);
11196 // Remap any references to RHS to use LHS.
11197 std::vector<Constant*> Elts;
11198 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11199 if (Mask[i] >= 2*e)
11200 Elts.push_back(UndefValue::get(Type::Int32Ty));
11202 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11203 (Mask[i] < e && isa<UndefValue>(LHS)))
11204 Mask[i] = 2*e; // Turn into undef.
11206 Mask[i] &= (e-1); // Force to LHS.
11207 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11210 SVI.setOperand(0, SVI.getOperand(1));
11211 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11212 SVI.setOperand(2, ConstantVector::get(Elts));
11213 LHS = SVI.getOperand(0);
11214 RHS = SVI.getOperand(1);
11218 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11219 bool isLHSID = true, isRHSID = true;
11221 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11222 if (Mask[i] >= e*2) continue; // Ignore undef values.
11223 // Is this an identity shuffle of the LHS value?
11224 isLHSID &= (Mask[i] == i);
11226 // Is this an identity shuffle of the RHS value?
11227 isRHSID &= (Mask[i]-e == i);
11230 // Eliminate identity shuffles.
11231 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11232 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11234 // If the LHS is a shufflevector itself, see if we can combine it with this
11235 // one without producing an unusual shuffle. Here we are really conservative:
11236 // we are absolutely afraid of producing a shuffle mask not in the input
11237 // program, because the code gen may not be smart enough to turn a merged
11238 // shuffle into two specific shuffles: it may produce worse code. As such,
11239 // we only merge two shuffles if the result is one of the two input shuffle
11240 // masks. In this case, merging the shuffles just removes one instruction,
11241 // which we know is safe. This is good for things like turning:
11242 // (splat(splat)) -> splat.
11243 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11244 if (isa<UndefValue>(RHS)) {
11245 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11247 std::vector<unsigned> NewMask;
11248 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11249 if (Mask[i] >= 2*e)
11250 NewMask.push_back(2*e);
11252 NewMask.push_back(LHSMask[Mask[i]]);
11254 // If the result mask is equal to the src shuffle or this shuffle mask, do
11255 // the replacement.
11256 if (NewMask == LHSMask || NewMask == Mask) {
11257 std::vector<Constant*> Elts;
11258 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11259 if (NewMask[i] >= e*2) {
11260 Elts.push_back(UndefValue::get(Type::Int32Ty));
11262 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11265 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11266 LHSSVI->getOperand(1),
11267 ConstantVector::get(Elts));
11272 return MadeChange ? &SVI : 0;
11278 /// TryToSinkInstruction - Try to move the specified instruction from its
11279 /// current block into the beginning of DestBlock, which can only happen if it's
11280 /// safe to move the instruction past all of the instructions between it and the
11281 /// end of its block.
11282 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11283 assert(I->hasOneUse() && "Invariants didn't hold!");
11285 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11286 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11289 // Do not sink alloca instructions out of the entry block.
11290 if (isa<AllocaInst>(I) && I->getParent() ==
11291 &DestBlock->getParent()->getEntryBlock())
11294 // We can only sink load instructions if there is nothing between the load and
11295 // the end of block that could change the value.
11296 if (I->mayReadFromMemory()) {
11297 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11299 if (Scan->mayWriteToMemory())
11303 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11305 I->moveBefore(InsertPos);
11311 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11312 /// all reachable code to the worklist.
11314 /// This has a couple of tricks to make the code faster and more powerful. In
11315 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11316 /// them to the worklist (this significantly speeds up instcombine on code where
11317 /// many instructions are dead or constant). Additionally, if we find a branch
11318 /// whose condition is a known constant, we only visit the reachable successors.
11320 static void AddReachableCodeToWorklist(BasicBlock *BB,
11321 SmallPtrSet<BasicBlock*, 64> &Visited,
11323 const TargetData *TD) {
11324 std::vector<BasicBlock*> Worklist;
11325 Worklist.push_back(BB);
11327 while (!Worklist.empty()) {
11328 BB = Worklist.back();
11329 Worklist.pop_back();
11331 // We have now visited this block! If we've already been here, ignore it.
11332 if (!Visited.insert(BB)) continue;
11334 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11335 Instruction *Inst = BBI++;
11337 // DCE instruction if trivially dead.
11338 if (isInstructionTriviallyDead(Inst)) {
11340 DOUT << "IC: DCE: " << *Inst;
11341 Inst->eraseFromParent();
11345 // ConstantProp instruction if trivially constant.
11346 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11347 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11348 Inst->replaceAllUsesWith(C);
11350 Inst->eraseFromParent();
11354 IC.AddToWorkList(Inst);
11357 // Recursively visit successors. If this is a branch or switch on a
11358 // constant, only visit the reachable successor.
11359 TerminatorInst *TI = BB->getTerminator();
11360 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11361 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11362 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11363 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11364 Worklist.push_back(ReachableBB);
11367 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11368 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11369 // See if this is an explicit destination.
11370 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11371 if (SI->getCaseValue(i) == Cond) {
11372 BasicBlock *ReachableBB = SI->getSuccessor(i);
11373 Worklist.push_back(ReachableBB);
11377 // Otherwise it is the default destination.
