1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 std::vector<Instruction*> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass((intptr_t)&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 Instruction *commonRemTransforms(BinaryOperator &I);
176 Instruction *commonIRemTransforms(BinaryOperator &I);
177 Instruction *commonDivTransforms(BinaryOperator &I);
178 Instruction *commonIDivTransforms(BinaryOperator &I);
179 Instruction *visitUDiv(BinaryOperator &I);
180 Instruction *visitSDiv(BinaryOperator &I);
181 Instruction *visitFDiv(BinaryOperator &I);
182 Instruction *visitAnd(BinaryOperator &I);
183 Instruction *visitOr (BinaryOperator &I);
184 Instruction *visitXor(BinaryOperator &I);
185 Instruction *visitShl(BinaryOperator &I);
186 Instruction *visitAShr(BinaryOperator &I);
187 Instruction *visitLShr(BinaryOperator &I);
188 Instruction *commonShiftTransforms(BinaryOperator &I);
189 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
191 Instruction *visitFCmpInst(FCmpInst &I);
192 Instruction *visitICmpInst(ICmpInst &I);
193 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
194 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
197 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
198 ConstantInt *DivRHS);
200 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
201 ICmpInst::Predicate Cond, Instruction &I);
202 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
204 Instruction *commonCastTransforms(CastInst &CI);
205 Instruction *commonIntCastTransforms(CastInst &CI);
206 Instruction *commonPointerCastTransforms(CastInst &CI);
207 Instruction *visitTrunc(TruncInst &CI);
208 Instruction *visitZExt(ZExtInst &CI);
209 Instruction *visitSExt(SExtInst &CI);
210 Instruction *visitFPTrunc(FPTruncInst &CI);
211 Instruction *visitFPExt(CastInst &CI);
212 Instruction *visitFPToUI(FPToUIInst &FI);
213 Instruction *visitFPToSI(FPToSIInst &FI);
214 Instruction *visitUIToFP(CastInst &CI);
215 Instruction *visitSIToFP(CastInst &CI);
216 Instruction *visitPtrToInt(CastInst &CI);
217 Instruction *visitIntToPtr(IntToPtrInst &CI);
218 Instruction *visitBitCast(BitCastInst &CI);
219 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
221 Instruction *visitSelectInst(SelectInst &CI);
222 Instruction *visitCallInst(CallInst &CI);
223 Instruction *visitInvokeInst(InvokeInst &II);
224 Instruction *visitPHINode(PHINode &PN);
225 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
226 Instruction *visitAllocationInst(AllocationInst &AI);
227 Instruction *visitFreeInst(FreeInst &FI);
228 Instruction *visitLoadInst(LoadInst &LI);
229 Instruction *visitStoreInst(StoreInst &SI);
230 Instruction *visitBranchInst(BranchInst &BI);
231 Instruction *visitSwitchInst(SwitchInst &SI);
232 Instruction *visitInsertElementInst(InsertElementInst &IE);
233 Instruction *visitExtractElementInst(ExtractElementInst &EI);
234 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
235 Instruction *visitExtractValueInst(ExtractValueInst &EV);
237 // visitInstruction - Specify what to return for unhandled instructions...
238 Instruction *visitInstruction(Instruction &I) { return 0; }
241 Instruction *visitCallSite(CallSite CS);
242 bool transformConstExprCastCall(CallSite CS);
243 Instruction *transformCallThroughTrampoline(CallSite CS);
244 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
245 bool DoXform = true);
246 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
249 // InsertNewInstBefore - insert an instruction New before instruction Old
250 // in the program. Add the new instruction to the worklist.
252 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
253 assert(New && New->getParent() == 0 &&
254 "New instruction already inserted into a basic block!");
255 BasicBlock *BB = Old.getParent();
256 BB->getInstList().insert(&Old, New); // Insert inst
261 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
262 /// This also adds the cast to the worklist. Finally, this returns the
264 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
266 if (V->getType() == Ty) return V;
268 if (Constant *CV = dyn_cast<Constant>(V))
269 return ConstantExpr::getCast(opc, CV, Ty);
271 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
276 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
277 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
281 // ReplaceInstUsesWith - This method is to be used when an instruction is
282 // found to be dead, replacable with another preexisting expression. Here
283 // we add all uses of I to the worklist, replace all uses of I with the new
284 // value, then return I, so that the inst combiner will know that I was
287 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
288 AddUsersToWorkList(I); // Add all modified instrs to worklist
290 I.replaceAllUsesWith(V);
293 // If we are replacing the instruction with itself, this must be in a
294 // segment of unreachable code, so just clobber the instruction.
295 I.replaceAllUsesWith(UndefValue::get(I.getType()));
300 // UpdateValueUsesWith - This method is to be used when an value is
301 // found to be replacable with another preexisting expression or was
302 // updated. Here we add all uses of I to the worklist, replace all uses of
303 // I with the new value (unless the instruction was just updated), then
304 // return true, so that the inst combiner will know that I was modified.
306 bool UpdateValueUsesWith(Value *Old, Value *New) {
307 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
309 Old->replaceAllUsesWith(New);
310 if (Instruction *I = dyn_cast<Instruction>(Old))
312 if (Instruction *I = dyn_cast<Instruction>(New))
317 // EraseInstFromFunction - When dealing with an instruction that has side
318 // effects or produces a void value, we can't rely on DCE to delete the
319 // instruction. Instead, visit methods should return the value returned by
321 Instruction *EraseInstFromFunction(Instruction &I) {
322 assert(I.use_empty() && "Cannot erase instruction that is used!");
323 AddUsesToWorkList(I);
324 RemoveFromWorkList(&I);
326 return 0; // Don't do anything with FI
329 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
330 APInt &KnownOne, unsigned Depth = 0) const {
331 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
334 bool MaskedValueIsZero(Value *V, const APInt &Mask,
335 unsigned Depth = 0) const {
336 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
338 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
339 return llvm::ComputeNumSignBits(Op, TD, Depth);
343 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
344 /// InsertBefore instruction. This is specialized a bit to avoid inserting
345 /// casts that are known to not do anything...
347 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
348 Value *V, const Type *DestTy,
349 Instruction *InsertBefore);
351 /// SimplifyCommutative - This performs a few simplifications for
352 /// commutative operators.
353 bool SimplifyCommutative(BinaryOperator &I);
355 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
356 /// most-complex to least-complex order.
357 bool SimplifyCompare(CmpInst &I);
359 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
360 /// on the demanded bits.
361 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
362 APInt& KnownZero, APInt& KnownOne,
365 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
366 uint64_t &UndefElts, unsigned Depth = 0);
368 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
369 // PHI node as operand #0, see if we can fold the instruction into the PHI
370 // (which is only possible if all operands to the PHI are constants).
371 Instruction *FoldOpIntoPhi(Instruction &I);
373 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
374 // operator and they all are only used by the PHI, PHI together their
375 // inputs, and do the operation once, to the result of the PHI.
376 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
377 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
380 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
381 ConstantInt *AndRHS, BinaryOperator &TheAnd);
383 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
384 bool isSub, Instruction &I);
385 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
386 bool isSigned, bool Inside, Instruction &IB);
387 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
388 Instruction *MatchBSwap(BinaryOperator &I);
389 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
390 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
391 Instruction *SimplifyMemSet(MemSetInst *MI);
394 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
396 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
398 int &NumCastsRemoved);
399 unsigned GetOrEnforceKnownAlignment(Value *V,
400 unsigned PrefAlign = 0);
405 char InstCombiner::ID = 0;
406 static RegisterPass<InstCombiner>
407 X("instcombine", "Combine redundant instructions");
409 // getComplexity: Assign a complexity or rank value to LLVM Values...
410 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
411 static unsigned getComplexity(Value *V) {
412 if (isa<Instruction>(V)) {
413 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
417 if (isa<Argument>(V)) return 3;
418 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
421 // isOnlyUse - Return true if this instruction will be deleted if we stop using
423 static bool isOnlyUse(Value *V) {
424 return V->hasOneUse() || isa<Constant>(V);
427 // getPromotedType - Return the specified type promoted as it would be to pass
428 // though a va_arg area...
429 static const Type *getPromotedType(const Type *Ty) {
430 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
431 if (ITy->getBitWidth() < 32)
432 return Type::Int32Ty;
437 /// getBitCastOperand - If the specified operand is a CastInst or a constant
438 /// expression bitcast, return the operand value, otherwise return null.
439 static Value *getBitCastOperand(Value *V) {
440 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
441 return I->getOperand(0);
442 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
443 if (CE->getOpcode() == Instruction::BitCast)
444 return CE->getOperand(0);
448 /// This function is a wrapper around CastInst::isEliminableCastPair. It
449 /// simply extracts arguments and returns what that function returns.
450 static Instruction::CastOps
451 isEliminableCastPair(
452 const CastInst *CI, ///< The first cast instruction
453 unsigned opcode, ///< The opcode of the second cast instruction
454 const Type *DstTy, ///< The target type for the second cast instruction
455 TargetData *TD ///< The target data for pointer size
458 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
459 const Type *MidTy = CI->getType(); // B from above
461 // Get the opcodes of the two Cast instructions
462 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
463 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
465 return Instruction::CastOps(
466 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
467 DstTy, TD->getIntPtrType()));
470 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
471 /// in any code being generated. It does not require codegen if V is simple
472 /// enough or if the cast can be folded into other casts.
473 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
474 const Type *Ty, TargetData *TD) {
475 if (V->getType() == Ty || isa<Constant>(V)) return false;
477 // If this is another cast that can be eliminated, it isn't codegen either.
478 if (const CastInst *CI = dyn_cast<CastInst>(V))
479 if (isEliminableCastPair(CI, opcode, Ty, TD))
484 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
485 /// InsertBefore instruction. This is specialized a bit to avoid inserting
486 /// casts that are known to not do anything...
488 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
489 Value *V, const Type *DestTy,
490 Instruction *InsertBefore) {
491 if (V->getType() == DestTy) return V;
492 if (Constant *C = dyn_cast<Constant>(V))
493 return ConstantExpr::getCast(opcode, C, DestTy);
495 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
498 // SimplifyCommutative - This performs a few simplifications for commutative
501 // 1. Order operands such that they are listed from right (least complex) to
502 // left (most complex). This puts constants before unary operators before
505 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
506 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
508 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
509 bool Changed = false;
510 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
511 Changed = !I.swapOperands();
513 if (!I.isAssociative()) return Changed;
514 Instruction::BinaryOps Opcode = I.getOpcode();
515 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
516 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
517 if (isa<Constant>(I.getOperand(1))) {
518 Constant *Folded = ConstantExpr::get(I.getOpcode(),
519 cast<Constant>(I.getOperand(1)),
520 cast<Constant>(Op->getOperand(1)));
521 I.setOperand(0, Op->getOperand(0));
522 I.setOperand(1, Folded);
524 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
525 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
526 isOnlyUse(Op) && isOnlyUse(Op1)) {
527 Constant *C1 = cast<Constant>(Op->getOperand(1));
528 Constant *C2 = cast<Constant>(Op1->getOperand(1));
530 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
531 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
532 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
536 I.setOperand(0, New);
537 I.setOperand(1, Folded);
544 /// SimplifyCompare - For a CmpInst this function just orders the operands
545 /// so that theyare listed from right (least complex) to left (most complex).
546 /// This puts constants before unary operators before binary operators.
547 bool InstCombiner::SimplifyCompare(CmpInst &I) {
548 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
551 // Compare instructions are not associative so there's nothing else we can do.
555 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
556 // if the LHS is a constant zero (which is the 'negate' form).
558 static inline Value *dyn_castNegVal(Value *V) {
559 if (BinaryOperator::isNeg(V))
560 return BinaryOperator::getNegArgument(V);
562 // Constants can be considered to be negated values if they can be folded.
563 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
564 return ConstantExpr::getNeg(C);
566 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
567 if (C->getType()->getElementType()->isInteger())
568 return ConstantExpr::getNeg(C);
573 static inline Value *dyn_castNotVal(Value *V) {
574 if (BinaryOperator::isNot(V))
575 return BinaryOperator::getNotArgument(V);
577 // Constants can be considered to be not'ed values...
578 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
579 return ConstantInt::get(~C->getValue());
583 // dyn_castFoldableMul - If this value is a multiply that can be folded into
584 // other computations (because it has a constant operand), return the
585 // non-constant operand of the multiply, and set CST to point to the multiplier.
586 // Otherwise, return null.
588 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
589 if (V->hasOneUse() && V->getType()->isInteger())
590 if (Instruction *I = dyn_cast<Instruction>(V)) {
591 if (I->getOpcode() == Instruction::Mul)
592 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
593 return I->getOperand(0);
594 if (I->getOpcode() == Instruction::Shl)
595 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
596 // The multiplier is really 1 << CST.
597 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
598 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
599 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
600 return I->getOperand(0);
606 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
607 /// expression, return it.
608 static User *dyn_castGetElementPtr(Value *V) {
609 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
611 if (CE->getOpcode() == Instruction::GetElementPtr)
612 return cast<User>(V);
616 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
617 /// opcode value. Otherwise return UserOp1.
618 static unsigned getOpcode(const Value *V) {
619 if (const Instruction *I = dyn_cast<Instruction>(V))
620 return I->getOpcode();
621 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
622 return CE->getOpcode();
623 // Use UserOp1 to mean there's no opcode.
624 return Instruction::UserOp1;
627 /// AddOne - Add one to a ConstantInt
628 static ConstantInt *AddOne(ConstantInt *C) {
629 APInt Val(C->getValue());
630 return ConstantInt::get(++Val);
632 /// SubOne - Subtract one from a ConstantInt
633 static ConstantInt *SubOne(ConstantInt *C) {
634 APInt Val(C->getValue());
635 return ConstantInt::get(--Val);
637 /// Add - Add two ConstantInts together
638 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
639 return ConstantInt::get(C1->getValue() + C2->getValue());
641 /// And - Bitwise AND two ConstantInts together
642 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
643 return ConstantInt::get(C1->getValue() & C2->getValue());
645 /// Subtract - Subtract one ConstantInt from another
646 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
647 return ConstantInt::get(C1->getValue() - C2->getValue());
649 /// Multiply - Multiply two ConstantInts together
650 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
651 return ConstantInt::get(C1->getValue() * C2->getValue());
653 /// MultiplyOverflows - True if the multiply can not be expressed in an int
655 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
656 uint32_t W = C1->getBitWidth();
657 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
666 APInt MulExt = LHSExt * RHSExt;
669 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
670 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
671 return MulExt.slt(Min) || MulExt.sgt(Max);
673 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
677 /// ShrinkDemandedConstant - Check to see if the specified operand of the
678 /// specified instruction is a constant integer. If so, check to see if there
679 /// are any bits set in the constant that are not demanded. If so, shrink the
680 /// constant and return true.
681 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
683 assert(I && "No instruction?");
684 assert(OpNo < I->getNumOperands() && "Operand index too large");
686 // If the operand is not a constant integer, nothing to do.
687 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
688 if (!OpC) return false;
690 // If there are no bits set that aren't demanded, nothing to do.
691 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
692 if ((~Demanded & OpC->getValue()) == 0)
695 // This instruction is producing bits that are not demanded. Shrink the RHS.
696 Demanded &= OpC->getValue();
697 I->setOperand(OpNo, ConstantInt::get(Demanded));
701 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
702 // set of known zero and one bits, compute the maximum and minimum values that
703 // could have the specified known zero and known one bits, returning them in
705 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
706 const APInt& KnownZero,
707 const APInt& KnownOne,
708 APInt& Min, APInt& Max) {
709 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
710 assert(KnownZero.getBitWidth() == BitWidth &&
711 KnownOne.getBitWidth() == BitWidth &&
712 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
713 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
714 APInt UnknownBits = ~(KnownZero|KnownOne);
716 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
717 // bit if it is unknown.
719 Max = KnownOne|UnknownBits;
721 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
723 Max.clear(BitWidth-1);
727 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
728 // a set of known zero and one bits, compute the maximum and minimum values that
729 // could have the specified known zero and known one bits, returning them in
731 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
732 const APInt &KnownZero,
733 const APInt &KnownOne,
734 APInt &Min, APInt &Max) {
735 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
736 assert(KnownZero.getBitWidth() == BitWidth &&
737 KnownOne.getBitWidth() == BitWidth &&
738 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
739 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
740 APInt UnknownBits = ~(KnownZero|KnownOne);
742 // The minimum value is when the unknown bits are all zeros.
744 // The maximum value is when the unknown bits are all ones.
745 Max = KnownOne|UnknownBits;
748 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
749 /// value based on the demanded bits. When this function is called, it is known
750 /// that only the bits set in DemandedMask of the result of V are ever used
751 /// downstream. Consequently, depending on the mask and V, it may be possible
752 /// to replace V with a constant or one of its operands. In such cases, this
753 /// function does the replacement and returns true. In all other cases, it
754 /// returns false after analyzing the expression and setting KnownOne and known
755 /// to be one in the expression. KnownZero contains all the bits that are known
756 /// to be zero in the expression. These are provided to potentially allow the
757 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
758 /// the expression. KnownOne and KnownZero always follow the invariant that
759 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
760 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
761 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
762 /// and KnownOne must all be the same.
763 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
764 APInt& KnownZero, APInt& KnownOne,
766 assert(V != 0 && "Null pointer of Value???");
767 assert(Depth <= 6 && "Limit Search Depth");
768 uint32_t BitWidth = DemandedMask.getBitWidth();
769 const IntegerType *VTy = cast<IntegerType>(V->getType());
770 assert(VTy->getBitWidth() == BitWidth &&
771 KnownZero.getBitWidth() == BitWidth &&
772 KnownOne.getBitWidth() == BitWidth &&
773 "Value *V, DemandedMask, KnownZero and KnownOne \
774 must have same BitWidth");
775 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
776 // We know all of the bits for a constant!
777 KnownOne = CI->getValue() & DemandedMask;
778 KnownZero = ~KnownOne & DemandedMask;
784 if (!V->hasOneUse()) { // Other users may use these bits.
785 if (Depth != 0) { // Not at the root.
786 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
787 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
790 // If this is the root being simplified, allow it to have multiple uses,
791 // just set the DemandedMask to all bits.
792 DemandedMask = APInt::getAllOnesValue(BitWidth);
793 } else if (DemandedMask == 0) { // Not demanding any bits from V.
794 if (V != UndefValue::get(VTy))
795 return UpdateValueUsesWith(V, UndefValue::get(VTy));
797 } else if (Depth == 6) { // Limit search depth.
801 Instruction *I = dyn_cast<Instruction>(V);
802 if (!I) return false; // Only analyze instructions.
804 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
805 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
806 switch (I->getOpcode()) {
808 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
810 case Instruction::And:
811 // If either the LHS or the RHS are Zero, the result is zero.
812 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
813 RHSKnownZero, RHSKnownOne, Depth+1))
815 assert((RHSKnownZero & RHSKnownOne) == 0 &&
816 "Bits known to be one AND zero?");
818 // If something is known zero on the RHS, the bits aren't demanded on the
820 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
821 LHSKnownZero, LHSKnownOne, Depth+1))
823 assert((LHSKnownZero & LHSKnownOne) == 0 &&
824 "Bits known to be one AND zero?");
826 // If all of the demanded bits are known 1 on one side, return the other.
827 // These bits cannot contribute to the result of the 'and'.
828 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
829 (DemandedMask & ~LHSKnownZero))
830 return UpdateValueUsesWith(I, I->getOperand(0));
831 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
832 (DemandedMask & ~RHSKnownZero))
833 return UpdateValueUsesWith(I, I->getOperand(1));
835 // If all of the demanded bits in the inputs are known zeros, return zero.
836 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
837 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
839 // If the RHS is a constant, see if we can simplify it.
840 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
841 return UpdateValueUsesWith(I, I);
843 // Output known-1 bits are only known if set in both the LHS & RHS.
844 RHSKnownOne &= LHSKnownOne;
845 // Output known-0 are known to be clear if zero in either the LHS | RHS.
846 RHSKnownZero |= LHSKnownZero;
848 case Instruction::Or:
849 // If either the LHS or the RHS are One, the result is One.
850 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
851 RHSKnownZero, RHSKnownOne, Depth+1))
853 assert((RHSKnownZero & RHSKnownOne) == 0 &&
854 "Bits known to be one AND zero?");
855 // If something is known one on the RHS, the bits aren't demanded on the
857 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
858 LHSKnownZero, LHSKnownOne, Depth+1))
860 assert((LHSKnownZero & LHSKnownOne) == 0 &&
861 "Bits known to be one AND zero?");
863 // If all of the demanded bits are known zero on one side, return the other.
864 // These bits cannot contribute to the result of the 'or'.
865 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
866 (DemandedMask & ~LHSKnownOne))
867 return UpdateValueUsesWith(I, I->getOperand(0));
868 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
869 (DemandedMask & ~RHSKnownOne))
870 return UpdateValueUsesWith(I, I->getOperand(1));
872 // If all of the potentially set bits on one side are known to be set on
873 // the other side, just use the 'other' side.
874 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
875 (DemandedMask & (~RHSKnownZero)))
876 return UpdateValueUsesWith(I, I->getOperand(0));
877 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
878 (DemandedMask & (~LHSKnownZero)))
879 return UpdateValueUsesWith(I, I->getOperand(1));
881 // If the RHS is a constant, see if we can simplify it.
882 if (ShrinkDemandedConstant(I, 1, DemandedMask))
883 return UpdateValueUsesWith(I, I);
885 // Output known-0 bits are only known if clear in both the LHS & RHS.
886 RHSKnownZero &= LHSKnownZero;
887 // Output known-1 are known to be set if set in either the LHS | RHS.
888 RHSKnownOne |= LHSKnownOne;
890 case Instruction::Xor: {
891 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
892 RHSKnownZero, RHSKnownOne, Depth+1))
894 assert((RHSKnownZero & RHSKnownOne) == 0 &&
895 "Bits known to be one AND zero?");
896 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
897 LHSKnownZero, LHSKnownOne, Depth+1))
899 assert((LHSKnownZero & LHSKnownOne) == 0 &&
900 "Bits known to be one AND zero?");
902 // If all of the demanded bits are known zero on one side, return the other.
903 // These bits cannot contribute to the result of the 'xor'.
904 if ((DemandedMask & RHSKnownZero) == DemandedMask)
905 return UpdateValueUsesWith(I, I->getOperand(0));
906 if ((DemandedMask & LHSKnownZero) == DemandedMask)
907 return UpdateValueUsesWith(I, I->getOperand(1));
909 // Output known-0 bits are known if clear or set in both the LHS & RHS.
910 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
911 (RHSKnownOne & LHSKnownOne);
912 // Output known-1 are known to be set if set in only one of the LHS, RHS.
913 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
914 (RHSKnownOne & LHSKnownZero);
916 // If all of the demanded bits are known to be zero on one side or the
917 // other, turn this into an *inclusive* or.
918 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
919 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
921 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
923 InsertNewInstBefore(Or, *I);
924 return UpdateValueUsesWith(I, Or);
927 // If all of the demanded bits on one side are known, and all of the set
928 // bits on that side are also known to be set on the other side, turn this
929 // into an AND, as we know the bits will be cleared.
930 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
931 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
933 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
934 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
936 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
937 InsertNewInstBefore(And, *I);
938 return UpdateValueUsesWith(I, And);
942 // If the RHS is a constant, see if we can simplify it.
943 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
944 if (ShrinkDemandedConstant(I, 1, DemandedMask))
945 return UpdateValueUsesWith(I, I);
947 RHSKnownZero = KnownZeroOut;
948 RHSKnownOne = KnownOneOut;
951 case Instruction::Select:
952 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
953 RHSKnownZero, RHSKnownOne, Depth+1))
955 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
956 LHSKnownZero, LHSKnownOne, Depth+1))
958 assert((RHSKnownZero & RHSKnownOne) == 0 &&
959 "Bits known to be one AND zero?");
960 assert((LHSKnownZero & LHSKnownOne) == 0 &&
961 "Bits known to be one AND zero?");
963 // If the operands are constants, see if we can simplify them.
964 if (ShrinkDemandedConstant(I, 1, DemandedMask))
965 return UpdateValueUsesWith(I, I);
966 if (ShrinkDemandedConstant(I, 2, DemandedMask))
967 return UpdateValueUsesWith(I, I);
969 // Only known if known in both the LHS and RHS.
970 RHSKnownOne &= LHSKnownOne;
971 RHSKnownZero &= LHSKnownZero;
973 case Instruction::Trunc: {
975 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
976 DemandedMask.zext(truncBf);
977 RHSKnownZero.zext(truncBf);
978 RHSKnownOne.zext(truncBf);
979 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
980 RHSKnownZero, RHSKnownOne, Depth+1))
982 DemandedMask.trunc(BitWidth);
983 RHSKnownZero.trunc(BitWidth);
984 RHSKnownOne.trunc(BitWidth);
985 assert((RHSKnownZero & RHSKnownOne) == 0 &&
986 "Bits known to be one AND zero?");
989 case Instruction::BitCast:
990 if (!I->getOperand(0)->getType()->isInteger())
993 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
994 RHSKnownZero, RHSKnownOne, Depth+1))
996 assert((RHSKnownZero & RHSKnownOne) == 0 &&
997 "Bits known to be one AND zero?");
999 case Instruction::ZExt: {
1000 // Compute the bits in the result that are not present in the input.
1001 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1002 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1004 DemandedMask.trunc(SrcBitWidth);
1005 RHSKnownZero.trunc(SrcBitWidth);
1006 RHSKnownOne.trunc(SrcBitWidth);
1007 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1008 RHSKnownZero, RHSKnownOne, Depth+1))
1010 DemandedMask.zext(BitWidth);
1011 RHSKnownZero.zext(BitWidth);
1012 RHSKnownOne.zext(BitWidth);
1013 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1014 "Bits known to be one AND zero?");
1015 // The top bits are known to be zero.
1016 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1019 case Instruction::SExt: {
1020 // Compute the bits in the result that are not present in the input.
1021 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1022 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1024 APInt InputDemandedBits = DemandedMask &
1025 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1027 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1028 // If any of the sign extended bits are demanded, we know that the sign
1030 if ((NewBits & DemandedMask) != 0)
1031 InputDemandedBits.set(SrcBitWidth-1);
1033 InputDemandedBits.trunc(SrcBitWidth);
1034 RHSKnownZero.trunc(SrcBitWidth);
1035 RHSKnownOne.trunc(SrcBitWidth);
1036 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1037 RHSKnownZero, RHSKnownOne, Depth+1))
1039 InputDemandedBits.zext(BitWidth);
1040 RHSKnownZero.zext(BitWidth);
1041 RHSKnownOne.zext(BitWidth);
1042 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1043 "Bits known to be one AND zero?");
1045 // If the sign bit of the input is known set or clear, then we know the
1046 // top bits of the result.
1048 // If the input sign bit is known zero, or if the NewBits are not demanded
1049 // convert this into a zero extension.
1050 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1052 // Convert to ZExt cast
1053 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1054 return UpdateValueUsesWith(I, NewCast);
1055 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1056 RHSKnownOne |= NewBits;
1060 case Instruction::Add: {
1061 // Figure out what the input bits are. If the top bits of the and result
1062 // are not demanded, then the add doesn't demand them from its input
1064 uint32_t NLZ = DemandedMask.countLeadingZeros();
1066 // If there is a constant on the RHS, there are a variety of xformations
1068 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1069 // If null, this should be simplified elsewhere. Some of the xforms here
1070 // won't work if the RHS is zero.
1074 // If the top bit of the output is demanded, demand everything from the
1075 // input. Otherwise, we demand all the input bits except NLZ top bits.
1076 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1078 // Find information about known zero/one bits in the input.
1079 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1080 LHSKnownZero, LHSKnownOne, Depth+1))
1083 // If the RHS of the add has bits set that can't affect the input, reduce
1085 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1086 return UpdateValueUsesWith(I, I);
1088 // Avoid excess work.
1089 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1092 // Turn it into OR if input bits are zero.
1093 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1095 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1097 InsertNewInstBefore(Or, *I);
1098 return UpdateValueUsesWith(I, Or);
1101 // We can say something about the output known-zero and known-one bits,
1102 // depending on potential carries from the input constant and the
1103 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1104 // bits set and the RHS constant is 0x01001, then we know we have a known
1105 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1107 // To compute this, we first compute the potential carry bits. These are
1108 // the bits which may be modified. I'm not aware of a better way to do
1110 const APInt& RHSVal = RHS->getValue();
1111 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1113 // Now that we know which bits have carries, compute the known-1/0 sets.
1115 // Bits are known one if they are known zero in one operand and one in the
1116 // other, and there is no input carry.
1117 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1118 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1120 // Bits are known zero if they are known zero in both operands and there
1121 // is no input carry.
1122 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1124 // If the high-bits of this ADD are not demanded, then it does not demand
1125 // the high bits of its LHS or RHS.
1126 if (DemandedMask[BitWidth-1] == 0) {
1127 // Right fill the mask of bits for this ADD to demand the most
1128 // significant bit and all those below it.
1129 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1130 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1131 LHSKnownZero, LHSKnownOne, Depth+1))
1133 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1134 LHSKnownZero, LHSKnownOne, Depth+1))
1140 case Instruction::Sub:
1141 // If the high-bits of this SUB are not demanded, then it does not demand
1142 // the high bits of its LHS or RHS.
1143 if (DemandedMask[BitWidth-1] == 0) {
1144 // Right fill the mask of bits for this SUB to demand the most
1145 // significant bit and all those below it.
1146 uint32_t NLZ = DemandedMask.countLeadingZeros();
1147 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1148 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1149 LHSKnownZero, LHSKnownOne, Depth+1))
1151 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1152 LHSKnownZero, LHSKnownOne, Depth+1))
1155 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1156 // the known zeros and ones.
1157 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1159 case Instruction::Shl:
1160 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1161 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1162 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1163 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1164 RHSKnownZero, RHSKnownOne, Depth+1))
1166 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1167 "Bits known to be one AND zero?");
1168 RHSKnownZero <<= ShiftAmt;
1169 RHSKnownOne <<= ShiftAmt;
1170 // low bits known zero.
1172 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1175 case Instruction::LShr:
1176 // For a logical shift right
1177 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1178 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1180 // Unsigned shift right.
1181 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1182 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1183 RHSKnownZero, RHSKnownOne, Depth+1))
1185 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1186 "Bits known to be one AND zero?");
1187 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1188 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1190 // Compute the new bits that are at the top now.
1191 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1192 RHSKnownZero |= HighBits; // high bits known zero.
1196 case Instruction::AShr:
1197 // If this is an arithmetic shift right and only the low-bit is set, we can
1198 // always convert this into a logical shr, even if the shift amount is
1199 // variable. The low bit of the shift cannot be an input sign bit unless
1200 // the shift amount is >= the size of the datatype, which is undefined.
1201 if (DemandedMask == 1) {
1202 // Perform the logical shift right.
1203 Value *NewVal = BinaryOperator::CreateLShr(
1204 I->getOperand(0), I->getOperand(1), I->getName());
1205 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1206 return UpdateValueUsesWith(I, NewVal);
1209 // If the sign bit is the only bit demanded by this ashr, then there is no
1210 // need to do it, the shift doesn't change the high bit.
1211 if (DemandedMask.isSignBit())
1212 return UpdateValueUsesWith(I, I->getOperand(0));
1214 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1215 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1217 // Signed shift right.
1218 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1219 // If any of the "high bits" are demanded, we should set the sign bit as
1221 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1222 DemandedMaskIn.set(BitWidth-1);
1223 if (SimplifyDemandedBits(I->getOperand(0),
1225 RHSKnownZero, RHSKnownOne, Depth+1))
1227 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1228 "Bits known to be one AND zero?");
1229 // Compute the new bits that are at the top now.
1230 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1231 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1232 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1234 // Handle the sign bits.
1235 APInt SignBit(APInt::getSignBit(BitWidth));
1236 // Adjust to where it is now in the mask.
1237 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1239 // If the input sign bit is known to be zero, or if none of the top bits
1240 // are demanded, turn this into an unsigned shift right.
1241 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1242 (HighBits & ~DemandedMask) == HighBits) {
1243 // Perform the logical shift right.
1244 Value *NewVal = BinaryOperator::CreateLShr(
1245 I->getOperand(0), SA, I->getName());
1246 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1247 return UpdateValueUsesWith(I, NewVal);
1248 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1249 RHSKnownOne |= HighBits;
1253 case Instruction::SRem:
1254 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1255 APInt RA = Rem->getValue();
1256 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1257 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1258 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1259 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1260 LHSKnownZero, LHSKnownOne, Depth+1))
1263 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1264 LHSKnownZero |= ~LowBits;
1265 else if (LHSKnownOne[BitWidth-1])
1266 LHSKnownOne |= ~LowBits;
1268 KnownZero |= LHSKnownZero & DemandedMask;
1269 KnownOne |= LHSKnownOne & DemandedMask;
1271 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1275 case Instruction::URem: {
1276 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1277 APInt RA = Rem->getValue();
1278 if (RA.isPowerOf2()) {
1279 APInt LowBits = (RA - 1);
1280 APInt Mask2 = LowBits & DemandedMask;
1281 KnownZero |= ~LowBits & DemandedMask;
1282 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1283 KnownZero, KnownOne, Depth+1))
1286 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1291 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1292 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1293 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1294 KnownZero2, KnownOne2, Depth+1))
1297 uint32_t Leaders = KnownZero2.countLeadingOnes();
1298 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1299 KnownZero2, KnownOne2, Depth+1))
1302 Leaders = std::max(Leaders,
1303 KnownZero2.countLeadingOnes());
1304 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1307 case Instruction::Call:
1308 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1309 switch (II->getIntrinsicID()) {
1311 case Intrinsic::bswap: {
1312 // If the only bits demanded come from one byte of the bswap result,
1313 // just shift the input byte into position to eliminate the bswap.
1314 unsigned NLZ = DemandedMask.countLeadingZeros();
1315 unsigned NTZ = DemandedMask.countTrailingZeros();
1317 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1318 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1319 // have 14 leading zeros, round to 8.
1322 // If we need exactly one byte, we can do this transformation.
1323 if (BitWidth-NLZ-NTZ == 8) {
1324 unsigned ResultBit = NTZ;
1325 unsigned InputBit = BitWidth-NTZ-8;
1327 // Replace this with either a left or right shift to get the byte into
1329 Instruction *NewVal;
1330 if (InputBit > ResultBit)
1331 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1332 ConstantInt::get(I->getType(), InputBit-ResultBit));
1334 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1335 ConstantInt::get(I->getType(), ResultBit-InputBit));
1336 NewVal->takeName(I);
1337 InsertNewInstBefore(NewVal, *I);
1338 return UpdateValueUsesWith(I, NewVal);
1341 // TODO: Could compute known zero/one bits based on the input.
1346 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1350 // If the client is only demanding bits that we know, return the known
1352 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1353 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1358 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1359 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1360 /// actually used by the caller. This method analyzes which elements of the
1361 /// operand are undef and returns that information in UndefElts.
1363 /// If the information about demanded elements can be used to simplify the
1364 /// operation, the operation is simplified, then the resultant value is
1365 /// returned. This returns null if no change was made.
1366 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1367 uint64_t &UndefElts,
1369 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1370 assert(VWidth <= 64 && "Vector too wide to analyze!");
1371 uint64_t EltMask = ~0ULL >> (64-VWidth);
1372 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1373 "Invalid DemandedElts!");
1375 if (isa<UndefValue>(V)) {
1376 // If the entire vector is undefined, just return this info.
