1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 std::vector<Instruction*> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass((intptr_t)&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 Instruction *commonRemTransforms(BinaryOperator &I);
176 Instruction *commonIRemTransforms(BinaryOperator &I);
177 Instruction *commonDivTransforms(BinaryOperator &I);
178 Instruction *commonIDivTransforms(BinaryOperator &I);
179 Instruction *visitUDiv(BinaryOperator &I);
180 Instruction *visitSDiv(BinaryOperator &I);
181 Instruction *visitFDiv(BinaryOperator &I);
182 Instruction *visitAnd(BinaryOperator &I);
183 Instruction *visitOr (BinaryOperator &I);
184 Instruction *visitXor(BinaryOperator &I);
185 Instruction *visitShl(BinaryOperator &I);
186 Instruction *visitAShr(BinaryOperator &I);
187 Instruction *visitLShr(BinaryOperator &I);
188 Instruction *commonShiftTransforms(BinaryOperator &I);
189 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
191 Instruction *visitFCmpInst(FCmpInst &I);
192 Instruction *visitICmpInst(ICmpInst &I);
193 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
194 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
197 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
198 ConstantInt *DivRHS);
200 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
201 ICmpInst::Predicate Cond, Instruction &I);
202 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
204 Instruction *commonCastTransforms(CastInst &CI);
205 Instruction *commonIntCastTransforms(CastInst &CI);
206 Instruction *commonPointerCastTransforms(CastInst &CI);
207 Instruction *visitTrunc(TruncInst &CI);
208 Instruction *visitZExt(ZExtInst &CI);
209 Instruction *visitSExt(SExtInst &CI);
210 Instruction *visitFPTrunc(FPTruncInst &CI);
211 Instruction *visitFPExt(CastInst &CI);
212 Instruction *visitFPToUI(FPToUIInst &FI);
213 Instruction *visitFPToSI(FPToSIInst &FI);
214 Instruction *visitUIToFP(CastInst &CI);
215 Instruction *visitSIToFP(CastInst &CI);
216 Instruction *visitPtrToInt(CastInst &CI);
217 Instruction *visitIntToPtr(IntToPtrInst &CI);
218 Instruction *visitBitCast(BitCastInst &CI);
219 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
221 Instruction *visitSelectInst(SelectInst &CI);
222 Instruction *visitCallInst(CallInst &CI);
223 Instruction *visitInvokeInst(InvokeInst &II);
224 Instruction *visitPHINode(PHINode &PN);
225 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
226 Instruction *visitAllocationInst(AllocationInst &AI);
227 Instruction *visitFreeInst(FreeInst &FI);
228 Instruction *visitLoadInst(LoadInst &LI);
229 Instruction *visitStoreInst(StoreInst &SI);
230 Instruction *visitBranchInst(BranchInst &BI);
231 Instruction *visitSwitchInst(SwitchInst &SI);
232 Instruction *visitInsertElementInst(InsertElementInst &IE);
233 Instruction *visitExtractElementInst(ExtractElementInst &EI);
234 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
235 Instruction *visitExtractValueInst(ExtractValueInst &EV);
237 // visitInstruction - Specify what to return for unhandled instructions...
238 Instruction *visitInstruction(Instruction &I) { return 0; }
241 Instruction *visitCallSite(CallSite CS);
242 bool transformConstExprCastCall(CallSite CS);
243 Instruction *transformCallThroughTrampoline(CallSite CS);
244 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
245 bool DoXform = true);
246 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
249 // InsertNewInstBefore - insert an instruction New before instruction Old
250 // in the program. Add the new instruction to the worklist.
252 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
253 assert(New && New->getParent() == 0 &&
254 "New instruction already inserted into a basic block!");
255 BasicBlock *BB = Old.getParent();
256 BB->getInstList().insert(&Old, New); // Insert inst
261 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
262 /// This also adds the cast to the worklist. Finally, this returns the
264 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
266 if (V->getType() == Ty) return V;
268 if (Constant *CV = dyn_cast<Constant>(V))
269 return ConstantExpr::getCast(opc, CV, Ty);
271 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
276 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
277 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
281 // ReplaceInstUsesWith - This method is to be used when an instruction is
282 // found to be dead, replacable with another preexisting expression. Here
283 // we add all uses of I to the worklist, replace all uses of I with the new
284 // value, then return I, so that the inst combiner will know that I was
287 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
288 AddUsersToWorkList(I); // Add all modified instrs to worklist
290 I.replaceAllUsesWith(V);
293 // If we are replacing the instruction with itself, this must be in a
294 // segment of unreachable code, so just clobber the instruction.
295 I.replaceAllUsesWith(UndefValue::get(I.getType()));
300 // UpdateValueUsesWith - This method is to be used when an value is
301 // found to be replacable with another preexisting expression or was
302 // updated. Here we add all uses of I to the worklist, replace all uses of
303 // I with the new value (unless the instruction was just updated), then
304 // return true, so that the inst combiner will know that I was modified.
306 bool UpdateValueUsesWith(Value *Old, Value *New) {
307 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
309 Old->replaceAllUsesWith(New);
310 if (Instruction *I = dyn_cast<Instruction>(Old))
312 if (Instruction *I = dyn_cast<Instruction>(New))
317 // EraseInstFromFunction - When dealing with an instruction that has side
318 // effects or produces a void value, we can't rely on DCE to delete the
319 // instruction. Instead, visit methods should return the value returned by
321 Instruction *EraseInstFromFunction(Instruction &I) {
322 assert(I.use_empty() && "Cannot erase instruction that is used!");
323 AddUsesToWorkList(I);
324 RemoveFromWorkList(&I);
326 return 0; // Don't do anything with FI
329 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
330 APInt &KnownOne, unsigned Depth = 0) const {
331 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
334 bool MaskedValueIsZero(Value *V, const APInt &Mask,
335 unsigned Depth = 0) const {
336 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
338 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
339 return llvm::ComputeNumSignBits(Op, TD, Depth);
343 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
344 /// InsertBefore instruction. This is specialized a bit to avoid inserting
345 /// casts that are known to not do anything...
347 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
348 Value *V, const Type *DestTy,
349 Instruction *InsertBefore);
351 /// SimplifyCommutative - This performs a few simplifications for
352 /// commutative operators.
353 bool SimplifyCommutative(BinaryOperator &I);
355 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
356 /// most-complex to least-complex order.
357 bool SimplifyCompare(CmpInst &I);
359 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
360 /// on the demanded bits.
361 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
362 APInt& KnownZero, APInt& KnownOne,
365 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
366 uint64_t &UndefElts, unsigned Depth = 0);
368 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
369 // PHI node as operand #0, see if we can fold the instruction into the PHI
370 // (which is only possible if all operands to the PHI are constants).
371 Instruction *FoldOpIntoPhi(Instruction &I);
373 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
374 // operator and they all are only used by the PHI, PHI together their
375 // inputs, and do the operation once, to the result of the PHI.
376 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
377 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
380 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
381 ConstantInt *AndRHS, BinaryOperator &TheAnd);
383 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
384 bool isSub, Instruction &I);
385 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
386 bool isSigned, bool Inside, Instruction &IB);
387 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
388 Instruction *MatchBSwap(BinaryOperator &I);
389 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
390 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
391 Instruction *SimplifyMemSet(MemSetInst *MI);
394 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
396 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
398 int &NumCastsRemoved);
399 unsigned GetOrEnforceKnownAlignment(Value *V,
400 unsigned PrefAlign = 0);
405 char InstCombiner::ID = 0;
406 static RegisterPass<InstCombiner>
407 X("instcombine", "Combine redundant instructions");
409 // getComplexity: Assign a complexity or rank value to LLVM Values...
410 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
411 static unsigned getComplexity(Value *V) {
412 if (isa<Instruction>(V)) {
413 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
417 if (isa<Argument>(V)) return 3;
418 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
421 // isOnlyUse - Return true if this instruction will be deleted if we stop using
423 static bool isOnlyUse(Value *V) {
424 return V->hasOneUse() || isa<Constant>(V);
427 // getPromotedType - Return the specified type promoted as it would be to pass
428 // though a va_arg area...
429 static const Type *getPromotedType(const Type *Ty) {
430 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
431 if (ITy->getBitWidth() < 32)
432 return Type::Int32Ty;
437 /// getBitCastOperand - If the specified operand is a CastInst or a constant
438 /// expression bitcast, return the operand value, otherwise return null.
439 static Value *getBitCastOperand(Value *V) {
440 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
441 return I->getOperand(0);
442 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
443 if (CE->getOpcode() == Instruction::BitCast)
444 return CE->getOperand(0);
448 /// This function is a wrapper around CastInst::isEliminableCastPair. It
449 /// simply extracts arguments and returns what that function returns.
450 static Instruction::CastOps
451 isEliminableCastPair(
452 const CastInst *CI, ///< The first cast instruction
453 unsigned opcode, ///< The opcode of the second cast instruction
454 const Type *DstTy, ///< The target type for the second cast instruction
455 TargetData *TD ///< The target data for pointer size
458 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
459 const Type *MidTy = CI->getType(); // B from above
461 // Get the opcodes of the two Cast instructions
462 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
463 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
465 return Instruction::CastOps(
466 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
467 DstTy, TD->getIntPtrType()));
470 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
471 /// in any code being generated. It does not require codegen if V is simple
472 /// enough or if the cast can be folded into other casts.
473 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
474 const Type *Ty, TargetData *TD) {
475 if (V->getType() == Ty || isa<Constant>(V)) return false;
477 // If this is another cast that can be eliminated, it isn't codegen either.
478 if (const CastInst *CI = dyn_cast<CastInst>(V))
479 if (isEliminableCastPair(CI, opcode, Ty, TD))
484 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
485 /// InsertBefore instruction. This is specialized a bit to avoid inserting
486 /// casts that are known to not do anything...
488 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
489 Value *V, const Type *DestTy,
490 Instruction *InsertBefore) {
491 if (V->getType() == DestTy) return V;
492 if (Constant *C = dyn_cast<Constant>(V))
493 return ConstantExpr::getCast(opcode, C, DestTy);
495 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
498 // SimplifyCommutative - This performs a few simplifications for commutative
501 // 1. Order operands such that they are listed from right (least complex) to
502 // left (most complex). This puts constants before unary operators before
505 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
506 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
508 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
509 bool Changed = false;
510 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
511 Changed = !I.swapOperands();
513 if (!I.isAssociative()) return Changed;
514 Instruction::BinaryOps Opcode = I.getOpcode();
515 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
516 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
517 if (isa<Constant>(I.getOperand(1))) {
518 Constant *Folded = ConstantExpr::get(I.getOpcode(),
519 cast<Constant>(I.getOperand(1)),
520 cast<Constant>(Op->getOperand(1)));
521 I.setOperand(0, Op->getOperand(0));
522 I.setOperand(1, Folded);
524 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
525 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
526 isOnlyUse(Op) && isOnlyUse(Op1)) {
527 Constant *C1 = cast<Constant>(Op->getOperand(1));
528 Constant *C2 = cast<Constant>(Op1->getOperand(1));
530 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
531 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
532 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
536 I.setOperand(0, New);
537 I.setOperand(1, Folded);
544 /// SimplifyCompare - For a CmpInst this function just orders the operands
545 /// so that theyare listed from right (least complex) to left (most complex).
546 /// This puts constants before unary operators before binary operators.
547 bool InstCombiner::SimplifyCompare(CmpInst &I) {
548 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
551 // Compare instructions are not associative so there's nothing else we can do.
555 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
556 // if the LHS is a constant zero (which is the 'negate' form).
558 static inline Value *dyn_castNegVal(Value *V) {
559 if (BinaryOperator::isNeg(V))
560 return BinaryOperator::getNegArgument(V);
562 // Constants can be considered to be negated values if they can be folded.
563 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
564 return ConstantExpr::getNeg(C);
566 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
567 if (C->getType()->getElementType()->isInteger())
568 return ConstantExpr::getNeg(C);
573 static inline Value *dyn_castNotVal(Value *V) {
574 if (BinaryOperator::isNot(V))
575 return BinaryOperator::getNotArgument(V);
577 // Constants can be considered to be not'ed values...
578 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
579 return ConstantInt::get(~C->getValue());
583 // dyn_castFoldableMul - If this value is a multiply that can be folded into
584 // other computations (because it has a constant operand), return the
585 // non-constant operand of the multiply, and set CST to point to the multiplier.
586 // Otherwise, return null.
588 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
589 if (V->hasOneUse() && V->getType()->isInteger())
590 if (Instruction *I = dyn_cast<Instruction>(V)) {
591 if (I->getOpcode() == Instruction::Mul)
592 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
593 return I->getOperand(0);
594 if (I->getOpcode() == Instruction::Shl)
595 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
596 // The multiplier is really 1 << CST.
597 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
598 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
599 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
600 return I->getOperand(0);
606 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
607 /// expression, return it.
608 static User *dyn_castGetElementPtr(Value *V) {
609 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
611 if (CE->getOpcode() == Instruction::GetElementPtr)
612 return cast<User>(V);
616 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
617 /// opcode value. Otherwise return UserOp1.
618 static unsigned getOpcode(const Value *V) {
619 if (const Instruction *I = dyn_cast<Instruction>(V))
620 return I->getOpcode();
621 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
622 return CE->getOpcode();
623 // Use UserOp1 to mean there's no opcode.
624 return Instruction::UserOp1;
627 /// AddOne - Add one to a ConstantInt
628 static ConstantInt *AddOne(ConstantInt *C) {
629 APInt Val(C->getValue());
630 return ConstantInt::get(++Val);
632 /// SubOne - Subtract one from a ConstantInt
633 static ConstantInt *SubOne(ConstantInt *C) {
634 APInt Val(C->getValue());
635 return ConstantInt::get(--Val);
637 /// Add - Add two ConstantInts together
638 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
639 return ConstantInt::get(C1->getValue() + C2->getValue());
641 /// And - Bitwise AND two ConstantInts together
642 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
643 return ConstantInt::get(C1->getValue() & C2->getValue());
645 /// Subtract - Subtract one ConstantInt from another
646 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
647 return ConstantInt::get(C1->getValue() - C2->getValue());
649 /// Multiply - Multiply two ConstantInts together
650 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
651 return ConstantInt::get(C1->getValue() * C2->getValue());
653 /// MultiplyOverflows - True if the multiply can not be expressed in an int
655 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
656 uint32_t W = C1->getBitWidth();
657 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
666 APInt MulExt = LHSExt * RHSExt;
669 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
670 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
671 return MulExt.slt(Min) || MulExt.sgt(Max);
673 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
677 /// ShrinkDemandedConstant - Check to see if the specified operand of the
678 /// specified instruction is a constant integer. If so, check to see if there
679 /// are any bits set in the constant that are not demanded. If so, shrink the
680 /// constant and return true.