11378 Worklist.push_back(SI->getSuccessor(0));
11383 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11384 Worklist.push_back(TI->getSuccessor(i));
11388 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11389 bool Changed = false;
11390 TD = &getAnalysis<TargetData>();
11392 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11393 << F.getNameStr() << "\n");
11396 // Do a depth-first traversal of the function, populate the worklist with
11397 // the reachable instructions. Ignore blocks that are not reachable. Keep
11398 // track of which blocks we visit.
11399 SmallPtrSet<BasicBlock*, 64> Visited;
11400 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11402 // Do a quick scan over the function. If we find any blocks that are
11403 // unreachable, remove any instructions inside of them. This prevents
11404 // the instcombine code from having to deal with some bad special cases.
11405 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11406 if (!Visited.count(BB)) {
11407 Instruction *Term = BB->getTerminator();
11408 while (Term != BB->begin()) { // Remove instrs bottom-up
11409 BasicBlock::iterator I = Term; --I;
11411 DOUT << "IC: DCE: " << *I;
11414 if (!I->use_empty())
11415 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11416 I->eraseFromParent();
11421 while (!Worklist.empty()) {
11422 Instruction *I = RemoveOneFromWorkList();
11423 if (I == 0) continue; // skip null values.
11425 // Check to see if we can DCE the instruction.
11426 if (isInstructionTriviallyDead(I)) {
11427 // Add operands to the worklist.
11428 if (I->getNumOperands() < 4)
11429 AddUsesToWorkList(*I);
11432 DOUT << "IC: DCE: " << *I;
11434 I->eraseFromParent();
11435 RemoveFromWorkList(I);
11439 // Instruction isn't dead, see if we can constant propagate it.
11440 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11441 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11443 // Add operands to the worklist.
11444 AddUsesToWorkList(*I);
11445 ReplaceInstUsesWith(*I, C);
11448 I->eraseFromParent();
11449 RemoveFromWorkList(I);
11453 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
11454 // See if we can constant fold its operands.
11455 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
11456 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
11457 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
11463 // See if we can trivially sink this instruction to a successor basic block.
11464 if (I->hasOneUse()) {
11465 BasicBlock *BB = I->getParent();
11466 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11467 if (UserParent != BB) {
11468 bool UserIsSuccessor = false;
11469 // See if the user is one of our successors.
11470 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11471 if (*SI == UserParent) {
11472 UserIsSuccessor = true;
11476 // If the user is one of our immediate successors, and if that successor
11477 // only has us as a predecessors (we'd have to split the critical edge
11478 // otherwise), we can keep going.
11479 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11480 next(pred_begin(UserParent)) == pred_end(UserParent))
11481 // Okay, the CFG is simple enough, try to sink this instruction.
11482 Changed |= TryToSinkInstruction(I, UserParent);
11486 // Now that we have an instruction, try combining it to simplify it...
11490 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11491 if (Instruction *Result = visit(*I)) {
11493 // Should we replace the old instruction with a new one?
11495 DOUT << "IC: Old = " << *I
11496 << " New = " << *Result;
11498 // Everything uses the new instruction now.
11499 I->replaceAllUsesWith(Result);
11501 // Push the new instruction and any users onto the worklist.
11502 AddToWorkList(Result);
11503 AddUsersToWorkList(*Result);
11505 // Move the name to the new instruction first.
11506 Result->takeName(I);
11508 // Insert the new instruction into the basic block...
11509 BasicBlock *InstParent = I->getParent();
11510 BasicBlock::iterator InsertPos = I;
11512 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11513 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11516 InstParent->getInstList().insert(InsertPos, Result);
11518 // Make sure that we reprocess all operands now that we reduced their
11520 AddUsesToWorkList(*I);
11522 // Instructions can end up on the worklist more than once. Make sure
11523 // we do not process an instruction that has been deleted.
11524 RemoveFromWorkList(I);
11526 // Erase the old instruction.
11527 InstParent->getInstList().erase(I);
11530 DOUT << "IC: Mod = " << OrigI
11531 << " New = " << *I;
11534 // If the instruction was modified, it's possible that it is now dead.
11535 // if so, remove it.
11536 if (isInstructionTriviallyDead(I)) {
11537 // Make sure we process all operands now that we are reducing their
11539 AddUsesToWorkList(*I);
11541 // Instructions may end up in the worklist more than once. Erase all
11542 // occurrences of this instruction.
11543 RemoveFromWorkList(I);
11544 I->eraseFromParent();
11547 AddUsersToWorkList(*I);
11554 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11556 // Do an explicit clear, this shrinks the map if needed.
11557 WorklistMap.clear();
11562 bool InstCombiner::runOnFunction(Function &F) {
11563 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11565 bool EverMadeChange = false;
11567 // Iterate while there is work to do.
11568 unsigned Iteration = 0;
11569 while (DoOneIteration(F, Iteration++))
11570 EverMadeChange = true;
11571 return EverMadeChange;
11574 FunctionPass *llvm::createInstructionCombiningPass() {
11575 return new InstCombiner();