1377 UndefElts = EltMask;
1379 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1380 UndefElts = EltMask;
1381 return UndefValue::get(V->getType());
1385 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1386 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1387 Constant *Undef = UndefValue::get(EltTy);
1389 std::vector<Constant*> Elts;
1390 for (unsigned i = 0; i != VWidth; ++i)
1391 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1392 Elts.push_back(Undef);
1393 UndefElts |= (1ULL << i);
1394 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1395 Elts.push_back(Undef);
1396 UndefElts |= (1ULL << i);
1397 } else { // Otherwise, defined.
1398 Elts.push_back(CP->getOperand(i));
1401 // If we changed the constant, return it.
1402 Constant *NewCP = ConstantVector::get(Elts);
1403 return NewCP != CP ? NewCP : 0;
1404 } else if (isa<ConstantAggregateZero>(V)) {
1405 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1407 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1408 Constant *Zero = Constant::getNullValue(EltTy);
1409 Constant *Undef = UndefValue::get(EltTy);
1410 std::vector<Constant*> Elts;
1411 for (unsigned i = 0; i != VWidth; ++i)
1412 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1413 UndefElts = DemandedElts ^ EltMask;
1414 return ConstantVector::get(Elts);
1417 if (!V->hasOneUse()) { // Other users may use these bits.
1418 if (Depth != 0) { // Not at the root.
1419 // TODO: Just compute the UndefElts information recursively.
1423 } else if (Depth == 10) { // Limit search depth.
1427 Instruction *I = dyn_cast<Instruction>(V);
1428 if (!I) return false; // Only analyze instructions.
1430 bool MadeChange = false;
1431 uint64_t UndefElts2;
1433 switch (I->getOpcode()) {
1436 case Instruction::InsertElement: {
1437 // If this is a variable index, we don't know which element it overwrites.
1438 // demand exactly the same input as we produce.
1439 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1441 // Note that we can't propagate undef elt info, because we don't know
1442 // which elt is getting updated.
1443 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1444 UndefElts2, Depth+1);
1445 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1449 // If this is inserting an element that isn't demanded, remove this
1451 unsigned IdxNo = Idx->getZExtValue();
1452 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1453 return AddSoonDeadInstToWorklist(*I, 0);
1455 // Otherwise, the element inserted overwrites whatever was there, so the
1456 // input demanded set is simpler than the output set.
1457 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1458 DemandedElts & ~(1ULL << IdxNo),
1459 UndefElts, Depth+1);
1460 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1462 // The inserted element is defined.
1463 UndefElts |= 1ULL << IdxNo;
1466 case Instruction::BitCast: {
1467 // Vector->vector casts only.
1468 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1470 unsigned InVWidth = VTy->getNumElements();
1471 uint64_t InputDemandedElts = 0;
1474 if (VWidth == InVWidth) {
1475 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1476 // elements as are demanded of us.
1478 InputDemandedElts = DemandedElts;
1479 } else if (VWidth > InVWidth) {
1483 // If there are more elements in the result than there are in the source,
1484 // then an input element is live if any of the corresponding output
1485 // elements are live.
1486 Ratio = VWidth/InVWidth;
1487 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1488 if (DemandedElts & (1ULL << OutIdx))
1489 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1495 // If there are more elements in the source than there are in the result,
1496 // then an input element is live if the corresponding output element is
1498 Ratio = InVWidth/VWidth;
1499 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1500 if (DemandedElts & (1ULL << InIdx/Ratio))
1501 InputDemandedElts |= 1ULL << InIdx;
1504 // div/rem demand all inputs, because they don't want divide by zero.
1505 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1506 UndefElts2, Depth+1);
1508 I->setOperand(0, TmpV);
1512 UndefElts = UndefElts2;
1513 if (VWidth > InVWidth) {
1514 assert(0 && "Unimp");
1515 // If there are more elements in the result than there are in the source,
1516 // then an output element is undef if the corresponding input element is
1518 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1519 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1520 UndefElts |= 1ULL << OutIdx;
1521 } else if (VWidth < InVWidth) {
1522 assert(0 && "Unimp");
1523 // If there are more elements in the source than there are in the result,
1524 // then a result element is undef if all of the corresponding input
1525 // elements are undef.
1526 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1527 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1528 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1529 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1533 case Instruction::And:
1534 case Instruction::Or:
1535 case Instruction::Xor:
1536 case Instruction::Add:
1537 case Instruction::Sub:
1538 case Instruction::Mul:
1539 // div/rem demand all inputs, because they don't want divide by zero.
1540 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1541 UndefElts, Depth+1);
1542 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1543 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1544 UndefElts2, Depth+1);
1545 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1547 // Output elements are undefined if both are undefined. Consider things
1548 // like undef&0. The result is known zero, not undef.
1549 UndefElts &= UndefElts2;
1552 case Instruction::Call: {
1553 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1555 switch (II->getIntrinsicID()) {
1558 // Binary vector operations that work column-wise. A dest element is a
1559 // function of the corresponding input elements from the two inputs.
1560 case Intrinsic::x86_sse_sub_ss:
1561 case Intrinsic::x86_sse_mul_ss:
1562 case Intrinsic::x86_sse_min_ss:
1563 case Intrinsic::x86_sse_max_ss:
1564 case Intrinsic::x86_sse2_sub_sd:
1565 case Intrinsic::x86_sse2_mul_sd:
1566 case Intrinsic::x86_sse2_min_sd:
1567 case Intrinsic::x86_sse2_max_sd:
1568 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1569 UndefElts, Depth+1);
1570 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1571 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1572 UndefElts2, Depth+1);
1573 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1575 // If only the low elt is demanded and this is a scalarizable intrinsic,
1576 // scalarize it now.
1577 if (DemandedElts == 1) {
1578 switch (II->getIntrinsicID()) {
1580 case Intrinsic::x86_sse_sub_ss:
1581 case Intrinsic::x86_sse_mul_ss:
1582 case Intrinsic::x86_sse2_sub_sd:
1583 case Intrinsic::x86_sse2_mul_sd:
1584 // TODO: Lower MIN/MAX/ABS/etc
1585 Value *LHS = II->getOperand(1);
1586 Value *RHS = II->getOperand(2);
1587 // Extract the element as scalars.
1588 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1589 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1591 switch (II->getIntrinsicID()) {
1592 default: assert(0 && "Case stmts out of sync!");
1593 case Intrinsic::x86_sse_sub_ss:
1594 case Intrinsic::x86_sse2_sub_sd:
1595 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1596 II->getName()), *II);
1598 case Intrinsic::x86_sse_mul_ss:
1599 case Intrinsic::x86_sse2_mul_sd:
1600 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1601 II->getName()), *II);
1606 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1608 InsertNewInstBefore(New, *II);
1609 AddSoonDeadInstToWorklist(*II, 0);
1614 // Output elements are undefined if both are undefined. Consider things
1615 // like undef&0. The result is known zero, not undef.
1616 UndefElts &= UndefElts2;
1622 return MadeChange ? I : 0;
1626 /// AssociativeOpt - Perform an optimization on an associative operator. This
1627 /// function is designed to check a chain of associative operators for a
1628 /// potential to apply a certain optimization. Since the optimization may be
1629 /// applicable if the expression was reassociated, this checks the chain, then
1630 /// reassociates the expression as necessary to expose the optimization
1631 /// opportunity. This makes use of a special Functor, which must define
1632 /// 'shouldApply' and 'apply' methods.
1634 template<typename Functor>
1635 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1636 unsigned Opcode = Root.getOpcode();
1637 Value *LHS = Root.getOperand(0);
1639 // Quick check, see if the immediate LHS matches...
1640 if (F.shouldApply(LHS))
1641 return F.apply(Root);
1643 // Otherwise, if the LHS is not of the same opcode as the root, return.
1644 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1645 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1646 // Should we apply this transform to the RHS?
1647 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1649 // If not to the RHS, check to see if we should apply to the LHS...
1650 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1651 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1655 // If the functor wants to apply the optimization to the RHS of LHSI,
1656 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1658 // Now all of the instructions are in the current basic block, go ahead
1659 // and perform the reassociation.
1660 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1662 // First move the selected RHS to the LHS of the root...
1663 Root.setOperand(0, LHSI->getOperand(1));
1665 // Make what used to be the LHS of the root be the user of the root...
1666 Value *ExtraOperand = TmpLHSI->getOperand(1);
1667 if (&Root == TmpLHSI) {
1668 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1671 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1672 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1673 BasicBlock::iterator ARI = &Root; ++ARI;
1674 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1677 // Now propagate the ExtraOperand down the chain of instructions until we
1679 while (TmpLHSI != LHSI) {
1680 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1681 // Move the instruction to immediately before the chain we are
1682 // constructing to avoid breaking dominance properties.
1683 NextLHSI->moveBefore(ARI);
1686 Value *NextOp = NextLHSI->getOperand(1);
1687 NextLHSI->setOperand(1, ExtraOperand);
1689 ExtraOperand = NextOp;
1692 // Now that the instructions are reassociated, have the functor perform
1693 // the transformation...
1694 return F.apply(Root);
1697 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1704 // AddRHS - Implements: X + X --> X << 1
1707 AddRHS(Value *rhs) : RHS(rhs) {}
1708 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1709 Instruction *apply(BinaryOperator &Add) const {
1710 return BinaryOperator::CreateShl(Add.getOperand(0),
1711 ConstantInt::get(Add.getType(), 1));
1715 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1717 struct AddMaskingAnd {
1719 AddMaskingAnd(Constant *c) : C2(c) {}
1720 bool shouldApply(Value *LHS) const {
1722 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1723 ConstantExpr::getAnd(C1, C2)->isNullValue();
1725 Instruction *apply(BinaryOperator &Add) const {
1726 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1732 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1734 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1735 if (Constant *SOC = dyn_cast<Constant>(SO))
1736 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1738 return IC->InsertNewInstBefore(CastInst::Create(
1739 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1742 // Figure out if the constant is the left or the right argument.
1743 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1744 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1746 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1748 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1749 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1752 Value *Op0 = SO, *Op1 = ConstOperand;
1754 std::swap(Op0, Op1);
1756 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1757 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1758 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1759 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1760 SO->getName()+".cmp");
1762 assert(0 && "Unknown binary instruction type!");
1765 return IC->InsertNewInstBefore(New, I);
1768 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1769 // constant as the other operand, try to fold the binary operator into the
1770 // select arguments. This also works for Cast instructions, which obviously do
1771 // not have a second operand.
1772 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1774 // Don't modify shared select instructions
1775 if (!SI->hasOneUse()) return 0;
1776 Value *TV = SI->getOperand(1);
1777 Value *FV = SI->getOperand(2);
1779 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1780 // Bool selects with constant operands can be folded to logical ops.
1781 if (SI->getType() == Type::Int1Ty) return 0;
1783 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1784 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1786 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1793 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1794 /// node as operand #0, see if we can fold the instruction into the PHI (which
1795 /// is only possible if all operands to the PHI are constants).
1796 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1797 PHINode *PN = cast<PHINode>(I.getOperand(0));
1798 unsigned NumPHIValues = PN->getNumIncomingValues();
1799 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1801 // Check to see if all of the operands of the PHI are constants. If there is
1802 // one non-constant value, remember the BB it is. If there is more than one
1803 // or if *it* is a PHI, bail out.
1804 BasicBlock *NonConstBB = 0;
1805 for (unsigned i = 0; i != NumPHIValues; ++i)
1806 if (!isa<Constant>(PN->getIncomingValue(i))) {
1807 if (NonConstBB) return 0; // More than one non-const value.
1808 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1809 NonConstBB = PN->getIncomingBlock(i);
1811 // If the incoming non-constant value is in I's block, we have an infinite
1813 if (NonConstBB == I.getParent())
1817 // If there is exactly one non-constant value, we can insert a copy of the
1818 // operation in that block. However, if this is a critical edge, we would be
1819 // inserting the computation one some other paths (e.g. inside a loop). Only
1820 // do this if the pred block is unconditionally branching into the phi block.
1822 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1823 if (!BI || !BI->isUnconditional()) return 0;
1826 // Okay, we can do the transformation: create the new PHI node.
1827 PHINode *NewPN = PHINode::Create(I.getType(), "");
1828 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1829 InsertNewInstBefore(NewPN, *PN);
1830 NewPN->takeName(PN);
1832 // Next, add all of the operands to the PHI.
1833 if (I.getNumOperands() == 2) {
1834 Constant *C = cast<Constant>(I.getOperand(1));
1835 for (unsigned i = 0; i != NumPHIValues; ++i) {
1837 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1838 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1839 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1841 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1843 assert(PN->getIncomingBlock(i) == NonConstBB);
1844 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1845 InV = BinaryOperator::Create(BO->getOpcode(),
1846 PN->getIncomingValue(i), C, "phitmp",
1847 NonConstBB->getTerminator());
1848 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1849 InV = CmpInst::Create(CI->getOpcode(),
1851 PN->getIncomingValue(i), C, "phitmp",
1852 NonConstBB->getTerminator());
1854 assert(0 && "Unknown binop!");
1856 AddToWorkList(cast<Instruction>(InV));
1858 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1861 CastInst *CI = cast<CastInst>(&I);
1862 const Type *RetTy = CI->getType();
1863 for (unsigned i = 0; i != NumPHIValues; ++i) {
1865 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1866 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1868 assert(PN->getIncomingBlock(i) == NonConstBB);
1869 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1870 I.getType(), "phitmp",
1871 NonConstBB->getTerminator());
1872 AddToWorkList(cast<Instruction>(InV));
1874 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1877 return ReplaceInstUsesWith(I, NewPN);
1881 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1882 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1883 /// This basically requires proving that the add in the original type would not
1884 /// overflow to change the sign bit or have a carry out.
1885 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1886 // There are different heuristics we can use for this. Here are some simple
1889 // Add has the property that adding any two 2's complement numbers can only
1890 // have one carry bit which can change a sign. As such, if LHS and RHS each
1891 // have at least two sign bits, we know that the addition of the two values will
1892 // sign extend fine.
1893 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1897 // If one of the operands only has one non-zero bit, and if the other operand
1898 // has a known-zero bit in a more significant place than it (not including the
1899 // sign bit) the ripple may go up to and fill the zero, but won't change the
1900 // sign. For example, (X & ~4) + 1.
1908 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1909 bool Changed = SimplifyCommutative(I);
1910 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1912 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1913 // X + undef -> undef
1914 if (isa<UndefValue>(RHS))
1915 return ReplaceInstUsesWith(I, RHS);
1918 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1919 if (RHSC->isNullValue())
1920 return ReplaceInstUsesWith(I, LHS);
1921 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1922 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1923 (I.getType())->getValueAPF()))
1924 return ReplaceInstUsesWith(I, LHS);
1927 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1928 // X + (signbit) --> X ^ signbit
1929 const APInt& Val = CI->getValue();
1930 uint32_t BitWidth = Val.getBitWidth();
1931 if (Val == APInt::getSignBit(BitWidth))
1932 return BinaryOperator::CreateXor(LHS, RHS);
1934 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1935 // (X & 254)+1 -> (X&254)|1
1936 if (!isa<VectorType>(I.getType())) {
1937 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1938 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1939 KnownZero, KnownOne))
1944 if (isa<PHINode>(LHS))
1945 if (Instruction *NV = FoldOpIntoPhi(I))
1948 ConstantInt *XorRHS = 0;
1950 if (isa<ConstantInt>(RHSC) &&
1951 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1952 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1953 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1955 uint32_t Size = TySizeBits / 2;
1956 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1957 APInt CFF80Val(-C0080Val);
1959 if (TySizeBits > Size) {
1960 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1961 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1962 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1963 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1964 // This is a sign extend if the top bits are known zero.
1965 if (!MaskedValueIsZero(XorLHS,
1966 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1967 Size = 0; // Not a sign ext, but can't be any others either.
1972 C0080Val = APIntOps::lshr(C0080Val, Size);
1973 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1974 } while (Size >= 1);
1976 // FIXME: This shouldn't be necessary. When the backends can handle types
1977 // with funny bit widths then this switch statement should be removed. It
1978 // is just here to get the size of the "middle" type back up to something
1979 // that the back ends can handle.
1980 const Type *MiddleType = 0;
1983 case 32: MiddleType = Type::Int32Ty; break;
1984 case 16: MiddleType = Type::Int16Ty; break;
1985 case 8: MiddleType = Type::Int8Ty; break;
1988 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1989 InsertNewInstBefore(NewTrunc, I);
1990 return new SExtInst(NewTrunc, I.getType(), I.getName());
1995 if (I.getType() == Type::Int1Ty)
1996 return BinaryOperator::CreateXor(LHS, RHS);
1999 if (I.getType()->isInteger()) {
2000 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2002 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2003 if (RHSI->getOpcode() == Instruction::Sub)
2004 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2005 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2007 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2008 if (LHSI->getOpcode() == Instruction::Sub)
2009 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2010 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2015 // -A + -B --> -(A + B)
2016 if (Value *LHSV = dyn_castNegVal(LHS)) {
2017 if (LHS->getType()->isIntOrIntVector()) {
2018 if (Value *RHSV = dyn_castNegVal(RHS)) {
2019 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2020 InsertNewInstBefore(NewAdd, I);
2021 return BinaryOperator::CreateNeg(NewAdd);
2025 return BinaryOperator::CreateSub(RHS, LHSV);
2029 if (!isa<Constant>(RHS))
2030 if (Value *V = dyn_castNegVal(RHS))
2031 return BinaryOperator::CreateSub(LHS, V);
2035 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2036 if (X == RHS) // X*C + X --> X * (C+1)
2037 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2039 // X*C1 + X*C2 --> X * (C1+C2)
2041 if (X == dyn_castFoldableMul(RHS, C1))
2042 return BinaryOperator::CreateMul(X, Add(C1, C2));
2045 // X + X*C --> X * (C+1)
2046 if (dyn_castFoldableMul(RHS, C2) == LHS)
2047 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2049 // X + ~X --> -1 since ~X = -X-1
2050 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2051 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2054 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2055 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2056 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2059 // A+B --> A|B iff A and B have no bits set in common.
2060 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2061 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2062 APInt LHSKnownOne(IT->getBitWidth(), 0);
2063 APInt LHSKnownZero(IT->getBitWidth(), 0);
2064 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2065 if (LHSKnownZero != 0) {
2066 APInt RHSKnownOne(IT->getBitWidth(), 0);
2067 APInt RHSKnownZero(IT->getBitWidth(), 0);
2068 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2070 // No bits in common -> bitwise or.
2071 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2072 return BinaryOperator::CreateOr(LHS, RHS);
2076 // W*X + Y*Z --> W * (X+Z) iff W == Y
2077 if (I.getType()->isIntOrIntVector()) {
2078 Value *W, *X, *Y, *Z;
2079 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2080 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2084 } else if (Y == X) {
2086 } else if (X == Z) {
2093 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2094 LHS->getName()), I);
2095 return BinaryOperator::CreateMul(W, NewAdd);
2100 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2102 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2103 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2105 // (X & FF00) + xx00 -> (X+xx00) & FF00
2106 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2107 Constant *Anded = And(CRHS, C2);
2108 if (Anded == CRHS) {
2109 // See if all bits from the first bit set in the Add RHS up are included
2110 // in the mask. First, get the rightmost bit.
2111 const APInt& AddRHSV = CRHS->getValue();
2113 // Form a mask of all bits from the lowest bit added through the top.
2114 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2116 // See if the and mask includes all of these bits.
2117 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2119 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2120 // Okay, the xform is safe. Insert the new add pronto.
2121 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2122 LHS->getName()), I);
2123 return BinaryOperator::CreateAnd(NewAdd, C2);
2128 // Try to fold constant add into select arguments.
2129 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2130 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2134 // add (cast *A to intptrtype) B ->
2135 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2137 CastInst *CI = dyn_cast<CastInst>(LHS);
2140 CI = dyn_cast<CastInst>(RHS);
2143 if (CI && CI->getType()->isSized() &&
2144 (CI->getType()->getPrimitiveSizeInBits() ==
2145 TD->getIntPtrType()->getPrimitiveSizeInBits())
2146 && isa<PointerType>(CI->getOperand(0)->getType())) {
2148 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2149 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2150 PointerType::get(Type::Int8Ty, AS), I);
2151 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2152 return new PtrToIntInst(I2, CI->getType());
2156 // add (select X 0 (sub n A)) A --> select X A n
2158 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2161 SI = dyn_cast<SelectInst>(RHS);
2164 if (SI && SI->hasOneUse()) {
2165 Value *TV = SI->getTrueValue();
2166 Value *FV = SI->getFalseValue();
2169 // Can we fold the add into the argument of the select?
2170 // We check both true and false select arguments for a matching subtract.
2171 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2172 A == Other) // Fold the add into the true select value.
2173 return SelectInst::Create(SI->getCondition(), N, A);
2174 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2175 A == Other) // Fold the add into the false select value.
2176 return SelectInst::Create(SI->getCondition(), A, N);
2180 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2181 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2182 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2183 return ReplaceInstUsesWith(I, LHS);
2185 // Check for (add (sext x), y), see if we can merge this into an
2186 // integer add followed by a sext.
2187 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2188 // (add (sext x), cst) --> (sext (add x, cst'))
2189 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2191 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2192 if (LHSConv->hasOneUse() &&
2193 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2194 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2195 // Insert the new, smaller add.
2196 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2198 InsertNewInstBefore(NewAdd, I);
2199 return new SExtInst(NewAdd, I.getType());
2203 // (add (sext x), (sext y)) --> (sext (add int x, y))
2204 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2205 // Only do this if x/y have the same type, if at last one of them has a
2206 // single use (so we don't increase the number of sexts), and if the
2207 // integer add will not overflow.
2208 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2209 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2210 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2211 RHSConv->getOperand(0))) {
2212 // Insert the new integer add.
2213 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2214 RHSConv->getOperand(0),
2216 InsertNewInstBefore(NewAdd, I);
2217 return new SExtInst(NewAdd, I.getType());
2222 // Check for (add double (sitofp x), y), see if we can merge this into an
2223 // integer add followed by a promotion.
2224 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2225 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2226 // ... if the constant fits in the integer value. This is useful for things
2227 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2228 // requires a constant pool load, and generally allows the add to be better
2230 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2232 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2233 if (LHSConv->hasOneUse() &&
2234 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2235 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2236 // Insert the new integer add.
2237 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2239 InsertNewInstBefore(NewAdd, I);
2240 return new SIToFPInst(NewAdd, I.getType());
2244 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2245 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2246 // Only do this if x/y have the same type, if at last one of them has a
2247 // single use (so we don't increase the number of int->fp conversions),
2248 // and if the integer add will not overflow.
2249 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2250 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2251 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2252 RHSConv->getOperand(0))) {
2253 // Insert the new integer add.
2254 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2255 RHSConv->getOperand(0),
2257 InsertNewInstBefore(NewAdd, I);
2258 return new SIToFPInst(NewAdd, I.getType());
2263 return Changed ? &I : 0;
2266 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2267 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2269 if (Op0 == Op1) // sub X, X -> 0
2270 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2272 // If this is a 'B = x-(-A)', change to B = x+A...
2273 if (Value *V = dyn_castNegVal(Op1))
2274 return BinaryOperator::CreateAdd(Op0, V);
2276 if (isa<UndefValue>(Op0))
2277 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2278 if (isa<UndefValue>(Op1))
2279 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2281 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2282 // Replace (-1 - A) with (~A)...
2283 if (C->isAllOnesValue())
2284 return BinaryOperator::CreateNot(Op1);
2286 // C - ~X == X + (1+C)
2288 if (match(Op1, m_Not(m_Value(X))))
2289 return BinaryOperator::CreateAdd(X, AddOne(C));
2291 // -(X >>u 31) -> (X >>s 31)
2292 // -(X >>s 31) -> (X >>u 31)
2294 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2295 if (SI->getOpcode() == Instruction::LShr) {
2296 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2297 // Check to see if we are shifting out everything but the sign bit.
2298 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2299 SI->getType()->getPrimitiveSizeInBits()-1) {
2300 // Ok, the transformation is safe. Insert AShr.
2301 return BinaryOperator::Create(Instruction::AShr,
2302 SI->getOperand(0), CU, SI->getName());
2306 else if (SI->getOpcode() == Instruction::AShr) {
2307 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2308 // Check to see if we are shifting out everything but the sign bit.
2309 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2310 SI->getType()->getPrimitiveSizeInBits()-1) {
2311 // Ok, the transformation is safe. Insert LShr.
2312 return BinaryOperator::CreateLShr(
2313 SI->getOperand(0), CU, SI->getName());
2320 // Try to fold constant sub into select arguments.
2321 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2322 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2325 if (isa<PHINode>(Op0))
2326 if (Instruction *NV = FoldOpIntoPhi(I))
2330 if (I.getType() == Type::Int1Ty)
2331 return BinaryOperator::CreateXor(Op0, Op1);
2333 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2334 if (Op1I->getOpcode() == Instruction::Add &&
2335 !Op0->getType()->isFPOrFPVector()) {
2336 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2337 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2338 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2339 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2340 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2341 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2342 // C1-(X+C2) --> (C1-C2)-X
2343 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2344 Op1I->getOperand(0));
2348 if (Op1I->hasOneUse()) {
2349 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2350 // is not used by anyone else...
2352 if (Op1I->getOpcode() == Instruction::Sub &&
2353 !Op1I->getType()->isFPOrFPVector()) {
2354 // Swap the two operands of the subexpr...
2355 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2356 Op1I->setOperand(0, IIOp1);
2357 Op1I->setOperand(1, IIOp0);
2359 // Create the new top level add instruction...
2360 return BinaryOperator::CreateAdd(Op0, Op1);
2363 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2365 if (Op1I->getOpcode() == Instruction::And &&
2366 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2367 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2370 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2371 return BinaryOperator::CreateAnd(Op0, NewNot);
2374 // 0 - (X sdiv C) -> (X sdiv -C)
2375 if (Op1I->getOpcode() == Instruction::SDiv)
2376 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2378 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2379 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2380 ConstantExpr::getNeg(DivRHS));
2382 // X - X*C --> X * (1-C)
2383 ConstantInt *C2 = 0;
2384 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2385 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2386 return BinaryOperator::CreateMul(Op0, CP1);
2389 // X - ((X / Y) * Y) --> X % Y
2390 if (Op1I->getOpcode() == Instruction::Mul)
2391 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2392 if (Op0 == I->getOperand(0) &&
2393 Op1I->getOperand(1) == I->getOperand(1)) {
2394 if (I->getOpcode() == Instruction::SDiv)
2395 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2396 if (I->getOpcode() == Instruction::UDiv)
2397 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2402 if (!Op0->getType()->isFPOrFPVector())
2403 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2404 if (Op0I->getOpcode() == Instruction::Add) {
2405 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2406 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2407 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2408 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2409 } else if (Op0I->getOpcode() == Instruction::Sub) {
2410 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2411 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2416 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2417 if (X == Op1) // X*C - X --> X * (C-1)
2418 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2420 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2421 if (X == dyn_castFoldableMul(Op1, C2))
2422 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2427 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2428 /// comparison only checks the sign bit. If it only checks the sign bit, set
2429 /// TrueIfSigned if the result of the comparison is true when the input value is
2431 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2432 bool &TrueIfSigned) {
2434 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2435 TrueIfSigned = true;
2436 return RHS->isZero();
2437 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2438 TrueIfSigned = true;
2439 return RHS->isAllOnesValue();
2440 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2441 TrueIfSigned = false;
2442 return RHS->isAllOnesValue();
2443 case ICmpInst::ICMP_UGT:
2444 // True if LHS u> RHS and RHS == high-bit-mask - 1
2445 TrueIfSigned = true;
2446 return RHS->getValue() ==
2447 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2448 case ICmpInst::ICMP_UGE:
2449 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2450 TrueIfSigned = true;
2451 return RHS->getValue().isSignBit();
2457 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2458 bool Changed = SimplifyCommutative(I);
2459 Value *Op0 = I.getOperand(0);
2461 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2462 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2464 // Simplify mul instructions with a constant RHS...
2465 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2466 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2468 // ((X << C1)*C2) == (X * (C2 << C1))
2469 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2470 if (SI->getOpcode() == Instruction::Shl)
2471 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2472 return BinaryOperator::CreateMul(SI->getOperand(0),
2473 ConstantExpr::getShl(CI, ShOp));
2476 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2477 if (CI->equalsInt(1)) // X * 1 == X
2478 return ReplaceInstUsesWith(I, Op0);
2479 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2480 return BinaryOperator::CreateNeg(Op0, I.getName());
2482 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2483 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2484 return BinaryOperator::CreateShl(Op0,
2485 ConstantInt::get(Op0->getType(), Val.logBase2()));
2487 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2488 if (Op1F->isNullValue())
2489 return ReplaceInstUsesWith(I, Op1);
2491 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2492 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2493 // We need a better interface for long double here.
2494 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2495 if (Op1F->isExactlyValue(1.0))
2496 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2499 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2500 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2501 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2502 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2503 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2505 InsertNewInstBefore(Add, I);
2506 Value *C1C2 = ConstantExpr::getMul(Op1,
2507 cast<Constant>(Op0I->getOperand(1)));
2508 return BinaryOperator::CreateAdd(Add, C1C2);
2512 // Try to fold constant mul into select arguments.
2513 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2514 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2517 if (isa<PHINode>(Op0))
2518 if (Instruction *NV = FoldOpIntoPhi(I))
2522 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2523 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2524 return BinaryOperator::CreateMul(Op0v, Op1v);
2526 if (I.getType() == Type::Int1Ty)
2527 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2529 // If one of the operands of the multiply is a cast from a boolean value, then
2530 // we know the bool is either zero or one, so this is a 'masking' multiply.
2531 // See if we can simplify things based on how the boolean was originally
2533 CastInst *BoolCast = 0;
2534 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2535 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2538 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2539 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2542 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2543 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2544 const Type *SCOpTy = SCIOp0->getType();
2547 // If the icmp is true iff the sign bit of X is set, then convert this
2548 // multiply into a shift/and combination.
2549 if (isa<ConstantInt>(SCIOp1) &&
2550 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2552 // Shift the X value right to turn it into "all signbits".
2553 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2554 SCOpTy->getPrimitiveSizeInBits()-1);
2556 InsertNewInstBefore(
2557 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2558 BoolCast->getOperand(0)->getName()+
2561 // If the multiply type is not the same as the source type, sign extend
2562 // or truncate to the multiply type.
2563 if (I.getType() != V->getType()) {
2564 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2565 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2566 Instruction::CastOps opcode =
2567 (SrcBits == DstBits ? Instruction::BitCast :
2568 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2569 V = InsertCastBefore(opcode, V, I.getType(), I);
2572 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2573 return BinaryOperator::CreateAnd(V, OtherOp);
2578 return Changed ? &I : 0;
2581 /// This function implements the transforms on div instructions that work
2582 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2583 /// used by the visitors to those instructions.
2584 /// @brief Transforms common to all three div instructions
2585 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2586 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2588 // undef / X -> 0 for integer.
2589 // undef / X -> undef for FP (the undef could be a snan).
2590 if (isa<UndefValue>(Op0)) {
2591 if (Op0->getType()->isFPOrFPVector())
2592 return ReplaceInstUsesWith(I, Op0);
2593 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2596 // X / undef -> undef
2597 if (isa<UndefValue>(Op1))
2598 return ReplaceInstUsesWith(I, Op1);
2600 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2601 // This does not apply for fdiv.
2602 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2603 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2604 // the same basic block, then we replace the select with Y, and the
2605 // condition of the select with false (if the cond value is in the same BB).
2606 // If the select has uses other than the div, this allows them to be
2607 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2608 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2609 if (ST->isNullValue()) {
2610 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2611 if (CondI && CondI->getParent() == I.getParent())
2612 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2613 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2614 I.setOperand(1, SI->getOperand(2));
2616 UpdateValueUsesWith(SI, SI->getOperand(2));
2620 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2621 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2622 if (ST->isNullValue()) {
2623 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2624 if (CondI && CondI->getParent() == I.getParent())
2625 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2626 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2627 I.setOperand(1, SI->getOperand(1));
2629 UpdateValueUsesWith(SI, SI->getOperand(1));
2637 /// This function implements the transforms common to both integer division
2638 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2639 /// division instructions.
2640 /// @brief Common integer divide transforms
2641 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2642 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2644 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2646 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2647 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2648 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2649 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2652 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2653 return ReplaceInstUsesWith(I, CI);
2656 if (Instruction *Common = commonDivTransforms(I))
2659 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2661 if (RHS->equalsInt(1))
2662 return ReplaceInstUsesWith(I, Op0);
2664 // (X / C1) / C2 -> X / (C1*C2)
2665 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2666 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2667 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2668 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2669 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2671 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2672 Multiply(RHS, LHSRHS));
2675 if (!RHS->isZero()) { // avoid X udiv 0
2676 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2677 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2679 if (isa<PHINode>(Op0))
2680 if (Instruction *NV = FoldOpIntoPhi(I))
2685 // 0 / X == 0, we don't need to preserve faults!
2686 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2687 if (LHS->equalsInt(0))
2688 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2690 // It can't be division by zero, hence it must be division by one.
2691 if (I.getType() == Type::Int1Ty)
2692 return ReplaceInstUsesWith(I, Op0);
2697 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2698 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2700 // Handle the integer div common cases
2701 if (Instruction *Common = commonIDivTransforms(I))
2704 // X udiv C^2 -> X >> C
2705 // Check to see if this is an unsigned division with an exact power of 2,
2706 // if so, convert to a right shift.
2707 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2708 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2709 return BinaryOperator::CreateLShr(Op0,
2710 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2713 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2714 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2715 if (RHSI->getOpcode() == Instruction::Shl &&
2716 isa<ConstantInt>(RHSI->getOperand(0))) {
2717 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2718 if (C1.isPowerOf2()) {
2719 Value *N = RHSI->getOperand(1);
2720 const Type *NTy = N->getType();
2721 if (uint32_t C2 = C1.logBase2()) {
2722 Constant *C2V = ConstantInt::get(NTy, C2);
2723 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2725 return BinaryOperator::CreateLShr(Op0, N);
2730 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2731 // where C1&C2 are powers of two.
2732 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2733 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2734 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2735 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2736 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2737 // Compute the shift amounts
2738 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2739 // Construct the "on true" case of the select
2740 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2741 Instruction *TSI = BinaryOperator::CreateLShr(
2742 Op0, TC, SI->getName()+".t");
2743 TSI = InsertNewInstBefore(TSI, I);
2745 // Construct the "on false" case of the select
2746 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2747 Instruction *FSI = BinaryOperator::CreateLShr(
2748 Op0, FC, SI->getName()+".f");
2749 FSI = InsertNewInstBefore(FSI, I);
2751 // construct the select instruction and return it.
2752 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2758 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2759 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2761 // Handle the integer div common cases
2762 if (Instruction *Common = commonIDivTransforms(I))
2765 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2767 if (RHS->isAllOnesValue())
2768 return BinaryOperator::CreateNeg(Op0);
2771 if (Value *LHSNeg = dyn_castNegVal(Op0))
2772 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2775 // If the sign bits of both operands are zero (i.e. we can prove they are
2776 // unsigned inputs), turn this into a udiv.
2777 if (I.getType()->isInteger()) {
2778 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2779 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2780 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2781 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2788 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2789 return commonDivTransforms(I);
2792 /// This function implements the transforms on rem instructions that work
2793 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2794 /// is used by the visitors to those instructions.
2795 /// @brief Transforms common to all three rem instructions
2796 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2797 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2799 // 0 % X == 0 for integer, we don't need to preserve faults!