681 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
683 assert(I && "No instruction?");
684 assert(OpNo < I->getNumOperands() && "Operand index too large");
686 // If the operand is not a constant integer, nothing to do.
687 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
688 if (!OpC) return false;
690 // If there are no bits set that aren't demanded, nothing to do.
691 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
692 if ((~Demanded & OpC->getValue()) == 0)
695 // This instruction is producing bits that are not demanded. Shrink the RHS.
696 Demanded &= OpC->getValue();
697 I->setOperand(OpNo, ConstantInt::get(Demanded));
701 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
702 // set of known zero and one bits, compute the maximum and minimum values that
703 // could have the specified known zero and known one bits, returning them in
705 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
706 const APInt& KnownZero,
707 const APInt& KnownOne,
708 APInt& Min, APInt& Max) {
709 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
710 assert(KnownZero.getBitWidth() == BitWidth &&
711 KnownOne.getBitWidth() == BitWidth &&
712 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
713 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
714 APInt UnknownBits = ~(KnownZero|KnownOne);
716 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
717 // bit if it is unknown.
719 Max = KnownOne|UnknownBits;
721 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
723 Max.clear(BitWidth-1);
727 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
728 // a set of known zero and one bits, compute the maximum and minimum values that
729 // could have the specified known zero and known one bits, returning them in
731 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
732 const APInt &KnownZero,
733 const APInt &KnownOne,
734 APInt &Min, APInt &Max) {
735 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
736 assert(KnownZero.getBitWidth() == BitWidth &&
737 KnownOne.getBitWidth() == BitWidth &&
738 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
739 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
740 APInt UnknownBits = ~(KnownZero|KnownOne);
742 // The minimum value is when the unknown bits are all zeros.
744 // The maximum value is when the unknown bits are all ones.
745 Max = KnownOne|UnknownBits;
748 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
749 /// value based on the demanded bits. When this function is called, it is known
750 /// that only the bits set in DemandedMask of the result of V are ever used
751 /// downstream. Consequently, depending on the mask and V, it may be possible
752 /// to replace V with a constant or one of its operands. In such cases, this
753 /// function does the replacement and returns true. In all other cases, it
754 /// returns false after analyzing the expression and setting KnownOne and known
755 /// to be one in the expression. KnownZero contains all the bits that are known
756 /// to be zero in the expression. These are provided to potentially allow the
757 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
758 /// the expression. KnownOne and KnownZero always follow the invariant that
759 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
760 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
761 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
762 /// and KnownOne must all be the same.
763 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
764 APInt& KnownZero, APInt& KnownOne,
766 assert(V != 0 && "Null pointer of Value???");
767 assert(Depth <= 6 && "Limit Search Depth");
768 uint32_t BitWidth = DemandedMask.getBitWidth();
769 const IntegerType *VTy = cast<IntegerType>(V->getType());
770 assert(VTy->getBitWidth() == BitWidth &&
771 KnownZero.getBitWidth() == BitWidth &&
772 KnownOne.getBitWidth() == BitWidth &&
773 "Value *V, DemandedMask, KnownZero and KnownOne \
774 must have same BitWidth");
775 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
776 // We know all of the bits for a constant!
777 KnownOne = CI->getValue() & DemandedMask;
778 KnownZero = ~KnownOne & DemandedMask;
784 if (!V->hasOneUse()) { // Other users may use these bits.
785 if (Depth != 0) { // Not at the root.
786 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
787 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
790 // If this is the root being simplified, allow it to have multiple uses,
791 // just set the DemandedMask to all bits.
792 DemandedMask = APInt::getAllOnesValue(BitWidth);
793 } else if (DemandedMask == 0) { // Not demanding any bits from V.
794 if (V != UndefValue::get(VTy))
795 return UpdateValueUsesWith(V, UndefValue::get(VTy));
797 } else if (Depth == 6) { // Limit search depth.
801 Instruction *I = dyn_cast<Instruction>(V);
802 if (!I) return false; // Only analyze instructions.
804 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
805 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
806 switch (I->getOpcode()) {
808 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
810 case Instruction::And:
811 // If either the LHS or the RHS are Zero, the result is zero.
812 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
813 RHSKnownZero, RHSKnownOne, Depth+1))
815 assert((RHSKnownZero & RHSKnownOne) == 0 &&
816 "Bits known to be one AND zero?");
818 // If something is known zero on the RHS, the bits aren't demanded on the
820 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
821 LHSKnownZero, LHSKnownOne, Depth+1))
823 assert((LHSKnownZero & LHSKnownOne) == 0 &&
824 "Bits known to be one AND zero?");
826 // If all of the demanded bits are known 1 on one side, return the other.
827 // These bits cannot contribute to the result of the 'and'.
828 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
829 (DemandedMask & ~LHSKnownZero))
830 return UpdateValueUsesWith(I, I->getOperand(0));
831 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
832 (DemandedMask & ~RHSKnownZero))
833 return UpdateValueUsesWith(I, I->getOperand(1));
835 // If all of the demanded bits in the inputs are known zeros, return zero.
836 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
837 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
839 // If the RHS is a constant, see if we can simplify it.
840 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
841 return UpdateValueUsesWith(I, I);
843 // Output known-1 bits are only known if set in both the LHS & RHS.
844 RHSKnownOne &= LHSKnownOne;
845 // Output known-0 are known to be clear if zero in either the LHS | RHS.
846 RHSKnownZero |= LHSKnownZero;
848 case Instruction::Or:
849 // If either the LHS or the RHS are One, the result is One.
850 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
851 RHSKnownZero, RHSKnownOne, Depth+1))
853 assert((RHSKnownZero & RHSKnownOne) == 0 &&
854 "Bits known to be one AND zero?");
855 // If something is known one on the RHS, the bits aren't demanded on the
857 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
858 LHSKnownZero, LHSKnownOne, Depth+1))
860 assert((LHSKnownZero & LHSKnownOne) == 0 &&
861 "Bits known to be one AND zero?");
863 // If all of the demanded bits are known zero on one side, return the other.
864 // These bits cannot contribute to the result of the 'or'.
865 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
866 (DemandedMask & ~LHSKnownOne))
867 return UpdateValueUsesWith(I, I->getOperand(0));
868 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
869 (DemandedMask & ~RHSKnownOne))
870 return UpdateValueUsesWith(I, I->getOperand(1));
872 // If all of the potentially set bits on one side are known to be set on
873 // the other side, just use the 'other' side.
874 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
875 (DemandedMask & (~RHSKnownZero)))
876 return UpdateValueUsesWith(I, I->getOperand(0));
877 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
878 (DemandedMask & (~LHSKnownZero)))
879 return UpdateValueUsesWith(I, I->getOperand(1));
881 // If the RHS is a constant, see if we can simplify it.
882 if (ShrinkDemandedConstant(I, 1, DemandedMask))
883 return UpdateValueUsesWith(I, I);
885 // Output known-0 bits are only known if clear in both the LHS & RHS.
886 RHSKnownZero &= LHSKnownZero;
887 // Output known-1 are known to be set if set in either the LHS | RHS.
888 RHSKnownOne |= LHSKnownOne;
890 case Instruction::Xor: {
891 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
892 RHSKnownZero, RHSKnownOne, Depth+1))
894 assert((RHSKnownZero & RHSKnownOne) == 0 &&
895 "Bits known to be one AND zero?");
896 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
897 LHSKnownZero, LHSKnownOne, Depth+1))
899 assert((LHSKnownZero & LHSKnownOne) == 0 &&
900 "Bits known to be one AND zero?");
902 // If all of the demanded bits are known zero on one side, return the other.
903 // These bits cannot contribute to the result of the 'xor'.
904 if ((DemandedMask & RHSKnownZero) == DemandedMask)
905 return UpdateValueUsesWith(I, I->getOperand(0));
906 if ((DemandedMask & LHSKnownZero) == DemandedMask)
907 return UpdateValueUsesWith(I, I->getOperand(1));
909 // Output known-0 bits are known if clear or set in both the LHS & RHS.
910 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
911 (RHSKnownOne & LHSKnownOne);
912 // Output known-1 are known to be set if set in only one of the LHS, RHS.
913 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
914 (RHSKnownOne & LHSKnownZero);
916 // If all of the demanded bits are known to be zero on one side or the
917 // other, turn this into an *inclusive* or.
918 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
919 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
921 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
923 InsertNewInstBefore(Or, *I);
924 return UpdateValueUsesWith(I, Or);
927 // If all of the demanded bits on one side are known, and all of the set
928 // bits on that side are also known to be set on the other side, turn this
929 // into an AND, as we know the bits will be cleared.
930 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
931 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
933 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
934 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
936 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
937 InsertNewInstBefore(And, *I);
938 return UpdateValueUsesWith(I, And);
942 // If the RHS is a constant, see if we can simplify it.
943 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
944 if (ShrinkDemandedConstant(I, 1, DemandedMask))
945 return UpdateValueUsesWith(I, I);
947 RHSKnownZero = KnownZeroOut;
948 RHSKnownOne = KnownOneOut;
951 case Instruction::Select:
952 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
953 RHSKnownZero, RHSKnownOne, Depth+1))
955 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
956 LHSKnownZero, LHSKnownOne, Depth+1))
958 assert((RHSKnownZero & RHSKnownOne) == 0 &&
959 "Bits known to be one AND zero?");
960 assert((LHSKnownZero & LHSKnownOne) == 0 &&
961 "Bits known to be one AND zero?");
963 // If the operands are constants, see if we can simplify them.
964 if (ShrinkDemandedConstant(I, 1, DemandedMask))
965 return UpdateValueUsesWith(I, I);
966 if (ShrinkDemandedConstant(I, 2, DemandedMask))
967 return UpdateValueUsesWith(I, I);
969 // Only known if known in both the LHS and RHS.
970 RHSKnownOne &= LHSKnownOne;
971 RHSKnownZero &= LHSKnownZero;
973 case Instruction::Trunc: {
975 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
976 DemandedMask.zext(truncBf);
977 RHSKnownZero.zext(truncBf);
978 RHSKnownOne.zext(truncBf);
979 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
980 RHSKnownZero, RHSKnownOne, Depth+1))
982 DemandedMask.trunc(BitWidth);
983 RHSKnownZero.trunc(BitWidth);
984 RHSKnownOne.trunc(BitWidth);
985 assert((RHSKnownZero & RHSKnownOne) == 0 &&
986 "Bits known to be one AND zero?");
989 case Instruction::BitCast:
990 if (!I->getOperand(0)->getType()->isInteger())
993 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
994 RHSKnownZero, RHSKnownOne, Depth+1))
996 assert((RHSKnownZero & RHSKnownOne) == 0 &&
997 "Bits known to be one AND zero?");
999 case Instruction::ZExt: {
1000 // Compute the bits in the result that are not present in the input.
1001 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1002 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1004 DemandedMask.trunc(SrcBitWidth);
1005 RHSKnownZero.trunc(SrcBitWidth);
1006 RHSKnownOne.trunc(SrcBitWidth);
1007 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1008 RHSKnownZero, RHSKnownOne, Depth+1))
1010 DemandedMask.zext(BitWidth);
1011 RHSKnownZero.zext(BitWidth);
1012 RHSKnownOne.zext(BitWidth);
1013 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1014 "Bits known to be one AND zero?");
1015 // The top bits are known to be zero.
1016 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1019 case Instruction::SExt: {
1020 // Compute the bits in the result that are not present in the input.
1021 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1022 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1024 APInt InputDemandedBits = DemandedMask &
1025 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1027 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1028 // If any of the sign extended bits are demanded, we know that the sign
1030 if ((NewBits & DemandedMask) != 0)
1031 InputDemandedBits.set(SrcBitWidth-1);
1033 InputDemandedBits.trunc(SrcBitWidth);
1034 RHSKnownZero.trunc(SrcBitWidth);
1035 RHSKnownOne.trunc(SrcBitWidth);
1036 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1037 RHSKnownZero, RHSKnownOne, Depth+1))
1039 InputDemandedBits.zext(BitWidth);
1040 RHSKnownZero.zext(BitWidth);
1041 RHSKnownOne.zext(BitWidth);
1042 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1043 "Bits known to be one AND zero?");
1045 // If the sign bit of the input is known set or clear, then we know the
1046 // top bits of the result.
1048 // If the input sign bit is known zero, or if the NewBits are not demanded
1049 // convert this into a zero extension.
1050 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1052 // Convert to ZExt cast
1053 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1054 return UpdateValueUsesWith(I, NewCast);
1055 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1056 RHSKnownOne |= NewBits;
1060 case Instruction::Add: {
1061 // Figure out what the input bits are. If the top bits of the and result
1062 // are not demanded, then the add doesn't demand them from its input
1064 uint32_t NLZ = DemandedMask.countLeadingZeros();
1066 // If there is a constant on the RHS, there are a variety of xformations
1068 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1069 // If null, this should be simplified elsewhere. Some of the xforms here
1070 // won't work if the RHS is zero.
1074 // If the top bit of the output is demanded, demand everything from the
1075 // input. Otherwise, we demand all the input bits except NLZ top bits.
1076 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1078 // Find information about known zero/one bits in the input.
1079 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1080 LHSKnownZero, LHSKnownOne, Depth+1))
1083 // If the RHS of the add has bits set that can't affect the input, reduce
1085 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1086 return UpdateValueUsesWith(I, I);
1088 // Avoid excess work.
1089 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1092 // Turn it into OR if input bits are zero.
1093 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1095 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1097 InsertNewInstBefore(Or, *I);
1098 return UpdateValueUsesWith(I, Or);
1101 // We can say something about the output known-zero and known-one bits,
1102 // depending on potential carries from the input constant and the
1103 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1104 // bits set and the RHS constant is 0x01001, then we know we have a known
1105 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1107 // To compute this, we first compute the potential carry bits. These are
1108 // the bits which may be modified. I'm not aware of a better way to do
1110 const APInt& RHSVal = RHS->getValue();
1111 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1113 // Now that we know which bits have carries, compute the known-1/0 sets.
1115 // Bits are known one if they are known zero in one operand and one in the
1116 // other, and there is no input carry.
1117 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1118 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1120 // Bits are known zero if they are known zero in both operands and there
1121 // is no input carry.
1122 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1124 // If the high-bits of this ADD are not demanded, then it does not demand
1125 // the high bits of its LHS or RHS.
1126 if (DemandedMask[BitWidth-1] == 0) {
1127 // Right fill the mask of bits for this ADD to demand the most
1128 // significant bit and all those below it.
1129 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1130 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1131 LHSKnownZero, LHSKnownOne, Depth+1))
1133 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1134 LHSKnownZero, LHSKnownOne, Depth+1))
1140 case Instruction::Sub:
1141 // If the high-bits of this SUB are not demanded, then it does not demand
1142 // the high bits of its LHS or RHS.
1143 if (DemandedMask[BitWidth-1] == 0) {
1144 // Right fill the mask of bits for this SUB to demand the most
1145 // significant bit and all those below it.
1146 uint32_t NLZ = DemandedMask.countLeadingZeros();
1147 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1148 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1149 LHSKnownZero, LHSKnownOne, Depth+1))
1151 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1152 LHSKnownZero, LHSKnownOne, Depth+1))
1155 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1156 // the known zeros and ones.
1157 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1159 case Instruction::Shl:
1160 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1161 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1162 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1163 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1164 RHSKnownZero, RHSKnownOne, Depth+1))
1166 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1167 "Bits known to be one AND zero?");
1168 RHSKnownZero <<= ShiftAmt;
1169 RHSKnownOne <<= ShiftAmt;
1170 // low bits known zero.
1172 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1175 case Instruction::LShr:
1176 // For a logical shift right
1177 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1178 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1180 // Unsigned shift right.
1181 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1182 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1183 RHSKnownZero, RHSKnownOne, Depth+1))
1185 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1186 "Bits known to be one AND zero?");
1187 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1188 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1190 // Compute the new bits that are at the top now.
1191 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1192 RHSKnownZero |= HighBits; // high bits known zero.
1196 case Instruction::AShr:
1197 // If this is an arithmetic shift right and only the low-bit is set, we can
1198 // always convert this into a logical shr, even if the shift amount is
1199 // variable. The low bit of the shift cannot be an input sign bit unless
1200 // the shift amount is >= the size of the datatype, which is undefined.
1201 if (DemandedMask == 1) {
1202 // Perform the logical shift right.
1203 Value *NewVal = BinaryOperator::CreateLShr(
1204 I->getOperand(0), I->getOperand(1), I->getName());
1205 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1206 return UpdateValueUsesWith(I, NewVal);
1209 // If the sign bit is the only bit demanded by this ashr, then there is no
1210 // need to do it, the shift doesn't change the high bit.
1211 if (DemandedMask.isSignBit())
1212 return UpdateValueUsesWith(I, I->getOperand(0));
1214 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1215 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1217 // Signed shift right.
1218 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1219 // If any of the "high bits" are demanded, we should set the sign bit as
1221 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1222 DemandedMaskIn.set(BitWidth-1);
1223 if (SimplifyDemandedBits(I->getOperand(0),
1225 RHSKnownZero, RHSKnownOne, Depth+1))
1227 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1228 "Bits known to be one AND zero?");
1229 // Compute the new bits that are at the top now.
1230 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1231 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1232 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1234 // Handle the sign bits.
1235 APInt SignBit(APInt::getSignBit(BitWidth));
1236 // Adjust to where it is now in the mask.
1237 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1239 // If the input sign bit is known to be zero, or if none of the top bits
1240 // are demanded, turn this into an unsigned shift right.
1241 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1242 (HighBits & ~DemandedMask) == HighBits) {
1243 // Perform the logical shift right.
1244 Value *NewVal = BinaryOperator::CreateLShr(
1245 I->getOperand(0), SA, I->getName());
1246 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1247 return UpdateValueUsesWith(I, NewVal);
1248 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1249 RHSKnownOne |= HighBits;
1253 case Instruction::SRem:
1254 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1255 APInt RA = Rem->getValue();
1256 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1257 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1258 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1259 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1260 LHSKnownZero, LHSKnownOne, Depth+1))
1263 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1264 LHSKnownZero |= ~LowBits;
1265 else if (LHSKnownOne[BitWidth-1])
1266 LHSKnownOne |= ~LowBits;
1268 KnownZero |= LHSKnownZero & DemandedMask;
1269 KnownOne |= LHSKnownOne & DemandedMask;
1271 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1275 case Instruction::URem: {
1276 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1277 APInt RA = Rem->getValue();
1278 if (RA.isPowerOf2()) {
1279 APInt LowBits = (RA - 1);
1280 APInt Mask2 = LowBits & DemandedMask;
1281 KnownZero |= ~LowBits & DemandedMask;
1282 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1283 KnownZero, KnownOne, Depth+1))
1286 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1291 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1292 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1293 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1294 KnownZero2, KnownOne2, Depth+1))
1297 uint32_t Leaders = KnownZero2.countLeadingOnes();
1298 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1299 KnownZero2, KnownOne2, Depth+1))
1302 Leaders = std::max(Leaders,
1303 KnownZero2.countLeadingOnes());
1304 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1307 case Instruction::Call:
1308 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1309 switch (II->getIntrinsicID()) {
1311 case Intrinsic::bswap: {
1312 // If the only bits demanded come from one byte of the bswap result,
1313 // just shift the input byte into position to eliminate the bswap.
1314 unsigned NLZ = DemandedMask.countLeadingZeros();
1315 unsigned NTZ = DemandedMask.countTrailingZeros();
1317 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1318 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1319 // have 14 leading zeros, round to 8.
1322 // If we need exactly one byte, we can do this transformation.
1323 if (BitWidth-NLZ-NTZ == 8) {
1324 unsigned ResultBit = NTZ;
1325 unsigned InputBit = BitWidth-NTZ-8;
1327 // Replace this with either a left or right shift to get the byte into
1329 Instruction *NewVal;
1330 if (InputBit > ResultBit)
1331 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1332 ConstantInt::get(I->getType(), InputBit-ResultBit));
1334 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1335 ConstantInt::get(I->getType(), ResultBit-InputBit));
1336 NewVal->takeName(I);
1337 InsertNewInstBefore(NewVal, *I);
1338 return UpdateValueUsesWith(I, NewVal);
1341 // TODO: Could compute known zero/one bits based on the input.
1346 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1350 // If the client is only demanding bits that we know, return the known
1352 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1353 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1358 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1359 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1360 /// actually used by the caller. This method analyzes which elements of the
1361 /// operand are undef and returns that information in UndefElts.
1363 /// If the information about demanded elements can be used to simplify the
1364 /// operation, the operation is simplified, then the resultant value is
1365 /// returned. This returns null if no change was made.
1366 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1367 uint64_t &UndefElts,
1369 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1370 assert(VWidth <= 64 && "Vector too wide to analyze!");
1371 uint64_t EltMask = ~0ULL >> (64-VWidth);
1372 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1373 "Invalid DemandedElts!");
1375 if (isa<UndefValue>(V)) {
1376 // If the entire vector is undefined, just return this info.
1377 UndefElts = EltMask;
1379 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1380 UndefElts = EltMask;
1381 return UndefValue::get(V->getType());
1385 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1386 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1387 Constant *Undef = UndefValue::get(EltTy);
1389 std::vector<Constant*> Elts;
1390 for (unsigned i = 0; i != VWidth; ++i)
1391 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1392 Elts.push_back(Undef);
1393 UndefElts |= (1ULL << i);
1394 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1395 Elts.push_back(Undef);
1396 UndefElts |= (1ULL << i);
1397 } else { // Otherwise, defined.
1398 Elts.push_back(CP->getOperand(i));
1401 // If we changed the constant, return it.
1402 Constant *NewCP = ConstantVector::get(Elts);
1403 return NewCP != CP ? NewCP : 0;
1404 } else if (isa<ConstantAggregateZero>(V)) {
1405 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1407 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1408 Constant *Zero = Constant::getNullValue(EltTy);
1409 Constant *Undef = UndefValue::get(EltTy);
1410 std::vector<Constant*> Elts;
1411 for (unsigned i = 0; i != VWidth; ++i)
1412 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1413 UndefElts = DemandedElts ^ EltMask;
1414 return ConstantVector::get(Elts);
1417 if (!V->hasOneUse()) { // Other users may use these bits.
1418 if (Depth != 0) { // Not at the root.
1419 // TODO: Just compute the UndefElts information recursively.
1423 } else if (Depth == 10) { // Limit search depth.
1427 Instruction *I = dyn_cast<Instruction>(V);
1428 if (!I) return false; // Only analyze instructions.
1430 bool MadeChange = false;
1431 uint64_t UndefElts2;
1433 switch (I->getOpcode()) {
1436 case Instruction::InsertElement: {
1437 // If this is a variable index, we don't know which element it overwrites.
1438 // demand exactly the same input as we produce.
1439 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1441 // Note that we can't propagate undef elt info, because we don't know
1442 // which elt is getting updated.
1443 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1444 UndefElts2, Depth+1);
1445 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1449 // If this is inserting an element that isn't demanded, remove this
1451 unsigned IdxNo = Idx->getZExtValue();
1452 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1453 return AddSoonDeadInstToWorklist(*I, 0);
1455 // Otherwise, the element inserted overwrites whatever was there, so the
1456 // input demanded set is simpler than the output set.
1457 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1458 DemandedElts & ~(1ULL << IdxNo),
1459 UndefElts, Depth+1);
1460 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1462 // The inserted element is defined.
1463 UndefElts |= 1ULL << IdxNo;
1466 case Instruction::BitCast: {
1467 // Vector->vector casts only.
1468 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1470 unsigned InVWidth = VTy->getNumElements();
1471 uint64_t InputDemandedElts = 0;
1474 if (VWidth == InVWidth) {
1475 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1476 // elements as are demanded of us.
1478 InputDemandedElts = DemandedElts;
1479 } else if (VWidth > InVWidth) {
1483 // If there are more elements in the result than there are in the source,
1484 // then an input element is live if any of the corresponding output
1485 // elements are live.
1486 Ratio = VWidth/InVWidth;
1487 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1488 if (DemandedElts & (1ULL << OutIdx))
1489 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1495 // If there are more elements in the source than there are in the result,
1496 // then an input element is live if the corresponding output element is
1498 Ratio = InVWidth/VWidth;
1499 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1500 if (DemandedElts & (1ULL << InIdx/Ratio))
1501 InputDemandedElts |= 1ULL << InIdx;
1504 // div/rem demand all inputs, because they don't want divide by zero.
1505 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1506 UndefElts2, Depth+1);
1508 I->setOperand(0, TmpV);
1512 UndefElts = UndefElts2;
1513 if (VWidth > InVWidth) {
1514 assert(0 && "Unimp");
1515 // If there are more elements in the result than there are in the source,
1516 // then an output element is undef if the corresponding input element is
1518 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1519 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1520 UndefElts |= 1ULL << OutIdx;
1521 } else if (VWidth < InVWidth) {
1522 assert(0 && "Unimp");
1523 // If there are more elements in the source than there are in the result,
1524 // then a result element is undef if all of the corresponding input
1525 // elements are undef.
1526 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1527 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1528 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1529 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1533 case Instruction::And:
1534 case Instruction::Or:
1535 case Instruction::Xor:
1536 case Instruction::Add:
1537 case Instruction::Sub:
1538 case Instruction::Mul:
1539 // div/rem demand all inputs, because they don't want divide by zero.
1540 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1541 UndefElts, Depth+1);
1542 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1543 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1544 UndefElts2, Depth+1);
1545 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1547 // Output elements are undefined if both are undefined. Consider things
1548 // like undef&0. The result is known zero, not undef.
1549 UndefElts &= UndefElts2;
1552 case Instruction::Call: {
1553 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1555 switch (II->getIntrinsicID()) {
1558 // Binary vector operations that work column-wise. A dest element is a
1559 // function of the corresponding input elements from the two inputs.
1560 case Intrinsic::x86_sse_sub_ss:
1561 case Intrinsic::x86_sse_mul_ss:
1562 case Intrinsic::x86_sse_min_ss:
1563 case Intrinsic::x86_sse_max_ss:
1564 case Intrinsic::x86_sse2_sub_sd:
1565 case Intrinsic::x86_sse2_mul_sd:
1566 case Intrinsic::x86_sse2_min_sd:
1567 case Intrinsic::x86_sse2_max_sd:
1568 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1569 UndefElts, Depth+1);
1570 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1571 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1572 UndefElts2, Depth+1);
1573 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1575 // If only the low elt is demanded and this is a scalarizable intrinsic,
1576 // scalarize it now.
1577 if (DemandedElts == 1) {
1578 switch (II->getIntrinsicID()) {
1580 case Intrinsic::x86_sse_sub_ss:
1581 case Intrinsic::x86_sse_mul_ss:
1582 case Intrinsic::x86_sse2_sub_sd:
1583 case Intrinsic::x86_sse2_mul_sd:
1584 // TODO: Lower MIN/MAX/ABS/etc
1585 Value *LHS = II->getOperand(1);
1586 Value *RHS = II->getOperand(2);
1587 // Extract the element as scalars.
1588 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1589 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1591 switch (II->getIntrinsicID()) {
1592 default: assert(0 && "Case stmts out of sync!");
1593 case Intrinsic::x86_sse_sub_ss:
1594 case Intrinsic::x86_sse2_sub_sd:
1595 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1596 II->getName()), *II);
1598 case Intrinsic::x86_sse_mul_ss:
1599 case Intrinsic::x86_sse2_mul_sd:
1600 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1601 II->getName()), *II);
1606 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1608 InsertNewInstBefore(New, *II);
1609 AddSoonDeadInstToWorklist(*II, 0);
1614 // Output elements are undefined if both are undefined. Consider things
1615 // like undef&0. The result is known zero, not undef.
1616 UndefElts &= UndefElts2;
1622 return MadeChange ? I : 0;
1626 /// AssociativeOpt - Perform an optimization on an associative operator. This
1627 /// function is designed to check a chain of associative operators for a
1628 /// potential to apply a certain optimization. Since the optimization may be
1629 /// applicable if the expression was reassociated, this checks the chain, then
1630 /// reassociates the expression as necessary to expose the optimization
1631 /// opportunity. This makes use of a special Functor, which must define
1632 /// 'shouldApply' and 'apply' methods.
1634 template<typename Functor>
1635 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1636 unsigned Opcode = Root.getOpcode();
1637 Value *LHS = Root.getOperand(0);
1639 // Quick check, see if the immediate LHS matches...
1640 if (F.shouldApply(LHS))
1641 return F.apply(Root);
1643 // Otherwise, if the LHS is not of the same opcode as the root, return.
1644 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1645 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1646 // Should we apply this transform to the RHS?
1647 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1649 // If not to the RHS, check to see if we should apply to the LHS...
1650 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1651 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1655 // If the functor wants to apply the optimization to the RHS of LHSI,
1656 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1658 BasicBlock *BB = Root.getParent();
1660 // Now all of the instructions are in the current basic block, go ahead
1661 // and perform the reassociation.
1662 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1664 // First move the selected RHS to the LHS of the root...
1665 Root.setOperand(0, LHSI->getOperand(1));
1667 // Make what used to be the LHS of the root be the user of the root...