2800 if (Constant *LHS = dyn_cast<Constant>(Op0))
2801 if (LHS->isNullValue())
2802 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2804 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2805 if (I.getType()->isFPOrFPVector())
2806 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2807 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2809 if (isa<UndefValue>(Op1))
2810 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2812 // Handle cases involving: rem X, (select Cond, Y, Z)
2813 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2814 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2815 // the same basic block, then we replace the select with Y, and the
2816 // condition of the select with false (if the cond value is in the same
2817 // BB). If the select has uses other than the div, this allows them to be
2819 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2820 if (ST->isNullValue()) {
2821 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2822 if (CondI && CondI->getParent() == I.getParent())
2823 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2824 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2825 I.setOperand(1, SI->getOperand(2));
2827 UpdateValueUsesWith(SI, SI->getOperand(2));
2830 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2831 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2832 if (ST->isNullValue()) {
2833 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2834 if (CondI && CondI->getParent() == I.getParent())
2835 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2836 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2837 I.setOperand(1, SI->getOperand(1));
2839 UpdateValueUsesWith(SI, SI->getOperand(1));
2847 /// This function implements the transforms common to both integer remainder
2848 /// instructions (urem and srem). It is called by the visitors to those integer
2849 /// remainder instructions.
2850 /// @brief Common integer remainder transforms
2851 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2852 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2854 if (Instruction *common = commonRemTransforms(I))
2857 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2858 // X % 0 == undef, we don't need to preserve faults!
2859 if (RHS->equalsInt(0))
2860 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2862 if (RHS->equalsInt(1)) // X % 1 == 0
2863 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2865 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2866 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2867 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2869 } else if (isa<PHINode>(Op0I)) {
2870 if (Instruction *NV = FoldOpIntoPhi(I))
2874 // See if we can fold away this rem instruction.
2875 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2876 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2877 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2878 KnownZero, KnownOne))
2886 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2887 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2889 if (Instruction *common = commonIRemTransforms(I))
2892 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2893 // X urem C^2 -> X and C
2894 // Check to see if this is an unsigned remainder with an exact power of 2,
2895 // if so, convert to a bitwise and.
2896 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2897 if (C->getValue().isPowerOf2())
2898 return BinaryOperator::CreateAnd(Op0, SubOne(C));
2901 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2902 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2903 if (RHSI->getOpcode() == Instruction::Shl &&
2904 isa<ConstantInt>(RHSI->getOperand(0))) {
2905 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2906 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2907 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
2909 return BinaryOperator::CreateAnd(Op0, Add);
2914 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2915 // where C1&C2 are powers of two.
2916 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2917 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2918 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2919 // STO == 0 and SFO == 0 handled above.
2920 if ((STO->getValue().isPowerOf2()) &&
2921 (SFO->getValue().isPowerOf2())) {
2922 Value *TrueAnd = InsertNewInstBefore(
2923 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2924 Value *FalseAnd = InsertNewInstBefore(
2925 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2926 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
2934 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2935 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2937 // Handle the integer rem common cases
2938 if (Instruction *common = commonIRemTransforms(I))
2941 if (Value *RHSNeg = dyn_castNegVal(Op1))
2942 if (!isa<ConstantInt>(RHSNeg) ||
2943 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2945 AddUsesToWorkList(I);
2946 I.setOperand(1, RHSNeg);
2950 // If the sign bits of both operands are zero (i.e. we can prove they are
2951 // unsigned inputs), turn this into a urem.
2952 if (I.getType()->isInteger()) {
2953 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2954 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2955 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2956 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
2963 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2964 return commonRemTransforms(I);
2967 // isMaxValueMinusOne - return true if this is Max-1
2968 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2969 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2971 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2972 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2975 // isMinValuePlusOne - return true if this is Min+1
2976 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2978 return C->getValue() == 1; // unsigned
2980 // Calculate 1111111111000000000000
2981 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2982 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2985 // isOneBitSet - Return true if there is exactly one bit set in the specified
2987 static bool isOneBitSet(const ConstantInt *CI) {
2988 return CI->getValue().isPowerOf2();
2991 // isHighOnes - Return true if the constant is of the form 1+0+.
2992 // This is the same as lowones(~X).
2993 static bool isHighOnes(const ConstantInt *CI) {
2994 return (~CI->getValue() + 1).isPowerOf2();
2997 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2998 /// are carefully arranged to allow folding of expressions such as:
3000 /// (A < B) | (A > B) --> (A != B)
3002 /// Note that this is only valid if the first and second predicates have the
3003 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3005 /// Three bits are used to represent the condition, as follows:
3010 /// <=> Value Definition
3011 /// 000 0 Always false
3018 /// 111 7 Always true
3020 static unsigned getICmpCode(const ICmpInst *ICI) {
3021 switch (ICI->getPredicate()) {
3023 case ICmpInst::ICMP_UGT: return 1; // 001
3024 case ICmpInst::ICMP_SGT: return 1; // 001
3025 case ICmpInst::ICMP_EQ: return 2; // 010
3026 case ICmpInst::ICMP_UGE: return 3; // 011
3027 case ICmpInst::ICMP_SGE: return 3; // 011
3028 case ICmpInst::ICMP_ULT: return 4; // 100
3029 case ICmpInst::ICMP_SLT: return 4; // 100
3030 case ICmpInst::ICMP_NE: return 5; // 101
3031 case ICmpInst::ICMP_ULE: return 6; // 110
3032 case ICmpInst::ICMP_SLE: return 6; // 110
3035 assert(0 && "Invalid ICmp predicate!");
3040 /// getICmpValue - This is the complement of getICmpCode, which turns an
3041 /// opcode and two operands into either a constant true or false, or a brand
3042 /// new ICmp instruction. The sign is passed in to determine which kind
3043 /// of predicate to use in new icmp instructions.
3044 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3046 default: assert(0 && "Illegal ICmp code!");
3047 case 0: return ConstantInt::getFalse();
3050 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3052 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3053 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3056 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3058 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3061 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3063 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3064 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3067 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3069 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3070 case 7: return ConstantInt::getTrue();
3074 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3075 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3076 (ICmpInst::isSignedPredicate(p1) &&
3077 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3078 (ICmpInst::isSignedPredicate(p2) &&
3079 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3083 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3084 struct FoldICmpLogical {
3087 ICmpInst::Predicate pred;
3088 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3089 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3090 pred(ICI->getPredicate()) {}
3091 bool shouldApply(Value *V) const {
3092 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3093 if (PredicatesFoldable(pred, ICI->getPredicate()))
3094 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3095 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3098 Instruction *apply(Instruction &Log) const {
3099 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3100 if (ICI->getOperand(0) != LHS) {
3101 assert(ICI->getOperand(1) == LHS);
3102 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3105 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3106 unsigned LHSCode = getICmpCode(ICI);
3107 unsigned RHSCode = getICmpCode(RHSICI);
3109 switch (Log.getOpcode()) {
3110 case Instruction::And: Code = LHSCode & RHSCode; break;
3111 case Instruction::Or: Code = LHSCode | RHSCode; break;
3112 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3113 default: assert(0 && "Illegal logical opcode!"); return 0;
3116 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3117 ICmpInst::isSignedPredicate(ICI->getPredicate());
3119 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3120 if (Instruction *I = dyn_cast<Instruction>(RV))
3122 // Otherwise, it's a constant boolean value...
3123 return IC.ReplaceInstUsesWith(Log, RV);
3126 } // end anonymous namespace
3128 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3129 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3130 // guaranteed to be a binary operator.
3131 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3133 ConstantInt *AndRHS,
3134 BinaryOperator &TheAnd) {
3135 Value *X = Op->getOperand(0);
3136 Constant *Together = 0;
3138 Together = And(AndRHS, OpRHS);
3140 switch (Op->getOpcode()) {
3141 case Instruction::Xor:
3142 if (Op->hasOneUse()) {
3143 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3144 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3145 InsertNewInstBefore(And, TheAnd);
3147 return BinaryOperator::CreateXor(And, Together);
3150 case Instruction::Or:
3151 if (Together == AndRHS) // (X | C) & C --> C
3152 return ReplaceInstUsesWith(TheAnd, AndRHS);
3154 if (Op->hasOneUse() && Together != OpRHS) {
3155 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3156 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3157 InsertNewInstBefore(Or, TheAnd);
3159 return BinaryOperator::CreateAnd(Or, AndRHS);
3162 case Instruction::Add:
3163 if (Op->hasOneUse()) {
3164 // Adding a one to a single bit bit-field should be turned into an XOR
3165 // of the bit. First thing to check is to see if this AND is with a
3166 // single bit constant.
3167 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3169 // If there is only one bit set...
3170 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3171 // Ok, at this point, we know that we are masking the result of the
3172 // ADD down to exactly one bit. If the constant we are adding has
3173 // no bits set below this bit, then we can eliminate the ADD.
3174 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3176 // Check to see if any bits below the one bit set in AndRHSV are set.
3177 if ((AddRHS & (AndRHSV-1)) == 0) {
3178 // If not, the only thing that can effect the output of the AND is
3179 // the bit specified by AndRHSV. If that bit is set, the effect of
3180 // the XOR is to toggle the bit. If it is clear, then the ADD has
3182 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3183 TheAnd.setOperand(0, X);
3186 // Pull the XOR out of the AND.
3187 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3188 InsertNewInstBefore(NewAnd, TheAnd);
3189 NewAnd->takeName(Op);
3190 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3197 case Instruction::Shl: {
3198 // We know that the AND will not produce any of the bits shifted in, so if
3199 // the anded constant includes them, clear them now!
3201 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3202 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3203 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3204 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3206 if (CI->getValue() == ShlMask) {
3207 // Masking out bits that the shift already masks
3208 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3209 } else if (CI != AndRHS) { // Reducing bits set in and.
3210 TheAnd.setOperand(1, CI);
3215 case Instruction::LShr:
3217 // We know that the AND will not produce any of the bits shifted in, so if
3218 // the anded constant includes them, clear them now! This only applies to
3219 // unsigned shifts, because a signed shr may bring in set bits!
3221 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3222 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3223 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3224 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3226 if (CI->getValue() == ShrMask) {
3227 // Masking out bits that the shift already masks.
3228 return ReplaceInstUsesWith(TheAnd, Op);
3229 } else if (CI != AndRHS) {
3230 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3235 case Instruction::AShr:
3237 // See if this is shifting in some sign extension, then masking it out
3239 if (Op->hasOneUse()) {
3240 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3241 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3242 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3243 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3244 if (C == AndRHS) { // Masking out bits shifted in.
3245 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3246 // Make the argument unsigned.
3247 Value *ShVal = Op->getOperand(0);
3248 ShVal = InsertNewInstBefore(
3249 BinaryOperator::CreateLShr(ShVal, OpRHS,
3250 Op->getName()), TheAnd);
3251 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3260 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3261 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3262 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3263 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3264 /// insert new instructions.
3265 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3266 bool isSigned, bool Inside,
3268 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3269 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3270 "Lo is not <= Hi in range emission code!");
3273 if (Lo == Hi) // Trivially false.
3274 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3276 // V >= Min && V < Hi --> V < Hi
3277 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3278 ICmpInst::Predicate pred = (isSigned ?
3279 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3280 return new ICmpInst(pred, V, Hi);
3283 // Emit V-Lo <u Hi-Lo
3284 Constant *NegLo = ConstantExpr::getNeg(Lo);
3285 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3286 InsertNewInstBefore(Add, IB);
3287 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3288 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3291 if (Lo == Hi) // Trivially true.
3292 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3294 // V < Min || V >= Hi -> V > Hi-1
3295 Hi = SubOne(cast<ConstantInt>(Hi));
3296 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3297 ICmpInst::Predicate pred = (isSigned ?
3298 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3299 return new ICmpInst(pred, V, Hi);
3302 // Emit V-Lo >u Hi-1-Lo
3303 // Note that Hi has already had one subtracted from it, above.
3304 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3305 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3306 InsertNewInstBefore(Add, IB);
3307 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3308 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3311 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3312 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3313 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3314 // not, since all 1s are not contiguous.
3315 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3316 const APInt& V = Val->getValue();
3317 uint32_t BitWidth = Val->getType()->getBitWidth();
3318 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3320 // look for the first zero bit after the run of ones
3321 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3322 // look for the first non-zero bit
3323 ME = V.getActiveBits();
3327 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3328 /// where isSub determines whether the operator is a sub. If we can fold one of
3329 /// the following xforms:
3331 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3332 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3333 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3335 /// return (A +/- B).
3337 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3338 ConstantInt *Mask, bool isSub,
3340 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3341 if (!LHSI || LHSI->getNumOperands() != 2 ||
3342 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3344 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3346 switch (LHSI->getOpcode()) {
3348 case Instruction::And:
3349 if (And(N, Mask) == Mask) {
3350 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3351 if ((Mask->getValue().countLeadingZeros() +
3352 Mask->getValue().countPopulation()) ==
3353 Mask->getValue().getBitWidth())
3356 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3357 // part, we don't need any explicit masks to take them out of A. If that
3358 // is all N is, ignore it.
3359 uint32_t MB = 0, ME = 0;
3360 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3361 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3362 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3363 if (MaskedValueIsZero(RHS, Mask))
3368 case Instruction::Or:
3369 case Instruction::Xor:
3370 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3371 if ((Mask->getValue().countLeadingZeros() +
3372 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3373 && And(N, Mask)->isZero())
3380 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3382 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3383 return InsertNewInstBefore(New, I);
3386 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3387 bool Changed = SimplifyCommutative(I);
3388 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3390 if (isa<UndefValue>(Op1)) // X & undef -> 0
3391 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3395 return ReplaceInstUsesWith(I, Op1);
3397 // See if we can simplify any instructions used by the instruction whose sole
3398 // purpose is to compute bits we don't care about.
3399 if (!isa<VectorType>(I.getType())) {
3400 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3401 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3402 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3403 KnownZero, KnownOne))
3406 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3407 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3408 return ReplaceInstUsesWith(I, I.getOperand(0));
3409 } else if (isa<ConstantAggregateZero>(Op1)) {
3410 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3414 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3415 const APInt& AndRHSMask = AndRHS->getValue();
3416 APInt NotAndRHS(~AndRHSMask);
3418 // Optimize a variety of ((val OP C1) & C2) combinations...
3419 if (isa<BinaryOperator>(Op0)) {
3420 Instruction *Op0I = cast<Instruction>(Op0);
3421 Value *Op0LHS = Op0I->getOperand(0);
3422 Value *Op0RHS = Op0I->getOperand(1);
3423 switch (Op0I->getOpcode()) {
3424 case Instruction::Xor:
3425 case Instruction::Or:
3426 // If the mask is only needed on one incoming arm, push it up.
3427 if (Op0I->hasOneUse()) {
3428 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3429 // Not masking anything out for the LHS, move to RHS.
3430 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3431 Op0RHS->getName()+".masked");
3432 InsertNewInstBefore(NewRHS, I);
3433 return BinaryOperator::Create(
3434 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3436 if (!isa<Constant>(Op0RHS) &&
3437 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3438 // Not masking anything out for the RHS, move to LHS.
3439 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3440 Op0LHS->getName()+".masked");
3441 InsertNewInstBefore(NewLHS, I);
3442 return BinaryOperator::Create(
3443 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3448 case Instruction::Add:
3449 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3450 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3451 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3452 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3453 return BinaryOperator::CreateAnd(V, AndRHS);
3454 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3455 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3458 case Instruction::Sub:
3459 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3460 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3461 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3462 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3463 return BinaryOperator::CreateAnd(V, AndRHS);
3467 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3468 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3470 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3471 // If this is an integer truncation or change from signed-to-unsigned, and
3472 // if the source is an and/or with immediate, transform it. This
3473 // frequently occurs for bitfield accesses.
3474 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3475 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3476 CastOp->getNumOperands() == 2)
3477 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3478 if (CastOp->getOpcode() == Instruction::And) {
3479 // Change: and (cast (and X, C1) to T), C2
3480 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3481 // This will fold the two constants together, which may allow
3482 // other simplifications.
3483 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3484 CastOp->getOperand(0), I.getType(),
3485 CastOp->getName()+".shrunk");
3486 NewCast = InsertNewInstBefore(NewCast, I);
3487 // trunc_or_bitcast(C1)&C2
3488 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3489 C3 = ConstantExpr::getAnd(C3, AndRHS);
3490 return BinaryOperator::CreateAnd(NewCast, C3);
3491 } else if (CastOp->getOpcode() == Instruction::Or) {
3492 // Change: and (cast (or X, C1) to T), C2
3493 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3494 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3495 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3496 return ReplaceInstUsesWith(I, AndRHS);
3502 // Try to fold constant and into select arguments.
3503 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3504 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3506 if (isa<PHINode>(Op0))
3507 if (Instruction *NV = FoldOpIntoPhi(I))
3511 Value *Op0NotVal = dyn_castNotVal(Op0);
3512 Value *Op1NotVal = dyn_castNotVal(Op1);
3514 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3515 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3517 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3518 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3519 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3520 I.getName()+".demorgan");
3521 InsertNewInstBefore(Or, I);
3522 return BinaryOperator::CreateNot(Or);
3526 Value *A = 0, *B = 0, *C = 0, *D = 0;
3527 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3528 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3529 return ReplaceInstUsesWith(I, Op1);
3531 // (A|B) & ~(A&B) -> A^B
3532 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3533 if ((A == C && B == D) || (A == D && B == C))
3534 return BinaryOperator::CreateXor(A, B);
3538 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3539 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3540 return ReplaceInstUsesWith(I, Op0);
3542 // ~(A&B) & (A|B) -> A^B
3543 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3544 if ((A == C && B == D) || (A == D && B == C))
3545 return BinaryOperator::CreateXor(A, B);
3549 if (Op0->hasOneUse() &&
3550 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3551 if (A == Op1) { // (A^B)&A -> A&(A^B)
3552 I.swapOperands(); // Simplify below
3553 std::swap(Op0, Op1);
3554 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3555 cast<BinaryOperator>(Op0)->swapOperands();
3556 I.swapOperands(); // Simplify below
3557 std::swap(Op0, Op1);
3560 if (Op1->hasOneUse() &&
3561 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3562 if (B == Op0) { // B&(A^B) -> B&(B^A)
3563 cast<BinaryOperator>(Op1)->swapOperands();
3566 if (A == Op0) { // A&(A^B) -> A & ~B
3567 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3568 InsertNewInstBefore(NotB, I);
3569 return BinaryOperator::CreateAnd(A, NotB);
3574 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3575 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3576 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3579 Value *LHSVal, *RHSVal;
3580 ConstantInt *LHSCst, *RHSCst;
3581 ICmpInst::Predicate LHSCC, RHSCC;
3582 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3583 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3584 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3585 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3586 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3587 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3588 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3589 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3591 // Don't try to fold ICMP_SLT + ICMP_ULT.
3592 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3593 ICmpInst::isSignedPredicate(LHSCC) ==
3594 ICmpInst::isSignedPredicate(RHSCC))) {
3595 // Ensure that the larger constant is on the RHS.
3596 ICmpInst::Predicate GT;
3597 if (ICmpInst::isSignedPredicate(LHSCC) ||
3598 (ICmpInst::isEquality(LHSCC) &&
3599 ICmpInst::isSignedPredicate(RHSCC)))
3600 GT = ICmpInst::ICMP_SGT;
3602 GT = ICmpInst::ICMP_UGT;
3604 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3605 ICmpInst *LHS = cast<ICmpInst>(Op0);
3606 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3607 std::swap(LHS, RHS);
3608 std::swap(LHSCst, RHSCst);
3609 std::swap(LHSCC, RHSCC);
3612 // At this point, we know we have have two icmp instructions
3613 // comparing a value against two constants and and'ing the result
3614 // together. Because of the above check, we know that we only have
3615 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3616 // (from the FoldICmpLogical check above), that the two constants
3617 // are not equal and that the larger constant is on the RHS
3618 assert(LHSCst != RHSCst && "Compares not folded above?");
3621 default: assert(0 && "Unknown integer condition code!");
3622 case ICmpInst::ICMP_EQ:
3624 default: assert(0 && "Unknown integer condition code!");
3625 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3626 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3627 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3628 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3629 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3630 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3631 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3632 return ReplaceInstUsesWith(I, LHS);
3634 case ICmpInst::ICMP_NE:
3636 default: assert(0 && "Unknown integer condition code!");
3637 case ICmpInst::ICMP_ULT:
3638 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3639 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3640 break; // (X != 13 & X u< 15) -> no change
3641 case ICmpInst::ICMP_SLT:
3642 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3643 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3644 break; // (X != 13 & X s< 15) -> no change
3645 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3646 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3647 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3648 return ReplaceInstUsesWith(I, RHS);
3649 case ICmpInst::ICMP_NE:
3650 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3651 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3652 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
3653 LHSVal->getName()+".off");
3654 InsertNewInstBefore(Add, I);
3655 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3656 ConstantInt::get(Add->getType(), 1));
3658 break; // (X != 13 & X != 15) -> no change
3661 case ICmpInst::ICMP_ULT:
3663 default: assert(0 && "Unknown integer condition code!");
3664 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3665 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3666 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3667 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3669 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3670 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3671 return ReplaceInstUsesWith(I, LHS);
3672 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3676 case ICmpInst::ICMP_SLT:
3678 default: assert(0 && "Unknown integer condition code!");
3679 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3680 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3681 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3682 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3684 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3685 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3686 return ReplaceInstUsesWith(I, LHS);
3687 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3691 case ICmpInst::ICMP_UGT:
3693 default: assert(0 && "Unknown integer condition code!");
3694 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3695 return ReplaceInstUsesWith(I, LHS);
3696 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3697 return ReplaceInstUsesWith(I, RHS);
3698 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3700 case ICmpInst::ICMP_NE:
3701 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3702 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3703 break; // (X u> 13 & X != 15) -> no change
3704 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3705 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3707 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3711 case ICmpInst::ICMP_SGT:
3713 default: assert(0 && "Unknown integer condition code!");
3714 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3715 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3716 return ReplaceInstUsesWith(I, RHS);
3717 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3719 case ICmpInst::ICMP_NE:
3720 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3721 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3722 break; // (X s> 13 & X != 15) -> no change
3723 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3724 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3726 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3734 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3735 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3736 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3737 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3738 const Type *SrcTy = Op0C->getOperand(0)->getType();
3739 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3740 // Only do this if the casts both really cause code to be generated.
3741 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3743 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3745 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3746 Op1C->getOperand(0),
3748 InsertNewInstBefore(NewOp, I);
3749 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3753 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3754 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3755 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3756 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3757 SI0->getOperand(1) == SI1->getOperand(1) &&
3758 (SI0->hasOneUse() || SI1->hasOneUse())) {
3759 Instruction *NewOp =
3760 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3762 SI0->getName()), I);
3763 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3764 SI1->getOperand(1));
3768 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3769 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3770 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3771 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3772 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3773 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3774 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3775 // If either of the constants are nans, then the whole thing returns
3777 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3778 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3779 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3780 RHS->getOperand(0));
3785 return Changed ? &I : 0;
3788 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3789 /// in the result. If it does, and if the specified byte hasn't been filled in
3790 /// yet, fill it in and return false.
3791 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3792 Instruction *I = dyn_cast<Instruction>(V);
3793 if (I == 0) return true;
3795 // If this is an or instruction, it is an inner node of the bswap.
3796 if (I->getOpcode() == Instruction::Or)
3797 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3798 CollectBSwapParts(I->getOperand(1), ByteValues);
3800 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3801 // If this is a shift by a constant int, and it is "24", then its operand
3802 // defines a byte. We only handle unsigned types here.
3803 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3804 // Not shifting the entire input by N-1 bytes?
3805 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3806 8*(ByteValues.size()-1))
3810 if (I->getOpcode() == Instruction::Shl) {
3811 // X << 24 defines the top byte with the lowest of the input bytes.
3812 DestNo = ByteValues.size()-1;
3814 // X >>u 24 defines the low byte with the highest of the input bytes.
3818 // If the destination byte value is already defined, the values are or'd
3819 // together, which isn't a bswap (unless it's an or of the same bits).
3820 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3822 ByteValues[DestNo] = I->getOperand(0);
3826 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3828 Value *Shift = 0, *ShiftLHS = 0;
3829 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3830 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3831 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3833 Instruction *SI = cast<Instruction>(Shift);
3835 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3836 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3837 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3840 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3842 if (AndAmt->getValue().getActiveBits() > 64)
3844 uint64_t AndAmtVal = AndAmt->getZExtValue();
3845 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3846 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3848 // Unknown mask for bswap.
3849 if (DestByte == ByteValues.size()) return true;
3851 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3853 if (SI->getOpcode() == Instruction::Shl)
3854 SrcByte = DestByte - ShiftBytes;
3856 SrcByte = DestByte + ShiftBytes;
3858 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3859 if (SrcByte != ByteValues.size()-DestByte-1)
3862 // If the destination byte value is already defined, the values are or'd
3863 // together, which isn't a bswap (unless it's an or of the same bits).
3864 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3866 ByteValues[DestByte] = SI->getOperand(0);
3870 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3871 /// If so, insert the new bswap intrinsic and return it.
3872 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3873 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3874 if (!ITy || ITy->getBitWidth() % 16)
3875 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3877 /// ByteValues - For each byte of the result, we keep track of which value
3878 /// defines each byte.
3879 SmallVector<Value*, 8> ByteValues;
3880 ByteValues.resize(ITy->getBitWidth()/8);
3882 // Try to find all the pieces corresponding to the bswap.
3883 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3884 CollectBSwapParts(I.getOperand(1), ByteValues))
3887 // Check to see if all of the bytes come from the same value.
3888 Value *V = ByteValues[0];
3889 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3891 // Check to make sure that all of the bytes come from the same value.
3892 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3893 if (ByteValues[i] != V)
3895 const Type *Tys[] = { ITy };
3896 Module *M = I.getParent()->getParent()->getParent();
3897 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3898 return CallInst::Create(F, V);
3902 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3903 bool Changed = SimplifyCommutative(I);
3904 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3906 if (isa<UndefValue>(Op1)) // X | undef -> -1
3907 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3911 return ReplaceInstUsesWith(I, Op0);
3913 // See if we can simplify any instructions used by the instruction whose sole
3914 // purpose is to compute bits we don't care about.
3915 if (!isa<VectorType>(I.getType())) {
3916 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3917 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3918 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3919 KnownZero, KnownOne))
3921 } else if (isa<ConstantAggregateZero>(Op1)) {
3922 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3923 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3924 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3925 return ReplaceInstUsesWith(I, I.getOperand(1));
3931 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3932 ConstantInt *C1 = 0; Value *X = 0;
3933 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3934 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3935 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3936 InsertNewInstBefore(Or, I);
3938 return BinaryOperator::CreateAnd(Or,
3939 ConstantInt::get(RHS->getValue() | C1->getValue()));
3942 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3943 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3944 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3945 InsertNewInstBefore(Or, I);
3947 return BinaryOperator::CreateXor(Or,
3948 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3951 // Try to fold constant and into select arguments.
3952 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3953 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3955 if (isa<PHINode>(Op0))
3956 if (Instruction *NV = FoldOpIntoPhi(I))
3960 Value *A = 0, *B = 0;
3961 ConstantInt *C1 = 0, *C2 = 0;
3963 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3964 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3965 return ReplaceInstUsesWith(I, Op1);
3966 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3967 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3968 return ReplaceInstUsesWith(I, Op0);
3970 // (A | B) | C and A | (B | C) -> bswap if possible.
3971 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3972 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3973 match(Op1, m_Or(m_Value(), m_Value())) ||
3974 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3975 match(Op1, m_Shift(m_Value(), m_Value())))) {
3976 if (Instruction *BSwap = MatchBSwap(I))
3980 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3981 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3982 MaskedValueIsZero(Op1, C1->getValue())) {
3983 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
3984 InsertNewInstBefore(NOr, I);
3986 return BinaryOperator::CreateXor(NOr, C1);
3989 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3990 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3991 MaskedValueIsZero(Op0, C1->getValue())) {
3992 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
3993 InsertNewInstBefore(NOr, I);
3995 return BinaryOperator::CreateXor(NOr, C1);
3999 Value *C = 0, *D = 0;
4000 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4001 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4002 Value *V1 = 0, *V2 = 0, *V3 = 0;
4003 C1 = dyn_cast<ConstantInt>(C);
4004 C2 = dyn_cast<ConstantInt>(D);
4005 if (C1 && C2) { // (A & C1)|(B & C2)
4006 // If we have: ((V + N) & C1) | (V & C2)
4007 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4008 // replace with V+N.
4009 if (C1->getValue() == ~C2->getValue()) {
4010 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4011 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4012 // Add commutes, try both ways.
4013 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4014 return ReplaceInstUsesWith(I, A);
4015 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4016 return ReplaceInstUsesWith(I, A);
4018 // Or commutes, try both ways.
4019 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4020 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4021 // Add commutes, try both ways.
4022 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4023 return ReplaceInstUsesWith(I, B);
4024 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4025 return ReplaceInstUsesWith(I, B);
4028 V1 = 0; V2 = 0; V3 = 0;
4031 // Check to see if we have any common things being and'ed. If so, find the
4032 // terms for V1 & (V2|V3).
4033 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4034 if (A == B) // (A & C)|(A & D) == A & (C|D)
4035 V1 = A, V2 = C, V3 = D;
4036 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4037 V1 = A, V2 = B, V3 = C;
4038 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4039 V1 = C, V2 = A, V3 = D;
4040 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4041 V1 = C, V2 = A, V3 = B;
4045 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4046 return BinaryOperator::CreateAnd(V1, Or);
4051 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4052 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4053 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4054 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4055 SI0->getOperand(1) == SI1->getOperand(1) &&
4056 (SI0->hasOneUse() || SI1->hasOneUse())) {
4057 Instruction *NewOp =
4058 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4060 SI0->getName()), I);
4061 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4062 SI1->getOperand(1));
4066 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4067 if (A == Op1) // ~A | A == -1
4068 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4072 // Note, A is still live here!
4073 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4075 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4077 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4078 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4079 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4080 I.getName()+".demorgan"), I);
4081 return BinaryOperator::CreateNot(And);
4085 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4086 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4087 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4090 Value *LHSVal, *RHSVal;
4091 ConstantInt *LHSCst, *RHSCst;
4092 ICmpInst::Predicate LHSCC, RHSCC;
4093 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4094 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4095 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4096 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4097 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4098 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4099 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4100 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4101 // We can't fold (ugt x, C) | (sgt x, C2).
4102 PredicatesFoldable(LHSCC, RHSCC)) {
4103 // Ensure that the larger constant is on the RHS.
4104 ICmpInst *LHS = cast<ICmpInst>(Op0);
4106 if (ICmpInst::isSignedPredicate(LHSCC))
4107 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4109 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4112 std::swap(LHS, RHS);
4113 std::swap(LHSCst, RHSCst);
4114 std::swap(LHSCC, RHSCC);
4117 // At this point, we know we have have two icmp instructions
4118 // comparing a value against two constants and or'ing the result
4119 // together. Because of the above check, we know that we only have
4120 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4121 // FoldICmpLogical check above), that the two constants are not
4123 assert(LHSCst != RHSCst && "Compares not folded above?");
4126 default: assert(0 && "Unknown integer condition code!");
4127 case ICmpInst::ICMP_EQ:
4129 default: assert(0 && "Unknown integer condition code!");
4130 case ICmpInst::ICMP_EQ:
4131 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4132 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4133 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4134 LHSVal->getName()+".off");
4135 InsertNewInstBefore(Add, I);
4136 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4137 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4139 break; // (X == 13 | X == 15) -> no change
4140 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4141 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4143 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4144 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4145 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4146 return ReplaceInstUsesWith(I, RHS);
4149 case ICmpInst::ICMP_NE:
4151 default: assert(0 && "Unknown integer condition code!");
4152 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4153 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4154 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4155 return ReplaceInstUsesWith(I, LHS);
4156 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4157 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4158 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4159 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4162 case ICmpInst::ICMP_ULT:
4164 default: assert(0 && "Unknown integer condition code!");
4165 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4167 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4168 // If RHSCst is [us]MAXINT, it is always false. Not handling
4169 // this can cause overflow.
4170 if (RHSCst->isMaxValue(false))
4171 return ReplaceInstUsesWith(I, LHS);
4172 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4174 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4176 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4177 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4178 return ReplaceInstUsesWith(I, RHS);
4179 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4183 case ICmpInst::ICMP_SLT:
4185 default: assert(0 && "Unknown integer condition code!");
4186 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4188 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4189 // If RHSCst is [us]MAXINT, it is always false. Not handling
4190 // this can cause overflow.
4191 if (RHSCst->isMaxValue(true))
4192 return ReplaceInstUsesWith(I, LHS);
4193 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4195 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4197 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4198 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4199 return ReplaceInstUsesWith(I, RHS);
4200 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4204 case ICmpInst::ICMP_UGT:
4206 default: assert(0 && "Unknown integer condition code!");
4207 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4208 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4209 return ReplaceInstUsesWith(I, LHS);
4210 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4212 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4213 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4214 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4215 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4219 case ICmpInst::ICMP_SGT:
4221 default: assert(0 && "Unknown integer condition code!");
4222 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4223 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4224 return ReplaceInstUsesWith(I, LHS);
4225 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4227 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4228 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4229 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4230 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4238 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4239 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4240 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4241 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4242 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4243 !isa<ICmpInst>(Op1C->getOperand(0))) {
4244 const Type *SrcTy = Op0C->getOperand(0)->getType();
4245 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4246 // Only do this if the casts both really cause code to be
4248 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4250 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4252 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4253 Op1C->getOperand(0),
4255 InsertNewInstBefore(NewOp, I);
4256 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4263 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4264 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4265 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4266 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4267 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4268 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4269 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4270 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4271 // If either of the constants are nans, then the whole thing returns
4273 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4274 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4276 // Otherwise, no need to compare the two constants, compare the
4278 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4279 RHS->getOperand(0));
4284 return Changed ? &I : 0;
4289 // XorSelf - Implements: X ^ X --> 0
4292 XorSelf(Value *rhs) : RHS(rhs) {}
4293 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4294 Instruction *apply(BinaryOperator &Xor) const {
4301 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4302 bool Changed = SimplifyCommutative(I);
4303 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4305 if (isa<UndefValue>(Op1)) {
4306 if (isa<UndefValue>(Op0))
4307 // Handle undef ^ undef -> 0 special case. This is a common
4309 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4310 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4313 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4314 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4315 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4316 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4319 // See if we can simplify any instructions used by the instruction whose sole
4320 // purpose is to compute bits we don't care about.
4321 if (!isa<VectorType>(I.getType())) {
4322 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4323 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4324 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4325 KnownZero, KnownOne))
4327 } else if (isa<ConstantAggregateZero>(Op1)) {
4328 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4331 // Is this a ~ operation?