1668 Value *ExtraOperand = TmpLHSI->getOperand(1);
1669 if (&Root == TmpLHSI) {
1670 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1673 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1674 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1675 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1676 BasicBlock::iterator ARI = &Root; ++ARI;
1677 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1680 // Now propagate the ExtraOperand down the chain of instructions until we
1682 while (TmpLHSI != LHSI) {
1683 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1684 // Move the instruction to immediately before the chain we are
1685 // constructing to avoid breaking dominance properties.
1686 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1687 BB->getInstList().insert(ARI, NextLHSI);
1690 Value *NextOp = NextLHSI->getOperand(1);
1691 NextLHSI->setOperand(1, ExtraOperand);
1693 ExtraOperand = NextOp;
1696 // Now that the instructions are reassociated, have the functor perform
1697 // the transformation...
1698 return F.apply(Root);
1701 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1708 // AddRHS - Implements: X + X --> X << 1
1711 AddRHS(Value *rhs) : RHS(rhs) {}
1712 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1713 Instruction *apply(BinaryOperator &Add) const {
1714 return BinaryOperator::CreateShl(Add.getOperand(0),
1715 ConstantInt::get(Add.getType(), 1));
1719 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1721 struct AddMaskingAnd {
1723 AddMaskingAnd(Constant *c) : C2(c) {}
1724 bool shouldApply(Value *LHS) const {
1726 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1727 ConstantExpr::getAnd(C1, C2)->isNullValue();
1729 Instruction *apply(BinaryOperator &Add) const {
1730 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1736 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1738 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1739 if (Constant *SOC = dyn_cast<Constant>(SO))
1740 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1742 return IC->InsertNewInstBefore(CastInst::Create(
1743 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1746 // Figure out if the constant is the left or the right argument.
1747 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1748 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1750 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1752 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1753 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1756 Value *Op0 = SO, *Op1 = ConstOperand;
1758 std::swap(Op0, Op1);
1760 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1761 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1762 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1763 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1764 SO->getName()+".cmp");
1766 assert(0 && "Unknown binary instruction type!");
1769 return IC->InsertNewInstBefore(New, I);
1772 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1773 // constant as the other operand, try to fold the binary operator into the
1774 // select arguments. This also works for Cast instructions, which obviously do
1775 // not have a second operand.
1776 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1778 // Don't modify shared select instructions
1779 if (!SI->hasOneUse()) return 0;
1780 Value *TV = SI->getOperand(1);
1781 Value *FV = SI->getOperand(2);
1783 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1784 // Bool selects with constant operands can be folded to logical ops.
1785 if (SI->getType() == Type::Int1Ty) return 0;
1787 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1788 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1790 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1797 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1798 /// node as operand #0, see if we can fold the instruction into the PHI (which
1799 /// is only possible if all operands to the PHI are constants).
1800 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1801 PHINode *PN = cast<PHINode>(I.getOperand(0));
1802 unsigned NumPHIValues = PN->getNumIncomingValues();
1803 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1805 // Check to see if all of the operands of the PHI are constants. If there is
1806 // one non-constant value, remember the BB it is. If there is more than one
1807 // or if *it* is a PHI, bail out.
1808 BasicBlock *NonConstBB = 0;
1809 for (unsigned i = 0; i != NumPHIValues; ++i)
1810 if (!isa<Constant>(PN->getIncomingValue(i))) {
1811 if (NonConstBB) return 0; // More than one non-const value.
1812 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1813 NonConstBB = PN->getIncomingBlock(i);
1815 // If the incoming non-constant value is in I's block, we have an infinite
1817 if (NonConstBB == I.getParent())
1821 // If there is exactly one non-constant value, we can insert a copy of the
1822 // operation in that block. However, if this is a critical edge, we would be
1823 // inserting the computation one some other paths (e.g. inside a loop). Only
1824 // do this if the pred block is unconditionally branching into the phi block.
1826 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1827 if (!BI || !BI->isUnconditional()) return 0;
1830 // Okay, we can do the transformation: create the new PHI node.
1831 PHINode *NewPN = PHINode::Create(I.getType(), "");
1832 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1833 InsertNewInstBefore(NewPN, *PN);
1834 NewPN->takeName(PN);
1836 // Next, add all of the operands to the PHI.
1837 if (I.getNumOperands() == 2) {
1838 Constant *C = cast<Constant>(I.getOperand(1));
1839 for (unsigned i = 0; i != NumPHIValues; ++i) {
1841 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1842 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1843 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1845 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1847 assert(PN->getIncomingBlock(i) == NonConstBB);
1848 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1849 InV = BinaryOperator::Create(BO->getOpcode(),
1850 PN->getIncomingValue(i), C, "phitmp",
1851 NonConstBB->getTerminator());
1852 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1853 InV = CmpInst::Create(CI->getOpcode(),
1855 PN->getIncomingValue(i), C, "phitmp",
1856 NonConstBB->getTerminator());
1858 assert(0 && "Unknown binop!");
1860 AddToWorkList(cast<Instruction>(InV));
1862 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1865 CastInst *CI = cast<CastInst>(&I);
1866 const Type *RetTy = CI->getType();
1867 for (unsigned i = 0; i != NumPHIValues; ++i) {
1869 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1870 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1872 assert(PN->getIncomingBlock(i) == NonConstBB);
1873 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1874 I.getType(), "phitmp",
1875 NonConstBB->getTerminator());
1876 AddToWorkList(cast<Instruction>(InV));
1878 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1881 return ReplaceInstUsesWith(I, NewPN);
1885 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1886 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1887 /// This basically requires proving that the add in the original type would not
1888 /// overflow to change the sign bit or have a carry out.
1889 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1890 // There are different heuristics we can use for this. Here are some simple
1893 // Add has the property that adding any two 2's complement numbers can only
1894 // have one carry bit which can change a sign. As such, if LHS and RHS each
1895 // have at least two sign bits, we know that the addition of the two values will
1896 // sign extend fine.
1897 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1901 // If one of the operands only has one non-zero bit, and if the other operand
1902 // has a known-zero bit in a more significant place than it (not including the
1903 // sign bit) the ripple may go up to and fill the zero, but won't change the
1904 // sign. For example, (X & ~4) + 1.
1912 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1913 bool Changed = SimplifyCommutative(I);
1914 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1916 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1917 // X + undef -> undef
1918 if (isa<UndefValue>(RHS))
1919 return ReplaceInstUsesWith(I, RHS);
1922 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1923 if (RHSC->isNullValue())
1924 return ReplaceInstUsesWith(I, LHS);
1925 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1926 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1927 (I.getType())->getValueAPF()))
1928 return ReplaceInstUsesWith(I, LHS);
1931 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1932 // X + (signbit) --> X ^ signbit
1933 const APInt& Val = CI->getValue();
1934 uint32_t BitWidth = Val.getBitWidth();
1935 if (Val == APInt::getSignBit(BitWidth))
1936 return BinaryOperator::CreateXor(LHS, RHS);
1938 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1939 // (X & 254)+1 -> (X&254)|1
1940 if (!isa<VectorType>(I.getType())) {
1941 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1942 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1943 KnownZero, KnownOne))
1948 if (isa<PHINode>(LHS))
1949 if (Instruction *NV = FoldOpIntoPhi(I))
1952 ConstantInt *XorRHS = 0;
1954 if (isa<ConstantInt>(RHSC) &&
1955 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1956 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1957 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1959 uint32_t Size = TySizeBits / 2;
1960 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1961 APInt CFF80Val(-C0080Val);
1963 if (TySizeBits > Size) {
1964 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1965 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1966 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1967 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1968 // This is a sign extend if the top bits are known zero.
1969 if (!MaskedValueIsZero(XorLHS,
1970 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1971 Size = 0; // Not a sign ext, but can't be any others either.
1976 C0080Val = APIntOps::lshr(C0080Val, Size);
1977 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1978 } while (Size >= 1);
1980 // FIXME: This shouldn't be necessary. When the backends can handle types
1981 // with funny bit widths then this switch statement should be removed. It
1982 // is just here to get the size of the "middle" type back up to something
1983 // that the back ends can handle.
1984 const Type *MiddleType = 0;
1987 case 32: MiddleType = Type::Int32Ty; break;
1988 case 16: MiddleType = Type::Int16Ty; break;
1989 case 8: MiddleType = Type::Int8Ty; break;
1992 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1993 InsertNewInstBefore(NewTrunc, I);
1994 return new SExtInst(NewTrunc, I.getType(), I.getName());
1999 if (I.getType() == Type::Int1Ty)
2000 return BinaryOperator::CreateXor(LHS, RHS);
2003 if (I.getType()->isInteger()) {
2004 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2006 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2007 if (RHSI->getOpcode() == Instruction::Sub)
2008 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2009 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2011 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2012 if (LHSI->getOpcode() == Instruction::Sub)
2013 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2014 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2019 // -A + -B --> -(A + B)
2020 if (Value *LHSV = dyn_castNegVal(LHS)) {
2021 if (LHS->getType()->isIntOrIntVector()) {
2022 if (Value *RHSV = dyn_castNegVal(RHS)) {
2023 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2024 InsertNewInstBefore(NewAdd, I);
2025 return BinaryOperator::CreateNeg(NewAdd);
2029 return BinaryOperator::CreateSub(RHS, LHSV);
2033 if (!isa<Constant>(RHS))
2034 if (Value *V = dyn_castNegVal(RHS))
2035 return BinaryOperator::CreateSub(LHS, V);
2039 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2040 if (X == RHS) // X*C + X --> X * (C+1)
2041 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2043 // X*C1 + X*C2 --> X * (C1+C2)
2045 if (X == dyn_castFoldableMul(RHS, C1))
2046 return BinaryOperator::CreateMul(X, Add(C1, C2));
2049 // X + X*C --> X * (C+1)
2050 if (dyn_castFoldableMul(RHS, C2) == LHS)
2051 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2053 // X + ~X --> -1 since ~X = -X-1
2054 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2055 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2058 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2059 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2060 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2063 // A+B --> A|B iff A and B have no bits set in common.
2064 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2065 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2066 APInt LHSKnownOne(IT->getBitWidth(), 0);
2067 APInt LHSKnownZero(IT->getBitWidth(), 0);
2068 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2069 if (LHSKnownZero != 0) {
2070 APInt RHSKnownOne(IT->getBitWidth(), 0);
2071 APInt RHSKnownZero(IT->getBitWidth(), 0);
2072 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2074 // No bits in common -> bitwise or.
2075 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2076 return BinaryOperator::CreateOr(LHS, RHS);
2080 // W*X + Y*Z --> W * (X+Z) iff W == Y
2081 if (I.getType()->isIntOrIntVector()) {
2082 Value *W, *X, *Y, *Z;
2083 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2084 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2088 } else if (Y == X) {
2090 } else if (X == Z) {
2097 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2098 LHS->getName()), I);
2099 return BinaryOperator::CreateMul(W, NewAdd);
2104 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2106 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2107 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2109 // (X & FF00) + xx00 -> (X+xx00) & FF00
2110 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2111 Constant *Anded = And(CRHS, C2);
2112 if (Anded == CRHS) {
2113 // See if all bits from the first bit set in the Add RHS up are included
2114 // in the mask. First, get the rightmost bit.
2115 const APInt& AddRHSV = CRHS->getValue();
2117 // Form a mask of all bits from the lowest bit added through the top.
2118 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2120 // See if the and mask includes all of these bits.
2121 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2123 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2124 // Okay, the xform is safe. Insert the new add pronto.
2125 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2126 LHS->getName()), I);
2127 return BinaryOperator::CreateAnd(NewAdd, C2);
2132 // Try to fold constant add into select arguments.
2133 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2134 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2138 // add (cast *A to intptrtype) B ->
2139 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2141 CastInst *CI = dyn_cast<CastInst>(LHS);
2144 CI = dyn_cast<CastInst>(RHS);
2147 if (CI && CI->getType()->isSized() &&
2148 (CI->getType()->getPrimitiveSizeInBits() ==
2149 TD->getIntPtrType()->getPrimitiveSizeInBits())
2150 && isa<PointerType>(CI->getOperand(0)->getType())) {
2152 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2153 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2154 PointerType::get(Type::Int8Ty, AS), I);
2155 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2156 return new PtrToIntInst(I2, CI->getType());
2160 // add (select X 0 (sub n A)) A --> select X A n
2162 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2165 SI = dyn_cast<SelectInst>(RHS);
2168 if (SI && SI->hasOneUse()) {
2169 Value *TV = SI->getTrueValue();
2170 Value *FV = SI->getFalseValue();
2173 // Can we fold the add into the argument of the select?
2174 // We check both true and false select arguments for a matching subtract.
2175 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2176 A == Other) // Fold the add into the true select value.
2177 return SelectInst::Create(SI->getCondition(), N, A);
2178 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2179 A == Other) // Fold the add into the false select value.
2180 return SelectInst::Create(SI->getCondition(), A, N);
2184 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2185 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2186 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2187 return ReplaceInstUsesWith(I, LHS);
2189 // Check for (add (sext x), y), see if we can merge this into an
2190 // integer add followed by a sext.
2191 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2192 // (add (sext x), cst) --> (sext (add x, cst'))
2193 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2195 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2196 if (LHSConv->hasOneUse() &&
2197 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2198 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2199 // Insert the new, smaller add.
2200 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2202 InsertNewInstBefore(NewAdd, I);
2203 return new SExtInst(NewAdd, I.getType());
2207 // (add (sext x), (sext y)) --> (sext (add int x, y))
2208 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2209 // Only do this if x/y have the same type, if at last one of them has a
2210 // single use (so we don't increase the number of sexts), and if the
2211 // integer add will not overflow.
2212 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2213 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2214 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2215 RHSConv->getOperand(0))) {
2216 // Insert the new integer add.
2217 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2218 RHSConv->getOperand(0),
2220 InsertNewInstBefore(NewAdd, I);
2221 return new SExtInst(NewAdd, I.getType());
2226 // Check for (add double (sitofp x), y), see if we can merge this into an
2227 // integer add followed by a promotion.
2228 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2229 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2230 // ... if the constant fits in the integer value. This is useful for things
2231 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2232 // requires a constant pool load, and generally allows the add to be better
2234 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2236 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2237 if (LHSConv->hasOneUse() &&
2238 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2239 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2240 // Insert the new integer add.
2241 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2243 InsertNewInstBefore(NewAdd, I);
2244 return new SIToFPInst(NewAdd, I.getType());
2248 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2249 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2250 // Only do this if x/y have the same type, if at last one of them has a
2251 // single use (so we don't increase the number of int->fp conversions),
2252 // and if the integer add will not overflow.
2253 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2254 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2255 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2256 RHSConv->getOperand(0))) {
2257 // Insert the new integer add.
2258 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2259 RHSConv->getOperand(0),
2261 InsertNewInstBefore(NewAdd, I);
2262 return new SIToFPInst(NewAdd, I.getType());
2267 return Changed ? &I : 0;
2270 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2271 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2273 if (Op0 == Op1) // sub X, X -> 0
2274 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2276 // If this is a 'B = x-(-A)', change to B = x+A...