4332 if (Value *NotOp = dyn_castNotVal(&I)) {
4333 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4334 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4335 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4336 if (Op0I->getOpcode() == Instruction::And ||
4337 Op0I->getOpcode() == Instruction::Or) {
4338 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4339 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4341 BinaryOperator::CreateNot(Op0I->getOperand(1),
4342 Op0I->getOperand(1)->getName()+".not");
4343 InsertNewInstBefore(NotY, I);
4344 if (Op0I->getOpcode() == Instruction::And)
4345 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4347 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4354 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4355 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4356 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4357 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4358 return new ICmpInst(ICI->getInversePredicate(),
4359 ICI->getOperand(0), ICI->getOperand(1));
4361 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4362 return new FCmpInst(FCI->getInversePredicate(),
4363 FCI->getOperand(0), FCI->getOperand(1));
4366 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4367 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4368 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4369 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4370 Instruction::CastOps Opcode = Op0C->getOpcode();
4371 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4372 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4373 Op0C->getDestTy())) {
4374 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4375 CI->getOpcode(), CI->getInversePredicate(),
4376 CI->getOperand(0), CI->getOperand(1)), I);
4377 NewCI->takeName(CI);
4378 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4385 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4386 // ~(c-X) == X-c-1 == X+(-c-1)
4387 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4388 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4389 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4390 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4391 ConstantInt::get(I.getType(), 1));
4392 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4395 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4396 if (Op0I->getOpcode() == Instruction::Add) {
4397 // ~(X-c) --> (-c-1)-X
4398 if (RHS->isAllOnesValue()) {
4399 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4400 return BinaryOperator::CreateSub(
4401 ConstantExpr::getSub(NegOp0CI,
4402 ConstantInt::get(I.getType(), 1)),
4403 Op0I->getOperand(0));
4404 } else if (RHS->getValue().isSignBit()) {
4405 // (X + C) ^ signbit -> (X + C + signbit)
4406 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4407 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4410 } else if (Op0I->getOpcode() == Instruction::Or) {
4411 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4412 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4413 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4414 // Anything in both C1 and C2 is known to be zero, remove it from
4416 Constant *CommonBits = And(Op0CI, RHS);
4417 NewRHS = ConstantExpr::getAnd(NewRHS,
4418 ConstantExpr::getNot(CommonBits));
4419 AddToWorkList(Op0I);
4420 I.setOperand(0, Op0I->getOperand(0));
4421 I.setOperand(1, NewRHS);
4428 // Try to fold constant and into select arguments.
4429 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4430 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4432 if (isa<PHINode>(Op0))
4433 if (Instruction *NV = FoldOpIntoPhi(I))
4437 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4439 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4441 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4443 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4446 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4449 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4450 if (A == Op0) { // B^(B|A) == (A|B)^B
4451 Op1I->swapOperands();
4453 std::swap(Op0, Op1);
4454 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4455 I.swapOperands(); // Simplified below.
4456 std::swap(Op0, Op1);
4458 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4459 if (Op0 == A) // A^(A^B) == B
4460 return ReplaceInstUsesWith(I, B);
4461 else if (Op0 == B) // A^(B^A) == B
4462 return ReplaceInstUsesWith(I, A);
4463 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4464 if (A == Op0) { // A^(A&B) -> A^(B&A)
4465 Op1I->swapOperands();
4468 if (B == Op0) { // A^(B&A) -> (B&A)^A
4469 I.swapOperands(); // Simplified below.
4470 std::swap(Op0, Op1);
4475 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4478 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4479 if (A == Op1) // (B|A)^B == (A|B)^B
4481 if (B == Op1) { // (A|B)^B == A & ~B
4483 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4484 return BinaryOperator::CreateAnd(A, NotB);
4486 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4487 if (Op1 == A) // (A^B)^A == B
4488 return ReplaceInstUsesWith(I, B);
4489 else if (Op1 == B) // (B^A)^A == B
4490 return ReplaceInstUsesWith(I, A);
4491 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4492 if (A == Op1) // (A&B)^A -> (B&A)^A
4494 if (B == Op1 && // (B&A)^A == ~B & A
4495 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4497 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4498 return BinaryOperator::CreateAnd(N, Op1);
4503 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4504 if (Op0I && Op1I && Op0I->isShift() &&
4505 Op0I->getOpcode() == Op1I->getOpcode() &&
4506 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4507 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4508 Instruction *NewOp =
4509 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4510 Op1I->getOperand(0),
4511 Op0I->getName()), I);
4512 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4513 Op1I->getOperand(1));
4517 Value *A, *B, *C, *D;
4518 // (A & B)^(A | B) -> A ^ B
4519 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4520 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4521 if ((A == C && B == D) || (A == D && B == C))
4522 return BinaryOperator::CreateXor(A, B);
4524 // (A | B)^(A & B) -> A ^ B
4525 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4526 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4527 if ((A == C && B == D) || (A == D && B == C))
4528 return BinaryOperator::CreateXor(A, B);
4532 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4533 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4534 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4535 // (X & Y)^(X & Y) -> (Y^Z) & X
4536 Value *X = 0, *Y = 0, *Z = 0;
4538 X = A, Y = B, Z = D;
4540 X = A, Y = B, Z = C;
4542 X = B, Y = A, Z = D;
4544 X = B, Y = A, Z = C;
4547 Instruction *NewOp =
4548 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4549 return BinaryOperator::CreateAnd(NewOp, X);
4554 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4555 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4556 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4559 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4560 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4561 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4562 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4563 const Type *SrcTy = Op0C->getOperand(0)->getType();
4564 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4565 // Only do this if the casts both really cause code to be generated.
4566 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4568 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4570 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4571 Op1C->getOperand(0),
4573 InsertNewInstBefore(NewOp, I);
4574 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4579 return Changed ? &I : 0;
4582 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4583 /// overflowed for this type.
4584 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4585 ConstantInt *In2, bool IsSigned = false) {
4586 Result = cast<ConstantInt>(Add(In1, In2));
4589 if (In2->getValue().isNegative())
4590 return Result->getValue().sgt(In1->getValue());
4592 return Result->getValue().slt(In1->getValue());
4594 return Result->getValue().ult(In1->getValue());
4597 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4598 /// code necessary to compute the offset from the base pointer (without adding
4599 /// in the base pointer). Return the result as a signed integer of intptr size.
4600 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4601 TargetData &TD = IC.getTargetData();
4602 gep_type_iterator GTI = gep_type_begin(GEP);
4603 const Type *IntPtrTy = TD.getIntPtrType();
4604 Value *Result = Constant::getNullValue(IntPtrTy);
4606 // Build a mask for high order bits.
4607 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4608 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4610 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
4613 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4614 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4615 if (OpC->isZero()) continue;
4617 // Handle a struct index, which adds its field offset to the pointer.
4618 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4619 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4621 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4622 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4624 Result = IC.InsertNewInstBefore(
4625 BinaryOperator::CreateAdd(Result,
4626 ConstantInt::get(IntPtrTy, Size),
4627 GEP->getName()+".offs"), I);
4631 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4632 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4633 Scale = ConstantExpr::getMul(OC, Scale);
4634 if (Constant *RC = dyn_cast<Constant>(Result))
4635 Result = ConstantExpr::getAdd(RC, Scale);
4637 // Emit an add instruction.
4638 Result = IC.InsertNewInstBefore(
4639 BinaryOperator::CreateAdd(Result, Scale,
4640 GEP->getName()+".offs"), I);
4644 // Convert to correct type.
4645 if (Op->getType() != IntPtrTy) {
4646 if (Constant *OpC = dyn_cast<Constant>(Op))
4647 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4649 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4650 Op->getName()+".c"), I);
4653 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4654 if (Constant *OpC = dyn_cast<Constant>(Op))
4655 Op = ConstantExpr::getMul(OpC, Scale);
4656 else // We'll let instcombine(mul) convert this to a shl if possible.
4657 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
4658 GEP->getName()+".idx"), I);
4661 // Emit an add instruction.
4662 if (isa<Constant>(Op) && isa<Constant>(Result))
4663 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4664 cast<Constant>(Result));
4666 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
4667 GEP->getName()+".offs"), I);
4673 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4674 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4675 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4676 /// complex, and scales are involved. The above expression would also be legal
4677 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4678 /// later form is less amenable to optimization though, and we are allowed to
4679 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4681 /// If we can't emit an optimized form for this expression, this returns null.
4683 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4685 TargetData &TD = IC.getTargetData();
4686 gep_type_iterator GTI = gep_type_begin(GEP);
4688 // Check to see if this gep only has a single variable index. If so, and if
4689 // any constant indices are a multiple of its scale, then we can compute this
4690 // in terms of the scale of the variable index. For example, if the GEP
4691 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4692 // because the expression will cross zero at the same point.
4693 unsigned i, e = GEP->getNumOperands();
4695 for (i = 1; i != e; ++i, ++GTI) {
4696 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4697 // Compute the aggregate offset of constant indices.
4698 if (CI->isZero()) continue;
4700 // Handle a struct index, which adds its field offset to the pointer.
4701 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4702 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4704 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4705 Offset += Size*CI->getSExtValue();
4708 // Found our variable index.
4713 // If there are no variable indices, we must have a constant offset, just
4714 // evaluate it the general way.
4715 if (i == e) return 0;
4717 Value *VariableIdx = GEP->getOperand(i);
4718 // Determine the scale factor of the variable element. For example, this is
4719 // 4 if the variable index is into an array of i32.
4720 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4722 // Verify that there are no other variable indices. If so, emit the hard way.
4723 for (++i, ++GTI; i != e; ++i, ++GTI) {
4724 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4727 // Compute the aggregate offset of constant indices.
4728 if (CI->isZero()) continue;
4730 // Handle a struct index, which adds its field offset to the pointer.
4731 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4732 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4734 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4735 Offset += Size*CI->getSExtValue();
4739 // Okay, we know we have a single variable index, which must be a
4740 // pointer/array/vector index. If there is no offset, life is simple, return
4742 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4744 // Cast to intptrty in case a truncation occurs. If an extension is needed,
4745 // we don't need to bother extending: the extension won't affect where the
4746 // computation crosses zero.
4747 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
4748 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
4749 VariableIdx->getNameStart(), &I);
4753 // Otherwise, there is an index. The computation we will do will be modulo
4754 // the pointer size, so get it.
4755 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4757 Offset &= PtrSizeMask;
4758 VariableScale &= PtrSizeMask;
4760 // To do this transformation, any constant index must be a multiple of the
4761 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
4762 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
4763 // multiple of the variable scale.
4764 int64_t NewOffs = Offset / (int64_t)VariableScale;
4765 if (Offset != NewOffs*(int64_t)VariableScale)
4768 // Okay, we can do this evaluation. Start by converting the index to intptr.
4769 const Type *IntPtrTy = TD.getIntPtrType();
4770 if (VariableIdx->getType() != IntPtrTy)
4771 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
4773 VariableIdx->getNameStart(), &I);
4774 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
4775 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
4779 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4780 /// else. At this point we know that the GEP is on the LHS of the comparison.
4781 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4782 ICmpInst::Predicate Cond,
4784 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4786 // Look through bitcasts.
4787 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
4788 RHS = BCI->getOperand(0);
4790 Value *PtrBase = GEPLHS->getOperand(0);
4791 if (PtrBase == RHS) {
4792 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4793 // This transformation (ignoring the base and scales) is valid because we
4794 // know pointers can't overflow. See if we can output an optimized form.
4795 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
4797 // If not, synthesize the offset the hard way.
4799 Offset = EmitGEPOffset(GEPLHS, I, *this);
4800 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4801 Constant::getNullValue(Offset->getType()));
4802 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4803 // If the base pointers are different, but the indices are the same, just
4804 // compare the base pointer.
4805 if (PtrBase != GEPRHS->getOperand(0)) {
4806 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4807 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4808 GEPRHS->getOperand(0)->getType();
4810 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4811 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4812 IndicesTheSame = false;
4816 // If all indices are the same, just compare the base pointers.
4818 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4819 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4821 // Otherwise, the base pointers are different and the indices are
4822 // different, bail out.
4826 // If one of the GEPs has all zero indices, recurse.
4827 bool AllZeros = true;
4828 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4829 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4830 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4835 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4836 ICmpInst::getSwappedPredicate(Cond), I);
4838 // If the other GEP has all zero indices, recurse.
4840 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4841 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4842 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4847 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4849 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4850 // If the GEPs only differ by one index, compare it.
4851 unsigned NumDifferences = 0; // Keep track of # differences.
4852 unsigned DiffOperand = 0; // The operand that differs.
4853 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4854 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4855 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4856 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4857 // Irreconcilable differences.
4861 if (NumDifferences++) break;
4866 if (NumDifferences == 0) // SAME GEP?
4867 return ReplaceInstUsesWith(I, // No comparison is needed here.
4868 ConstantInt::get(Type::Int1Ty,
4869 ICmpInst::isTrueWhenEqual(Cond)));
4871 else if (NumDifferences == 1) {
4872 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4873 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4874 // Make sure we do a signed comparison here.
4875 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4879 // Only lower this if the icmp is the only user of the GEP or if we expect
4880 // the result to fold to a constant!
4881 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4882 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4883 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4884 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4885 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4886 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4892 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
4894 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
4897 if (!isa<ConstantFP>(RHSC)) return 0;
4898 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4900 // Get the width of the mantissa. We don't want to hack on conversions that
4901 // might lose information from the integer, e.g. "i64 -> float"
4902 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4903 if (MantissaWidth == -1) return 0; // Unknown.
4905 // Check to see that the input is converted from an integer type that is small
4906 // enough that preserves all bits. TODO: check here for "known" sign bits.
4907 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4908 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
4910 // If this is a uitofp instruction, we need an extra bit to hold the sign.
4911 if (isa<UIToFPInst>(LHSI))
4914 // If the conversion would lose info, don't hack on this.
4915 if ((int)InputSize > MantissaWidth)
4918 // Otherwise, we can potentially simplify the comparison. We know that it
4919 // will always come through as an integer value and we know the constant is
4920 // not a NAN (it would have been previously simplified).
4921 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4923 ICmpInst::Predicate Pred;
4924 switch (I.getPredicate()) {
4925 default: assert(0 && "Unexpected predicate!");
4926 case FCmpInst::FCMP_UEQ:
4927 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
4928 case FCmpInst::FCMP_UGT:
4929 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
4930 case FCmpInst::FCMP_UGE:
4931 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
4932 case FCmpInst::FCMP_ULT:
4933 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
4934 case FCmpInst::FCMP_ULE:
4935 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
4936 case FCmpInst::FCMP_UNE:
4937 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
4938 case FCmpInst::FCMP_ORD:
4939 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4940 case FCmpInst::FCMP_UNO:
4941 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4944 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4946 // Now we know that the APFloat is a normal number, zero or inf.
4948 // See if the FP constant is too large for the integer. For example,
4949 // comparing an i8 to 300.0.
4950 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
4952 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4953 // and large values.
4954 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
4955 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4956 APFloat::rmNearestTiesToEven);
4957 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4958 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4959 Pred == ICmpInst::ICMP_SLE)
4960 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4961 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4964 // See if the RHS value is < SignedMin.
4965 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
4966 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4967 APFloat::rmNearestTiesToEven);
4968 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4969 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4970 Pred == ICmpInst::ICMP_SGE)
4971 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4972 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4975 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
4976 // it may still be fractional. See if it is fractional by casting the FP
4977 // value to the integer value and back, checking for equality. Don't do this
4978 // for zero, because -0.0 is not fractional.
4979 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
4980 if (!RHS.isZero() &&
4981 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
4982 // If we had a comparison against a fractional value, we have to adjust
4983 // the compare predicate and sometimes the value. RHSC is rounded towards
4984 // zero at this point.
4986 default: assert(0 && "Unexpected integer comparison!");
4987 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
4988 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4989 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
4990 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4991 case ICmpInst::ICMP_SLE:
4992 // (float)int <= 4.4 --> int <= 4
4993 // (float)int <= -4.4 --> int < -4
4994 if (RHS.isNegative())
4995 Pred = ICmpInst::ICMP_SLT;
4997 case ICmpInst::ICMP_SLT:
4998 // (float)int < -4.4 --> int < -4
4999 // (float)int < 4.4 --> int <= 4
5000 if (!RHS.isNegative())
5001 Pred = ICmpInst::ICMP_SLE;
5003 case ICmpInst::ICMP_SGT:
5004 // (float)int > 4.4 --> int > 4
5005 // (float)int > -4.4 --> int >= -4
5006 if (RHS.isNegative())
5007 Pred = ICmpInst::ICMP_SGE;
5009 case ICmpInst::ICMP_SGE:
5010 // (float)int >= -4.4 --> int >= -4
5011 // (float)int >= 4.4 --> int > 4
5012 if (!RHS.isNegative())
5013 Pred = ICmpInst::ICMP_SGT;
5018 // Lower this FP comparison into an appropriate integer version of the
5020 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5023 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5024 bool Changed = SimplifyCompare(I);
5025 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5027 // Fold trivial predicates.
5028 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5029 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5030 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5031 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5033 // Simplify 'fcmp pred X, X'
5035 switch (I.getPredicate()) {
5036 default: assert(0 && "Unknown predicate!");
5037 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5038 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5039 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5040 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5041 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5042 case FCmpInst::FCMP_OLT: // True if ordered and less than
5043 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5044 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5046 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5047 case FCmpInst::FCMP_ULT: // True if unordered or less than
5048 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5049 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5050 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5051 I.setPredicate(FCmpInst::FCMP_UNO);
5052 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5055 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5056 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5057 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5058 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5059 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5060 I.setPredicate(FCmpInst::FCMP_ORD);
5061 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5066 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5067 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5069 // Handle fcmp with constant RHS
5070 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5071 // If the constant is a nan, see if we can fold the comparison based on it.
5072 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5073 if (CFP->getValueAPF().isNaN()) {
5074 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5075 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5076 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5077 "Comparison must be either ordered or unordered!");
5078 // True if unordered.
5079 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5083 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5084 switch (LHSI->getOpcode()) {
5085 case Instruction::PHI:
5086 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5087 // block. If in the same block, we're encouraging jump threading. If
5088 // not, we are just pessimizing the code by making an i1 phi.
5089 if (LHSI->getParent() == I.getParent())
5090 if (Instruction *NV = FoldOpIntoPhi(I))
5093 case Instruction::SIToFP:
5094 case Instruction::UIToFP:
5095 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5098 case Instruction::Select:
5099 // If either operand of the select is a constant, we can fold the
5100 // comparison into the select arms, which will cause one to be
5101 // constant folded and the select turned into a bitwise or.
5102 Value *Op1 = 0, *Op2 = 0;
5103 if (LHSI->hasOneUse()) {
5104 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5105 // Fold the known value into the constant operand.
5106 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5107 // Insert a new FCmp of the other select operand.
5108 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5109 LHSI->getOperand(2), RHSC,
5111 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5112 // Fold the known value into the constant operand.
5113 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5114 // Insert a new FCmp of the other select operand.
5115 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5116 LHSI->getOperand(1), RHSC,
5122 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5127 return Changed ? &I : 0;
5130 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5131 bool Changed = SimplifyCompare(I);
5132 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5133 const Type *Ty = Op0->getType();
5137 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5138 I.isTrueWhenEqual()));
5140 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5141 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5143 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5144 // addresses never equal each other! We already know that Op0 != Op1.
5145 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5146 isa<ConstantPointerNull>(Op0)) &&
5147 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5148 isa<ConstantPointerNull>(Op1)))
5149 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5150 !I.isTrueWhenEqual()));
5152 // icmp's with boolean values can always be turned into bitwise operations
5153 if (Ty == Type::Int1Ty) {
5154 switch (I.getPredicate()) {
5155 default: assert(0 && "Invalid icmp instruction!");
5156 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5157 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5158 InsertNewInstBefore(Xor, I);
5159 return BinaryOperator::CreateNot(Xor);
5161 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5162 return BinaryOperator::CreateXor(Op0, Op1);
5164 case ICmpInst::ICMP_UGT:
5165 case ICmpInst::ICMP_SGT:
5166 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5168 case ICmpInst::ICMP_ULT:
5169 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5170 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5171 InsertNewInstBefore(Not, I);
5172 return BinaryOperator::CreateAnd(Not, Op1);
5174 case ICmpInst::ICMP_UGE:
5175 case ICmpInst::ICMP_SGE:
5176 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5178 case ICmpInst::ICMP_ULE:
5179 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5180 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5181 InsertNewInstBefore(Not, I);
5182 return BinaryOperator::CreateOr(Not, Op1);
5187 // See if we are doing a comparison between a constant and an instruction that
5188 // can be folded into the comparison.
5189 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5192 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5193 if (I.isEquality() && CI->isNullValue() &&
5194 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5195 // (icmp cond A B) if cond is equality
5196 return new ICmpInst(I.getPredicate(), A, B);
5199 switch (I.getPredicate()) {
5201 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5202 if (CI->isMinValue(false))
5203 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5204 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5205 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5206 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5207 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5208 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5209 if (CI->isMinValue(true))
5210 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5211 ConstantInt::getAllOnesValue(Op0->getType()));
5215 case ICmpInst::ICMP_SLT:
5216 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5217 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5218 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5219 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5220 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5221 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5224 case ICmpInst::ICMP_UGT:
5225 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5226 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5227 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5228 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5229 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5230 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5232 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5233 if (CI->isMaxValue(true))
5234 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5235 ConstantInt::getNullValue(Op0->getType()));
5238 case ICmpInst::ICMP_SGT:
5239 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5240 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5241 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5242 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5243 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5244 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5247 case ICmpInst::ICMP_ULE:
5248 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5249 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5250 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5251 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5252 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5253 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5256 case ICmpInst::ICMP_SLE:
5257 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5258 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5259 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5260 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5261 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5262 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5265 case ICmpInst::ICMP_UGE:
5266 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5267 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5268 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5269 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5270 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5271 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5274 case ICmpInst::ICMP_SGE:
5275 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5276 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5277 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5278 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5279 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5280 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5284 // If we still have a icmp le or icmp ge instruction, turn it into the
5285 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5286 // already been handled above, this requires little checking.
5288 switch (I.getPredicate()) {
5290 case ICmpInst::ICMP_ULE:
5291 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5292 case ICmpInst::ICMP_SLE:
5293 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5294 case ICmpInst::ICMP_UGE:
5295 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5296 case ICmpInst::ICMP_SGE:
5297 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5300 // See if we can fold the comparison based on bits known to be zero or one
5301 // in the input. If this comparison is a normal comparison, it demands all
5302 // bits, if it is a sign bit comparison, it only demands the sign bit.
5305 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5307 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5308 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5309 if (SimplifyDemandedBits(Op0,
5310 isSignBit ? APInt::getSignBit(BitWidth)
5311 : APInt::getAllOnesValue(BitWidth),
5312 KnownZero, KnownOne, 0))
5315 // Given the known and unknown bits, compute a range that the LHS could be
5317 if ((KnownOne | KnownZero) != 0) {
5318 // Compute the Min, Max and RHS values based on the known bits. For the
5319 // EQ and NE we use unsigned values.
5320 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5321 const APInt& RHSVal = CI->getValue();
5322 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5323 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5326 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5329 switch (I.getPredicate()) { // LE/GE have been folded already.
5330 default: assert(0 && "Unknown icmp opcode!");
5331 case ICmpInst::ICMP_EQ:
5332 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5333 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5335 case ICmpInst::ICMP_NE:
5336 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5337 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5339 case ICmpInst::ICMP_ULT:
5340 if (Max.ult(RHSVal))
5341 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5342 if (Min.uge(RHSVal))
5343 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5345 case ICmpInst::ICMP_UGT:
5346 if (Min.ugt(RHSVal))
5347 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5348 if (Max.ule(RHSVal))
5349 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5351 case ICmpInst::ICMP_SLT:
5352 if (Max.slt(RHSVal))
5353 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5354 if (Min.sgt(RHSVal))
5355 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5357 case ICmpInst::ICMP_SGT:
5358 if (Min.sgt(RHSVal))
5359 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5360 if (Max.sle(RHSVal))
5361 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5366 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5367 // instruction, see if that instruction also has constants so that the
5368 // instruction can be folded into the icmp
5369 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5370 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5374 // Handle icmp with constant (but not simple integer constant) RHS
5375 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5376 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5377 switch (LHSI->getOpcode()) {
5378 case Instruction::GetElementPtr:
5379 if (RHSC->isNullValue()) {
5380 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5381 bool isAllZeros = true;
5382 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5383 if (!isa<Constant>(LHSI->getOperand(i)) ||
5384 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5389 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5390 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5394 case Instruction::PHI:
5395 // Only fold icmp into the PHI if the phi and fcmp are in the same
5396 // block. If in the same block, we're encouraging jump threading. If
5397 // not, we are just pessimizing the code by making an i1 phi.
5398 if (LHSI->getParent() == I.getParent())
5399 if (Instruction *NV = FoldOpIntoPhi(I))
5402 case Instruction::Select: {
5403 // If either operand of the select is a constant, we can fold the
5404 // comparison into the select arms, which will cause one to be
5405 // constant folded and the select turned into a bitwise or.
5406 Value *Op1 = 0, *Op2 = 0;
5407 if (LHSI->hasOneUse()) {
5408 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5409 // Fold the known value into the constant operand.
5410 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5411 // Insert a new ICmp of the other select operand.
5412 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5413 LHSI->getOperand(2), RHSC,
5415 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5416 // Fold the known value into the constant operand.
5417 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5418 // Insert a new ICmp of the other select operand.
5419 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5420 LHSI->getOperand(1), RHSC,
5426 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5429 case Instruction::Malloc:
5430 // If we have (malloc != null), and if the malloc has a single use, we
5431 // can assume it is successful and remove the malloc.
5432 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5433 AddToWorkList(LHSI);
5434 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5435 !I.isTrueWhenEqual()));
5441 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5442 if (User *GEP = dyn_castGetElementPtr(Op0))
5443 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5445 if (User *GEP = dyn_castGetElementPtr(Op1))
5446 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5447 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5450 // Test to see if the operands of the icmp are casted versions of other
5451 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5453 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5454 if (isa<PointerType>(Op0->getType()) &&
5455 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5456 // We keep moving the cast from the left operand over to the right
5457 // operand, where it can often be eliminated completely.
5458 Op0 = CI->getOperand(0);
5460 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5461 // so eliminate it as well.
5462 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5463 Op1 = CI2->getOperand(0);
5465 // If Op1 is a constant, we can fold the cast into the constant.
5466 if (Op0->getType() != Op1->getType()) {
5467 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5468 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5470 // Otherwise, cast the RHS right before the icmp
5471 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5474 return new ICmpInst(I.getPredicate(), Op0, Op1);
5478 if (isa<CastInst>(Op0)) {
5479 // Handle the special case of: icmp (cast bool to X), <cst>
5480 // This comes up when you have code like
5483 // For generality, we handle any zero-extension of any operand comparison
5484 // with a constant or another cast from the same type.
5485 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5486 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5490 // ~x < ~y --> y < x
5492 if (match(Op0, m_Not(m_Value(A))) &&
5493 match(Op1, m_Not(m_Value(B))))
5494 return new ICmpInst(I.getPredicate(), B, A);
5497 if (I.isEquality()) {
5498 Value *A, *B, *C, *D;
5500 // -x == -y --> x == y
5501 if (match(Op0, m_Neg(m_Value(A))) &&
5502 match(Op1, m_Neg(m_Value(B))))
5503 return new ICmpInst(I.getPredicate(), A, B);
5505 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5506 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5507 Value *OtherVal = A == Op1 ? B : A;
5508 return new ICmpInst(I.getPredicate(), OtherVal,
5509 Constant::getNullValue(A->getType()));
5512 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5513 // A^c1 == C^c2 --> A == C^(c1^c2)
5514 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5515 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5516 if (Op1->hasOneUse()) {
5517 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5518 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5519 return new ICmpInst(I.getPredicate(), A,
5520 InsertNewInstBefore(Xor, I));
5523 // A^B == A^D -> B == D
5524 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5525 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5526 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5527 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5531 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5532 (A == Op0 || B == Op0)) {
5533 // A == (A^B) -> B == 0
5534 Value *OtherVal = A == Op0 ? B : A;
5535 return new ICmpInst(I.getPredicate(), OtherVal,
5536 Constant::getNullValue(A->getType()));
5538 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5539 // (A-B) == A -> B == 0
5540 return new ICmpInst(I.getPredicate(), B,
5541 Constant::getNullValue(B->getType()));
5543 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5544 // A == (A-B) -> B == 0
5545 return new ICmpInst(I.getPredicate(), B,
5546 Constant::getNullValue(B->getType()));
5549 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5550 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5551 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5552 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5553 Value *X = 0, *Y = 0, *Z = 0;
5556 X = B; Y = D; Z = A;
5557 } else if (A == D) {
5558 X = B; Y = C; Z = A;
5559 } else if (B == C) {
5560 X = A; Y = D; Z = B;
5561 } else if (B == D) {
5562 X = A; Y = C; Z = B;
5565 if (X) { // Build (X^Y) & Z
5566 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
5567 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
5568 I.setOperand(0, Op1);
5569 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5574 return Changed ? &I : 0;
5578 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5579 /// and CmpRHS are both known to be integer constants.
5580 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5581 ConstantInt *DivRHS) {
5582 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5583 const APInt &CmpRHSV = CmpRHS->getValue();
5585 // FIXME: If the operand types don't match the type of the divide
5586 // then don't attempt this transform. The code below doesn't have the
5587 // logic to deal with a signed divide and an unsigned compare (and
5588 // vice versa). This is because (x /s C1) <s C2 produces different
5589 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5590 // (x /u C1) <u C2. Simply casting the operands and result won't
5591 // work. :( The if statement below tests that condition and bails
5593 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5594 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5596 if (DivRHS->isZero())
5597 return 0; // The ProdOV computation fails on divide by zero.
5599 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5600 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5601 // C2 (CI). By solving for X we can turn this into a range check
5602 // instead of computing a divide.
5603 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5605 // Determine if the product overflows by seeing if the product is
5606 // not equal to the divide. Make sure we do the same kind of divide
5607 // as in the LHS instruction that we're folding.
5608 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5609 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5611 // Get the ICmp opcode
5612 ICmpInst::Predicate Pred = ICI.getPredicate();
5614 // Figure out the interval that is being checked. For example, a comparison
5615 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5616 // Compute this interval based on the constants involved and the signedness of
5617 // the compare/divide. This computes a half-open interval, keeping track of
5618 // whether either value in the interval overflows. After analysis each
5619 // overflow variable is set to 0 if it's corresponding bound variable is valid
5620 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5621 int LoOverflow = 0, HiOverflow = 0;
5622 ConstantInt *LoBound = 0, *HiBound = 0;
5625 if (!DivIsSigned) { // udiv
5626 // e.g. X/5 op 3 --> [15, 20)
5628 HiOverflow = LoOverflow = ProdOV;
5630 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5631 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5632 if (CmpRHSV == 0) { // (X / pos) op 0
5633 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5634 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5636 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5637 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5638 HiOverflow = LoOverflow = ProdOV;
5640 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5641 } else { // (X / pos) op neg
5642 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5643 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5644 LoOverflow = AddWithOverflow(LoBound, Prod,
5645 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5646 HiBound = AddOne(Prod);
5647 HiOverflow = ProdOV ? -1 : 0;
5649 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5650 if (CmpRHSV == 0) { // (X / neg) op 0
5651 // e.g. X/-5 op 0 --> [-4, 5)
5652 LoBound = AddOne(DivRHS);
5653 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5654 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5655 HiOverflow = 1; // [INTMIN+1, overflow)
5656 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5658 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5659 // e.g. X/-5 op 3 --> [-19, -14)
5660 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5662 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5663 HiBound = AddOne(Prod);
5664 } else { // (X / neg) op neg
5665 // e.g. X/-5 op -3 --> [15, 20)
5667 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5668 HiBound = Subtract(Prod, DivRHS);
5671 // Dividing by a negative swaps the condition. LT <-> GT
5672 Pred = ICmpInst::getSwappedPredicate(Pred);
5675 Value *X = DivI->getOperand(0);
5677 default: assert(0 && "Unhandled icmp opcode!");
5678 case ICmpInst::ICMP_EQ:
5679 if (LoOverflow && HiOverflow)
5680 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5681 else if (HiOverflow)
5682 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5683 ICmpInst::ICMP_UGE, X, LoBound);
5684 else if (LoOverflow)
5685 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5686 ICmpInst::ICMP_ULT, X, HiBound);
5688 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5689 case ICmpInst::ICMP_NE:
5690 if (LoOverflow && HiOverflow)
5691 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5692 else if (HiOverflow)
5693 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5694 ICmpInst::ICMP_ULT, X, LoBound);
5695 else if (LoOverflow)
5696 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5697 ICmpInst::ICMP_UGE, X, HiBound);
5699 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5700 case ICmpInst::ICMP_ULT:
5701 case ICmpInst::ICMP_SLT:
5702 if (LoOverflow == +1) // Low bound is greater than input range.
5703 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5704 if (LoOverflow == -1) // Low bound is less than input range.
5705 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5706 return new ICmpInst(Pred, X, LoBound);
5707 case ICmpInst::ICMP_UGT:
5708 case ICmpInst::ICMP_SGT:
5709 if (HiOverflow == +1) // High bound greater than input range.
5710 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5711 else if (HiOverflow == -1) // High bound less than input range.
5712 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5713 if (Pred == ICmpInst::ICMP_UGT)
5714 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5716 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5721 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5723 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5726 const APInt &RHSV = RHS->getValue();
5728 switch (LHSI->getOpcode()) {
5729 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5730 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5731 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5733 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5734 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5735 Value *CompareVal = LHSI->getOperand(0);
5737 // If the sign bit of the XorCST is not set, there is no change to
5738 // the operation, just stop using the Xor.
5739 if (!XorCST->getValue().isNegative()) {
5740 ICI.setOperand(0, CompareVal);
5741 AddToWorkList(LHSI);
5745 // Was the old condition true if the operand is positive?
5746 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5748 // If so, the new one isn't.
5749 isTrueIfPositive ^= true;
5751 if (isTrueIfPositive)
5752 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5754 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5758 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5759 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5760 LHSI->getOperand(0)->hasOneUse()) {
5761 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5763 // If the LHS is an AND of a truncating cast, we can widen the
5764 // and/compare to be the input width without changing the value
5765 // produced, eliminating a cast.
5766 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5767 // We can do this transformation if either the AND constant does not
5768 // have its sign bit set or if it is an equality comparison.
5769 // Extending a relational comparison when we're checking the sign
5770 // bit would not work.
5771 if (Cast->hasOneUse() &&
5772 (ICI.isEquality() ||
5773 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5775 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5776 APInt NewCST = AndCST->getValue();
5777 NewCST.zext(BitWidth);
5779 NewCI.zext(BitWidth);
5780 Instruction *NewAnd =
5781 BinaryOperator::CreateAnd(Cast->getOperand(0),
5782 ConstantInt::get(NewCST),LHSI->getName());
5783 InsertNewInstBefore(NewAnd, ICI);
5784 return new ICmpInst(ICI.getPredicate(), NewAnd,
5785 ConstantInt::get(NewCI));
5789 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5790 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5791 // happens a LOT in code produced by the C front-end, for bitfield
5793 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5794 if (Shift && !Shift->isShift())
5798 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5799 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5800 const Type *AndTy = AndCST->getType(); // Type of the and.
5802 // We can fold this as long as we can't shift unknown bits
5803 // into the mask. This can only happen with signed shift
5804 // rights, as they sign-extend.
5806 bool CanFold = Shift->isLogicalShift();
5808 // To test for the bad case of the signed shr, see if any
5809 // of the bits shifted in could be tested after the mask.
5810 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5811 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5813 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5814 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5815 AndCST->getValue()) == 0)
5821 if (Shift->getOpcode() == Instruction::Shl)
5822 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5824 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5826 // Check to see if we are shifting out any of the bits being
5828 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5829 // If we shifted bits out, the fold is not going to work out.
5830 // As a special case, check to see if this means that the
5831 // result is always true or false now.
5832 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5833 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5834 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5835 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5837 ICI.setOperand(1, NewCst);
5838 Constant *NewAndCST;
5839 if (Shift->getOpcode() == Instruction::Shl)
5840 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5842 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5843 LHSI->setOperand(1, NewAndCST);
5844 LHSI->setOperand(0, Shift->getOperand(0));
5845 AddToWorkList(Shift); // Shift is dead.
5846 AddUsesToWorkList(ICI);
5852 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5853 // preferable because it allows the C<<Y expression to be hoisted out
5854 // of a loop if Y is invariant and X is not.