2277 if (Value *V = dyn_castNegVal(Op1))
2278 return BinaryOperator::CreateAdd(Op0, V);
2280 if (isa<UndefValue>(Op0))
2281 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2282 if (isa<UndefValue>(Op1))
2283 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2285 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2286 // Replace (-1 - A) with (~A)...
2287 if (C->isAllOnesValue())
2288 return BinaryOperator::CreateNot(Op1);
2290 // C - ~X == X + (1+C)
2292 if (match(Op1, m_Not(m_Value(X))))
2293 return BinaryOperator::CreateAdd(X, AddOne(C));
2295 // -(X >>u 31) -> (X >>s 31)
2296 // -(X >>s 31) -> (X >>u 31)
2298 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2299 if (SI->getOpcode() == Instruction::LShr) {
2300 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2301 // Check to see if we are shifting out everything but the sign bit.
2302 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2303 SI->getType()->getPrimitiveSizeInBits()-1) {
2304 // Ok, the transformation is safe. Insert AShr.
2305 return BinaryOperator::Create(Instruction::AShr,
2306 SI->getOperand(0), CU, SI->getName());
2310 else if (SI->getOpcode() == Instruction::AShr) {
2311 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2312 // Check to see if we are shifting out everything but the sign bit.
2313 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2314 SI->getType()->getPrimitiveSizeInBits()-1) {
2315 // Ok, the transformation is safe. Insert LShr.
2316 return BinaryOperator::CreateLShr(
2317 SI->getOperand(0), CU, SI->getName());
2324 // Try to fold constant sub into select arguments.
2325 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2326 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2329 if (isa<PHINode>(Op0))
2330 if (Instruction *NV = FoldOpIntoPhi(I))
2334 if (I.getType() == Type::Int1Ty)
2335 return BinaryOperator::CreateXor(Op0, Op1);
2337 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2338 if (Op1I->getOpcode() == Instruction::Add &&
2339 !Op0->getType()->isFPOrFPVector()) {
2340 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2341 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2342 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2343 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2344 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2345 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2346 // C1-(X+C2) --> (C1-C2)-X
2347 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2348 Op1I->getOperand(0));
2352 if (Op1I->hasOneUse()) {
2353 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2354 // is not used by anyone else...
2356 if (Op1I->getOpcode() == Instruction::Sub &&
2357 !Op1I->getType()->isFPOrFPVector()) {
2358 // Swap the two operands of the subexpr...
2359 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2360 Op1I->setOperand(0, IIOp1);
2361 Op1I->setOperand(1, IIOp0);
2363 // Create the new top level add instruction...
2364 return BinaryOperator::CreateAdd(Op0, Op1);
2367 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2369 if (Op1I->getOpcode() == Instruction::And &&
2370 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2371 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2374 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2375 return BinaryOperator::CreateAnd(Op0, NewNot);
2378 // 0 - (X sdiv C) -> (X sdiv -C)
2379 if (Op1I->getOpcode() == Instruction::SDiv)
2380 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2382 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2383 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2384 ConstantExpr::getNeg(DivRHS));
2386 // X - X*C --> X * (1-C)
2387 ConstantInt *C2 = 0;
2388 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2389 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2390 return BinaryOperator::CreateMul(Op0, CP1);
2393 // X - ((X / Y) * Y) --> X % Y
2394 if (Op1I->getOpcode() == Instruction::Mul)
2395 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2396 if (Op0 == I->getOperand(0) &&
2397 Op1I->getOperand(1) == I->getOperand(1)) {
2398 if (I->getOpcode() == Instruction::SDiv)
2399 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2400 if (I->getOpcode() == Instruction::UDiv)
2401 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2406 if (!Op0->getType()->isFPOrFPVector())
2407 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2408 if (Op0I->getOpcode() == Instruction::Add) {
2409 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2410 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2411 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2412 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2413 } else if (Op0I->getOpcode() == Instruction::Sub) {
2414 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2415 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2420 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2421 if (X == Op1) // X*C - X --> X * (C-1)
2422 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2424 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2425 if (X == dyn_castFoldableMul(Op1, C2))
2426 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2431 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2432 /// comparison only checks the sign bit. If it only checks the sign bit, set
2433 /// TrueIfSigned if the result of the comparison is true when the input value is
2435 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2436 bool &TrueIfSigned) {
2438 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2439 TrueIfSigned = true;
2440 return RHS->isZero();
2441 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2442 TrueIfSigned = true;
2443 return RHS->isAllOnesValue();
2444 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2445 TrueIfSigned = false;
2446 return RHS->isAllOnesValue();
2447 case ICmpInst::ICMP_UGT:
2448 // True if LHS u> RHS and RHS == high-bit-mask - 1
2449 TrueIfSigned = true;
2450 return RHS->getValue() ==
2451 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2452 case ICmpInst::ICMP_UGE:
2453 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2454 TrueIfSigned = true;
2455 return RHS->getValue().isSignBit();
2461 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2462 bool Changed = SimplifyCommutative(I);
2463 Value *Op0 = I.getOperand(0);
2465 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2466 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2468 // Simplify mul instructions with a constant RHS...
2469 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2470 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2472 // ((X << C1)*C2) == (X * (C2 << C1))
2473 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2474 if (SI->getOpcode() == Instruction::Shl)
2475 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2476 return BinaryOperator::CreateMul(SI->getOperand(0),
2477 ConstantExpr::getShl(CI, ShOp));
2480 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2481 if (CI->equalsInt(1)) // X * 1 == X
2482 return ReplaceInstUsesWith(I, Op0);
2483 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2484 return BinaryOperator::CreateNeg(Op0, I.getName());
2486 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2487 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2488 return BinaryOperator::CreateShl(Op0,
2489 ConstantInt::get(Op0->getType(), Val.logBase2()));
2491 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2492 if (Op1F->isNullValue())
2493 return ReplaceInstUsesWith(I, Op1);
2495 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2496 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2497 // We need a better interface for long double here.
2498 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2499 if (Op1F->isExactlyValue(1.0))
2500 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2503 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2504 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2505 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2506 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2507 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2509 InsertNewInstBefore(Add, I);
2510 Value *C1C2 = ConstantExpr::getMul(Op1,
2511 cast<Constant>(Op0I->getOperand(1)));
2512 return BinaryOperator::CreateAdd(Add, C1C2);
2516 // Try to fold constant mul into select arguments.
2517 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2518 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2521 if (isa<PHINode>(Op0))
2522 if (Instruction *NV = FoldOpIntoPhi(I))
2526 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2527 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2528 return BinaryOperator::CreateMul(Op0v, Op1v);
2530 if (I.getType() == Type::Int1Ty)
2531 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2533 // If one of the operands of the multiply is a cast from a boolean value, then
2534 // we know the bool is either zero or one, so this is a 'masking' multiply.
2535 // See if we can simplify things based on how the boolean was originally
2537 CastInst *BoolCast = 0;
2538 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2539 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2542 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2543 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2546 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2547 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2548 const Type *SCOpTy = SCIOp0->getType();
2551 // If the icmp is true iff the sign bit of X is set, then convert this
2552 // multiply into a shift/and combination.
2553 if (isa<ConstantInt>(SCIOp1) &&
2554 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2556 // Shift the X value right to turn it into "all signbits".
2557 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2558 SCOpTy->getPrimitiveSizeInBits()-1);
2560 InsertNewInstBefore(
2561 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2562 BoolCast->getOperand(0)->getName()+
2565 // If the multiply type is not the same as the source type, sign extend
2566 // or truncate to the multiply type.
2567 if (I.getType() != V->getType()) {
2568 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2569 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2570 Instruction::CastOps opcode =
2571 (SrcBits == DstBits ? Instruction::BitCast :
2572 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2573 V = InsertCastBefore(opcode, V, I.getType(), I);
2576 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2577 return BinaryOperator::CreateAnd(V, OtherOp);
2582 return Changed ? &I : 0;
2585 /// This function implements the transforms on div instructions that work
2586 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2587 /// used by the visitors to those instructions.
2588 /// @brief Transforms common to all three div instructions
2589 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2590 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2592 // undef / X -> 0 for integer.
2593 // undef / X -> undef for FP (the undef could be a snan).
2594 if (isa<UndefValue>(Op0)) {
2595 if (Op0->getType()->isFPOrFPVector())
2596 return ReplaceInstUsesWith(I, Op0);
2597 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2600 // X / undef -> undef
2601 if (isa<UndefValue>(Op1))
2602 return ReplaceInstUsesWith(I, Op1);
2604 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2605 // This does not apply for fdiv.
2606 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2607 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2608 // the same basic block, then we replace the select with Y, and the
2609 // condition of the select with false (if the cond value is in the same BB).
2610 // If the select has uses other than the div, this allows them to be
2611 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2612 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2613 if (ST->isNullValue()) {
2614 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2615 if (CondI && CondI->getParent() == I.getParent())
2616 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2617 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2618 I.setOperand(1, SI->getOperand(2));
2620 UpdateValueUsesWith(SI, SI->getOperand(2));
2624 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2625 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2626 if (ST->isNullValue()) {
2627 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2628 if (CondI && CondI->getParent() == I.getParent())
2629 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2630 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2631 I.setOperand(1, SI->getOperand(1));
2633 UpdateValueUsesWith(SI, SI->getOperand(1));
2641 /// This function implements the transforms common to both integer division
2642 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2643 /// division instructions.
2644 /// @brief Common integer divide transforms
2645 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2646 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2648 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2650 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2651 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2652 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2653 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2656 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2657 return ReplaceInstUsesWith(I, CI);
2660 if (Instruction *Common = commonDivTransforms(I))
2663 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2665 if (RHS->equalsInt(1))
2666 return ReplaceInstUsesWith(I, Op0);
2668 // (X / C1) / C2 -> X / (C1*C2)
2669 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2670 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2671 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2672 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2673 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2675 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2676 Multiply(RHS, LHSRHS));
2679 if (!RHS->isZero()) { // avoid X udiv 0
2680 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2681 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2683 if (isa<PHINode>(Op0))
2684 if (Instruction *NV = FoldOpIntoPhi(I))
2689 // 0 / X == 0, we don't need to preserve faults!
2690 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2691 if (LHS->equalsInt(0))
2692 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2694 // It can't be division by zero, hence it must be division by one.
2695 if (I.getType() == Type::Int1Ty)
2696 return ReplaceInstUsesWith(I, Op0);
2701 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2702 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2704 // Handle the integer div common cases
2705 if (Instruction *Common = commonIDivTransforms(I))
2708 // X udiv C^2 -> X >> C
2709 // Check to see if this is an unsigned division with an exact power of 2,
2710 // if so, convert to a right shift.
2711 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2712 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2713 return BinaryOperator::CreateLShr(Op0,
2714 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2717 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2718 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2719 if (RHSI->getOpcode() == Instruction::Shl &&
2720 isa<ConstantInt>(RHSI->getOperand(0))) {
2721 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2722 if (C1.isPowerOf2()) {
2723 Value *N = RHSI->getOperand(1);
2724 const Type *NTy = N->getType();
2725 if (uint32_t C2 = C1.logBase2()) {
2726 Constant *C2V = ConstantInt::get(NTy, C2);
2727 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2729 return BinaryOperator::CreateLShr(Op0, N);
2734 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2735 // where C1&C2 are powers of two.
2736 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2737 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2738 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2739 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2740 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2741 // Compute the shift amounts
2742 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2743 // Construct the "on true" case of the select
2744 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2745 Instruction *TSI = BinaryOperator::CreateLShr(
2746 Op0, TC, SI->getName()+".t");
2747 TSI = InsertNewInstBefore(TSI, I);
2749 // Construct the "on false" case of the select
2750 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2751 Instruction *FSI = BinaryOperator::CreateLShr(
2752 Op0, FC, SI->getName()+".f");
2753 FSI = InsertNewInstBefore(FSI, I);
2755 // construct the select instruction and return it.
2756 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2762 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2763 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2765 // Handle the integer div common cases
2766 if (Instruction *Common = commonIDivTransforms(I))
2769 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2771 if (RHS->isAllOnesValue())
2772 return BinaryOperator::CreateNeg(Op0);
2775 if (Value *LHSNeg = dyn_castNegVal(Op0))
2776 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2779 // If the sign bits of both operands are zero (i.e. we can prove they are
2780 // unsigned inputs), turn this into a udiv.
2781 if (I.getType()->isInteger()) {
2782 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2783 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2784 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2785 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2792 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2793 return commonDivTransforms(I);
2796 /// This function implements the transforms on rem instructions that work
2797 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2798 /// is used by the visitors to those instructions.
2799 /// @brief Transforms common to all three rem instructions
2800 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2801 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2803 // 0 % X == 0 for integer, we don't need to preserve faults!
2804 if (Constant *LHS = dyn_cast<Constant>(Op0))
2805 if (LHS->isNullValue())
2806 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2808 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2809 if (I.getType()->isFPOrFPVector())
2810 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2811 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2813 if (isa<UndefValue>(Op1))
2814 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2816 // Handle cases involving: rem X, (select Cond, Y, Z)
2817 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2818 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2819 // the same basic block, then we replace the select with Y, and the
2820 // condition of the select with false (if the cond value is in the same
2821 // BB). If the select has uses other than the div, this allows them to be
2823 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2824 if (ST->isNullValue()) {
2825 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2826 if (CondI && CondI->getParent() == I.getParent())
2827 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2828 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2829 I.setOperand(1, SI->getOperand(2));
2831 UpdateValueUsesWith(SI, SI->getOperand(2));
2834 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2835 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2836 if (ST->isNullValue()) {
2837 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2838 if (CondI && CondI->getParent() == I.getParent())
2839 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2840 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2841 I.setOperand(1, SI->getOperand(1));
2843 UpdateValueUsesWith(SI, SI->getOperand(1));
2851 /// This function implements the transforms common to both integer remainder
2852 /// instructions (urem and srem). It is called by the visitors to those integer
2853 /// remainder instructions.
2854 /// @brief Common integer remainder transforms
2855 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2856 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2858 if (Instruction *common = commonRemTransforms(I))
2861 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2862 // X % 0 == undef, we don't need to preserve faults!
2863 if (RHS->equalsInt(0))
2864 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2866 if (RHS->equalsInt(1)) // X % 1 == 0
2867 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2869 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2870 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2871 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2873 } else if (isa<PHINode>(Op0I)) {
2874 if (Instruction *NV = FoldOpIntoPhi(I))
2878 // See if we can fold away this rem instruction.