5855 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5856 ICI.isEquality() && !Shift->isArithmeticShift() &&
5857 isa<Instruction>(Shift->getOperand(0))) {
5860 if (Shift->getOpcode() == Instruction::LShr) {
5861 NS = BinaryOperator::CreateShl(AndCST,
5862 Shift->getOperand(1), "tmp");
5864 // Insert a logical shift.
5865 NS = BinaryOperator::CreateLShr(AndCST,
5866 Shift->getOperand(1), "tmp");
5868 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5870 // Compute X & (C << Y).
5871 Instruction *NewAnd =
5872 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
5873 InsertNewInstBefore(NewAnd, ICI);
5875 ICI.setOperand(0, NewAnd);
5881 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5882 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5885 uint32_t TypeBits = RHSV.getBitWidth();
5887 // Check that the shift amount is in range. If not, don't perform
5888 // undefined shifts. When the shift is visited it will be
5890 if (ShAmt->uge(TypeBits))
5893 if (ICI.isEquality()) {
5894 // If we are comparing against bits always shifted out, the
5895 // comparison cannot succeed.
5897 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5898 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5899 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5900 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5901 return ReplaceInstUsesWith(ICI, Cst);
5904 if (LHSI->hasOneUse()) {
5905 // Otherwise strength reduce the shift into an and.
5906 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5908 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5911 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5912 Mask, LHSI->getName()+".mask");
5913 Value *And = InsertNewInstBefore(AndI, ICI);
5914 return new ICmpInst(ICI.getPredicate(), And,
5915 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5919 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5920 bool TrueIfSigned = false;
5921 if (LHSI->hasOneUse() &&
5922 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5923 // (X << 31) <s 0 --> (X&1) != 0
5924 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5925 (TypeBits-ShAmt->getZExtValue()-1));
5927 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5928 Mask, LHSI->getName()+".mask");
5929 Value *And = InsertNewInstBefore(AndI, ICI);
5931 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5932 And, Constant::getNullValue(And->getType()));
5937 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5938 case Instruction::AShr: {
5939 // Only handle equality comparisons of shift-by-constant.
5940 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5941 if (!ShAmt || !ICI.isEquality()) break;
5943 // Check that the shift amount is in range. If not, don't perform
5944 // undefined shifts. When the shift is visited it will be
5946 uint32_t TypeBits = RHSV.getBitWidth();
5947 if (ShAmt->uge(TypeBits))
5950 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5952 // If we are comparing against bits always shifted out, the
5953 // comparison cannot succeed.
5954 APInt Comp = RHSV << ShAmtVal;
5955 if (LHSI->getOpcode() == Instruction::LShr)
5956 Comp = Comp.lshr(ShAmtVal);
5958 Comp = Comp.ashr(ShAmtVal);
5960 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5961 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5962 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5963 return ReplaceInstUsesWith(ICI, Cst);
5966 // Otherwise, check to see if the bits shifted out are known to be zero.
5967 // If so, we can compare against the unshifted value:
5968 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
5969 if (LHSI->hasOneUse() &&
5970 MaskedValueIsZero(LHSI->getOperand(0),
5971 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
5972 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
5973 ConstantExpr::getShl(RHS, ShAmt));
5976 if (LHSI->hasOneUse()) {
5977 // Otherwise strength reduce the shift into an and.
5978 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5979 Constant *Mask = ConstantInt::get(Val);
5982 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5983 Mask, LHSI->getName()+".mask");
5984 Value *And = InsertNewInstBefore(AndI, ICI);
5985 return new ICmpInst(ICI.getPredicate(), And,
5986 ConstantExpr::getShl(RHS, ShAmt));
5991 case Instruction::SDiv:
5992 case Instruction::UDiv:
5993 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5994 // Fold this div into the comparison, producing a range check.
5995 // Determine, based on the divide type, what the range is being
5996 // checked. If there is an overflow on the low or high side, remember
5997 // it, otherwise compute the range [low, hi) bounding the new value.
5998 // See: InsertRangeTest above for the kinds of replacements possible.
5999 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6000 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6005 case Instruction::Add:
6006 // Fold: icmp pred (add, X, C1), C2
6008 if (!ICI.isEquality()) {
6009 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6011 const APInt &LHSV = LHSC->getValue();
6013 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6016 if (ICI.isSignedPredicate()) {
6017 if (CR.getLower().isSignBit()) {
6018 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6019 ConstantInt::get(CR.getUpper()));
6020 } else if (CR.getUpper().isSignBit()) {
6021 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6022 ConstantInt::get(CR.getLower()));
6025 if (CR.getLower().isMinValue()) {
6026 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6027 ConstantInt::get(CR.getUpper()));
6028 } else if (CR.getUpper().isMinValue()) {
6029 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6030 ConstantInt::get(CR.getLower()));
6037 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6038 if (ICI.isEquality()) {
6039 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6041 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6042 // the second operand is a constant, simplify a bit.
6043 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6044 switch (BO->getOpcode()) {
6045 case Instruction::SRem:
6046 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6047 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6048 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6049 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6050 Instruction *NewRem =
6051 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6053 InsertNewInstBefore(NewRem, ICI);
6054 return new ICmpInst(ICI.getPredicate(), NewRem,
6055 Constant::getNullValue(BO->getType()));
6059 case Instruction::Add:
6060 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6061 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6062 if (BO->hasOneUse())
6063 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6064 Subtract(RHS, BOp1C));
6065 } else if (RHSV == 0) {
6066 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6067 // efficiently invertible, or if the add has just this one use.
6068 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6070 if (Value *NegVal = dyn_castNegVal(BOp1))
6071 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6072 else if (Value *NegVal = dyn_castNegVal(BOp0))
6073 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6074 else if (BO->hasOneUse()) {
6075 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6076 InsertNewInstBefore(Neg, ICI);
6078 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6082 case Instruction::Xor:
6083 // For the xor case, we can xor two constants together, eliminating
6084 // the explicit xor.
6085 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6086 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6087 ConstantExpr::getXor(RHS, BOC));
6090 case Instruction::Sub:
6091 // Replace (([sub|xor] A, B) != 0) with (A != B)
6093 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6097 case Instruction::Or:
6098 // If bits are being or'd in that are not present in the constant we
6099 // are comparing against, then the comparison could never succeed!
6100 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6101 Constant *NotCI = ConstantExpr::getNot(RHS);
6102 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6103 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6108 case Instruction::And:
6109 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6110 // If bits are being compared against that are and'd out, then the
6111 // comparison can never succeed!
6112 if ((RHSV & ~BOC->getValue()) != 0)
6113 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6116 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6117 if (RHS == BOC && RHSV.isPowerOf2())
6118 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6119 ICmpInst::ICMP_NE, LHSI,
6120 Constant::getNullValue(RHS->getType()));
6122 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6123 if (BOC->getValue().isSignBit()) {
6124 Value *X = BO->getOperand(0);
6125 Constant *Zero = Constant::getNullValue(X->getType());
6126 ICmpInst::Predicate pred = isICMP_NE ?
6127 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6128 return new ICmpInst(pred, X, Zero);
6131 // ((X & ~7) == 0) --> X < 8
6132 if (RHSV == 0 && isHighOnes(BOC)) {
6133 Value *X = BO->getOperand(0);
6134 Constant *NegX = ConstantExpr::getNeg(BOC);
6135 ICmpInst::Predicate pred = isICMP_NE ?
6136 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6137 return new ICmpInst(pred, X, NegX);
6142 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6143 // Handle icmp {eq|ne} <intrinsic>, intcst.
6144 if (II->getIntrinsicID() == Intrinsic::bswap) {
6146 ICI.setOperand(0, II->getOperand(1));
6147 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6151 } else { // Not a ICMP_EQ/ICMP_NE
6152 // If the LHS is a cast from an integral value of the same size,
6153 // then since we know the RHS is a constant, try to simlify.
6154 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6155 Value *CastOp = Cast->getOperand(0);
6156 const Type *SrcTy = CastOp->getType();
6157 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6158 if (SrcTy->isInteger() &&
6159 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6160 // If this is an unsigned comparison, try to make the comparison use
6161 // smaller constant values.
6162 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6163 // X u< 128 => X s> -1
6164 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6165 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6166 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6167 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6168 // X u> 127 => X s< 0
6169 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6170 Constant::getNullValue(SrcTy));
6178 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6179 /// We only handle extending casts so far.
6181 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6182 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6183 Value *LHSCIOp = LHSCI->getOperand(0);
6184 const Type *SrcTy = LHSCIOp->getType();
6185 const Type *DestTy = LHSCI->getType();
6188 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6189 // integer type is the same size as the pointer type.
6190 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6191 getTargetData().getPointerSizeInBits() ==
6192 cast<IntegerType>(DestTy)->getBitWidth()) {
6194 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6195 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6196 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6197 RHSOp = RHSC->getOperand(0);
6198 // If the pointer types don't match, insert a bitcast.
6199 if (LHSCIOp->getType() != RHSOp->getType())
6200 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6204 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6207 // The code below only handles extension cast instructions, so far.
6209 if (LHSCI->getOpcode() != Instruction::ZExt &&
6210 LHSCI->getOpcode() != Instruction::SExt)
6213 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6214 bool isSignedCmp = ICI.isSignedPredicate();
6216 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6217 // Not an extension from the same type?
6218 RHSCIOp = CI->getOperand(0);
6219 if (RHSCIOp->getType() != LHSCIOp->getType())
6222 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6223 // and the other is a zext), then we can't handle this.
6224 if (CI->getOpcode() != LHSCI->getOpcode())
6227 // Deal with equality cases early.
6228 if (ICI.isEquality())
6229 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6231 // A signed comparison of sign extended values simplifies into a
6232 // signed comparison.
6233 if (isSignedCmp && isSignedExt)
6234 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6236 // The other three cases all fold into an unsigned comparison.
6237 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6240 // If we aren't dealing with a constant on the RHS, exit early
6241 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6245 // Compute the constant that would happen if we truncated to SrcTy then
6246 // reextended to DestTy.
6247 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6248 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6250 // If the re-extended constant didn't change...
6252 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6253 // For example, we might have:
6254 // %A = sext short %X to uint
6255 // %B = icmp ugt uint %A, 1330
6256 // It is incorrect to transform this into
6257 // %B = icmp ugt short %X, 1330
6258 // because %A may have negative value.
6260 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6261 // OR operation is EQ/NE.
6262 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6263 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6268 // The re-extended constant changed so the constant cannot be represented
6269 // in the shorter type. Consequently, we cannot emit a simple comparison.
6271 // First, handle some easy cases. We know the result cannot be equal at this
6272 // point so handle the ICI.isEquality() cases
6273 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6274 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6275 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6276 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6278 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6279 // should have been folded away previously and not enter in here.
6282 // We're performing a signed comparison.
6283 if (cast<ConstantInt>(CI)->getValue().isNegative())
6284 Result = ConstantInt::getFalse(); // X < (small) --> false
6286 Result = ConstantInt::getTrue(); // X < (large) --> true
6288 // We're performing an unsigned comparison.
6290 // We're performing an unsigned comp with a sign extended value.
6291 // This is true if the input is >= 0. [aka >s -1]
6292 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6293 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6294 NegOne, ICI.getName()), ICI);
6296 // Unsigned extend & unsigned compare -> always true.
6297 Result = ConstantInt::getTrue();
6301 // Finally, return the value computed.
6302 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6303 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6304 return ReplaceInstUsesWith(ICI, Result);
6306 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6307 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6308 "ICmp should be folded!");
6309 if (Constant *CI = dyn_cast<Constant>(Result))
6310 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6312 return BinaryOperator::CreateNot(Result);
6316 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6317 return commonShiftTransforms(I);
6320 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6321 return commonShiftTransforms(I);
6324 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6325 if (Instruction *R = commonShiftTransforms(I))
6328 Value *Op0 = I.getOperand(0);
6330 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6331 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6332 if (CSI->isAllOnesValue())
6333 return ReplaceInstUsesWith(I, CSI);
6335 // See if we can turn a signed shr into an unsigned shr.
6336 if (MaskedValueIsZero(Op0,
6337 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6338 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6343 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6344 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6345 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6347 // shl X, 0 == X and shr X, 0 == X
6348 // shl 0, X == 0 and shr 0, X == 0
6349 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6350 Op0 == Constant::getNullValue(Op0->getType()))
6351 return ReplaceInstUsesWith(I, Op0);
6353 if (isa<UndefValue>(Op0)) {
6354 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6355 return ReplaceInstUsesWith(I, Op0);
6356 else // undef << X -> 0, undef >>u X -> 0
6357 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6359 if (isa<UndefValue>(Op1)) {
6360 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6361 return ReplaceInstUsesWith(I, Op0);
6362 else // X << undef, X >>u undef -> 0
6363 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6366 // Try to fold constant and into select arguments.
6367 if (isa<Constant>(Op0))
6368 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6369 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6372 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6373 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6378 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6379 BinaryOperator &I) {
6380 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6382 // See if we can simplify any instructions used by the instruction whose sole
6383 // purpose is to compute bits we don't care about.
6384 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6385 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6386 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6387 KnownZero, KnownOne))
6390 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6391 // of a signed value.
6393 if (Op1->uge(TypeBits)) {
6394 if (I.getOpcode() != Instruction::AShr)
6395 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6397 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6402 // ((X*C1) << C2) == (X * (C1 << C2))
6403 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6404 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6405 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6406 return BinaryOperator::CreateMul(BO->getOperand(0),
6407 ConstantExpr::getShl(BOOp, Op1));
6409 // Try to fold constant and into select arguments.
6410 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6411 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6413 if (isa<PHINode>(Op0))
6414 if (Instruction *NV = FoldOpIntoPhi(I))
6417 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6418 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6419 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6420 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6421 // place. Don't try to do this transformation in this case. Also, we
6422 // require that the input operand is a shift-by-constant so that we have
6423 // confidence that the shifts will get folded together. We could do this
6424 // xform in more cases, but it is unlikely to be profitable.
6425 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6426 isa<ConstantInt>(TrOp->getOperand(1))) {
6427 // Okay, we'll do this xform. Make the shift of shift.
6428 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6429 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6431 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6433 // For logical shifts, the truncation has the effect of making the high
6434 // part of the register be zeros. Emulate this by inserting an AND to
6435 // clear the top bits as needed. This 'and' will usually be zapped by
6436 // other xforms later if dead.
6437 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6438 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6439 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6441 // The mask we constructed says what the trunc would do if occurring
6442 // between the shifts. We want to know the effect *after* the second
6443 // shift. We know that it is a logical shift by a constant, so adjust the
6444 // mask as appropriate.
6445 if (I.getOpcode() == Instruction::Shl)
6446 MaskV <<= Op1->getZExtValue();
6448 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6449 MaskV = MaskV.lshr(Op1->getZExtValue());
6452 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6454 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6456 // Return the value truncated to the interesting size.
6457 return new TruncInst(And, I.getType());
6461 if (Op0->hasOneUse()) {
6462 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6463 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6466 switch (Op0BO->getOpcode()) {
6468 case Instruction::Add:
6469 case Instruction::And:
6470 case Instruction::Or:
6471 case Instruction::Xor: {
6472 // These operators commute.
6473 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6474 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6475 match(Op0BO->getOperand(1),
6476 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6477 Instruction *YS = BinaryOperator::CreateShl(
6478 Op0BO->getOperand(0), Op1,
6480 InsertNewInstBefore(YS, I); // (Y << C)
6482 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6483 Op0BO->getOperand(1)->getName());
6484 InsertNewInstBefore(X, I); // (X + (Y << C))
6485 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6486 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6487 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6490 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6491 Value *Op0BOOp1 = Op0BO->getOperand(1);
6492 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6494 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6495 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6497 Instruction *YS = BinaryOperator::CreateShl(
6498 Op0BO->getOperand(0), Op1,
6500 InsertNewInstBefore(YS, I); // (Y << C)
6502 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6503 V1->getName()+".mask");
6504 InsertNewInstBefore(XM, I); // X & (CC << C)
6506 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6511 case Instruction::Sub: {
6512 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6513 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6514 match(Op0BO->getOperand(0),
6515 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6516 Instruction *YS = BinaryOperator::CreateShl(
6517 Op0BO->getOperand(1), Op1,
6519 InsertNewInstBefore(YS, I); // (Y << C)
6521 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6522 Op0BO->getOperand(0)->getName());
6523 InsertNewInstBefore(X, I); // (X + (Y << C))
6524 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6525 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6526 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6529 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6530 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6531 match(Op0BO->getOperand(0),
6532 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6533 m_ConstantInt(CC))) && V2 == Op1 &&
6534 cast<BinaryOperator>(Op0BO->getOperand(0))
6535 ->getOperand(0)->hasOneUse()) {
6536 Instruction *YS = BinaryOperator::CreateShl(
6537 Op0BO->getOperand(1), Op1,
6539 InsertNewInstBefore(YS, I); // (Y << C)
6541 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6542 V1->getName()+".mask");
6543 InsertNewInstBefore(XM, I); // X & (CC << C)
6545 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
6553 // If the operand is an bitwise operator with a constant RHS, and the
6554 // shift is the only use, we can pull it out of the shift.
6555 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6556 bool isValid = true; // Valid only for And, Or, Xor
6557 bool highBitSet = false; // Transform if high bit of constant set?
6559 switch (Op0BO->getOpcode()) {
6560 default: isValid = false; break; // Do not perform transform!
6561 case Instruction::Add:
6562 isValid = isLeftShift;
6564 case Instruction::Or:
6565 case Instruction::Xor:
6568 case Instruction::And:
6573 // If this is a signed shift right, and the high bit is modified
6574 // by the logical operation, do not perform the transformation.
6575 // The highBitSet boolean indicates the value of the high bit of
6576 // the constant which would cause it to be modified for this
6579 if (isValid && I.getOpcode() == Instruction::AShr)
6580 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6583 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6585 Instruction *NewShift =
6586 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6587 InsertNewInstBefore(NewShift, I);
6588 NewShift->takeName(Op0BO);
6590 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
6597 // Find out if this is a shift of a shift by a constant.
6598 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6599 if (ShiftOp && !ShiftOp->isShift())
6602 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6603 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6604 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6605 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6606 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6607 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6608 Value *X = ShiftOp->getOperand(0);
6610 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6611 if (AmtSum > TypeBits)
6614 const IntegerType *Ty = cast<IntegerType>(I.getType());
6616 // Check for (X << c1) << c2 and (X >> c1) >> c2
6617 if (I.getOpcode() == ShiftOp->getOpcode()) {
6618 return BinaryOperator::Create(I.getOpcode(), X,
6619 ConstantInt::get(Ty, AmtSum));
6620 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6621 I.getOpcode() == Instruction::AShr) {
6622 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6623 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
6624 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6625 I.getOpcode() == Instruction::LShr) {
6626 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6627 Instruction *Shift =
6628 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
6629 InsertNewInstBefore(Shift, I);
6631 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6632 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6635 // Okay, if we get here, one shift must be left, and the other shift must be
6636 // right. See if the amounts are equal.
6637 if (ShiftAmt1 == ShiftAmt2) {
6638 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6639 if (I.getOpcode() == Instruction::Shl) {
6640 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6641 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6643 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6644 if (I.getOpcode() == Instruction::LShr) {
6645 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6646 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6648 // We can simplify ((X << C) >>s C) into a trunc + sext.
6649 // NOTE: we could do this for any C, but that would make 'unusual' integer
6650 // types. For now, just stick to ones well-supported by the code
6652 const Type *SExtType = 0;
6653 switch (Ty->getBitWidth() - ShiftAmt1) {
6660 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6665 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6666 InsertNewInstBefore(NewTrunc, I);
6667 return new SExtInst(NewTrunc, Ty);
6669 // Otherwise, we can't handle it yet.
6670 } else if (ShiftAmt1 < ShiftAmt2) {
6671 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6673 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6674 if (I.getOpcode() == Instruction::Shl) {
6675 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6676 ShiftOp->getOpcode() == Instruction::AShr);
6677 Instruction *Shift =
6678 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6679 InsertNewInstBefore(Shift, I);
6681 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6682 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6685 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6686 if (I.getOpcode() == Instruction::LShr) {
6687 assert(ShiftOp->getOpcode() == Instruction::Shl);
6688 Instruction *Shift =
6689 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
6690 InsertNewInstBefore(Shift, I);
6692 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6693 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6696 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6698 assert(ShiftAmt2 < ShiftAmt1);
6699 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6701 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6702 if (I.getOpcode() == Instruction::Shl) {
6703 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6704 ShiftOp->getOpcode() == Instruction::AShr);
6705 Instruction *Shift =
6706 BinaryOperator::Create(ShiftOp->getOpcode(), X,
6707 ConstantInt::get(Ty, ShiftDiff));
6708 InsertNewInstBefore(Shift, I);
6710 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6711 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6714 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6715 if (I.getOpcode() == Instruction::LShr) {
6716 assert(ShiftOp->getOpcode() == Instruction::Shl);
6717 Instruction *Shift =
6718 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6719 InsertNewInstBefore(Shift, I);
6721 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6722 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6725 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6732 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6733 /// expression. If so, decompose it, returning some value X, such that Val is
6736 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6738 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6739 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6740 Offset = CI->getZExtValue();
6742 return ConstantInt::get(Type::Int32Ty, 0);
6743 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6744 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6745 if (I->getOpcode() == Instruction::Shl) {
6746 // This is a value scaled by '1 << the shift amt'.
6747 Scale = 1U << RHS->getZExtValue();
6749 return I->getOperand(0);
6750 } else if (I->getOpcode() == Instruction::Mul) {
6751 // This value is scaled by 'RHS'.
6752 Scale = RHS->getZExtValue();
6754 return I->getOperand(0);
6755 } else if (I->getOpcode() == Instruction::Add) {
6756 // We have X+C. Check to see if we really have (X*C2)+C1,
6757 // where C1 is divisible by C2.
6760 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6761 Offset += RHS->getZExtValue();
6768 // Otherwise, we can't look past this.
6775 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6776 /// try to eliminate the cast by moving the type information into the alloc.
6777 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6778 AllocationInst &AI) {
6779 const PointerType *PTy = cast<PointerType>(CI.getType());
6781 // Remove any uses of AI that are dead.
6782 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6784 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6785 Instruction *User = cast<Instruction>(*UI++);
6786 if (isInstructionTriviallyDead(User)) {
6787 while (UI != E && *UI == User)
6788 ++UI; // If this instruction uses AI more than once, don't break UI.
6791 DOUT << "IC: DCE: " << *User;
6792 EraseInstFromFunction(*User);
6796 // Get the type really allocated and the type casted to.
6797 const Type *AllocElTy = AI.getAllocatedType();
6798 const Type *CastElTy = PTy->getElementType();
6799 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6801 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6802 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6803 if (CastElTyAlign < AllocElTyAlign) return 0;
6805 // If the allocation has multiple uses, only promote it if we are strictly
6806 // increasing the alignment of the resultant allocation. If we keep it the
6807 // same, we open the door to infinite loops of various kinds.
6808 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6810 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6811 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6812 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6814 // See if we can satisfy the modulus by pulling a scale out of the array
6816 unsigned ArraySizeScale;
6818 Value *NumElements = // See if the array size is a decomposable linear expr.
6819 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6821 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6823 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6824 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6826 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6831 // If the allocation size is constant, form a constant mul expression
6832 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6833 if (isa<ConstantInt>(NumElements))
6834 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6835 // otherwise multiply the amount and the number of elements
6836 else if (Scale != 1) {
6837 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
6838 Amt = InsertNewInstBefore(Tmp, AI);
6842 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6843 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6844 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
6845 Amt = InsertNewInstBefore(Tmp, AI);
6848 AllocationInst *New;
6849 if (isa<MallocInst>(AI))
6850 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6852 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6853 InsertNewInstBefore(New, AI);
6856 // If the allocation has multiple uses, insert a cast and change all things
6857 // that used it to use the new cast. This will also hack on CI, but it will
6859 if (!AI.hasOneUse()) {
6860 AddUsesToWorkList(AI);
6861 // New is the allocation instruction, pointer typed. AI is the original
6862 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6863 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6864 InsertNewInstBefore(NewCast, AI);
6865 AI.replaceAllUsesWith(NewCast);
6867 return ReplaceInstUsesWith(CI, New);
6870 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6871 /// and return it as type Ty without inserting any new casts and without
6872 /// changing the computed value. This is used by code that tries to decide
6873 /// whether promoting or shrinking integer operations to wider or smaller types
6874 /// will allow us to eliminate a truncate or extend.
6876 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6877 /// extension operation if Ty is larger.
6879 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
6880 /// should return true if trunc(V) can be computed by computing V in the smaller
6881 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
6882 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
6883 /// efficiently truncated.
6885 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
6886 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
6887 /// the final result.
6888 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6890 int &NumCastsRemoved) {
6891 // We can always evaluate constants in another type.
6892 if (isa<ConstantInt>(V))
6895 Instruction *I = dyn_cast<Instruction>(V);
6896 if (!I) return false;
6898 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6900 // If this is an extension or truncate, we can often eliminate it.
6901 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6902 // If this is a cast from the destination type, we can trivially eliminate
6903 // it, and this will remove a cast overall.
6904 if (I->getOperand(0)->getType() == Ty) {
6905 // If the first operand is itself a cast, and is eliminable, do not count
6906 // this as an eliminable cast. We would prefer to eliminate those two
6908 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
6914 // We can't extend or shrink something that has multiple uses: doing so would
6915 // require duplicating the instruction in general, which isn't profitable.
6916 if (!I->hasOneUse()) return false;
6918 switch (I->getOpcode()) {
6919 case Instruction::Add:
6920 case Instruction::Sub:
6921 case Instruction::And:
6922 case Instruction::Or:
6923 case Instruction::Xor:
6924 // These operators can all arbitrarily be extended or truncated.
6925 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6927 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6930 case Instruction::Mul:
6931 // A multiply can be truncated by truncating its operands.
6932 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6933 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6935 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6938 case Instruction::Shl:
6939 // If we are truncating the result of this SHL, and if it's a shift of a
6940 // constant amount, we can always perform a SHL in a smaller type.
6941 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6942 uint32_t BitWidth = Ty->getBitWidth();
6943 if (BitWidth < OrigTy->getBitWidth() &&
6944 CI->getLimitedValue(BitWidth) < BitWidth)
6945 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6949 case Instruction::LShr:
6950 // If this is a truncate of a logical shr, we can truncate it to a smaller
6951 // lshr iff we know that the bits we would otherwise be shifting in are
6953 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6954 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6955 uint32_t BitWidth = Ty->getBitWidth();
6956 if (BitWidth < OrigBitWidth &&
6957 MaskedValueIsZero(I->getOperand(0),
6958 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6959 CI->getLimitedValue(BitWidth) < BitWidth) {
6960 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6965 case Instruction::ZExt:
6966 case Instruction::SExt:
6967 case Instruction::Trunc:
6968 // If this is the same kind of case as our original (e.g. zext+zext), we
6969 // can safely replace it. Note that replacing it does not reduce the number
6970 // of casts in the input.
6971 if (I->getOpcode() == CastOpc)
6975 case Instruction::PHI: {
6976 // We can change a phi if we can change all operands.
6977 PHINode *PN = cast<PHINode>(I);
6978 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
6979 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
6985 // TODO: Can handle more cases here.
6992 /// EvaluateInDifferentType - Given an expression that
6993 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6994 /// evaluate the expression.
6995 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6997 if (Constant *C = dyn_cast<Constant>(V))
6998 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7000 // Otherwise, it must be an instruction.
7001 Instruction *I = cast<Instruction>(V);
7002 Instruction *Res = 0;
7003 switch (I->getOpcode()) {
7004 case Instruction::Add:
7005 case Instruction::Sub:
7006 case Instruction::Mul:
7007 case Instruction::And:
7008 case Instruction::Or:
7009 case Instruction::Xor:
7010 case Instruction::AShr:
7011 case Instruction::LShr:
7012 case Instruction::Shl: {
7013 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7014 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7015 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7019 case Instruction::Trunc:
7020 case Instruction::ZExt:
7021 case Instruction::SExt:
7022 // If the source type of the cast is the type we're trying for then we can
7023 // just return the source. There's no need to insert it because it is not
7025 if (I->getOperand(0)->getType() == Ty)
7026 return I->getOperand(0);
7028 // Otherwise, must be the same type of cast, so just reinsert a new one.
7029 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7032 case Instruction::PHI: {
7033 PHINode *OPN = cast<PHINode>(I);
7034 PHINode *NPN = PHINode::Create(Ty);
7035 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7036 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7037 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7043 // TODO: Can handle more cases here.
7044 assert(0 && "Unreachable!");
7049 return InsertNewInstBefore(Res, *I);
7052 /// @brief Implement the transforms common to all CastInst visitors.
7053 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7054 Value *Src = CI.getOperand(0);
7056 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7057 // eliminate it now.
7058 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7059 if (Instruction::CastOps opc =
7060 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7061 // The first cast (CSrc) is eliminable so we need to fix up or replace
7062 // the second cast (CI). CSrc will then have a good chance of being dead.
7063 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7067 // If we are casting a select then fold the cast into the select
7068 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7069 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7072 // If we are casting a PHI then fold the cast into the PHI
7073 if (isa<PHINode>(Src))
7074 if (Instruction *NV = FoldOpIntoPhi(CI))
7080 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7081 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7082 Value *Src = CI.getOperand(0);
7084 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7085 // If casting the result of a getelementptr instruction with no offset, turn
7086 // this into a cast of the original pointer!
7087 if (GEP->hasAllZeroIndices()) {
7088 // Changing the cast operand is usually not a good idea but it is safe
7089 // here because the pointer operand is being replaced with another
7090 // pointer operand so the opcode doesn't need to change.
7092 CI.setOperand(0, GEP->getOperand(0));
7096 // If the GEP has a single use, and the base pointer is a bitcast, and the
7097 // GEP computes a constant offset, see if we can convert these three
7098 // instructions into fewer. This typically happens with unions and other
7099 // non-type-safe code.
7100 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7101 if (GEP->hasAllConstantIndices()) {
7102 // We are guaranteed to get a constant from EmitGEPOffset.
7103 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7104 int64_t Offset = OffsetV->getSExtValue();
7106 // Get the base pointer input of the bitcast, and the type it points to.
7107 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7108 const Type *GEPIdxTy =
7109 cast<PointerType>(OrigBase->getType())->getElementType();
7110 if (GEPIdxTy->isSized()) {
7111 SmallVector<Value*, 8> NewIndices;
7113 // Start with the index over the outer type. Note that the type size
7114 // might be zero (even if the offset isn't zero) if the indexed type
7115 // is something like [0 x {int, int}]
7116 const Type *IntPtrTy = TD->getIntPtrType();
7117 int64_t FirstIdx = 0;
7118 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7119 FirstIdx = Offset/TySize;
7122 // Handle silly modulus not returning values values [0..TySize).
7126 assert(Offset >= 0);
7128 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7131 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7133 // Index into the types. If we fail, set OrigBase to null.
7135 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7136 const StructLayout *SL = TD->getStructLayout(STy);
7137 if (Offset < (int64_t)SL->getSizeInBytes()) {
7138 unsigned Elt = SL->getElementContainingOffset(Offset);
7139 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7141 Offset -= SL->getElementOffset(Elt);
7142 GEPIdxTy = STy->getElementType(Elt);
7144 // Otherwise, we can't index into this, bail out.
7148 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7149 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7150 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7151 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7154 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7156 GEPIdxTy = STy->getElementType();
7158 // Otherwise, we can't index into this, bail out.
7164 // If we were able to index down into an element, create the GEP
7165 // and bitcast the result. This eliminates one bitcast, potentially
7167 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7169 NewIndices.end(), "");
7170 InsertNewInstBefore(NGEP, CI);
7171 NGEP->takeName(GEP);
7173 if (isa<BitCastInst>(CI))
7174 return new BitCastInst(NGEP, CI.getType());
7175 assert(isa<PtrToIntInst>(CI));
7176 return new PtrToIntInst(NGEP, CI.getType());
7183 return commonCastTransforms(CI);
7188 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7189 /// integer types. This function implements the common transforms for all those
7191 /// @brief Implement the transforms common to CastInst with integer operands
7192 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7193 if (Instruction *Result = commonCastTransforms(CI))
7196 Value *Src = CI.getOperand(0);
7197 const Type *SrcTy = Src->getType();
7198 const Type *DestTy = CI.getType();
7199 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7200 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7202 // See if we can simplify any instructions used by the LHS whose sole
7203 // purpose is to compute bits we don't care about.
7204 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7205 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7206 KnownZero, KnownOne))
7209 // If the source isn't an instruction or has more than one use then we
7210 // can't do anything more.
7211 Instruction *SrcI = dyn_cast<Instruction>(Src);
7212 if (!SrcI || !Src->hasOneUse())
7215 // Attempt to propagate the cast into the instruction for int->int casts.
7216 int NumCastsRemoved = 0;
7217 if (!isa<BitCastInst>(CI) &&
7218 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7219 CI.getOpcode(), NumCastsRemoved)) {
7220 // If this cast is a truncate, evaluting in a different type always
7221 // eliminates the cast, so it is always a win. If this is a zero-extension,
7222 // we need to do an AND to maintain the clear top-part of the computation,
7223 // so we require that the input have eliminated at least one cast. If this
7224 // is a sign extension, we insert two new casts (to do the extension) so we
7225 // require that two casts have been eliminated.
7227 switch (CI.getOpcode()) {
7229 // All the others use floating point so we shouldn't actually
7230 // get here because of the check above.
7231 assert(0 && "Unknown cast type");
7232 case Instruction::Trunc:
7235 case Instruction::ZExt:
7236 DoXForm = NumCastsRemoved >= 1;
7238 case Instruction::SExt:
7239 DoXForm = NumCastsRemoved >= 2;
7244 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7245 CI.getOpcode() == Instruction::SExt);
7246 assert(Res->getType() == DestTy);
7247 switch (CI.getOpcode()) {
7248 default: assert(0 && "Unknown cast type!");
7249 case Instruction::Trunc:
7250 case Instruction::BitCast:
7251 // Just replace this cast with the result.
7252 return ReplaceInstUsesWith(CI, Res);
7253 case Instruction::ZExt: {
7254 // We need to emit an AND to clear the high bits.
7255 assert(SrcBitSize < DestBitSize && "Not a zext?");
7256 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7258 return BinaryOperator::CreateAnd(Res, C);
7260 case Instruction::SExt:
7261 // We need to emit a cast to truncate, then a cast to sext.
7262 return CastInst::Create(Instruction::SExt,
7263 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7269 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7270 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7272 switch (SrcI->getOpcode()) {
7273 case Instruction::Add:
7274 case Instruction::Mul:
7275 case Instruction::And:
7276 case Instruction::Or:
7277 case Instruction::Xor:
7278 // If we are discarding information, rewrite.
7279 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7280 // Don't insert two casts if they cannot be eliminated. We allow
7281 // two casts to be inserted if the sizes are the same. This could
7282 // only be converting signedness, which is a noop.
7283 if (DestBitSize == SrcBitSize ||
7284 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7285 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7286 Instruction::CastOps opcode = CI.getOpcode();
7287 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7288 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7289 return BinaryOperator::Create(
7290 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7294 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7295 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7296 SrcI->getOpcode() == Instruction::Xor &&
7297 Op1 == ConstantInt::getTrue() &&
7298 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7299 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7300 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7303 case Instruction::SDiv:
7304 case Instruction::UDiv:
7305 case Instruction::SRem:
7306 case Instruction::URem:
7307 // If we are just changing the sign, rewrite.
7308 if (DestBitSize == SrcBitSize) {
7309 // Don't insert two casts if they cannot be eliminated. We allow
7310 // two casts to be inserted if the sizes are the same. This could
7311 // only be converting signedness, which is a noop.
7312 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7313 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7314 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7316 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7318 return BinaryOperator::Create(
7319 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7324 case Instruction::Shl:
7325 // Allow changing the sign of the source operand. Do not allow
7326 // changing the size of the shift, UNLESS the shift amount is a
7327 // constant. We must not change variable sized shifts to a smaller
7328 // size, because it is undefined to shift more bits out than exist
7330 if (DestBitSize == SrcBitSize ||
7331 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7332 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7333 Instruction::BitCast : Instruction::Trunc);
7334 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7335 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7336 return BinaryOperator::CreateShl(Op0c, Op1c);
7339 case Instruction::AShr:
7340 // If this is a signed shr, and if all bits shifted in are about to be
7341 // truncated off, turn it into an unsigned shr to allow greater
7343 if (DestBitSize < SrcBitSize &&
7344 isa<ConstantInt>(Op1)) {
7345 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7346 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7347 // Insert the new logical shift right.
7348 return BinaryOperator::CreateLShr(Op0, Op1);
7356 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7357 if (Instruction *Result = commonIntCastTransforms(CI))
7360 Value *Src = CI.getOperand(0);
7361 const Type *Ty = CI.getType();
7362 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7363 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7365 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7366 switch (SrcI->getOpcode()) {
7368 case Instruction::LShr:
7369 // We can shrink lshr to something smaller if we know the bits shifted in
7370 // are already zeros.
7371 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7372 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7374 // Get a mask for the bits shifting in.
7375 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7376 Value* SrcIOp0 = SrcI->getOperand(0);
7377 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7378 if (ShAmt >= DestBitWidth) // All zeros.
7379 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7381 // Okay, we can shrink this. Truncate the input, then return a new
7383 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7384 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7386 return BinaryOperator::CreateLShr(V1, V2);
7388 } else { // This is a variable shr.
7390 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7391 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7392 // loop-invariant and CSE'd.
7393 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7394 Value *One = ConstantInt::get(SrcI->getType(), 1);
7396 Value *V = InsertNewInstBefore(
7397 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7399 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7400 SrcI->getOperand(0),
7402 Value *Zero = Constant::getNullValue(V->getType());
7403 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7413 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7414 /// in order to eliminate the icmp.
7415 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7417 // If we are just checking for a icmp eq of a single bit and zext'ing it
7418 // to an integer, then shift the bit to the appropriate place and then
7419 // cast to integer to avoid the comparison.
7420 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7421 const APInt &Op1CV = Op1C->getValue();
7423 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7424 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7425 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7426 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7427 if (!DoXform) return ICI;
7429 Value *In = ICI->getOperand(0);
7430 Value *Sh = ConstantInt::get(In->getType(),
7431 In->getType()->getPrimitiveSizeInBits()-1);
7432 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7433 In->getName()+".lobit"),
7435 if (In->getType() != CI.getType())
7436 In = CastInst::CreateIntegerCast(In, CI.getType(),
7437 false/*ZExt*/, "tmp", &CI);
7439 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7440 Constant *One = ConstantInt::get(In->getType(), 1);
7441 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7442 In->getName()+".not"),
7446 return ReplaceInstUsesWith(CI, In);
7451 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7452 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7453 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7454 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7455 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7456 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7457 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7458 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7459 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7460 // This only works for EQ and NE
7461 ICI->isEquality()) {
7462 // If Op1C some other power of two, convert:
7463 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7464 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7465 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7466 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7468 APInt KnownZeroMask(~KnownZero);
7469 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7470 if (!DoXform) return ICI;
7472 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7473 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7474 // (X&4) == 2 --> false
7475 // (X&4) != 2 --> true
7476 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7477 Res = ConstantExpr::getZExt(Res, CI.getType());
7478 return ReplaceInstUsesWith(CI, Res);
7481 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7482 Value *In = ICI->getOperand(0);
7484 // Perform a logical shr by shiftamt.
7485 // Insert the shift to put the result in the low bit.
7486 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7487 ConstantInt::get(In->getType(), ShiftAmt),
7488 In->getName()+".lobit"), CI);
7491 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7492 Constant *One = ConstantInt::get(In->getType(), 1);
7493 In = BinaryOperator::CreateXor(In, One, "tmp");
7494 InsertNewInstBefore(cast<Instruction>(In), CI);
7497 if (CI.getType() == In->getType())
7498 return ReplaceInstUsesWith(CI, In);
7500 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7508 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7509 // If one of the common conversion will work ..
7510 if (Instruction *Result = commonIntCastTransforms(CI))
7513 Value *Src = CI.getOperand(0);
7515 // If this is a cast of a cast
7516 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7517 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7518 // types and if the sizes are just right we can convert this into a logical
7519 // 'and' which will be much cheaper than the pair of casts.
7520 if (isa<TruncInst>(CSrc)) {
7521 // Get the sizes of the types involved
7522 Value *A = CSrc->getOperand(0);
7523 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7524 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7525 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7526 // If we're actually extending zero bits and the trunc is a no-op
7527 if (MidSize < DstSize && SrcSize == DstSize) {
7528 // Replace both of the casts with an And of the type mask.
7529 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7530 Constant *AndConst = ConstantInt::get(AndValue);
7532 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
7533 // Unfortunately, if the type changed, we need to cast it back.
7534 if (And->getType() != CI.getType()) {
7535 And->setName(CSrc->getName()+".mask");
7536 InsertNewInstBefore(And, CI);
7537 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
7544 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7545 return transformZExtICmp(ICI, CI);
7547 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7548 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7549 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7550 // of the (zext icmp) will be transformed.
7551 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7552 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7553 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7554 (transformZExtICmp(LHS, CI, false) ||
7555 transformZExtICmp(RHS, CI, false))) {
7556 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7557 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7558 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
7565 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7566 if (Instruction *I = commonIntCastTransforms(CI))
7569 Value *Src = CI.getOperand(0);
7571 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7572 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7573 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7574 // If we are just checking for a icmp eq of a single bit and zext'ing it
7575 // to an integer, then shift the bit to the appropriate place and then
7576 // cast to integer to avoid the comparison.
7577 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7578 const APInt &Op1CV = Op1C->getValue();
7580 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7581 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7582 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7583 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7584 Value *In = ICI->getOperand(0);
7585 Value *Sh = ConstantInt::get(In->getType(),
7586 In->getType()->getPrimitiveSizeInBits()-1);
7587 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
7588 In->getName()+".lobit"),
7590 if (In->getType() != CI.getType())
7591 In = CastInst::CreateIntegerCast(In, CI.getType(),
7592 true/*SExt*/, "tmp", &CI);
7594 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7595 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
7596 In->getName()+".not"), CI);
7598 return ReplaceInstUsesWith(CI, In);
7603 // See if the value being truncated is already sign extended. If so, just
7604 // eliminate the trunc/sext pair.
7605 if (getOpcode(Src) == Instruction::Trunc) {
7606 Value *Op = cast<User>(Src)->getOperand(0);
7607 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
7608 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
7609 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
7610 unsigned NumSignBits = ComputeNumSignBits(Op);
7612 if (OpBits == DestBits) {
7613 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7614 // bits, it is already ready.
7615 if (NumSignBits > DestBits-MidBits)
7616 return ReplaceInstUsesWith(CI, Op);
7617 } else if (OpBits < DestBits) {
7618 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7619 // bits, just sext from i32.
7620 if (NumSignBits > OpBits-MidBits)
7621 return new SExtInst(Op, CI.getType(), "tmp");
7623 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7624 // bits, just truncate to i32.
7625 if (NumSignBits > OpBits-MidBits)
7626 return new TruncInst(Op, CI.getType(), "tmp");
7633 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7634 /// in the specified FP type without changing its value.
7635 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7636 APFloat F = CFP->getValueAPF();
7637 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7638 return ConstantFP::get(F);
7642 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7643 /// through it until we get the source value.
7644 static Value *LookThroughFPExtensions(Value *V) {
7645 if (Instruction *I = dyn_cast<Instruction>(V))
7646 if (I->getOpcode() == Instruction::FPExt)
7647 return LookThroughFPExtensions(I->getOperand(0));
7649 // If this value is a constant, return the constant in the smallest FP type
7650 // that can accurately represent it. This allows us to turn
7651 // (float)((double)X+2.0) into x+2.0f.
7652 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7653 if (CFP->getType() == Type::PPC_FP128Ty)
7654 return V; // No constant folding of this.
7655 // See if the value can be truncated to float and then reextended.
7656 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7658 if (CFP->getType() == Type::DoubleTy)
7659 return V; // Won't shrink.
7660 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7662 // Don't try to shrink to various long double types.
7668 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7669 if (Instruction *I = commonCastTransforms(CI))
7672 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7673 // smaller than the destination type, we can eliminate the truncate by doing
7674 // the add as the smaller type. This applies to add/sub/mul/div as well as
7675 // many builtins (sqrt, etc).
7676 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7677 if (OpI && OpI->hasOneUse()) {
7678 switch (OpI->getOpcode()) {
7680 case Instruction::Add:
7681 case Instruction::Sub:
7682 case Instruction::Mul:
7683 case Instruction::FDiv:
7684 case Instruction::FRem:
7685 const Type *SrcTy = OpI->getType();
7686 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7687 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7688 if (LHSTrunc->getType() != SrcTy &&
7689 RHSTrunc->getType() != SrcTy) {
7690 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7691 // If the source types were both smaller than the destination type of
7692 // the cast, do this xform.
7693 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7694 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7695 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7697 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7699 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7708 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7709 return commonCastTransforms(CI);
7712 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
7713 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
7714 // mantissa to accurately represent all values of X. For example, do not
7715 // do this with i64->float->i64.
7716 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
7717 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7718 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
7719 SrcI->getType()->getFPMantissaWidth())
7720 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7722 return commonCastTransforms(FI);
7725 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
7726 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
7727 // mantissa to accurately represent all values of X. For example, do not
7728 // do this with i64->float->i64.
7729 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
7730 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7731 (int)FI.getType()->getPrimitiveSizeInBits() <=
7732 SrcI->getType()->getFPMantissaWidth())
7733 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7735 return commonCastTransforms(FI);
7738 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7739 return commonCastTransforms(CI);
7742 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7743 return commonCastTransforms(CI);
7746 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7747 return commonPointerCastTransforms(CI);
7750 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7751 if (Instruction *I = commonCastTransforms(CI))
7754 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7755 if (!DestPointee->isSized()) return 0;
7757 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7760 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7761 m_ConstantInt(Cst)))) {
7762 // If the source and destination operands have the same type, see if this
7763 // is a single-index GEP.
7764 if (X->getType() == CI.getType()) {
7765 // Get the size of the pointee type.
7766 uint64_t Size = TD->getABITypeSize(DestPointee);
7768 // Convert the constant to intptr type.
7769 APInt Offset = Cst->getValue();
7770 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7772 // If Offset is evenly divisible by Size, we can do this xform.
7773 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7774 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7775 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7778 // TODO: Could handle other cases, e.g. where add is indexing into field of
7780 } else if (CI.getOperand(0)->hasOneUse() &&
7781 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7782 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7783 // "inttoptr+GEP" instead of "add+intptr".
7785 // Get the size of the pointee type.
7786 uint64_t Size = TD->getABITypeSize(DestPointee);
7788 // Convert the constant to intptr type.
7789 APInt Offset = Cst->getValue();
7790 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7792 // If Offset is evenly divisible by Size, we can do this xform.
7793 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7794 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7796 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7798 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7804 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7805 // If the operands are integer typed then apply the integer transforms,
7806 // otherwise just apply the common ones.
7807 Value *Src = CI.getOperand(0);
7808 const Type *SrcTy = Src->getType();
7809 const Type *DestTy = CI.getType();
7811 if (SrcTy->isInteger() && DestTy->isInteger()) {
7812 if (Instruction *Result = commonIntCastTransforms(CI))
7814 } else if (isa<PointerType>(SrcTy)) {
7815 if (Instruction *I = commonPointerCastTransforms(CI))
7818 if (Instruction *Result = commonCastTransforms(CI))
7823 // Get rid of casts from one type to the same type. These are useless and can
7824 // be replaced by the operand.
7825 if (DestTy == Src->getType())
7826 return ReplaceInstUsesWith(CI, Src);
7828 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7829 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7830 const Type *DstElTy = DstPTy->getElementType();
7831 const Type *SrcElTy = SrcPTy->getElementType();
7833 // If the address spaces don't match, don't eliminate the bitcast, which is
7834 // required for changing types.
7835 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7838 // If we are casting a malloc or alloca to a pointer to a type of the same
7839 // size, rewrite the allocation instruction to allocate the "right" type.
7840 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7841 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7844 // If the source and destination are pointers, and this cast is equivalent
7845 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7846 // This can enhance SROA and other transforms that want type-safe pointers.
7847 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7848 unsigned NumZeros = 0;
7849 while (SrcElTy != DstElTy &&
7850 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7851 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7852 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7856 // If we found a path from the src to dest, create the getelementptr now.
7857 if (SrcElTy == DstElTy) {
7858 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7859 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7860 ((Instruction*) NULL));
7864 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7865 if (SVI->hasOneUse()) {
7866 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7867 // a bitconvert to a vector with the same # elts.
7868 if (isa<VectorType>(DestTy) &&
7869 cast<VectorType>(DestTy)->getNumElements() ==
7870 SVI->getType()->getNumElements()) {
7872 // If either of the operands is a cast from CI.getType(), then
7873 // evaluating the shuffle in the casted destination's type will allow
7874 // us to eliminate at least one cast.
7875 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7876 Tmp->getOperand(0)->getType() == DestTy) ||
7877 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7878 Tmp->getOperand(0)->getType() == DestTy)) {
7879 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7880 SVI->getOperand(0), DestTy, &CI);
7881 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7882 SVI->getOperand(1), DestTy, &CI);
7883 // Return a new shuffle vector. Use the same element ID's, as we
7884 // know the vector types match #elts.
7885 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7893 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7895 /// %D = select %cond, %C, %A
7897 /// %C = select %cond, %B, 0
7900 /// Assuming that the specified instruction is an operand to the select, return
7901 /// a bitmask indicating which operands of this instruction are foldable if they
7902 /// equal the other incoming value of the select.
7904 static unsigned GetSelectFoldableOperands(Instruction *I) {
7905 switch (I->getOpcode()) {
7906 case Instruction::Add:
7907 case Instruction::Mul:
7908 case Instruction::And:
7909 case Instruction::Or:
7910 case Instruction::Xor:
7911 return 3; // Can fold through either operand.
7912 case Instruction::Sub: // Can only fold on the amount subtracted.
7913 case Instruction::Shl: // Can only fold on the shift amount.
7914 case Instruction::LShr:
7915 case Instruction::AShr:
7918 return 0; // Cannot fold
7922 /// GetSelectFoldableConstant - For the same transformation as the previous
7923 /// function, return the identity constant that goes into the select.
7924 static Constant *GetSelectFoldableConstant(Instruction *I) {
7925 switch (I->getOpcode()) {
7926 default: assert(0 && "This cannot happen!"); abort();
7927 case Instruction::Add:
7928 case Instruction::Sub:
7929 case Instruction::Or:
7930 case Instruction::Xor:
7931 case Instruction::Shl:
7932 case Instruction::LShr:
7933 case Instruction::AShr:
7934 return Constant::getNullValue(I->getType());
7935 case Instruction::And:
7936 return Constant::getAllOnesValue(I->getType());
7937 case Instruction::Mul:
7938 return ConstantInt::get(I->getType(), 1);
7942 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7943 /// have the same opcode and only one use each. Try to simplify this.
7944 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7946 if (TI->getNumOperands() == 1) {
7947 // If this is a non-volatile load or a cast from the same type,
7950 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7953 return 0; // unknown unary op.
7956 // Fold this by inserting a select from the input values.
7957 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
7958 FI->getOperand(0), SI.getName()+".v");
7959 InsertNewInstBefore(NewSI, SI);
7960 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
7964 // Only handle binary operators here.
7965 if (!isa<BinaryOperator>(TI))
7968 // Figure out if the operations have any operands in common.
7969 Value *MatchOp, *OtherOpT, *OtherOpF;
7971 if (TI->getOperand(0) == FI->getOperand(0)) {
7972 MatchOp = TI->getOperand(0);
7973 OtherOpT = TI->getOperand(1);
7974 OtherOpF = FI->getOperand(1);
7975 MatchIsOpZero = true;
7976 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7977 MatchOp = TI->getOperand(1);
7978 OtherOpT = TI->getOperand(0);
7979 OtherOpF = FI->getOperand(0);
7980 MatchIsOpZero = false;
7981 } else if (!TI->isCommutative()) {
7983 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7984 MatchOp = TI->getOperand(0);
7985 OtherOpT = TI->getOperand(1);
7986 OtherOpF = FI->getOperand(0);
7987 MatchIsOpZero = true;
7988 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7989 MatchOp = TI->getOperand(1);
7990 OtherOpT = TI->getOperand(0);
7991 OtherOpF = FI->getOperand(1);
7992 MatchIsOpZero = true;
7997 // If we reach here, they do have operations in common.
7998 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
7999 OtherOpF, SI.getName()+".v");
8000 InsertNewInstBefore(NewSI, SI);
8002 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8004 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8006 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8008 assert(0 && "Shouldn't get here");
8012 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8013 Value *CondVal = SI.getCondition();
8014 Value *TrueVal = SI.getTrueValue();
8015 Value *FalseVal = SI.getFalseValue();
8017 // select true, X, Y -> X
8018 // select false, X, Y -> Y
8019 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8020 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8022 // select C, X, X -> X
8023 if (TrueVal == FalseVal)
8024 return ReplaceInstUsesWith(SI, TrueVal);
8026 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8027 return ReplaceInstUsesWith(SI, FalseVal);
8028 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8029 return ReplaceInstUsesWith(SI, TrueVal);
8030 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8031 if (isa<Constant>(TrueVal))
8032 return ReplaceInstUsesWith(SI, TrueVal);
8034 return ReplaceInstUsesWith(SI, FalseVal);
8037 if (SI.getType() == Type::Int1Ty) {
8038 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8039 if (C->getZExtValue()) {
8040 // Change: A = select B, true, C --> A = or B, C
8041 return BinaryOperator::CreateOr(CondVal, FalseVal);
8043 // Change: A = select B, false, C --> A = and !B, C
8045 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8046 "not."+CondVal->getName()), SI);
8047 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8049 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8050 if (C->getZExtValue() == false) {
8051 // Change: A = select B, C, false --> A = and B, C
8052 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8054 // Change: A = select B, C, true --> A = or !B, C
8056 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8057 "not."+CondVal->getName()), SI);
8058 return BinaryOperator::CreateOr(NotCond, TrueVal);
8062 // select a, b, a -> a&b
8063 // select a, a, b -> a|b
8064 if (CondVal == TrueVal)
8065 return BinaryOperator::CreateOr(CondVal, FalseVal);
8066 else if (CondVal == FalseVal)
8067 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8070 // Selecting between two integer constants?
8071 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8072 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8073 // select C, 1, 0 -> zext C to int
8074 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8075 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8076 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8077 // select C, 0, 1 -> zext !C to int
8079 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8080 "not."+CondVal->getName()), SI);
8081 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8084 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8086 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8088 // (x <s 0) ? -1 : 0 -> ashr x, 31
8089 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8090 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8091 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8092 // The comparison constant and the result are not neccessarily the
8093 // same width. Make an all-ones value by inserting a AShr.
8094 Value *X = IC->getOperand(0);
8095 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8096 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8097 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8099 InsertNewInstBefore(SRA, SI);
8101 // Finally, convert to the type of the select RHS. We figure out
8102 // if this requires a SExt, Trunc or BitCast based on the sizes.
8103 Instruction::CastOps opc = Instruction::BitCast;
8104 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8105 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8106 if (SRASize < SISize)
8107 opc = Instruction::SExt;
8108 else if (SRASize > SISize)
8109 opc = Instruction::Trunc;
8110 return CastInst::Create(opc, SRA, SI.getType());
8115 // If one of the constants is zero (we know they can't both be) and we
8116 // have an icmp instruction with zero, and we have an 'and' with the
8117 // non-constant value, eliminate this whole mess. This corresponds to
8118 // cases like this: ((X & 27) ? 27 : 0)
8119 if (TrueValC->isZero() || FalseValC->isZero())
8120 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8121 cast<Constant>(IC->getOperand(1))->isNullValue())
8122 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8123 if (ICA->getOpcode() == Instruction::And &&
8124 isa<ConstantInt>(ICA->getOperand(1)) &&
8125 (ICA->getOperand(1) == TrueValC ||
8126 ICA->getOperand(1) == FalseValC) &&
8127 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8128 // Okay, now we know that everything is set up, we just don't
8129 // know whether we have a icmp_ne or icmp_eq and whether the
8130 // true or false val is the zero.
8131 bool ShouldNotVal = !TrueValC->isZero();
8132 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8135 V = InsertNewInstBefore(BinaryOperator::Create(
8136 Instruction::Xor, V, ICA->getOperand(1)), SI);
8137 return ReplaceInstUsesWith(SI, V);
8142 // See if we are selecting two values based on a comparison of the two values.
8143 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8144 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8145 // Transform (X == Y) ? X : Y -> Y
8146 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8147 // This is not safe in general for floating point:
8148 // consider X== -0, Y== +0.
8149 // It becomes safe if either operand is a nonzero constant.
8150 ConstantFP *CFPt, *CFPf;
8151 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8152 !CFPt->getValueAPF().isZero()) ||
8153 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8154 !CFPf->getValueAPF().isZero()))
8155 return ReplaceInstUsesWith(SI, FalseVal);
8157 // Transform (X != Y) ? X : Y -> X
8158 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8159 return ReplaceInstUsesWith(SI, TrueVal);
8160 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8162 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8163 // Transform (X == Y) ? Y : X -> X
8164 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8165 // This is not safe in general for floating point:
8166 // consider X== -0, Y== +0.
8167 // It becomes safe if either operand is a nonzero constant.
8168 ConstantFP *CFPt, *CFPf;
8169 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8170 !CFPt->getValueAPF().isZero()) ||
8171 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8172 !CFPf->getValueAPF().isZero()))
8173 return ReplaceInstUsesWith(SI, FalseVal);
8175 // Transform (X != Y) ? Y : X -> Y
8176 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8177 return ReplaceInstUsesWith(SI, TrueVal);
8178 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8182 // See if we are selecting two values based on a comparison of the two values.
8183 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8184 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8185 // Transform (X == Y) ? X : Y -> Y
8186 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8187 return ReplaceInstUsesWith(SI, FalseVal);
8188 // Transform (X != Y) ? X : Y -> X
8189 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8190 return ReplaceInstUsesWith(SI, TrueVal);
8191 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8193 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8194 // Transform (X == Y) ? Y : X -> X
8195 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8196 return ReplaceInstUsesWith(SI, FalseVal);
8197 // Transform (X != Y) ? Y : X -> Y
8198 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8199 return ReplaceInstUsesWith(SI, TrueVal);
8200 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8204 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8205 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8206 if (TI->hasOneUse() && FI->hasOneUse()) {
8207 Instruction *AddOp = 0, *SubOp = 0;
8209 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8210 if (TI->getOpcode() == FI->getOpcode())
8211 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8214 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8215 // even legal for FP.
8216 if (TI->getOpcode() == Instruction::Sub &&
8217 FI->getOpcode() == Instruction::Add) {
8218 AddOp = FI; SubOp = TI;
8219 } else if (FI->getOpcode() == Instruction::Sub &&
8220 TI->getOpcode() == Instruction::Add) {
8221 AddOp = TI; SubOp = FI;
8225 Value *OtherAddOp = 0;
8226 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8227 OtherAddOp = AddOp->getOperand(1);
8228 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8229 OtherAddOp = AddOp->getOperand(0);
8233 // So at this point we know we have (Y -> OtherAddOp):
8234 // select C, (add X, Y), (sub X, Z)
8235 Value *NegVal; // Compute -Z
8236 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8237 NegVal = ConstantExpr::getNeg(C);
8239 NegVal = InsertNewInstBefore(
8240 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8243 Value *NewTrueOp = OtherAddOp;
8244 Value *NewFalseOp = NegVal;
8246 std::swap(NewTrueOp, NewFalseOp);
8247 Instruction *NewSel =
8248 SelectInst::Create(CondVal, NewTrueOp,
8249 NewFalseOp, SI.getName() + ".p");
8251 NewSel = InsertNewInstBefore(NewSel, SI);
8252 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8257 // See if we can fold the select into one of our operands.
8258 if (SI.getType()->isInteger()) {
8259 // See the comment above GetSelectFoldableOperands for a description of the
8260 // transformation we are doing here.
8261 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8262 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8263 !isa<Constant>(FalseVal))
8264 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8265 unsigned OpToFold = 0;
8266 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8268 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8273 Constant *C = GetSelectFoldableConstant(TVI);
8274 Instruction *NewSel =
8275 SelectInst::Create(SI.getCondition(),
8276 TVI->getOperand(2-OpToFold), C);
8277 InsertNewInstBefore(NewSel, SI);
8278 NewSel->takeName(TVI);
8279 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8280 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8282 assert(0 && "Unknown instruction!!");
8287 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8288 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8289 !isa<Constant>(TrueVal))
8290 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8291 unsigned OpToFold = 0;
8292 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8294 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8299 Constant *C = GetSelectFoldableConstant(FVI);
8300 Instruction *NewSel =
8301 SelectInst::Create(SI.getCondition(), C,
8302 FVI->getOperand(2-OpToFold));
8303 InsertNewInstBefore(NewSel, SI);
8304 NewSel->takeName(FVI);
8305 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8306 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8308 assert(0 && "Unknown instruction!!");
8313 if (BinaryOperator::isNot(CondVal)) {
8314 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8315 SI.setOperand(1, FalseVal);
8316 SI.setOperand(2, TrueVal);
8323 /// EnforceKnownAlignment - If the specified pointer points to an object that
8324 /// we control, modify the object's alignment to PrefAlign. This isn't
8325 /// often possible though. If alignment is important, a more reliable approach
8326 /// is to simply align all global variables and allocation instructions to
8327 /// their preferred alignment from the beginning.
8329 static unsigned EnforceKnownAlignment(Value *V,
8330 unsigned Align, unsigned PrefAlign) {
8332 User *U = dyn_cast<User>(V);
8333 if (!U) return Align;
8335 switch (getOpcode(U)) {
8337 case Instruction::BitCast:
8338 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8339 case Instruction::GetElementPtr: {
8340 // If all indexes are zero, it is just the alignment of the base pointer.
8341 bool AllZeroOperands = true;
8342 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8343 if (!isa<Constant>(*i) ||
8344 !cast<Constant>(*i)->isNullValue()) {
8345 AllZeroOperands = false;
8349 if (AllZeroOperands) {
8350 // Treat this like a bitcast.
8351 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8357 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8358 // If there is a large requested alignment and we can, bump up the alignment
8360 if (!GV->isDeclaration()) {
8361 GV->setAlignment(PrefAlign);
8364 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8365 // If there is a requested alignment and if this is an alloca, round up. We
8366 // don't do this for malloc, because some systems can't respect the request.
8367 if (isa<AllocaInst>(AI)) {
8368 AI->setAlignment(PrefAlign);
8376 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8377 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8378 /// and it is more than the alignment of the ultimate object, see if we can
8379 /// increase the alignment of the ultimate object, making this check succeed.
8380 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8381 unsigned PrefAlign) {
8382 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8383 sizeof(PrefAlign) * CHAR_BIT;
8384 APInt Mask = APInt::getAllOnesValue(BitWidth);
8385 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8386 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8387 unsigned TrailZ = KnownZero.countTrailingOnes();
8388 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8390 if (PrefAlign > Align)
8391 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8393 // We don't need to make any adjustment.
8397 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8398 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8399 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8400 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8401 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8403 if (CopyAlign < MinAlign) {
8404 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8408 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8410 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8411 if (MemOpLength == 0) return 0;
8413 // Source and destination pointer types are always "i8*" for intrinsic. See
8414 // if the size is something we can handle with a single primitive load/store.
8415 // A single load+store correctly handles overlapping memory in the memmove
8417 unsigned Size = MemOpLength->getZExtValue();
8418 if (Size == 0) return MI; // Delete this mem transfer.
8420 if (Size > 8 || (Size&(Size-1)))
8421 return 0; // If not 1/2/4/8 bytes, exit.
8423 // Use an integer load+store unless we can find something better.
8424 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8426 // Memcpy forces the use of i8* for the source and destination. That means
8427 // that if you're using memcpy to move one double around, you'll get a cast
8428 // from double* to i8*. We'd much rather use a double load+store rather than
8429 // an i64 load+store, here because this improves the odds that the source or
8430 // dest address will be promotable. See if we can find a better type than the
8431 // integer datatype.
8432 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8433 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8434 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8435 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8436 // down through these levels if so.
8437 while (!SrcETy->isSingleValueType()) {
8438 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8439 if (STy->getNumElements() == 1)
8440 SrcETy = STy->getElementType(0);
8443 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8444 if (ATy->getNumElements() == 1)
8445 SrcETy = ATy->getElementType();
8452 if (SrcETy->isSingleValueType())
8453 NewPtrTy = PointerType::getUnqual(SrcETy);
8458 // If the memcpy/memmove provides better alignment info than we can
8460 SrcAlign = std::max(SrcAlign, CopyAlign);
8461 DstAlign = std::max(DstAlign, CopyAlign);
8463 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8464 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8465 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8466 InsertNewInstBefore(L, *MI);
8467 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8469 // Set the size of the copy to 0, it will be deleted on the next iteration.
8470 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8474 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8475 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8476 if (MI->getAlignment()->getZExtValue() < Alignment) {
8477 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8481 // Extract the length and alignment and fill if they are constant.
8482 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8483 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8484 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8486 uint64_t Len = LenC->getZExtValue();
8487 Alignment = MI->getAlignment()->getZExtValue();
8489 // If the length is zero, this is a no-op
8490 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8492 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8493 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8494 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8496 Value *Dest = MI->getDest();
8497 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8499 // Alignment 0 is identity for alignment 1 for memset, but not store.
8500 if (Alignment == 0) Alignment = 1;
8502 // Extract the fill value and store.
8503 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8504 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8507 // Set the size of the copy to 0, it will be deleted on the next iteration.
8508 MI->setLength(Constant::getNullValue(LenC->getType()));
8516 /// visitCallInst - CallInst simplification. This mostly only handles folding
8517 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8518 /// the heavy lifting.
8520 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8521 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8522 if (!II) return visitCallSite(&CI);
8524 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8526 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8527 bool Changed = false;
8529 // memmove/cpy/set of zero bytes is a noop.
8530 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8531 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8533 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8534 if (CI->getZExtValue() == 1) {
8535 // Replace the instruction with just byte operations. We would
8536 // transform other cases to loads/stores, but we don't know if
8537 // alignment is sufficient.
8541 // If we have a memmove and the source operation is a constant global,
8542 // then the source and dest pointers can't alias, so we can change this
8543 // into a call to memcpy.
8544 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8545 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8546 if (GVSrc->isConstant()) {
8547 Module *M = CI.getParent()->getParent()->getParent();
8548 Intrinsic::ID MemCpyID;
8549 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8550 MemCpyID = Intrinsic::memcpy_i32;
8552 MemCpyID = Intrinsic::memcpy_i64;
8553 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8557 // memmove(x,x,size) -> noop.
8558 if (MMI->getSource() == MMI->getDest())
8559 return EraseInstFromFunction(CI);
8562 // If we can determine a pointer alignment that is bigger than currently
8563 // set, update the alignment.
8564 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8565 if (Instruction *I = SimplifyMemTransfer(MI))
8567 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8568 if (Instruction *I = SimplifyMemSet(MSI))
8572 if (Changed) return II;
8575 switch (II->getIntrinsicID()) {
8577 case Intrinsic::bswap:
8578 // bswap(bswap(x)) -> x
8579 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
8580 if (Operand->getIntrinsicID() == Intrinsic::bswap)
8581 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
8583 case Intrinsic::ppc_altivec_lvx:
8584 case Intrinsic::ppc_altivec_lvxl:
8585 case Intrinsic::x86_sse_loadu_ps:
8586 case Intrinsic::x86_sse2_loadu_pd:
8587 case Intrinsic::x86_sse2_loadu_dq:
8588 // Turn PPC lvx -> load if the pointer is known aligned.
8589 // Turn X86 loadups -> load if the pointer is known aligned.
8590 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8591 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8592 PointerType::getUnqual(II->getType()),
8594 return new LoadInst(Ptr);
8597 case Intrinsic::ppc_altivec_stvx:
8598 case Intrinsic::ppc_altivec_stvxl:
8599 // Turn stvx -> store if the pointer is known aligned.
8600 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8601 const Type *OpPtrTy =
8602 PointerType::getUnqual(II->getOperand(1)->getType());
8603 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8604 return new StoreInst(II->getOperand(1), Ptr);
8607 case Intrinsic::x86_sse_storeu_ps:
8608 case Intrinsic::x86_sse2_storeu_pd:
8609 case Intrinsic::x86_sse2_storeu_dq:
8610 case Intrinsic::x86_sse2_storel_dq:
8611 // Turn X86 storeu -> store if the pointer is known aligned.
8612 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8613 const Type *OpPtrTy =
8614 PointerType::getUnqual(II->getOperand(2)->getType());
8615 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8616 return new StoreInst(II->getOperand(2), Ptr);
8620 case Intrinsic::x86_sse_cvttss2si: {
8621 // These intrinsics only demands the 0th element of its input vector. If
8622 // we can simplify the input based on that, do so now.
8624 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8626 II->setOperand(1, V);
8632 case Intrinsic::ppc_altivec_vperm:
8633 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8634 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8635 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8637 // Check that all of the elements are integer constants or undefs.
8638 bool AllEltsOk = true;
8639 for (unsigned i = 0; i != 16; ++i) {
8640 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8641 !isa<UndefValue>(Mask->getOperand(i))) {
8648 // Cast the input vectors to byte vectors.
8649 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8650 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8651 Value *Result = UndefValue::get(Op0->getType());
8653 // Only extract each element once.
8654 Value *ExtractedElts[32];
8655 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8657 for (unsigned i = 0; i != 16; ++i) {
8658 if (isa<UndefValue>(Mask->getOperand(i)))
8660 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8661 Idx &= 31; // Match the hardware behavior.
8663 if (ExtractedElts[Idx] == 0) {
8665 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8666 InsertNewInstBefore(Elt, CI);
8667 ExtractedElts[Idx] = Elt;
8670 // Insert this value into the result vector.
8671 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
8673 InsertNewInstBefore(cast<Instruction>(Result), CI);
8675 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
8680 case Intrinsic::stackrestore: {
8681 // If the save is right next to the restore, remove the restore. This can
8682 // happen when variable allocas are DCE'd.
8683 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8684 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8685 BasicBlock::iterator BI = SS;
8687 return EraseInstFromFunction(CI);
8691 // Scan down this block to see if there is another stack restore in the
8692 // same block without an intervening call/alloca.
8693 BasicBlock::iterator BI = II;
8694 TerminatorInst *TI = II->getParent()->getTerminator();
8695 bool CannotRemove = false;
8696 for (++BI; &*BI != TI; ++BI) {
8697 if (isa<AllocaInst>(BI)) {
8698 CannotRemove = true;
8701 if (isa<CallInst>(BI)) {
8702 if (!isa<IntrinsicInst>(BI)) {
8703 CannotRemove = true;
8706 // If there is a stackrestore below this one, remove this one.