2879 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2880 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2881 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2882 KnownZero, KnownOne))
2890 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2891 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2893 if (Instruction *common = commonIRemTransforms(I))
2896 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2897 // X urem C^2 -> X and C
2898 // Check to see if this is an unsigned remainder with an exact power of 2,
2899 // if so, convert to a bitwise and.
2900 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2901 if (C->getValue().isPowerOf2())
2902 return BinaryOperator::CreateAnd(Op0, SubOne(C));
2905 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2906 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2907 if (RHSI->getOpcode() == Instruction::Shl &&
2908 isa<ConstantInt>(RHSI->getOperand(0))) {
2909 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2910 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2911 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
2913 return BinaryOperator::CreateAnd(Op0, Add);
2918 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2919 // where C1&C2 are powers of two.
2920 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2921 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2922 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2923 // STO == 0 and SFO == 0 handled above.
2924 if ((STO->getValue().isPowerOf2()) &&
2925 (SFO->getValue().isPowerOf2())) {
2926 Value *TrueAnd = InsertNewInstBefore(
2927 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2928 Value *FalseAnd = InsertNewInstBefore(
2929 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2930 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
2938 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2939 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2941 // Handle the integer rem common cases
2942 if (Instruction *common = commonIRemTransforms(I))
2945 if (Value *RHSNeg = dyn_castNegVal(Op1))
2946 if (!isa<ConstantInt>(RHSNeg) ||
2947 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2949 AddUsesToWorkList(I);
2950 I.setOperand(1, RHSNeg);
2954 // If the sign bits of both operands are zero (i.e. we can prove they are
2955 // unsigned inputs), turn this into a urem.
2956 if (I.getType()->isInteger()) {
2957 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2958 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2959 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2960 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
2967 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2968 return commonRemTransforms(I);
2971 // isMaxValueMinusOne - return true if this is Max-1
2972 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2973 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2975 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2976 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2979 // isMinValuePlusOne - return true if this is Min+1
2980 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2982 return C->getValue() == 1; // unsigned
2984 // Calculate 1111111111000000000000
2985 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2986 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2989 // isOneBitSet - Return true if there is exactly one bit set in the specified
2991 static bool isOneBitSet(const ConstantInt *CI) {
2992 return CI->getValue().isPowerOf2();
2995 // isHighOnes - Return true if the constant is of the form 1+0+.
2996 // This is the same as lowones(~X).
2997 static bool isHighOnes(const ConstantInt *CI) {
2998 return (~CI->getValue() + 1).isPowerOf2();
3001 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3002 /// are carefully arranged to allow folding of expressions such as:
3004 /// (A < B) | (A > B) --> (A != B)
3006 /// Note that this is only valid if the first and second predicates have the
3007 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3009 /// Three bits are used to represent the condition, as follows:
3014 /// <=> Value Definition
3015 /// 000 0 Always false
3022 /// 111 7 Always true
3024 static unsigned getICmpCode(const ICmpInst *ICI) {
3025 switch (ICI->getPredicate()) {
3027 case ICmpInst::ICMP_UGT: return 1; // 001
3028 case ICmpInst::ICMP_SGT: return 1; // 001
3029 case ICmpInst::ICMP_EQ: return 2; // 010
3030 case ICmpInst::ICMP_UGE: return 3; // 011
3031 case ICmpInst::ICMP_SGE: return 3; // 011
3032 case ICmpInst::ICMP_ULT: return 4; // 100
3033 case ICmpInst::ICMP_SLT: return 4; // 100
3034 case ICmpInst::ICMP_NE: return 5; // 101
3035 case ICmpInst::ICMP_ULE: return 6; // 110
3036 case ICmpInst::ICMP_SLE: return 6; // 110
3039 assert(0 && "Invalid ICmp predicate!");
3044 /// getICmpValue - This is the complement of getICmpCode, which turns an
3045 /// opcode and two operands into either a constant true or false, or a brand
3046 /// new ICmp instruction. The sign is passed in to determine which kind
3047 /// of predicate to use in new icmp instructions.
3048 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3050 default: assert(0 && "Illegal ICmp code!");
3051 case 0: return ConstantInt::getFalse();
3054 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3056 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3057 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3060 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3062 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3065 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3067 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3068 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3071 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3073 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3074 case 7: return ConstantInt::getTrue();
3078 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3079 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3080 (ICmpInst::isSignedPredicate(p1) &&
3081 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3082 (ICmpInst::isSignedPredicate(p2) &&
3083 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3087 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3088 struct FoldICmpLogical {
3091 ICmpInst::Predicate pred;
3092 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3093 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3094 pred(ICI->getPredicate()) {}
3095 bool shouldApply(Value *V) const {
3096 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3097 if (PredicatesFoldable(pred, ICI->getPredicate()))
3098 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3099 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3102 Instruction *apply(Instruction &Log) const {
3103 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3104 if (ICI->getOperand(0) != LHS) {
3105 assert(ICI->getOperand(1) == LHS);
3106 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3109 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3110 unsigned LHSCode = getICmpCode(ICI);
3111 unsigned RHSCode = getICmpCode(RHSICI);
3113 switch (Log.getOpcode()) {
3114 case Instruction::And: Code = LHSCode & RHSCode; break;
3115 case Instruction::Or: Code = LHSCode | RHSCode; break;
3116 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3117 default: assert(0 && "Illegal logical opcode!"); return 0;
3120 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3121 ICmpInst::isSignedPredicate(ICI->getPredicate());
3123 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3124 if (Instruction *I = dyn_cast<Instruction>(RV))
3126 // Otherwise, it's a constant boolean value...
3127 return IC.ReplaceInstUsesWith(Log, RV);
3130 } // end anonymous namespace
3132 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3133 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3134 // guaranteed to be a binary operator.
3135 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3137 ConstantInt *AndRHS,
3138 BinaryOperator &TheAnd) {
3139 Value *X = Op->getOperand(0);
3140 Constant *Together = 0;
3142 Together = And(AndRHS, OpRHS);
3144 switch (Op->getOpcode()) {
3145 case Instruction::Xor:
3146 if (Op->hasOneUse()) {
3147 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3148 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3149 InsertNewInstBefore(And, TheAnd);
3151 return BinaryOperator::CreateXor(And, Together);
3154 case Instruction::Or:
3155 if (Together == AndRHS) // (X | C) & C --> C
3156 return ReplaceInstUsesWith(TheAnd, AndRHS);
3158 if (Op->hasOneUse() && Together != OpRHS) {
3159 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3160 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3161 InsertNewInstBefore(Or, TheAnd);
3163 return BinaryOperator::CreateAnd(Or, AndRHS);
3166 case Instruction::Add:
3167 if (Op->hasOneUse()) {
3168 // Adding a one to a single bit bit-field should be turned into an XOR
3169 // of the bit. First thing to check is to see if this AND is with a
3170 // single bit constant.
3171 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3173 // If there is only one bit set...
3174 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3175 // Ok, at this point, we know that we are masking the result of the
3176 // ADD down to exactly one bit. If the constant we are adding has
3177 // no bits set below this bit, then we can eliminate the ADD.
3178 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3180 // Check to see if any bits below the one bit set in AndRHSV are set.
3181 if ((AddRHS & (AndRHSV-1)) == 0) {
3182 // If not, the only thing that can effect the output of the AND is
3183 // the bit specified by AndRHSV. If that bit is set, the effect of
3184 // the XOR is to toggle the bit. If it is clear, then the ADD has
3186 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3187 TheAnd.setOperand(0, X);
3190 // Pull the XOR out of the AND.
3191 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3192 InsertNewInstBefore(NewAnd, TheAnd);
3193 NewAnd->takeName(Op);
3194 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3201 case Instruction::Shl: {
3202 // We know that the AND will not produce any of the bits shifted in, so if
3203 // the anded constant includes them, clear them now!
3205 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3206 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3207 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3208 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3210 if (CI->getValue() == ShlMask) {
3211 // Masking out bits that the shift already masks
3212 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3213 } else if (CI != AndRHS) { // Reducing bits set in and.
3214 TheAnd.setOperand(1, CI);
3219 case Instruction::LShr:
3221 // We know that the AND will not produce any of the bits shifted in, so if
3222 // the anded constant includes them, clear them now! This only applies to
3223 // unsigned shifts, because a signed shr may bring in set bits!
3225 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3226 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3227 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3228 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3230 if (CI->getValue() == ShrMask) {
3231 // Masking out bits that the shift already masks.
3232 return ReplaceInstUsesWith(TheAnd, Op);
3233 } else if (CI != AndRHS) {
3234 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3239 case Instruction::AShr:
3241 // See if this is shifting in some sign extension, then masking it out
3243 if (Op->hasOneUse()) {
3244 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3245 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3246 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3247 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3248 if (C == AndRHS) { // Masking out bits shifted in.
3249 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3250 // Make the argument unsigned.
3251 Value *ShVal = Op->getOperand(0);
3252 ShVal = InsertNewInstBefore(
3253 BinaryOperator::CreateLShr(ShVal, OpRHS,
3254 Op->getName()), TheAnd);
3255 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3264 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3265 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3266 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3267 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3268 /// insert new instructions.
3269 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3270 bool isSigned, bool Inside,
3272 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3273 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3274 "Lo is not <= Hi in range emission code!");
3277 if (Lo == Hi) // Trivially false.
3278 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3280 // V >= Min && V < Hi --> V < Hi
3281 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3282 ICmpInst::Predicate pred = (isSigned ?
3283 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3284 return new ICmpInst(pred, V, Hi);
3287 // Emit V-Lo <u Hi-Lo
3288 Constant *NegLo = ConstantExpr::getNeg(Lo);
3289 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3290 InsertNewInstBefore(Add, IB);
3291 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3292 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3295 if (Lo == Hi) // Trivially true.
3296 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3298 // V < Min || V >= Hi -> V > Hi-1
3299 Hi = SubOne(cast<ConstantInt>(Hi));
3300 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3301 ICmpInst::Predicate pred = (isSigned ?
3302 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3303 return new ICmpInst(pred, V, Hi);
3306 // Emit V-Lo >u Hi-1-Lo
3307 // Note that Hi has already had one subtracted from it, above.
3308 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3309 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3310 InsertNewInstBefore(Add, IB);
3311 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3312 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3315 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3316 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3317 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3318 // not, since all 1s are not contiguous.
3319 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3320 const APInt& V = Val->getValue();
3321 uint32_t BitWidth = Val->getType()->getBitWidth();
3322 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3324 // look for the first zero bit after the run of ones
3325 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3326 // look for the first non-zero bit
3327 ME = V.getActiveBits();
3331 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3332 /// where isSub determines whether the operator is a sub. If we can fold one of
3333 /// the following xforms:
3335 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3336 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3337 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3339 /// return (A +/- B).
3341 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3342 ConstantInt *Mask, bool isSub,
3344 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3345 if (!LHSI || LHSI->getNumOperands() != 2 ||
3346 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3348 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3350 switch (LHSI->getOpcode()) {
3352 case Instruction::And:
3353 if (And(N, Mask) == Mask) {
3354 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3355 if ((Mask->getValue().countLeadingZeros() +
3356 Mask->getValue().countPopulation()) ==
3357 Mask->getValue().getBitWidth())
3360 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3361 // part, we don't need any explicit masks to take them out of A. If that
3362 // is all N is, ignore it.
3363 uint32_t MB = 0, ME = 0;
3364 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3365 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3366 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3367 if (MaskedValueIsZero(RHS, Mask))
3372 case Instruction::Or:
3373 case Instruction::Xor:
3374 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3375 if ((Mask->getValue().countLeadingZeros() +
3376 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3377 && And(N, Mask)->isZero())
3384 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3386 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3387 return InsertNewInstBefore(New, I);
3390 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3391 bool Changed = SimplifyCommutative(I);
3392 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3394 if (isa<UndefValue>(Op1)) // X & undef -> 0
3395 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3399 return ReplaceInstUsesWith(I, Op1);
3401 // See if we can simplify any instructions used by the instruction whose sole
3402 // purpose is to compute bits we don't care about.
3403 if (!isa<VectorType>(I.getType())) {
3404 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3405 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3406 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3407 KnownZero, KnownOne))
3410 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3411 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3412 return ReplaceInstUsesWith(I, I.getOperand(0));
3413 } else if (isa<ConstantAggregateZero>(Op1)) {
3414 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3418 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3419 const APInt& AndRHSMask = AndRHS->getValue();
3420 APInt NotAndRHS(~AndRHSMask);
3422 // Optimize a variety of ((val OP C1) & C2) combinations...
3423 if (isa<BinaryOperator>(Op0)) {
3424 Instruction *Op0I = cast<Instruction>(Op0);
3425 Value *Op0LHS = Op0I->getOperand(0);
3426 Value *Op0RHS = Op0I->getOperand(1);
3427 switch (Op0I->getOpcode()) {
3428 case Instruction::Xor:
3429 case Instruction::Or:
3430 // If the mask is only needed on one incoming arm, push it up.
3431 if (Op0I->hasOneUse()) {
3432 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3433 // Not masking anything out for the LHS, move to RHS.
3434 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3435 Op0RHS->getName()+".masked");
3436 InsertNewInstBefore(NewRHS, I);
3437 return BinaryOperator::Create(
3438 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3440 if (!isa<Constant>(Op0RHS) &&
3441 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3442 // Not masking anything out for the RHS, move to LHS.
3443 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3444 Op0LHS->getName()+".masked");
3445 InsertNewInstBefore(NewLHS, I);
3446 return BinaryOperator::Create(
3447 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3452 case Instruction::Add:
3453 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3454 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3455 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3456 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3457 return BinaryOperator::CreateAnd(V, AndRHS);
3458 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3459 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3462 case Instruction::Sub:
3463 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3464 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3465 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3466 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3467 return BinaryOperator::CreateAnd(V, AndRHS);
3471 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3472 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3474 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3475 // If this is an integer truncation or change from signed-to-unsigned, and
3476 // if the source is an and/or with immediate, transform it. This
3477 // frequently occurs for bitfield accesses.
3478 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3479 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3480 CastOp->getNumOperands() == 2)
3481 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3482 if (CastOp->getOpcode() == Instruction::And) {
3483 // Change: and (cast (and X, C1) to T), C2
3484 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3485 // This will fold the two constants together, which may allow
3486 // other simplifications.
3487 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3488 CastOp->getOperand(0), I.getType(),
3489 CastOp->getName()+".shrunk");
3490 NewCast = InsertNewInstBefore(NewCast, I);
3491 // trunc_or_bitcast(C1)&C2
3492 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3493 C3 = ConstantExpr::getAnd(C3, AndRHS);
3494 return BinaryOperator::CreateAnd(NewCast, C3);
3495 } else if (CastOp->getOpcode() == Instruction::Or) {
3496 // Change: and (cast (or X, C1) to T), C2
3497 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3498 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3499 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3500 return ReplaceInstUsesWith(I, AndRHS);
3506 // Try to fold constant and into select arguments.