8707 return EraseInstFromFunction(CI);
8711 // If the stack restore is in a return/unwind block and if there are no
8712 // allocas or calls between the restore and the return, nuke the restore.
8713 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8714 return EraseInstFromFunction(CI);
8719 return visitCallSite(II);
8722 // InvokeInst simplification
8724 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8725 return visitCallSite(&II);
8728 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8729 /// passed through the varargs area, we can eliminate the use of the cast.
8730 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8731 const CastInst * const CI,
8732 const TargetData * const TD,
8734 if (!CI->isLosslessCast())
8737 // The size of ByVal arguments is derived from the type, so we
8738 // can't change to a type with a different size. If the size were
8739 // passed explicitly we could avoid this check.
8740 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
8744 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
8745 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
8746 if (!SrcTy->isSized() || !DstTy->isSized())
8748 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
8753 // visitCallSite - Improvements for call and invoke instructions.
8755 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8756 bool Changed = false;
8758 // If the callee is a constexpr cast of a function, attempt to move the cast
8759 // to the arguments of the call/invoke.
8760 if (transformConstExprCastCall(CS)) return 0;
8762 Value *Callee = CS.getCalledValue();
8764 if (Function *CalleeF = dyn_cast<Function>(Callee))
8765 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8766 Instruction *OldCall = CS.getInstruction();
8767 // If the call and callee calling conventions don't match, this call must
8768 // be unreachable, as the call is undefined.
8769 new StoreInst(ConstantInt::getTrue(),
8770 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8772 if (!OldCall->use_empty())
8773 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8774 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8775 return EraseInstFromFunction(*OldCall);
8779 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8780 // This instruction is not reachable, just remove it. We insert a store to
8781 // undef so that we know that this code is not reachable, despite the fact
8782 // that we can't modify the CFG here.
8783 new StoreInst(ConstantInt::getTrue(),
8784 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8785 CS.getInstruction());
8787 if (!CS.getInstruction()->use_empty())
8788 CS.getInstruction()->
8789 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8791 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8792 // Don't break the CFG, insert a dummy cond branch.
8793 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8794 ConstantInt::getTrue(), II);
8796 return EraseInstFromFunction(*CS.getInstruction());
8799 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8800 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8801 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8802 return transformCallThroughTrampoline(CS);
8804 const PointerType *PTy = cast<PointerType>(Callee->getType());
8805 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8806 if (FTy->isVarArg()) {
8807 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
8808 // See if we can optimize any arguments passed through the varargs area of
8810 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8811 E = CS.arg_end(); I != E; ++I, ++ix) {
8812 CastInst *CI = dyn_cast<CastInst>(*I);
8813 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
8814 *I = CI->getOperand(0);
8820 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8821 // Inline asm calls cannot throw - mark them 'nounwind'.
8822 CS.setDoesNotThrow();
8826 return Changed ? CS.getInstruction() : 0;
8829 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8830 // attempt to move the cast to the arguments of the call/invoke.
8832 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8833 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8834 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8835 if (CE->getOpcode() != Instruction::BitCast ||
8836 !isa<Function>(CE->getOperand(0)))
8838 Function *Callee = cast<Function>(CE->getOperand(0));
8839 Instruction *Caller = CS.getInstruction();
8840 const PAListPtr &CallerPAL = CS.getParamAttrs();
8842 // Okay, this is a cast from a function to a different type. Unless doing so
8843 // would cause a type conversion of one of our arguments, change this call to
8844 // be a direct call with arguments casted to the appropriate types.
8846 const FunctionType *FT = Callee->getFunctionType();
8847 const Type *OldRetTy = Caller->getType();
8848 const Type *NewRetTy = FT->getReturnType();
8850 if (isa<StructType>(NewRetTy))
8851 return false; // TODO: Handle multiple return values.
8853 // Check to see if we are changing the return type...
8854 if (OldRetTy != NewRetTy) {
8855 if (Callee->isDeclaration() &&
8856 // Conversion is ok if changing from one pointer type to another or from
8857 // a pointer to an integer of the same size.
8858 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
8859 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
8860 return false; // Cannot transform this return value.
8862 if (!Caller->use_empty() &&
8863 // void -> non-void is handled specially
8864 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
8865 return false; // Cannot transform this return value.
8867 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8868 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8869 if (RAttrs & ParamAttr::typeIncompatible(NewRetTy))
8870 return false; // Attribute not compatible with transformed value.
8873 // If the callsite is an invoke instruction, and the return value is used by
8874 // a PHI node in a successor, we cannot change the return type of the call
8875 // because there is no place to put the cast instruction (without breaking
8876 // the critical edge). Bail out in this case.
8877 if (!Caller->use_empty())
8878 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8879 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8881 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8882 if (PN->getParent() == II->getNormalDest() ||
8883 PN->getParent() == II->getUnwindDest())
8887 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8888 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8890 CallSite::arg_iterator AI = CS.arg_begin();
8891 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8892 const Type *ParamTy = FT->getParamType(i);
8893 const Type *ActTy = (*AI)->getType();
8895 if (!CastInst::isCastable(ActTy, ParamTy))
8896 return false; // Cannot transform this parameter value.
8898 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8899 return false; // Attribute not compatible with transformed value.
8901 // Converting from one pointer type to another or between a pointer and an
8902 // integer of the same size is safe even if we do not have a body.
8903 bool isConvertible = ActTy == ParamTy ||
8904 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
8905 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
8906 if (Callee->isDeclaration() && !isConvertible) return false;
8909 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8910 Callee->isDeclaration())
8911 return false; // Do not delete arguments unless we have a function body.
8913 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
8914 !CallerPAL.isEmpty())
8915 // In this case we have more arguments than the new function type, but we
8916 // won't be dropping them. Check that these extra arguments have attributes
8917 // that are compatible with being a vararg call argument.
8918 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
8919 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
8921 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
8922 if (PAttrs & ParamAttr::VarArgsIncompatible)
8926 // Okay, we decided that this is a safe thing to do: go ahead and start
8927 // inserting cast instructions as necessary...
8928 std::vector<Value*> Args;
8929 Args.reserve(NumActualArgs);
8930 SmallVector<ParamAttrsWithIndex, 8> attrVec;
8931 attrVec.reserve(NumCommonArgs);
8933 // Get any return attributes.
8934 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8936 // If the return value is not being used, the type may not be compatible
8937 // with the existing attributes. Wipe out any problematic attributes.
8938 RAttrs &= ~ParamAttr::typeIncompatible(NewRetTy);
8940 // Add the new return attributes.
8942 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8944 AI = CS.arg_begin();
8945 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8946 const Type *ParamTy = FT->getParamType(i);
8947 if ((*AI)->getType() == ParamTy) {
8948 Args.push_back(*AI);
8950 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8951 false, ParamTy, false);
8952 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
8953 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8956 // Add any parameter attributes.
8957 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8958 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8961 // If the function takes more arguments than the call was taking, add them
8963 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8964 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8966 // If we are removing arguments to the function, emit an obnoxious warning...
8967 if (FT->getNumParams() < NumActualArgs) {
8968 if (!FT->isVarArg()) {
8969 cerr << "WARNING: While resolving call to function '"
8970 << Callee->getName() << "' arguments were dropped!\n";
8972 // Add all of the arguments in their promoted form to the arg list...
8973 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8974 const Type *PTy = getPromotedType((*AI)->getType());
8975 if (PTy != (*AI)->getType()) {
8976 // Must promote to pass through va_arg area!
8977 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8979 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
8980 InsertNewInstBefore(Cast, *Caller);
8981 Args.push_back(Cast);
8983 Args.push_back(*AI);
8986 // Add any parameter attributes.
8987 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8988 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8993 if (NewRetTy == Type::VoidTy)
8994 Caller->setName(""); // Void type should not have a name.
8996 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
8999 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9000 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9001 Args.begin(), Args.end(),
9002 Caller->getName(), Caller);
9003 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9004 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9006 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9007 Caller->getName(), Caller);
9008 CallInst *CI = cast<CallInst>(Caller);
9009 if (CI->isTailCall())
9010 cast<CallInst>(NC)->setTailCall();
9011 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9012 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9015 // Insert a cast of the return type as necessary.
9017 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9018 if (NV->getType() != Type::VoidTy) {
9019 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9021 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9023 // If this is an invoke instruction, we should insert it after the first
9024 // non-phi, instruction in the normal successor block.
9025 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9026 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9027 InsertNewInstBefore(NC, *I);
9029 // Otherwise, it's a call, just insert cast right after the call instr
9030 InsertNewInstBefore(NC, *Caller);
9032 AddUsersToWorkList(*Caller);
9034 NV = UndefValue::get(Caller->getType());
9038 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9039 Caller->replaceAllUsesWith(NV);
9040 Caller->eraseFromParent();
9041 RemoveFromWorkList(Caller);
9045 // transformCallThroughTrampoline - Turn a call to a function created by the
9046 // init_trampoline intrinsic into a direct call to the underlying function.
9048 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9049 Value *Callee = CS.getCalledValue();
9050 const PointerType *PTy = cast<PointerType>(Callee->getType());
9051 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9052 const PAListPtr &Attrs = CS.getParamAttrs();
9054 // If the call already has the 'nest' attribute somewhere then give up -
9055 // otherwise 'nest' would occur twice after splicing in the chain.
9056 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9059 IntrinsicInst *Tramp =
9060 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9062 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9063 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9064 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9066 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9067 if (!NestAttrs.isEmpty()) {
9068 unsigned NestIdx = 1;
9069 const Type *NestTy = 0;
9070 ParameterAttributes NestAttr = ParamAttr::None;
9072 // Look for a parameter marked with the 'nest' attribute.
9073 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9074 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9075 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9076 // Record the parameter type and any other attributes.
9078 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9083 Instruction *Caller = CS.getInstruction();
9084 std::vector<Value*> NewArgs;
9085 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9087 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9088 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9090 // Insert the nest argument into the call argument list, which may
9091 // mean appending it. Likewise for attributes.
9093 // Add any function result attributes.
9094 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9095 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9099 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9101 if (Idx == NestIdx) {
9102 // Add the chain argument and attributes.
9103 Value *NestVal = Tramp->getOperand(3);
9104 if (NestVal->getType() != NestTy)
9105 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9106 NewArgs.push_back(NestVal);
9107 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9113 // Add the original argument and attributes.
9114 NewArgs.push_back(*I);
9115 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9117 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9123 // The trampoline may have been bitcast to a bogus type (FTy).
9124 // Handle this by synthesizing a new function type, equal to FTy
9125 // with the chain parameter inserted.
9127 std::vector<const Type*> NewTypes;
9128 NewTypes.reserve(FTy->getNumParams()+1);
9130 // Insert the chain's type into the list of parameter types, which may
9131 // mean appending it.
9134 FunctionType::param_iterator I = FTy->param_begin(),
9135 E = FTy->param_end();
9139 // Add the chain's type.
9140 NewTypes.push_back(NestTy);
9145 // Add the original type.
9146 NewTypes.push_back(*I);
9152 // Replace the trampoline call with a direct call. Let the generic
9153 // code sort out any function type mismatches.
9154 FunctionType *NewFTy =
9155 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9156 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9157 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9158 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9160 Instruction *NewCaller;
9161 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9162 NewCaller = InvokeInst::Create(NewCallee,
9163 II->getNormalDest(), II->getUnwindDest(),
9164 NewArgs.begin(), NewArgs.end(),
9165 Caller->getName(), Caller);
9166 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9167 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9169 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9170 Caller->getName(), Caller);
9171 if (cast<CallInst>(Caller)->isTailCall())
9172 cast<CallInst>(NewCaller)->setTailCall();
9173 cast<CallInst>(NewCaller)->
9174 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9175 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9177 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9178 Caller->replaceAllUsesWith(NewCaller);
9179 Caller->eraseFromParent();
9180 RemoveFromWorkList(Caller);
9185 // Replace the trampoline call with a direct call. Since there is no 'nest'
9186 // parameter, there is no need to adjust the argument list. Let the generic
9187 // code sort out any function type mismatches.
9188 Constant *NewCallee =
9189 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9190 CS.setCalledFunction(NewCallee);
9191 return CS.getInstruction();
9194 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9195 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9196 /// and a single binop.
9197 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9198 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9199 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9200 isa<CmpInst>(FirstInst));
9201 unsigned Opc = FirstInst->getOpcode();
9202 Value *LHSVal = FirstInst->getOperand(0);
9203 Value *RHSVal = FirstInst->getOperand(1);
9205 const Type *LHSType = LHSVal->getType();
9206 const Type *RHSType = RHSVal->getType();
9208 // Scan to see if all operands are the same opcode, all have one use, and all
9209 // kill their operands (i.e. the operands have one use).
9210 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9211 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9212 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9213 // Verify type of the LHS matches so we don't fold cmp's of different
9214 // types or GEP's with different index types.
9215 I->getOperand(0)->getType() != LHSType ||
9216 I->getOperand(1)->getType() != RHSType)
9219 // If they are CmpInst instructions, check their predicates
9220 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9221 if (cast<CmpInst>(I)->getPredicate() !=
9222 cast<CmpInst>(FirstInst)->getPredicate())
9225 // Keep track of which operand needs a phi node.
9226 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9227 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9230 // Otherwise, this is safe to transform, determine if it is profitable.
9232 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9233 // Indexes are often folded into load/store instructions, so we don't want to
9234 // hide them behind a phi.
9235 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9238 Value *InLHS = FirstInst->getOperand(0);
9239 Value *InRHS = FirstInst->getOperand(1);
9240 PHINode *NewLHS = 0, *NewRHS = 0;
9242 NewLHS = PHINode::Create(LHSType,
9243 FirstInst->getOperand(0)->getName() + ".pn");
9244 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9245 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9246 InsertNewInstBefore(NewLHS, PN);
9251 NewRHS = PHINode::Create(RHSType,
9252 FirstInst->getOperand(1)->getName() + ".pn");
9253 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9254 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9255 InsertNewInstBefore(NewRHS, PN);
9259 // Add all operands to the new PHIs.
9260 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9262 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9263 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9266 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9267 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9271 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9272 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9273 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9274 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9277 assert(isa<GetElementPtrInst>(FirstInst));
9278 return GetElementPtrInst::Create(LHSVal, RHSVal);
9282 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9283 /// of the block that defines it. This means that it must be obvious the value
9284 /// of the load is not changed from the point of the load to the end of the
9287 /// Finally, it is safe, but not profitable, to sink a load targetting a
9288 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9290 static bool isSafeToSinkLoad(LoadInst *L) {
9291 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9293 for (++BBI; BBI != E; ++BBI)
9294 if (BBI->mayWriteToMemory())
9297 // Check for non-address taken alloca. If not address-taken already, it isn't
9298 // profitable to do this xform.
9299 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9300 bool isAddressTaken = false;
9301 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9303 if (isa<LoadInst>(UI)) continue;
9304 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9305 // If storing TO the alloca, then the address isn't taken.
9306 if (SI->getOperand(1) == AI) continue;
9308 isAddressTaken = true;
9312 if (!isAddressTaken)
9320 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9321 // operator and they all are only used by the PHI, PHI together their
9322 // inputs, and do the operation once, to the result of the PHI.
9323 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9324 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9326 // Scan the instruction, looking for input operations that can be folded away.
9327 // If all input operands to the phi are the same instruction (e.g. a cast from
9328 // the same type or "+42") we can pull the operation through the PHI, reducing
9329 // code size and simplifying code.
9330 Constant *ConstantOp = 0;
9331 const Type *CastSrcTy = 0;
9332 bool isVolatile = false;
9333 if (isa<CastInst>(FirstInst)) {
9334 CastSrcTy = FirstInst->getOperand(0)->getType();
9335 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9336 // Can fold binop, compare or shift here if the RHS is a constant,
9337 // otherwise call FoldPHIArgBinOpIntoPHI.
9338 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9339 if (ConstantOp == 0)
9340 return FoldPHIArgBinOpIntoPHI(PN);
9341 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9342 isVolatile = LI->isVolatile();
9343 // We can't sink the load if the loaded value could be modified between the
9344 // load and the PHI.
9345 if (LI->getParent() != PN.getIncomingBlock(0) ||
9346 !isSafeToSinkLoad(LI))
9348 } else if (isa<GetElementPtrInst>(FirstInst)) {
9349 if (FirstInst->getNumOperands() == 2)
9350 return FoldPHIArgBinOpIntoPHI(PN);
9351 // Can't handle general GEPs yet.
9354 return 0; // Cannot fold this operation.
9357 // Check to see if all arguments are the same operation.
9358 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9359 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9360 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9361 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9364 if (I->getOperand(0)->getType() != CastSrcTy)
9365 return 0; // Cast operation must match.
9366 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9367 // We can't sink the load if the loaded value could be modified between
9368 // the load and the PHI.
9369 if (LI->isVolatile() != isVolatile ||
9370 LI->getParent() != PN.getIncomingBlock(i) ||
9371 !isSafeToSinkLoad(LI))
9374 // If the PHI is volatile and its block has multiple successors, sinking
9375 // it would remove a load of the volatile value from the path through the
9378 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9382 } else if (I->getOperand(1) != ConstantOp) {
9387 // Okay, they are all the same operation. Create a new PHI node of the
9388 // correct type, and PHI together all of the LHS's of the instructions.
9389 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9390 PN.getName()+".in");
9391 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9393 Value *InVal = FirstInst->getOperand(0);
9394 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9396 // Add all operands to the new PHI.
9397 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9398 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9399 if (NewInVal != InVal)
9401 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9406 // The new PHI unions all of the same values together. This is really
9407 // common, so we handle it intelligently here for compile-time speed.
9411 InsertNewInstBefore(NewPN, PN);
9415 // Insert and return the new operation.
9416 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9417 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9418 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9419 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9420 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9421 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9422 PhiVal, ConstantOp);
9423 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9425 // If this was a volatile load that we are merging, make sure to loop through
9426 // and mark all the input loads as non-volatile. If we don't do this, we will
9427 // insert a new volatile load and the old ones will not be deletable.
9429 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9430 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9432 return new LoadInst(PhiVal, "", isVolatile);
9435 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9437 static bool DeadPHICycle(PHINode *PN,
9438 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9439 if (PN->use_empty()) return true;
9440 if (!PN->hasOneUse()) return false;
9442 // Remember this node, and if we find the cycle, return.
9443 if (!PotentiallyDeadPHIs.insert(PN))
9446 // Don't scan crazily complex things.
9447 if (PotentiallyDeadPHIs.size() == 16)
9450 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9451 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9456 /// PHIsEqualValue - Return true if this phi node is always equal to
9457 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9458 /// z = some value; x = phi (y, z); y = phi (x, z)
9459 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9460 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9461 // See if we already saw this PHI node.
9462 if (!ValueEqualPHIs.insert(PN))
9465 // Don't scan crazily complex things.
9466 if (ValueEqualPHIs.size() == 16)
9469 // Scan the operands to see if they are either phi nodes or are equal to
9471 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9472 Value *Op = PN->getIncomingValue(i);
9473 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9474 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9476 } else if (Op != NonPhiInVal)
9484 // PHINode simplification
9486 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9487 // If LCSSA is around, don't mess with Phi nodes
9488 if (MustPreserveLCSSA) return 0;
9490 if (Value *V = PN.hasConstantValue())
9491 return ReplaceInstUsesWith(PN, V);
9493 // If all PHI operands are the same operation, pull them through the PHI,
9494 // reducing code size.
9495 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9496 PN.getIncomingValue(0)->hasOneUse())
9497 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9500 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9501 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9502 // PHI)... break the cycle.
9503 if (PN.hasOneUse()) {
9504 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9505 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9506 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9507 PotentiallyDeadPHIs.insert(&PN);
9508 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9509 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9512 // If this phi has a single use, and if that use just computes a value for
9513 // the next iteration of a loop, delete the phi. This occurs with unused
9514 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9515 // common case here is good because the only other things that catch this
9516 // are induction variable analysis (sometimes) and ADCE, which is only run
9518 if (PHIUser->hasOneUse() &&
9519 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9520 PHIUser->use_back() == &PN) {
9521 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9525 // We sometimes end up with phi cycles that non-obviously end up being the
9526 // same value, for example:
9527 // z = some value; x = phi (y, z); y = phi (x, z)
9528 // where the phi nodes don't necessarily need to be in the same block. Do a
9529 // quick check to see if the PHI node only contains a single non-phi value, if
9530 // so, scan to see if the phi cycle is actually equal to that value.
9532 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9533 // Scan for the first non-phi operand.
9534 while (InValNo != NumOperandVals &&
9535 isa<PHINode>(PN.getIncomingValue(InValNo)))
9538 if (InValNo != NumOperandVals) {
9539 Value *NonPhiInVal = PN.getOperand(InValNo);
9541 // Scan the rest of the operands to see if there are any conflicts, if so
9542 // there is no need to recursively scan other phis.
9543 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9544 Value *OpVal = PN.getIncomingValue(InValNo);
9545 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9549 // If we scanned over all operands, then we have one unique value plus
9550 // phi values. Scan PHI nodes to see if they all merge in each other or
9552 if (InValNo == NumOperandVals) {
9553 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9554 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9555 return ReplaceInstUsesWith(PN, NonPhiInVal);
9562 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9563 Instruction *InsertPoint,
9565 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9566 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9567 // We must cast correctly to the pointer type. Ensure that we
9568 // sign extend the integer value if it is smaller as this is
9569 // used for address computation.
9570 Instruction::CastOps opcode =
9571 (VTySize < PtrSize ? Instruction::SExt :
9572 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9573 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9577 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9578 Value *PtrOp = GEP.getOperand(0);
9579 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9580 // If so, eliminate the noop.
9581 if (GEP.getNumOperands() == 1)
9582 return ReplaceInstUsesWith(GEP, PtrOp);
9584 if (isa<UndefValue>(GEP.getOperand(0)))
9585 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9587 bool HasZeroPointerIndex = false;
9588 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9589 HasZeroPointerIndex = C->isNullValue();
9591 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9592 return ReplaceInstUsesWith(GEP, PtrOp);
9594 // Eliminate unneeded casts for indices.
9595 bool MadeChange = false;
9597 gep_type_iterator GTI = gep_type_begin(GEP);
9598 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
9599 i != e; ++i, ++GTI) {
9600 if (isa<SequentialType>(*GTI)) {
9601 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
9602 if (CI->getOpcode() == Instruction::ZExt ||
9603 CI->getOpcode() == Instruction::SExt) {
9604 const Type *SrcTy = CI->getOperand(0)->getType();
9605 // We can eliminate a cast from i32 to i64 iff the target
9606 // is a 32-bit pointer target.
9607 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9609 *i = CI->getOperand(0);
9613 // If we are using a wider index than needed for this platform, shrink it
9614 // to what we need. If the incoming value needs a cast instruction,
9615 // insert it. This explicit cast can make subsequent optimizations more
9618 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9619 if (Constant *C = dyn_cast<Constant>(Op)) {
9620 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
9623 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9631 if (MadeChange) return &GEP;
9633 // If this GEP instruction doesn't move the pointer, and if the input operand
9634 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9635 // real input to the dest type.
9636 if (GEP.hasAllZeroIndices()) {
9637 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9638 // If the bitcast is of an allocation, and the allocation will be
9639 // converted to match the type of the cast, don't touch this.
9640 if (isa<AllocationInst>(BCI->getOperand(0))) {
9641 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9642 if (Instruction *I = visitBitCast(*BCI)) {
9645 BCI->getParent()->getInstList().insert(BCI, I);
9646 ReplaceInstUsesWith(*BCI, I);
9651 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9655 // Combine Indices - If the source pointer to this getelementptr instruction
9656 // is a getelementptr instruction, combine the indices of the two
9657 // getelementptr instructions into a single instruction.
9659 SmallVector<Value*, 8> SrcGEPOperands;
9660 if (User *Src = dyn_castGetElementPtr(PtrOp))
9661 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9663 if (!SrcGEPOperands.empty()) {
9664 // Note that if our source is a gep chain itself that we wait for that
9665 // chain to be resolved before we perform this transformation. This
9666 // avoids us creating a TON of code in some cases.
9668 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9669 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9670 return 0; // Wait until our source is folded to completion.
9672 SmallVector<Value*, 8> Indices;
9674 // Find out whether the last index in the source GEP is a sequential idx.
9675 bool EndsWithSequential = false;
9676 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9677 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9678 EndsWithSequential = !isa<StructType>(*I);
9680 // Can we combine the two pointer arithmetics offsets?
9681 if (EndsWithSequential) {
9682 // Replace: gep (gep %P, long B), long A, ...
9683 // With: T = long A+B; gep %P, T, ...
9685 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9686 if (SO1 == Constant::getNullValue(SO1->getType())) {
9688 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9691 // If they aren't the same type, convert both to an integer of the
9692 // target's pointer size.
9693 if (SO1->getType() != GO1->getType()) {
9694 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9695 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9696 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9697 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9699 unsigned PS = TD->getPointerSizeInBits();
9700 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9701 // Convert GO1 to SO1's type.
9702 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9704 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9705 // Convert SO1 to GO1's type.
9706 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9708 const Type *PT = TD->getIntPtrType();
9709 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9710 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9714 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9715 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9717 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
9718 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9722 // Recycle the GEP we already have if possible.
9723 if (SrcGEPOperands.size() == 2) {
9724 GEP.setOperand(0, SrcGEPOperands[0]);
9725 GEP.setOperand(1, Sum);
9728 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9729 SrcGEPOperands.end()-1);
9730 Indices.push_back(Sum);
9731 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9733 } else if (isa<Constant>(*GEP.idx_begin()) &&
9734 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9735 SrcGEPOperands.size() != 1) {
9736 // Otherwise we can do the fold if the first index of the GEP is a zero
9737 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9738 SrcGEPOperands.end());
9739 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9742 if (!Indices.empty())
9743 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9744 Indices.end(), GEP.getName());
9746 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9747 // GEP of global variable. If all of the indices for this GEP are
9748 // constants, we can promote this to a constexpr instead of an instruction.
9750 // Scan for nonconstants...
9751 SmallVector<Constant*, 8> Indices;
9752 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9753 for (; I != E && isa<Constant>(*I); ++I)
9754 Indices.push_back(cast<Constant>(*I));
9756 if (I == E) { // If they are all constants...
9757 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9758 &Indices[0],Indices.size());
9760 // Replace all uses of the GEP with the new constexpr...
9761 return ReplaceInstUsesWith(GEP, CE);
9763 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9764 if (!isa<PointerType>(X->getType())) {
9765 // Not interesting. Source pointer must be a cast from pointer.
9766 } else if (HasZeroPointerIndex) {
9767 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9768 // into : GEP [10 x i8]* X, i32 0, ...
9770 // This occurs when the program declares an array extern like "int X[];"
9772 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9773 const PointerType *XTy = cast<PointerType>(X->getType());
9774 if (const ArrayType *XATy =
9775 dyn_cast<ArrayType>(XTy->getElementType()))
9776 if (const ArrayType *CATy =
9777 dyn_cast<ArrayType>(CPTy->getElementType()))
9778 if (CATy->getElementType() == XATy->getElementType()) {
9779 // At this point, we know that the cast source type is a pointer
9780 // to an array of the same type as the destination pointer
9781 // array. Because the array type is never stepped over (there
9782 // is a leading zero) we can fold the cast into this GEP.
9783 GEP.setOperand(0, X);
9786 } else if (GEP.getNumOperands() == 2) {
9787 // Transform things like:
9788 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9789 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9790 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9791 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9792 if (isa<ArrayType>(SrcElTy) &&
9793 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9794 TD->getABITypeSize(ResElTy)) {
9796 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9797 Idx[1] = GEP.getOperand(1);
9798 Value *V = InsertNewInstBefore(
9799 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9800 // V and GEP are both pointer types --> BitCast
9801 return new BitCastInst(V, GEP.getType());
9804 // Transform things like:
9805 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9806 // (where tmp = 8*tmp2) into:
9807 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9809 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9810 uint64_t ArrayEltSize =
9811 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9813 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9814 // allow either a mul, shift, or constant here.
9816 ConstantInt *Scale = 0;
9817 if (ArrayEltSize == 1) {
9818 NewIdx = GEP.getOperand(1);
9819 Scale = ConstantInt::get(NewIdx->getType(), 1);
9820 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9821 NewIdx = ConstantInt::get(CI->getType(), 1);
9823 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9824 if (Inst->getOpcode() == Instruction::Shl &&
9825 isa<ConstantInt>(Inst->getOperand(1))) {
9826 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9827 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9828 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9829 NewIdx = Inst->getOperand(0);
9830 } else if (Inst->getOpcode() == Instruction::Mul &&
9831 isa<ConstantInt>(Inst->getOperand(1))) {
9832 Scale = cast<ConstantInt>(Inst->getOperand(1));
9833 NewIdx = Inst->getOperand(0);
9837 // If the index will be to exactly the right offset with the scale taken
9838 // out, perform the transformation. Note, we don't know whether Scale is
9839 // signed or not. We'll use unsigned version of division/modulo
9840 // operation after making sure Scale doesn't have the sign bit set.
9841 if (Scale && Scale->getSExtValue() >= 0LL &&
9842 Scale->getZExtValue() % ArrayEltSize == 0) {
9843 Scale = ConstantInt::get(Scale->getType(),
9844 Scale->getZExtValue() / ArrayEltSize);
9845 if (Scale->getZExtValue() != 1) {
9846 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9848 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
9849 NewIdx = InsertNewInstBefore(Sc, GEP);
9852 // Insert the new GEP instruction.
9854 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9856 Instruction *NewGEP =
9857 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9858 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9859 // The NewGEP must be pointer typed, so must the old one -> BitCast
9860 return new BitCastInst(NewGEP, GEP.getType());
9869 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9870 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9871 if (AI.isArrayAllocation()) { // Check C != 1
9872 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9874 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9875 AllocationInst *New = 0;
9877 // Create and insert the replacement instruction...
9878 if (isa<MallocInst>(AI))
9879 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9881 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9882 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9885 InsertNewInstBefore(New, AI);
9887 // Scan to the end of the allocation instructions, to skip over a block of
9888 // allocas if possible...
9890 BasicBlock::iterator It = New;
9891 while (isa<AllocationInst>(*It)) ++It;
9893 // Now that I is pointing to the first non-allocation-inst in the block,
9894 // insert our getelementptr instruction...
9896 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9900 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9901 New->getName()+".sub", It);
9903 // Now make everything use the getelementptr instead of the original
9905 return ReplaceInstUsesWith(AI, V);
9906 } else if (isa<UndefValue>(AI.getArraySize())) {
9907 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9911 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9912 // Note that we only do this for alloca's, because malloc should allocate and
9913 // return a unique pointer, even for a zero byte allocation.
9914 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9915 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9916 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9921 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9922 Value *Op = FI.getOperand(0);
9924 // free undef -> unreachable.
9925 if (isa<UndefValue>(Op)) {
9926 // Insert a new store to null because we cannot modify the CFG here.
9927 new StoreInst(ConstantInt::getTrue(),
9928 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9929 return EraseInstFromFunction(FI);
9932 // If we have 'free null' delete the instruction. This can happen in stl code
9933 // when lots of inlining happens.
9934 if (isa<ConstantPointerNull>(Op))
9935 return EraseInstFromFunction(FI);
9937 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9938 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9939 FI.setOperand(0, CI->getOperand(0));
9943 // Change free (gep X, 0,0,0,0) into free(X)
9944 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9945 if (GEPI->hasAllZeroIndices()) {
9946 AddToWorkList(GEPI);
9947 FI.setOperand(0, GEPI->getOperand(0));
9952 // Change free(malloc) into nothing, if the malloc has a single use.
9953 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9954 if (MI->hasOneUse()) {
9955 EraseInstFromFunction(FI);
9956 return EraseInstFromFunction(*MI);
9963 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9964 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9965 const TargetData *TD) {
9966 User *CI = cast<User>(LI.getOperand(0));
9967 Value *CastOp = CI->getOperand(0);
9969 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9970 // Instead of loading constant c string, use corresponding integer value
9971 // directly if string length is small enough.
9972 const std::string &Str = CE->getOperand(0)->getStringValue();
9974 unsigned len = Str.length();
9975 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9976 unsigned numBits = Ty->getPrimitiveSizeInBits();
9977 // Replace LI with immediate integer store.
9978 if ((numBits >> 3) == len + 1) {
9979 APInt StrVal(numBits, 0);
9980 APInt SingleChar(numBits, 0);
9981 if (TD->isLittleEndian()) {
9982 for (signed i = len-1; i >= 0; i--) {
9983 SingleChar = (uint64_t) Str[i];
9984 StrVal = (StrVal << 8) | SingleChar;
9987 for (unsigned i = 0; i < len; i++) {
9988 SingleChar = (uint64_t) Str[i];
9989 StrVal = (StrVal << 8) | SingleChar;
9991 // Append NULL at the end.
9993 StrVal = (StrVal << 8) | SingleChar;
9995 Value *NL = ConstantInt::get(StrVal);
9996 return IC.ReplaceInstUsesWith(LI, NL);
10001 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10002 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10003 const Type *SrcPTy = SrcTy->getElementType();
10005 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10006 isa<VectorType>(DestPTy)) {
10007 // If the source is an array, the code below will not succeed. Check to
10008 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10010 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10011 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10012 if (ASrcTy->getNumElements() != 0) {
10014 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10015 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10016 SrcTy = cast<PointerType>(CastOp->getType());
10017 SrcPTy = SrcTy->getElementType();
10020 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10021 isa<VectorType>(SrcPTy)) &&
10022 // Do not allow turning this into a load of an integer, which is then
10023 // casted to a pointer, this pessimizes pointer analysis a lot.
10024 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10025 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10026 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10028 // Okay, we are casting from one integer or pointer type to another of
10029 // the same size. Instead of casting the pointer before the load, cast
10030 // the result of the loaded value.
10031 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10033 LI.isVolatile()),LI);
10034 // Now cast the result of the load.
10035 return new BitCastInst(NewLoad, LI.getType());
10042 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10043 /// from this value cannot trap. If it is not obviously safe to load from the
10044 /// specified pointer, we do a quick local scan of the basic block containing
10045 /// ScanFrom, to determine if the address is already accessed.
10046 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10047 // If it is an alloca it is always safe to load from.
10048 if (isa<AllocaInst>(V)) return true;
10050 // If it is a global variable it is mostly safe to load from.
10051 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10052 // Don't try to evaluate aliases. External weak GV can be null.
10053 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10055 // Otherwise, be a little bit agressive by scanning the local block where we
10056 // want to check to see if the pointer is already being loaded or stored
10057 // from/to. If so, the previous load or store would have already trapped,
10058 // so there is no harm doing an extra load (also, CSE will later eliminate
10059 // the load entirely).
10060 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10065 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10066 if (LI->getOperand(0) == V) return true;
10067 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10068 if (SI->getOperand(1) == V) return true;
10074 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10075 /// until we find the underlying object a pointer is referring to or something
10076 /// we don't understand. Note that the returned pointer may be offset from the
10077 /// input, because we ignore GEP indices.
10078 static Value *GetUnderlyingObject(Value *Ptr) {
10080 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10081 if (CE->getOpcode() == Instruction::BitCast ||
10082 CE->getOpcode() == Instruction::GetElementPtr)
10083 Ptr = CE->getOperand(0);
10086 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10087 Ptr = BCI->getOperand(0);
10088 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10089 Ptr = GEP->getOperand(0);
10096 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10097 Value *Op = LI.getOperand(0);
10099 // Attempt to improve the alignment.