3507 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3508 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3510 if (isa<PHINode>(Op0))
3511 if (Instruction *NV = FoldOpIntoPhi(I))
3515 Value *Op0NotVal = dyn_castNotVal(Op0);
3516 Value *Op1NotVal = dyn_castNotVal(Op1);
3518 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3519 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3521 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3522 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3523 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3524 I.getName()+".demorgan");
3525 InsertNewInstBefore(Or, I);
3526 return BinaryOperator::CreateNot(Or);
3530 Value *A = 0, *B = 0, *C = 0, *D = 0;
3531 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3532 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3533 return ReplaceInstUsesWith(I, Op1);
3535 // (A|B) & ~(A&B) -> A^B
3536 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3537 if ((A == C && B == D) || (A == D && B == C))
3538 return BinaryOperator::CreateXor(A, B);
3542 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3543 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3544 return ReplaceInstUsesWith(I, Op0);
3546 // ~(A&B) & (A|B) -> A^B
3547 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3548 if ((A == C && B == D) || (A == D && B == C))
3549 return BinaryOperator::CreateXor(A, B);
3553 if (Op0->hasOneUse() &&
3554 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3555 if (A == Op1) { // (A^B)&A -> A&(A^B)
3556 I.swapOperands(); // Simplify below
3557 std::swap(Op0, Op1);
3558 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3559 cast<BinaryOperator>(Op0)->swapOperands();
3560 I.swapOperands(); // Simplify below
3561 std::swap(Op0, Op1);
3564 if (Op1->hasOneUse() &&
3565 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3566 if (B == Op0) { // B&(A^B) -> B&(B^A)
3567 cast<BinaryOperator>(Op1)->swapOperands();
3570 if (A == Op0) { // A&(A^B) -> A & ~B
3571 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3572 InsertNewInstBefore(NotB, I);
3573 return BinaryOperator::CreateAnd(A, NotB);
3578 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3579 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3580 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3583 Value *LHSVal, *RHSVal;
3584 ConstantInt *LHSCst, *RHSCst;
3585 ICmpInst::Predicate LHSCC, RHSCC;
3586 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3587 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3588 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3589 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3590 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3591 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3592 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3593 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3595 // Don't try to fold ICMP_SLT + ICMP_ULT.
3596 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3597 ICmpInst::isSignedPredicate(LHSCC) ==
3598 ICmpInst::isSignedPredicate(RHSCC))) {
3599 // Ensure that the larger constant is on the RHS.
3600 ICmpInst::Predicate GT;
3601 if (ICmpInst::isSignedPredicate(LHSCC) ||
3602 (ICmpInst::isEquality(LHSCC) &&
3603 ICmpInst::isSignedPredicate(RHSCC)))
3604 GT = ICmpInst::ICMP_SGT;
3606 GT = ICmpInst::ICMP_UGT;
3608 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3609 ICmpInst *LHS = cast<ICmpInst>(Op0);
3610 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3611 std::swap(LHS, RHS);
3612 std::swap(LHSCst, RHSCst);
3613 std::swap(LHSCC, RHSCC);
3616 // At this point, we know we have have two icmp instructions
3617 // comparing a value against two constants and and'ing the result
3618 // together. Because of the above check, we know that we only have
3619 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3620 // (from the FoldICmpLogical check above), that the two constants
3621 // are not equal and that the larger constant is on the RHS
3622 assert(LHSCst != RHSCst && "Compares not folded above?");
3625 default: assert(0 && "Unknown integer condition code!");
3626 case ICmpInst::ICMP_EQ:
3628 default: assert(0 && "Unknown integer condition code!");
3629 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3630 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3631 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3632 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3633 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3634 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3635 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3636 return ReplaceInstUsesWith(I, LHS);
3638 case ICmpInst::ICMP_NE:
3640 default: assert(0 && "Unknown integer condition code!");
3641 case ICmpInst::ICMP_ULT:
3642 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3643 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3644 break; // (X != 13 & X u< 15) -> no change
3645 case ICmpInst::ICMP_SLT:
3646 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3647 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3648 break; // (X != 13 & X s< 15) -> no change
3649 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3650 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3651 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3652 return ReplaceInstUsesWith(I, RHS);
3653 case ICmpInst::ICMP_NE:
3654 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3655 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3656 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
3657 LHSVal->getName()+".off");
3658 InsertNewInstBefore(Add, I);
3659 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3660 ConstantInt::get(Add->getType(), 1));
3662 break; // (X != 13 & X != 15) -> no change
3665 case ICmpInst::ICMP_ULT:
3667 default: assert(0 && "Unknown integer condition code!");
3668 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3669 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3670 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3671 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3673 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3674 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3675 return ReplaceInstUsesWith(I, LHS);
3676 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3680 case ICmpInst::ICMP_SLT:
3682 default: assert(0 && "Unknown integer condition code!");
3683 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3684 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3685 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3686 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3688 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3689 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3690 return ReplaceInstUsesWith(I, LHS);
3691 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3695 case ICmpInst::ICMP_UGT:
3697 default: assert(0 && "Unknown integer condition code!");
3698 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3699 return ReplaceInstUsesWith(I, LHS);
3700 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3701 return ReplaceInstUsesWith(I, RHS);
3702 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3704 case ICmpInst::ICMP_NE:
3705 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3706 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3707 break; // (X u> 13 & X != 15) -> no change
3708 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3709 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3711 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3715 case ICmpInst::ICMP_SGT:
3717 default: assert(0 && "Unknown integer condition code!");
3718 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3719 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3720 return ReplaceInstUsesWith(I, RHS);
3721 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3723 case ICmpInst::ICMP_NE:
3724 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3725 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3726 break; // (X s> 13 & X != 15) -> no change
3727 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3728 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3730 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3738 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3739 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3740 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3741 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3742 const Type *SrcTy = Op0C->getOperand(0)->getType();
3743 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3744 // Only do this if the casts both really cause code to be generated.
3745 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3747 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3749 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3750 Op1C->getOperand(0),
3752 InsertNewInstBefore(NewOp, I);
3753 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3757 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3758 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3759 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3760 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3761 SI0->getOperand(1) == SI1->getOperand(1) &&
3762 (SI0->hasOneUse() || SI1->hasOneUse())) {
3763 Instruction *NewOp =
3764 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3766 SI0->getName()), I);
3767 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3768 SI1->getOperand(1));
3772 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3773 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3774 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3775 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3776 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3777 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3778 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3779 // If either of the constants are nans, then the whole thing returns
3781 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3782 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3783 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3784 RHS->getOperand(0));
3789 return Changed ? &I : 0;
3792 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3793 /// in the result. If it does, and if the specified byte hasn't been filled in
3794 /// yet, fill it in and return false.
3795 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3796 Instruction *I = dyn_cast<Instruction>(V);
3797 if (I == 0) return true;
3799 // If this is an or instruction, it is an inner node of the bswap.
3800 if (I->getOpcode() == Instruction::Or)
3801 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3802 CollectBSwapParts(I->getOperand(1), ByteValues);
3804 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3805 // If this is a shift by a constant int, and it is "24", then its operand
3806 // defines a byte. We only handle unsigned types here.
3807 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3808 // Not shifting the entire input by N-1 bytes?
3809 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3810 8*(ByteValues.size()-1))
3814 if (I->getOpcode() == Instruction::Shl) {
3815 // X << 24 defines the top byte with the lowest of the input bytes.
3816 DestNo = ByteValues.size()-1;
3818 // X >>u 24 defines the low byte with the highest of the input bytes.
3822 // If the destination byte value is already defined, the values are or'd
3823 // together, which isn't a bswap (unless it's an or of the same bits).
3824 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3826 ByteValues[DestNo] = I->getOperand(0);
3830 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3832 Value *Shift = 0, *ShiftLHS = 0;
3833 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3834 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3835 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3837 Instruction *SI = cast<Instruction>(Shift);
3839 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3840 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3841 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3844 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3846 if (AndAmt->getValue().getActiveBits() > 64)
3848 uint64_t AndAmtVal = AndAmt->getZExtValue();
3849 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3850 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3852 // Unknown mask for bswap.
3853 if (DestByte == ByteValues.size()) return true;
3855 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3857 if (SI->getOpcode() == Instruction::Shl)
3858 SrcByte = DestByte - ShiftBytes;
3860 SrcByte = DestByte + ShiftBytes;
3862 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3863 if (SrcByte != ByteValues.size()-DestByte-1)
3866 // If the destination byte value is already defined, the values are or'd
3867 // together, which isn't a bswap (unless it's an or of the same bits).
3868 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3870 ByteValues[DestByte] = SI->getOperand(0);
3874 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3875 /// If so, insert the new bswap intrinsic and return it.
3876 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3877 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3878 if (!ITy || ITy->getBitWidth() % 16)
3879 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3881 /// ByteValues - For each byte of the result, we keep track of which value
3882 /// defines each byte.
3883 SmallVector<Value*, 8> ByteValues;
3884 ByteValues.resize(ITy->getBitWidth()/8);
3886 // Try to find all the pieces corresponding to the bswap.
3887 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3888 CollectBSwapParts(I.getOperand(1), ByteValues))
3891 // Check to see if all of the bytes come from the same value.
3892 Value *V = ByteValues[0];
3893 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3895 // Check to make sure that all of the bytes come from the same value.
3896 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3897 if (ByteValues[i] != V)
3899 const Type *Tys[] = { ITy };
3900 Module *M = I.getParent()->getParent()->getParent();
3901 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3902 return CallInst::Create(F, V);
3906 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3907 bool Changed = SimplifyCommutative(I);
3908 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3910 if (isa<UndefValue>(Op1)) // X | undef -> -1
3911 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3915 return ReplaceInstUsesWith(I, Op0);
3917 // See if we can simplify any instructions used by the instruction whose sole
3918 // purpose is to compute bits we don't care about.
3919 if (!isa<VectorType>(I.getType())) {
3920 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3921 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3922 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3923 KnownZero, KnownOne))
3925 } else if (isa<ConstantAggregateZero>(Op1)) {
3926 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3927 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3928 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3929 return ReplaceInstUsesWith(I, I.getOperand(1));
3935 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3936 ConstantInt *C1 = 0; Value *X = 0;
3937 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3938 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3939 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3940 InsertNewInstBefore(Or, I);
3942 return BinaryOperator::CreateAnd(Or,
3943 ConstantInt::get(RHS->getValue() | C1->getValue()));
3946 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3947 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3948 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
3949 InsertNewInstBefore(Or, I);
3951 return BinaryOperator::CreateXor(Or,
3952 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3955 // Try to fold constant and into select arguments.
3956 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3957 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3959 if (isa<PHINode>(Op0))
3960 if (Instruction *NV = FoldOpIntoPhi(I))
3964 Value *A = 0, *B = 0;
3965 ConstantInt *C1 = 0, *C2 = 0;
3967 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3968 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3969 return ReplaceInstUsesWith(I, Op1);
3970 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3971 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3972 return ReplaceInstUsesWith(I, Op0);
3974 // (A | B) | C and A | (B | C) -> bswap if possible.
3975 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3976 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3977 match(Op1, m_Or(m_Value(), m_Value())) ||
3978 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3979 match(Op1, m_Shift(m_Value(), m_Value())))) {
3980 if (Instruction *BSwap = MatchBSwap(I))
3984 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3985 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3986 MaskedValueIsZero(Op1, C1->getValue())) {
3987 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
3988 InsertNewInstBefore(NOr, I);
3990 return BinaryOperator::CreateXor(NOr, C1);
3993 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3994 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3995 MaskedValueIsZero(Op0, C1->getValue())) {
3996 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
3997 InsertNewInstBefore(NOr, I);
3999 return BinaryOperator::CreateXor(NOr, C1);
4003 Value *C = 0, *D = 0;
4004 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4005 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4006 Value *V1 = 0, *V2 = 0, *V3 = 0;
4007 C1 = dyn_cast<ConstantInt>(C);
4008 C2 = dyn_cast<ConstantInt>(D);
4009 if (C1 && C2) { // (A & C1)|(B & C2)
4010 // If we have: ((V + N) & C1) | (V & C2)
4011 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4012 // replace with V+N.
4013 if (C1->getValue() == ~C2->getValue()) {
4014 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4015 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4016 // Add commutes, try both ways.
4017 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4018 return ReplaceInstUsesWith(I, A);
4019 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4020 return ReplaceInstUsesWith(I, A);
4022 // Or commutes, try both ways.
4023 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4024 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4025 // Add commutes, try both ways.
4026 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4027 return ReplaceInstUsesWith(I, B);
4028 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4029 return ReplaceInstUsesWith(I, B);
4032 V1 = 0; V2 = 0; V3 = 0;
4035 // Check to see if we have any common things being and'ed. If so, find the
4036 // terms for V1 & (V2|V3).
4037 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4038 if (A == B) // (A & C)|(A & D) == A & (C|D)
4039 V1 = A, V2 = C, V3 = D;
4040 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4041 V1 = A, V2 = B, V3 = C;
4042 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4043 V1 = C, V2 = A, V3 = D;
4044 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4045 V1 = C, V2 = A, V3 = B;
4049 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4050 return BinaryOperator::CreateAnd(V1, Or);
4055 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4056 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4057 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4058 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4059 SI0->getOperand(1) == SI1->getOperand(1) &&
4060 (SI0->hasOneUse() || SI1->hasOneUse())) {
4061 Instruction *NewOp =
4062 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4064 SI0->getName()), I);
4065 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4066 SI1->getOperand(1));
4070 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4071 if (A == Op1) // ~A | A == -1
4072 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4076 // Note, A is still live here!
4077 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4079 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4081 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4082 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4083 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4084 I.getName()+".demorgan"), I);
4085 return BinaryOperator::CreateNot(And);
4089 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4090 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4091 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4094 Value *LHSVal, *RHSVal;
4095 ConstantInt *LHSCst, *RHSCst;
4096 ICmpInst::Predicate LHSCC, RHSCC;
4097 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4098 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4099 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4100 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4101 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4102 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4103 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4104 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4105 // We can't fold (ugt x, C) | (sgt x, C2).
4106 PredicatesFoldable(LHSCC, RHSCC)) {
4107 // Ensure that the larger constant is on the RHS.