10100 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10102 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10103 LI.getAlignment()))
10104 LI.setAlignment(KnownAlign);
10106 // load (cast X) --> cast (load X) iff safe
10107 if (isa<CastInst>(Op))
10108 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10111 // None of the following transforms are legal for volatile loads.
10112 if (LI.isVolatile()) return 0;
10114 if (&LI.getParent()->front() != &LI) {
10115 BasicBlock::iterator BBI = &LI; --BBI;
10116 // If the instruction immediately before this is a store to the same
10117 // address, do a simple form of store->load forwarding.
10118 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10119 if (SI->getOperand(1) == LI.getOperand(0))
10120 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10121 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10122 if (LIB->getOperand(0) == LI.getOperand(0))
10123 return ReplaceInstUsesWith(LI, LIB);
10126 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10127 const Value *GEPI0 = GEPI->getOperand(0);
10128 // TODO: Consider a target hook for valid address spaces for this xform.
10129 if (isa<ConstantPointerNull>(GEPI0) &&
10130 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10131 // Insert a new store to null instruction before the load to indicate
10132 // that this code is not reachable. We do this instead of inserting
10133 // an unreachable instruction directly because we cannot modify the
10135 new StoreInst(UndefValue::get(LI.getType()),
10136 Constant::getNullValue(Op->getType()), &LI);
10137 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10141 if (Constant *C = dyn_cast<Constant>(Op)) {
10142 // load null/undef -> undef
10143 // TODO: Consider a target hook for valid address spaces for this xform.
10144 if (isa<UndefValue>(C) || (C->isNullValue() &&
10145 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10146 // Insert a new store to null instruction before the load to indicate that
10147 // this code is not reachable. We do this instead of inserting an
10148 // unreachable instruction directly because we cannot modify the CFG.
10149 new StoreInst(UndefValue::get(LI.getType()),
10150 Constant::getNullValue(Op->getType()), &LI);
10151 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10154 // Instcombine load (constant global) into the value loaded.
10155 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10156 if (GV->isConstant() && !GV->isDeclaration())
10157 return ReplaceInstUsesWith(LI, GV->getInitializer());
10159 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10160 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10161 if (CE->getOpcode() == Instruction::GetElementPtr) {
10162 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10163 if (GV->isConstant() && !GV->isDeclaration())
10165 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10166 return ReplaceInstUsesWith(LI, V);
10167 if (CE->getOperand(0)->isNullValue()) {
10168 // Insert a new store to null instruction before the load to indicate
10169 // that this code is not reachable. We do this instead of inserting
10170 // an unreachable instruction directly because we cannot modify the
10172 new StoreInst(UndefValue::get(LI.getType()),
10173 Constant::getNullValue(Op->getType()), &LI);
10174 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10177 } else if (CE->isCast()) {
10178 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10184 // If this load comes from anywhere in a constant global, and if the global
10185 // is all undef or zero, we know what it loads.
10186 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10187 if (GV->isConstant() && GV->hasInitializer()) {
10188 if (GV->getInitializer()->isNullValue())
10189 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10190 else if (isa<UndefValue>(GV->getInitializer()))
10191 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10195 if (Op->hasOneUse()) {
10196 // Change select and PHI nodes to select values instead of addresses: this
10197 // helps alias analysis out a lot, allows many others simplifications, and
10198 // exposes redundancy in the code.
10200 // Note that we cannot do the transformation unless we know that the
10201 // introduced loads cannot trap! Something like this is valid as long as
10202 // the condition is always false: load (select bool %C, int* null, int* %G),
10203 // but it would not be valid if we transformed it to load from null
10204 // unconditionally.
10206 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10207 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10208 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10209 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10210 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10211 SI->getOperand(1)->getName()+".val"), LI);
10212 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10213 SI->getOperand(2)->getName()+".val"), LI);
10214 return SelectInst::Create(SI->getCondition(), V1, V2);
10217 // load (select (cond, null, P)) -> load P
10218 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10219 if (C->isNullValue()) {
10220 LI.setOperand(0, SI->getOperand(2));
10224 // load (select (cond, P, null)) -> load P
10225 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10226 if (C->isNullValue()) {
10227 LI.setOperand(0, SI->getOperand(1));
10235 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10237 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10238 User *CI = cast<User>(SI.getOperand(1));
10239 Value *CastOp = CI->getOperand(0);
10241 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10242 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10243 const Type *SrcPTy = SrcTy->getElementType();
10245 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10246 // If the source is an array, the code below will not succeed. Check to
10247 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10249 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10250 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10251 if (ASrcTy->getNumElements() != 0) {
10253 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10254 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10255 SrcTy = cast<PointerType>(CastOp->getType());
10256 SrcPTy = SrcTy->getElementType();
10259 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10260 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10261 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10263 // Okay, we are casting from one integer or pointer type to another of
10264 // the same size. Instead of casting the pointer before
10265 // the store, cast the value to be stored.
10267 Value *SIOp0 = SI.getOperand(0);
10268 Instruction::CastOps opcode = Instruction::BitCast;
10269 const Type* CastSrcTy = SIOp0->getType();
10270 const Type* CastDstTy = SrcPTy;
10271 if (isa<PointerType>(CastDstTy)) {
10272 if (CastSrcTy->isInteger())
10273 opcode = Instruction::IntToPtr;
10274 } else if (isa<IntegerType>(CastDstTy)) {
10275 if (isa<PointerType>(SIOp0->getType()))
10276 opcode = Instruction::PtrToInt;
10278 if (Constant *C = dyn_cast<Constant>(SIOp0))
10279 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10281 NewCast = IC.InsertNewInstBefore(
10282 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10284 return new StoreInst(NewCast, CastOp);
10291 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10292 Value *Val = SI.getOperand(0);
10293 Value *Ptr = SI.getOperand(1);
10295 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10296 EraseInstFromFunction(SI);
10301 // If the RHS is an alloca with a single use, zapify the store, making the
10303 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10304 if (isa<AllocaInst>(Ptr)) {
10305 EraseInstFromFunction(SI);
10310 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10311 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10312 GEP->getOperand(0)->hasOneUse()) {
10313 EraseInstFromFunction(SI);
10319 // Attempt to improve the alignment.
10320 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10322 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10323 SI.getAlignment()))
10324 SI.setAlignment(KnownAlign);
10326 // Do really simple DSE, to catch cases where there are several consequtive
10327 // stores to the same location, separated by a few arithmetic operations. This
10328 // situation often occurs with bitfield accesses.
10329 BasicBlock::iterator BBI = &SI;
10330 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10334 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10335 // Prev store isn't volatile, and stores to the same location?
10336 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10339 EraseInstFromFunction(*PrevSI);
10345 // If this is a load, we have to stop. However, if the loaded value is from
10346 // the pointer we're loading and is producing the pointer we're storing,
10347 // then *this* store is dead (X = load P; store X -> P).
10348 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10349 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10350 EraseInstFromFunction(SI);
10354 // Otherwise, this is a load from some other location. Stores before it
10355 // may not be dead.
10359 // Don't skip over loads or things that can modify memory.
10360 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10365 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10367 // store X, null -> turns into 'unreachable' in SimplifyCFG
10368 if (isa<ConstantPointerNull>(Ptr)) {
10369 if (!isa<UndefValue>(Val)) {
10370 SI.setOperand(0, UndefValue::get(Val->getType()));
10371 if (Instruction *U = dyn_cast<Instruction>(Val))
10372 AddToWorkList(U); // Dropped a use.
10375 return 0; // Do not modify these!
10378 // store undef, Ptr -> noop
10379 if (isa<UndefValue>(Val)) {
10380 EraseInstFromFunction(SI);
10385 // If the pointer destination is a cast, see if we can fold the cast into the
10387 if (isa<CastInst>(Ptr))
10388 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10390 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10392 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10396 // If this store is the last instruction in the basic block, and if the block
10397 // ends with an unconditional branch, try to move it to the successor block.
10399 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10400 if (BI->isUnconditional())
10401 if (SimplifyStoreAtEndOfBlock(SI))
10402 return 0; // xform done!
10407 /// SimplifyStoreAtEndOfBlock - Turn things like:
10408 /// if () { *P = v1; } else { *P = v2 }
10409 /// into a phi node with a store in the successor.
10411 /// Simplify things like:
10412 /// *P = v1; if () { *P = v2; }
10413 /// into a phi node with a store in the successor.
10415 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10416 BasicBlock *StoreBB = SI.getParent();
10418 // Check to see if the successor block has exactly two incoming edges. If
10419 // so, see if the other predecessor contains a store to the same location.
10420 // if so, insert a PHI node (if needed) and move the stores down.
10421 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10423 // Determine whether Dest has exactly two predecessors and, if so, compute
10424 // the other predecessor.
10425 pred_iterator PI = pred_begin(DestBB);
10426 BasicBlock *OtherBB = 0;
10427 if (*PI != StoreBB)
10430 if (PI == pred_end(DestBB))
10433 if (*PI != StoreBB) {
10438 if (++PI != pred_end(DestBB))
10441 // Bail out if all the relevant blocks aren't distinct (this can happen,
10442 // for example, if SI is in an infinite loop)
10443 if (StoreBB == DestBB || OtherBB == DestBB)
10446 // Verify that the other block ends in a branch and is not otherwise empty.
10447 BasicBlock::iterator BBI = OtherBB->getTerminator();
10448 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10449 if (!OtherBr || BBI == OtherBB->begin())
10452 // If the other block ends in an unconditional branch, check for the 'if then
10453 // else' case. there is an instruction before the branch.
10454 StoreInst *OtherStore = 0;
10455 if (OtherBr->isUnconditional()) {
10456 // If this isn't a store, or isn't a store to the same location, bail out.
10458 OtherStore = dyn_cast<StoreInst>(BBI);
10459 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10462 // Otherwise, the other block ended with a conditional branch. If one of the
10463 // destinations is StoreBB, then we have the if/then case.
10464 if (OtherBr->getSuccessor(0) != StoreBB &&
10465 OtherBr->getSuccessor(1) != StoreBB)
10468 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10469 // if/then triangle. See if there is a store to the same ptr as SI that
10470 // lives in OtherBB.
10472 // Check to see if we find the matching store.
10473 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10474 if (OtherStore->getOperand(1) != SI.getOperand(1))
10478 // If we find something that may be using or overwriting the stored
10479 // value, or if we run out of instructions, we can't do the xform.
10480 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
10481 BBI == OtherBB->begin())
10485 // In order to eliminate the store in OtherBr, we have to
10486 // make sure nothing reads or overwrites the stored value in
10488 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10489 // FIXME: This should really be AA driven.
10490 if (I->mayReadFromMemory() || I->mayWriteToMemory())
10495 // Insert a PHI node now if we need it.
10496 Value *MergedVal = OtherStore->getOperand(0);
10497 if (MergedVal != SI.getOperand(0)) {
10498 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10499 PN->reserveOperandSpace(2);
10500 PN->addIncoming(SI.getOperand(0), SI.getParent());
10501 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10502 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10505 // Advance to a place where it is safe to insert the new store and
10507 BBI = DestBB->getFirstNonPHI();
10508 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10509 OtherStore->isVolatile()), *BBI);
10511 // Nuke the old stores.
10512 EraseInstFromFunction(SI);
10513 EraseInstFromFunction(*OtherStore);
10519 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10520 // Change br (not X), label True, label False to: br X, label False, True
10522 BasicBlock *TrueDest;
10523 BasicBlock *FalseDest;
10524 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10525 !isa<Constant>(X)) {
10526 // Swap Destinations and condition...
10527 BI.setCondition(X);
10528 BI.setSuccessor(0, FalseDest);
10529 BI.setSuccessor(1, TrueDest);
10533 // Cannonicalize fcmp_one -> fcmp_oeq
10534 FCmpInst::Predicate FPred; Value *Y;
10535 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10536 TrueDest, FalseDest)))
10537 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10538 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10539 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10540 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10541 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10542 NewSCC->takeName(I);
10543 // Swap Destinations and condition...
10544 BI.setCondition(NewSCC);
10545 BI.setSuccessor(0, FalseDest);
10546 BI.setSuccessor(1, TrueDest);
10547 RemoveFromWorkList(I);
10548 I->eraseFromParent();
10549 AddToWorkList(NewSCC);
10553 // Cannonicalize icmp_ne -> icmp_eq
10554 ICmpInst::Predicate IPred;
10555 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10556 TrueDest, FalseDest)))
10557 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10558 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10559 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10560 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10561 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10562 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10563 NewSCC->takeName(I);
10564 // Swap Destinations and condition...
10565 BI.setCondition(NewSCC);
10566 BI.setSuccessor(0, FalseDest);
10567 BI.setSuccessor(1, TrueDest);
10568 RemoveFromWorkList(I);
10569 I->eraseFromParent();;
10570 AddToWorkList(NewSCC);
10577 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10578 Value *Cond = SI.getCondition();
10579 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10580 if (I->getOpcode() == Instruction::Add)
10581 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10582 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10583 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10584 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10586 SI.setOperand(0, I->getOperand(0));
10594 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
10595 // See if we are trying to extract a known value. If so, use that instead.
10596 if (Value *Elt = FindInsertedValue(EV.getOperand(0), EV.idx_begin(),
10597 EV.idx_end(), &EV))
10598 return ReplaceInstUsesWith(EV, Elt);
10604 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10605 /// is to leave as a vector operation.
10606 static bool CheapToScalarize(Value *V, bool isConstant) {
10607 if (isa<ConstantAggregateZero>(V))
10609 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10610 if (isConstant) return true;
10611 // If all elts are the same, we can extract.
10612 Constant *Op0 = C->getOperand(0);
10613 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10614 if (C->getOperand(i) != Op0)
10618 Instruction *I = dyn_cast<Instruction>(V);
10619 if (!I) return false;
10621 // Insert element gets simplified to the inserted element or is deleted if
10622 // this is constant idx extract element and its a constant idx insertelt.
10623 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10624 isa<ConstantInt>(I->getOperand(2)))
10626 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10628 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10629 if (BO->hasOneUse() &&
10630 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10631 CheapToScalarize(BO->getOperand(1), isConstant)))
10633 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10634 if (CI->hasOneUse() &&
10635 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10636 CheapToScalarize(CI->getOperand(1), isConstant)))
10642 /// Read and decode a shufflevector mask.
10644 /// It turns undef elements into values that are larger than the number of
10645 /// elements in the input.
10646 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10647 unsigned NElts = SVI->getType()->getNumElements();
10648 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10649 return std::vector<unsigned>(NElts, 0);
10650 if (isa<UndefValue>(SVI->getOperand(2)))
10651 return std::vector<unsigned>(NElts, 2*NElts);
10653 std::vector<unsigned> Result;
10654 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10655 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
10656 if (isa<UndefValue>(*i))
10657 Result.push_back(NElts*2); // undef -> 8
10659 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
10663 /// FindScalarElement - Given a vector and an element number, see if the scalar
10664 /// value is already around as a register, for example if it were inserted then
10665 /// extracted from the vector.
10666 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10667 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10668 const VectorType *PTy = cast<VectorType>(V->getType());
10669 unsigned Width = PTy->getNumElements();
10670 if (EltNo >= Width) // Out of range access.
10671 return UndefValue::get(PTy->getElementType());
10673 if (isa<UndefValue>(V))
10674 return UndefValue::get(PTy->getElementType());
10675 else if (isa<ConstantAggregateZero>(V))
10676 return Constant::getNullValue(PTy->getElementType());
10677 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10678 return CP->getOperand(EltNo);
10679 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10680 // If this is an insert to a variable element, we don't know what it is.
10681 if (!isa<ConstantInt>(III->getOperand(2)))
10683 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10685 // If this is an insert to the element we are looking for, return the
10687 if (EltNo == IIElt)
10688 return III->getOperand(1);
10690 // Otherwise, the insertelement doesn't modify the value, recurse on its
10692 return FindScalarElement(III->getOperand(0), EltNo);
10693 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10694 unsigned InEl = getShuffleMask(SVI)[EltNo];
10696 return FindScalarElement(SVI->getOperand(0), InEl);
10697 else if (InEl < Width*2)
10698 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10700 return UndefValue::get(PTy->getElementType());
10703 // Otherwise, we don't know.
10707 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10708 // If vector val is undef, replace extract with scalar undef.
10709 if (isa<UndefValue>(EI.getOperand(0)))
10710 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10712 // If vector val is constant 0, replace extract with scalar 0.
10713 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10714 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10716 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10717 // If vector val is constant with all elements the same, replace EI with
10718 // that element. When the elements are not identical, we cannot replace yet
10719 // (we do that below, but only when the index is constant).
10720 Constant *op0 = C->getOperand(0);
10721 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10722 if (C->getOperand(i) != op0) {
10727 return ReplaceInstUsesWith(EI, op0);
10730 // If extracting a specified index from the vector, see if we can recursively
10731 // find a previously computed scalar that was inserted into the vector.
10732 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10733 unsigned IndexVal = IdxC->getZExtValue();
10734 unsigned VectorWidth =
10735 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10737 // If this is extracting an invalid index, turn this into undef, to avoid
10738 // crashing the code below.
10739 if (IndexVal >= VectorWidth)
10740 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10742 // This instruction only demands the single element from the input vector.
10743 // If the input vector has a single use, simplify it based on this use
10745 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10746 uint64_t UndefElts;
10747 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10750 EI.setOperand(0, V);
10755 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10756 return ReplaceInstUsesWith(EI, Elt);
10758 // If the this extractelement is directly using a bitcast from a vector of
10759 // the same number of elements, see if we can find the source element from
10760 // it. In this case, we will end up needing to bitcast the scalars.
10761 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10762 if (const VectorType *VT =
10763 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10764 if (VT->getNumElements() == VectorWidth)
10765 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10766 return new BitCastInst(Elt, EI.getType());
10770 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10771 if (I->hasOneUse()) {
10772 // Push extractelement into predecessor operation if legal and
10773 // profitable to do so
10774 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10775 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10776 if (CheapToScalarize(BO, isConstantElt)) {
10777 ExtractElementInst *newEI0 =
10778 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10779 EI.getName()+".lhs");
10780 ExtractElementInst *newEI1 =
10781 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10782 EI.getName()+".rhs");
10783 InsertNewInstBefore(newEI0, EI);
10784 InsertNewInstBefore(newEI1, EI);
10785 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
10787 } else if (isa<LoadInst>(I)) {
10789 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10790 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10791 PointerType::get(EI.getType(), AS),EI);
10792 GetElementPtrInst *GEP =
10793 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
10794 InsertNewInstBefore(GEP, EI);
10795 return new LoadInst(GEP);
10798 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10799 // Extracting the inserted element?
10800 if (IE->getOperand(2) == EI.getOperand(1))
10801 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10802 // If the inserted and extracted elements are constants, they must not
10803 // be the same value, extract from the pre-inserted value instead.
10804 if (isa<Constant>(IE->getOperand(2)) &&
10805 isa<Constant>(EI.getOperand(1))) {
10806 AddUsesToWorkList(EI);
10807 EI.setOperand(0, IE->getOperand(0));
10810 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10811 // If this is extracting an element from a shufflevector, figure out where
10812 // it came from and extract from the appropriate input element instead.
10813 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10814 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10816 if (SrcIdx < SVI->getType()->getNumElements())
10817 Src = SVI->getOperand(0);
10818 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10819 SrcIdx -= SVI->getType()->getNumElements();
10820 Src = SVI->getOperand(1);
10822 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10824 return new ExtractElementInst(Src, SrcIdx);
10831 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10832 /// elements from either LHS or RHS, return the shuffle mask and true.
10833 /// Otherwise, return false.
10834 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10835 std::vector<Constant*> &Mask) {
10836 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10837 "Invalid CollectSingleShuffleElements");
10838 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10840 if (isa<UndefValue>(V)) {
10841 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10843 } else if (V == LHS) {
10844 for (unsigned i = 0; i != NumElts; ++i)
10845 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10847 } else if (V == RHS) {
10848 for (unsigned i = 0; i != NumElts; ++i)
10849 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10851 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10852 // If this is an insert of an extract from some other vector, include it.
10853 Value *VecOp = IEI->getOperand(0);
10854 Value *ScalarOp = IEI->getOperand(1);
10855 Value *IdxOp = IEI->getOperand(2);
10857 if (!isa<ConstantInt>(IdxOp))
10859 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10861 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10862 // Okay, we can handle this if the vector we are insertinting into is
10863 // transitively ok.
10864 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10865 // If so, update the mask to reflect the inserted undef.
10866 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10869 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10870 if (isa<ConstantInt>(EI->getOperand(1)) &&
10871 EI->getOperand(0)->getType() == V->getType()) {
10872 unsigned ExtractedIdx =
10873 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10875 // This must be extracting from either LHS or RHS.
10876 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10877 // Okay, we can handle this if the vector we are insertinting into is
10878 // transitively ok.
10879 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10880 // If so, update the mask to reflect the inserted value.
10881 if (EI->getOperand(0) == LHS) {
10882 Mask[InsertedIdx & (NumElts-1)] =
10883 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10885 assert(EI->getOperand(0) == RHS);
10886 Mask[InsertedIdx & (NumElts-1)] =
10887 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10896 // TODO: Handle shufflevector here!
10901 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10902 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10903 /// that computes V and the LHS value of the shuffle.
10904 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10906 assert(isa<VectorType>(V->getType()) &&
10907 (RHS == 0 || V->getType() == RHS->getType()) &&
10908 "Invalid shuffle!");
10909 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10911 if (isa<UndefValue>(V)) {
10912 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10914 } else if (isa<ConstantAggregateZero>(V)) {
10915 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10917 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10918 // If this is an insert of an extract from some other vector, include it.
10919 Value *VecOp = IEI->getOperand(0);
10920 Value *ScalarOp = IEI->getOperand(1);
10921 Value *IdxOp = IEI->getOperand(2);
10923 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10924 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10925 EI->getOperand(0)->getType() == V->getType()) {
10926 unsigned ExtractedIdx =
10927 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10928 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10930 // Either the extracted from or inserted into vector must be RHSVec,
10931 // otherwise we'd end up with a shuffle of three inputs.
10932 if (EI->getOperand(0) == RHS || RHS == 0) {
10933 RHS = EI->getOperand(0);
10934 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10935 Mask[InsertedIdx & (NumElts-1)] =
10936 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10940 if (VecOp == RHS) {
10941 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10942 // Everything but the extracted element is replaced with the RHS.
10943 for (unsigned i = 0; i != NumElts; ++i) {
10944 if (i != InsertedIdx)
10945 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10950 // If this insertelement is a chain that comes from exactly these two
10951 // vectors, return the vector and the effective shuffle.
10952 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10953 return EI->getOperand(0);
10958 // TODO: Handle shufflevector here!
10960 // Otherwise, can't do anything fancy. Return an identity vector.
10961 for (unsigned i = 0; i != NumElts; ++i)
10962 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10966 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10967 Value *VecOp = IE.getOperand(0);
10968 Value *ScalarOp = IE.getOperand(1);
10969 Value *IdxOp = IE.getOperand(2);
10971 // Inserting an undef or into an undefined place, remove this.
10972 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10973 ReplaceInstUsesWith(IE, VecOp);
10975 // If the inserted element was extracted from some other vector, and if the
10976 // indexes are constant, try to turn this into a shufflevector operation.
10977 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10978 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10979 EI->getOperand(0)->getType() == IE.getType()) {
10980 unsigned NumVectorElts = IE.getType()->getNumElements();
10981 unsigned ExtractedIdx =
10982 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10983 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10985 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10986 return ReplaceInstUsesWith(IE, VecOp);
10988 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10989 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10991 // If we are extracting a value from a vector, then inserting it right
10992 // back into the same place, just use the input vector.
10993 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10994 return ReplaceInstUsesWith(IE, VecOp);
10996 // We could theoretically do this for ANY input. However, doing so could
10997 // turn chains of insertelement instructions into a chain of shufflevector
10998 // instructions, and right now we do not merge shufflevectors. As such,
10999 // only do this in a situation where it is clear that there is benefit.
11000 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11001 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11002 // the values of VecOp, except then one read from EIOp0.
11003 // Build a new shuffle mask.
11004 std::vector<Constant*> Mask;
11005 if (isa<UndefValue>(VecOp))
11006 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11008 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11009 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11012 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11013 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11014 ConstantVector::get(Mask));
11017 // If this insertelement isn't used by some other insertelement, turn it
11018 // (and any insertelements it points to), into one big shuffle.
11019 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11020 std::vector<Constant*> Mask;
11022 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11023 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11024 // We now have a shuffle of LHS, RHS, Mask.
11025 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11034 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11035 Value *LHS = SVI.getOperand(0);
11036 Value *RHS = SVI.getOperand(1);
11037 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11039 bool MadeChange = false;
11041 // Undefined shuffle mask -> undefined value.
11042 if (isa<UndefValue>(SVI.getOperand(2)))
11043 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11045 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11046 // the undef, change them to undefs.
11047 if (isa<UndefValue>(SVI.getOperand(1))) {
11048 // Scan to see if there are any references to the RHS. If so, replace them
11049 // with undef element refs and set MadeChange to true.
11050 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11051 if (Mask[i] >= e && Mask[i] != 2*e) {
11058 // Remap any references to RHS to use LHS.
11059 std::vector<Constant*> Elts;
11060 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11061 if (Mask[i] == 2*e)
11062 Elts.push_back(UndefValue::get(Type::Int32Ty));
11064 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11066 SVI.setOperand(2, ConstantVector::get(Elts));
11070 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11071 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11072 if (LHS == RHS || isa<UndefValue>(LHS)) {
11073 if (isa<UndefValue>(LHS) && LHS == RHS) {
11074 // shuffle(undef,undef,mask) -> undef.
11075 return ReplaceInstUsesWith(SVI, LHS);
11078 // Remap any references to RHS to use LHS.
11079 std::vector<Constant*> Elts;
11080 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11081 if (Mask[i] >= 2*e)
11082 Elts.push_back(UndefValue::get(Type::Int32Ty));
11084 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11085 (Mask[i] < e && isa<UndefValue>(LHS)))
11086 Mask[i] = 2*e; // Turn into undef.
11088 Mask[i] &= (e-1); // Force to LHS.
11089 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11092 SVI.setOperand(0, SVI.getOperand(1));
11093 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11094 SVI.setOperand(2, ConstantVector::get(Elts));
11095 LHS = SVI.getOperand(0);
11096 RHS = SVI.getOperand(1);
11100 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11101 bool isLHSID = true, isRHSID = true;
11103 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11104 if (Mask[i] >= e*2) continue; // Ignore undef values.
11105 // Is this an identity shuffle of the LHS value?
11106 isLHSID &= (Mask[i] == i);
11108 // Is this an identity shuffle of the RHS value?
11109 isRHSID &= (Mask[i]-e == i);
11112 // Eliminate identity shuffles.
11113 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11114 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11116 // If the LHS is a shufflevector itself, see if we can combine it with this
11117 // one without producing an unusual shuffle. Here we are really conservative:
11118 // we are absolutely afraid of producing a shuffle mask not in the input
11119 // program, because the code gen may not be smart enough to turn a merged
11120 // shuffle into two specific shuffles: it may produce worse code. As such,
11121 // we only merge two shuffles if the result is one of the two input shuffle
11122 // masks. In this case, merging the shuffles just removes one instruction,
11123 // which we know is safe. This is good for things like turning:
11124 // (splat(splat)) -> splat.
11125 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11126 if (isa<UndefValue>(RHS)) {
11127 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11129 std::vector<unsigned> NewMask;
11130 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11131 if (Mask[i] >= 2*e)
11132 NewMask.push_back(2*e);
11134 NewMask.push_back(LHSMask[Mask[i]]);
11136 // If the result mask is equal to the src shuffle or this shuffle mask, do
11137 // the replacement.
11138 if (NewMask == LHSMask || NewMask == Mask) {
11139 std::vector<Constant*> Elts;
11140 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11141 if (NewMask[i] >= e*2) {
11142 Elts.push_back(UndefValue::get(Type::Int32Ty));
11144 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11147 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11148 LHSSVI->getOperand(1),
11149 ConstantVector::get(Elts));
11154 return MadeChange ? &SVI : 0;
11160 /// TryToSinkInstruction - Try to move the specified instruction from its
11161 /// current block into the beginning of DestBlock, which can only happen if it's
11162 /// safe to move the instruction past all of the instructions between it and the
11163 /// end of its block.
11164 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11165 assert(I->hasOneUse() && "Invariants didn't hold!");
11167 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11168 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11171 // Do not sink alloca instructions out of the entry block.
11172 if (isa<AllocaInst>(I) && I->getParent() ==
11173 &DestBlock->getParent()->getEntryBlock())
11176 // We can only sink load instructions if there is nothing between the load and
11177 // the end of block that could change the value.
11178 if (I->mayReadFromMemory()) {
11179 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11181 if (Scan->mayWriteToMemory())
11185 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11187 I->moveBefore(InsertPos);
11193 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11194 /// all reachable code to the worklist.
11196 /// This has a couple of tricks to make the code faster and more powerful. In
11197 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11198 /// them to the worklist (this significantly speeds up instcombine on code where
11199 /// many instructions are dead or constant). Additionally, if we find a branch
11200 /// whose condition is a known constant, we only visit the reachable successors.
11202 static void AddReachableCodeToWorklist(BasicBlock *BB,
11203 SmallPtrSet<BasicBlock*, 64> &Visited,
11205 const TargetData *TD) {
11206 std::vector<BasicBlock*> Worklist;
11207 Worklist.push_back(BB);
11209 while (!Worklist.empty()) {
11210 BB = Worklist.back();
11211 Worklist.pop_back();
11213 // We have now visited this block! If we've already been here, ignore it.
11214 if (!Visited.insert(BB)) continue;
11216 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11217 Instruction *Inst = BBI++;
11219 // DCE instruction if trivially dead.
11220 if (isInstructionTriviallyDead(Inst)) {
11222 DOUT << "IC: DCE: " << *Inst;
11223 Inst->eraseFromParent();
11227 // ConstantProp instruction if trivially constant.
11228 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11229 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11230 Inst->replaceAllUsesWith(C);
11232 Inst->eraseFromParent();
11236 IC.AddToWorkList(Inst);
11239 // Recursively visit successors. If this is a branch or switch on a
11240 // constant, only visit the reachable successor.
11241 TerminatorInst *TI = BB->getTerminator();
11242 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11243 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11244 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11245 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11246 Worklist.push_back(ReachableBB);
11249 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11250 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11251 // See if this is an explicit destination.
11252 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11253 if (SI->getCaseValue(i) == Cond) {
11254 BasicBlock *ReachableBB = SI->getSuccessor(i);
11255 Worklist.push_back(ReachableBB);
11259 // Otherwise it is the default destination.
11260 Worklist.push_back(SI->getSuccessor(0));
11265 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11266 Worklist.push_back(TI->getSuccessor(i));
11270 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11271 bool Changed = false;
11272 TD = &getAnalysis<TargetData>();
11274 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11275 << F.getNameStr() << "\n");
11278 // Do a depth-first traversal of the function, populate the worklist with
11279 // the reachable instructions. Ignore blocks that are not reachable. Keep
11280 // track of which blocks we visit.
11281 SmallPtrSet<BasicBlock*, 64> Visited;
11282 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11284 // Do a quick scan over the function. If we find any blocks that are
11285 // unreachable, remove any instructions inside of them. This prevents
11286 // the instcombine code from having to deal with some bad special cases.
11287 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11288 if (!Visited.count(BB)) {
11289 Instruction *Term = BB->getTerminator();
11290 while (Term != BB->begin()) { // Remove instrs bottom-up
11291 BasicBlock::iterator I = Term; --I;
11293 DOUT << "IC: DCE: " << *I;
11296 if (!I->use_empty())
11297 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11298 I->eraseFromParent();
11303 while (!Worklist.empty()) {
11304 Instruction *I = RemoveOneFromWorkList();
11305 if (I == 0) continue; // skip null values.
11307 // Check to see if we can DCE the instruction.
11308 if (isInstructionTriviallyDead(I)) {
11309 // Add operands to the worklist.
11310 if (I->getNumOperands() < 4)
11311 AddUsesToWorkList(*I);
11314 DOUT << "IC: DCE: " << *I;
11316 I->eraseFromParent();
11317 RemoveFromWorkList(I);
11321 // Instruction isn't dead, see if we can constant propagate it.
11322 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11323 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11325 // Add operands to the worklist.
11326 AddUsesToWorkList(*I);
11327 ReplaceInstUsesWith(*I, C);
11330 I->eraseFromParent();
11331 RemoveFromWorkList(I);
11335 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
11336 // See if we can constant fold its operands.
11337 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
11338 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
11339 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
11345 // See if we can trivially sink this instruction to a successor basic block.
11346 // FIXME: Remove GetResultInst test when first class support for aggregates
11348 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11349 BasicBlock *BB = I->getParent();
11350 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11351 if (UserParent != BB) {
11352 bool UserIsSuccessor = false;
11353 // See if the user is one of our successors.
11354 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11355 if (*SI == UserParent) {
11356 UserIsSuccessor = true;
11360 // If the user is one of our immediate successors, and if that successor
11361 // only has us as a predecessors (we'd have to split the critical edge
11362 // otherwise), we can keep going.
11363 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11364 next(pred_begin(UserParent)) == pred_end(UserParent))
11365 // Okay, the CFG is simple enough, try to sink this instruction.
11366 Changed |= TryToSinkInstruction(I, UserParent);
11370 // Now that we have an instruction, try combining it to simplify it...
11374 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11375 if (Instruction *Result = visit(*I)) {
11377 // Should we replace the old instruction with a new one?
11379 DOUT << "IC: Old = " << *I
11380 << " New = " << *Result;
11382 // Everything uses the new instruction now.
11383 I->replaceAllUsesWith(Result);
11385 // Push the new instruction and any users onto the worklist.
11386 AddToWorkList(Result);
11387 AddUsersToWorkList(*Result);
11389 // Move the name to the new instruction first.
11390 Result->takeName(I);
11392 // Insert the new instruction into the basic block...
11393 BasicBlock *InstParent = I->getParent();
11394 BasicBlock::iterator InsertPos = I;
11396 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11397 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11400 InstParent->getInstList().insert(InsertPos, Result);
11402 // Make sure that we reprocess all operands now that we reduced their
11404 AddUsesToWorkList(*I);
11406 // Instructions can end up on the worklist more than once. Make sure
11407 // we do not process an instruction that has been deleted.
11408 RemoveFromWorkList(I);
11410 // Erase the old instruction.
11411 InstParent->getInstList().erase(I);
11414 DOUT << "IC: Mod = " << OrigI
11415 << " New = " << *I;
11418 // If the instruction was modified, it's possible that it is now dead.
11419 // if so, remove it.
11420 if (isInstructionTriviallyDead(I)) {
11421 // Make sure we process all operands now that we are reducing their
11423 AddUsesToWorkList(*I);
11425 // Instructions may end up in the worklist more than once. Erase all
11426 // occurrences of this instruction.
11427 RemoveFromWorkList(I);
11428 I->eraseFromParent();
11431 AddUsersToWorkList(*I);
11438 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11440 // Do an explicit clear, this shrinks the map if needed.
11441 WorklistMap.clear();
11446 bool InstCombiner::runOnFunction(Function &F) {
11447 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11449 bool EverMadeChange = false;
11451 // Iterate while there is work to do.
11452 unsigned Iteration = 0;
11453 while (DoOneIteration(F, Iteration++))
11454 EverMadeChange = true;
11455 return EverMadeChange;
11458 FunctionPass *llvm::createInstructionCombiningPass() {
11459 return new InstCombiner();