4108 ICmpInst *LHS = cast<ICmpInst>(Op0);
4110 if (ICmpInst::isSignedPredicate(LHSCC))
4111 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4113 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4116 std::swap(LHS, RHS);
4117 std::swap(LHSCst, RHSCst);
4118 std::swap(LHSCC, RHSCC);
4121 // At this point, we know we have have two icmp instructions
4122 // comparing a value against two constants and or'ing the result
4123 // together. Because of the above check, we know that we only have
4124 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4125 // FoldICmpLogical check above), that the two constants are not
4127 assert(LHSCst != RHSCst && "Compares not folded above?");
4130 default: assert(0 && "Unknown integer condition code!");
4131 case ICmpInst::ICMP_EQ:
4133 default: assert(0 && "Unknown integer condition code!");
4134 case ICmpInst::ICMP_EQ:
4135 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4136 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4137 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4138 LHSVal->getName()+".off");
4139 InsertNewInstBefore(Add, I);
4140 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4141 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4143 break; // (X == 13 | X == 15) -> no change
4144 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4145 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4147 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4148 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4149 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4150 return ReplaceInstUsesWith(I, RHS);
4153 case ICmpInst::ICMP_NE:
4155 default: assert(0 && "Unknown integer condition code!");
4156 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4157 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4158 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4159 return ReplaceInstUsesWith(I, LHS);
4160 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4161 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4162 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4163 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4166 case ICmpInst::ICMP_ULT:
4168 default: assert(0 && "Unknown integer condition code!");
4169 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4171 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4172 // If RHSCst is [us]MAXINT, it is always false. Not handling
4173 // this can cause overflow.
4174 if (RHSCst->isMaxValue(false))
4175 return ReplaceInstUsesWith(I, LHS);
4176 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4178 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4180 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4181 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4182 return ReplaceInstUsesWith(I, RHS);
4183 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4187 case ICmpInst::ICMP_SLT:
4189 default: assert(0 && "Unknown integer condition code!");
4190 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4192 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4193 // If RHSCst is [us]MAXINT, it is always false. Not handling
4194 // this can cause overflow.
4195 if (RHSCst->isMaxValue(true))
4196 return ReplaceInstUsesWith(I, LHS);
4197 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4199 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4201 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4202 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4203 return ReplaceInstUsesWith(I, RHS);
4204 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4208 case ICmpInst::ICMP_UGT:
4210 default: assert(0 && "Unknown integer condition code!");
4211 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4212 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4213 return ReplaceInstUsesWith(I, LHS);
4214 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4216 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4217 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4218 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4219 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4223 case ICmpInst::ICMP_SGT:
4225 default: assert(0 && "Unknown integer condition code!");
4226 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4227 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4228 return ReplaceInstUsesWith(I, LHS);
4229 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4231 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4232 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4233 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4234 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4242 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4243 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4244 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4245 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4246 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4247 !isa<ICmpInst>(Op1C->getOperand(0))) {
4248 const Type *SrcTy = Op0C->getOperand(0)->getType();
4249 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4250 // Only do this if the casts both really cause code to be
4252 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4254 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4256 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4257 Op1C->getOperand(0),
4259 InsertNewInstBefore(NewOp, I);
4260 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4267 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4268 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4269 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4270 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4271 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4272 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4273 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4274 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4275 // If either of the constants are nans, then the whole thing returns
4277 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4278 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4280 // Otherwise, no need to compare the two constants, compare the
4282 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4283 RHS->getOperand(0));
4288 return Changed ? &I : 0;
4293 // XorSelf - Implements: X ^ X --> 0
4296 XorSelf(Value *rhs) : RHS(rhs) {}
4297 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4298 Instruction *apply(BinaryOperator &Xor) const {
4305 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4306 bool Changed = SimplifyCommutative(I);
4307 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4309 if (isa<UndefValue>(Op1)) {
4310 if (isa<UndefValue>(Op0))
4311 // Handle undef ^ undef -> 0 special case. This is a common
4313 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4314 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4317 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4318 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4319 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4320 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4323 // See if we can simplify any instructions used by the instruction whose sole
4324 // purpose is to compute bits we don't care about.
4325 if (!isa<VectorType>(I.getType())) {
4326 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4327 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4328 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4329 KnownZero, KnownOne))
4331 } else if (isa<ConstantAggregateZero>(Op1)) {
4332 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4335 // Is this a ~ operation?
4336 if (Value *NotOp = dyn_castNotVal(&I)) {
4337 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4338 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4339 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4340 if (Op0I->getOpcode() == Instruction::And ||
4341 Op0I->getOpcode() == Instruction::Or) {
4342 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4343 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4345 BinaryOperator::CreateNot(Op0I->getOperand(1),
4346 Op0I->getOperand(1)->getName()+".not");
4347 InsertNewInstBefore(NotY, I);
4348 if (Op0I->getOpcode() == Instruction::And)
4349 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4351 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4358 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4359 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4360 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4361 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4362 return new ICmpInst(ICI->getInversePredicate(),
4363 ICI->getOperand(0), ICI->getOperand(1));
4365 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4366 return new FCmpInst(FCI->getInversePredicate(),
4367 FCI->getOperand(0), FCI->getOperand(1));
4370 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4371 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4372 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4373 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4374 Instruction::CastOps Opcode = Op0C->getOpcode();
4375 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4376 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4377 Op0C->getDestTy())) {
4378 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4379 CI->getOpcode(), CI->getInversePredicate(),
4380 CI->getOperand(0), CI->getOperand(1)), I);
4381 NewCI->takeName(CI);
4382 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4389 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4390 // ~(c-X) == X-c-1 == X+(-c-1)
4391 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4392 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4393 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4394 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4395 ConstantInt::get(I.getType(), 1));
4396 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4399 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4400 if (Op0I->getOpcode() == Instruction::Add) {
4401 // ~(X-c) --> (-c-1)-X
4402 if (RHS->isAllOnesValue()) {
4403 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4404 return BinaryOperator::CreateSub(
4405 ConstantExpr::getSub(NegOp0CI,
4406 ConstantInt::get(I.getType(), 1)),
4407 Op0I->getOperand(0));
4408 } else if (RHS->getValue().isSignBit()) {
4409 // (X + C) ^ signbit -> (X + C + signbit)
4410 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4411 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4414 } else if (Op0I->getOpcode() == Instruction::Or) {
4415 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4416 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4417 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4418 // Anything in both C1 and C2 is known to be zero, remove it from
4420 Constant *CommonBits = And(Op0CI, RHS);
4421 NewRHS = ConstantExpr::getAnd(NewRHS,
4422 ConstantExpr::getNot(CommonBits));
4423 AddToWorkList(Op0I);
4424 I.setOperand(0, Op0I->getOperand(0));
4425 I.setOperand(1, NewRHS);
4432 // Try to fold constant and into select arguments.
4433 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4434 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4436 if (isa<PHINode>(Op0))
4437 if (Instruction *NV = FoldOpIntoPhi(I))
4441 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4443 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4445 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4447 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4450 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4453 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4454 if (A == Op0) { // B^(B|A) == (A|B)^B
4455 Op1I->swapOperands();
4457 std::swap(Op0, Op1);
4458 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4459 I.swapOperands(); // Simplified below.
4460 std::swap(Op0, Op1);
4462 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4463 if (Op0 == A) // A^(A^B) == B
4464 return ReplaceInstUsesWith(I, B);
4465 else if (Op0 == B) // A^(B^A) == B
4466 return ReplaceInstUsesWith(I, A);
4467 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4468 if (A == Op0) { // A^(A&B) -> A^(B&A)
4469 Op1I->swapOperands();
4472 if (B == Op0) { // A^(B&A) -> (B&A)^A
4473 I.swapOperands(); // Simplified below.
4474 std::swap(Op0, Op1);
4479 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4482 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4483 if (A == Op1) // (B|A)^B == (A|B)^B
4485 if (B == Op1) { // (A|B)^B == A & ~B
4487 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4488 return BinaryOperator::CreateAnd(A, NotB);
4490 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4491 if (Op1 == A) // (A^B)^A == B
4492 return ReplaceInstUsesWith(I, B);
4493 else if (Op1 == B) // (B^A)^A == B
4494 return ReplaceInstUsesWith(I, A);
4495 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4496 if (A == Op1) // (A&B)^A -> (B&A)^A
4498 if (B == Op1 && // (B&A)^A == ~B & A
4499 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4501 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4502 return BinaryOperator::CreateAnd(N, Op1);
4507 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4508 if (Op0I && Op1I && Op0I->isShift() &&
4509 Op0I->getOpcode() == Op1I->getOpcode() &&
4510 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4511 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4512 Instruction *NewOp =
4513 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4514 Op1I->getOperand(0),
4515 Op0I->getName()), I);
4516 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4517 Op1I->getOperand(1));
4521 Value *A, *B, *C, *D;
4522 // (A & B)^(A | B) -> A ^ B
4523 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4524 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4525 if ((A == C && B == D) || (A == D && B == C))
4526 return BinaryOperator::CreateXor(A, B);
4528 // (A | B)^(A & B) -> A ^ B
4529 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4530 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4531 if ((A == C && B == D) || (A == D && B == C))
4532 return BinaryOperator::CreateXor(A, B);
4536 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4537 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4538 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4539 // (X & Y)^(X & Y) -> (Y^Z) & X
4540 Value *X = 0, *Y = 0, *Z = 0;
4542 X = A, Y = B, Z = D;
4544 X = A, Y = B, Z = C;
4546 X = B, Y = A, Z = D;
4548 X = B, Y = A, Z = C;
4551 Instruction *NewOp =
4552 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4553 return BinaryOperator::CreateAnd(NewOp, X);
4558 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4559 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4560 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4563 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4564 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4565 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4566 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4567 const Type *SrcTy = Op0C->getOperand(0)->getType();
4568 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4569 // Only do this if the casts both really cause code to be generated.
4570 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4572 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4574 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4575 Op1C->getOperand(0),
4577 InsertNewInstBefore(NewOp, I);
4578 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4583 return Changed ? &I : 0;
4586 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4587 /// overflowed for this type.
4588 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4589 ConstantInt *In2, bool IsSigned = false) {
4590 Result = cast<ConstantInt>(Add(In1, In2));
4593 if (In2->getValue().isNegative())
4594 return Result->getValue().sgt(In1->getValue());
4596 return Result->getValue().slt(In1->getValue());
4598 return Result->getValue().ult(In1->getValue());
4601 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4602 /// code necessary to compute the offset from the base pointer (without adding
4603 /// in the base pointer). Return the result as a signed integer of intptr size.
4604 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4605 TargetData &TD = IC.getTargetData();
4606 gep_type_iterator GTI = gep_type_begin(GEP);
4607 const Type *IntPtrTy = TD.getIntPtrType();
4608 Value *Result = Constant::getNullValue(IntPtrTy);
4610 // Build a mask for high order bits.
4611 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4612 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4614 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
4617 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4618 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4619 if (OpC->isZero()) continue;
4621 // Handle a struct index, which adds its field offset to the pointer.
4622 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4623 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4625 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4626 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4628 Result = IC.InsertNewInstBefore(
4629 BinaryOperator::CreateAdd(Result,
4630 ConstantInt::get(IntPtrTy, Size),
4631 GEP->getName()+".offs"), I);
4635 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4636 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4637 Scale = ConstantExpr::getMul(OC, Scale);
4638 if (Constant *RC = dyn_cast<Constant>(Result))
4639 Result = ConstantExpr::getAdd(RC, Scale);
4641 // Emit an add instruction.
4642 Result = IC.InsertNewInstBefore(
4643 BinaryOperator::CreateAdd(Result, Scale,
4644 GEP->getName()+".offs"), I);
4648 // Convert to correct type.
4649 if (Op->getType() != IntPtrTy) {
4650 if (Constant *OpC = dyn_cast<Constant>(Op))
4651 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4653 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4654 Op->getName()+".c"), I);
4657 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4658 if (Constant *OpC = dyn_cast<Constant>(Op))
4659 Op = ConstantExpr::getMul(OpC, Scale);
4660 else // We'll let instcombine(mul) convert this to a shl if possible.
4661 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
4662 GEP->getName()+".idx"), I);
4665 // Emit an add instruction.
4666 if (isa<Constant>(Op) && isa<Constant>(Result))
4667 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4668 cast<Constant>(Result));
4670 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
4671 GEP->getName()+".offs"), I);
4677 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4678 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4679 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4680 /// complex, and scales are involved. The above expression would also be legal
4681 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4682 /// later form is less amenable to optimization though, and we are allowed to
4683 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4685 /// If we can't emit an optimized form for this expression, this returns null.
4687 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4689 TargetData &TD = IC.getTargetData();
4690 gep_type_iterator GTI = gep_type_begin(GEP);
4692 // Check to see if this gep only has a single variable index. If so, and if
4693 // any constant indices are a multiple of its scale, then we can compute this
4694 // in terms of the scale of the variable index. For example, if the GEP
4695 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4696 // because the expression will cross zero at the same point.
4697 unsigned i, e = GEP->getNumOperands();
4699 for (i = 1; i != e; ++i, ++GTI) {
4700 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4701 // Compute the aggregate offset of constant indices.
4702 if (CI->isZero()) continue;
4704 // Handle a struct index, which adds its field offset to the pointer.
4705 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4706 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4708 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4709 Offset += Size*CI->getSExtValue();
4712 // Found our variable index.
4717 // If there are no variable indices, we must have a constant offset, just
4718 // evaluate it the general way.
4719 if (i == e) return 0;
4721 Value *VariableIdx = GEP->getOperand(i);
4722 // Determine the scale factor of the variable element. For example, this is
4723 // 4 if the variable index is into an array of i32.
4724 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4726 // Verify that there are no other variable indices. If so, emit the hard way.
4727 for (++i, ++GTI; i != e; ++i, ++GTI) {
4728 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4731 // Compute the aggregate offset of constant indices.
4732 if (CI->isZero()) continue;
4734 // Handle a struct index, which adds its field offset to the pointer.
4735 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4736 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4738 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4739 Offset += Size*CI->getSExtValue();
4743 // Okay, we know we have a single variable index, which must be a
4744 // pointer/array/vector index. If there is no offset, life is simple, return
4746 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4748 // Cast to intptrty in case a truncation occurs. If an extension is needed,
4749 // we don't need to bother extending: the extension won't affect where the
4750 // computation crosses zero.
4751 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
4752 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
4753 VariableIdx->getNameStart(), &I);
4757 // Otherwise, there is an index. The computation we will do will be modulo
4758 // the pointer size, so get it.
4759 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4761 Offset &= PtrSizeMask;
4762 VariableScale &= PtrSizeMask;
4764 // To do this transformation, any constant index must be a multiple of the
4765 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
4766 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
4767 // multiple of the variable scale.
4768 int64_t NewOffs = Offset / (int64_t)VariableScale;
4769 if (Offset != NewOffs*(int64_t)VariableScale)
4772 // Okay, we can do this evaluation. Start by converting the index to intptr.
4773 const Type *IntPtrTy = TD.getIntPtrType();
4774 if (VariableIdx->getType() != IntPtrTy)
4775 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
4777 VariableIdx->getNameStart(), &I);
4778 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
4779 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
4783 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4784 /// else. At this point we know that the GEP is on the LHS of the comparison.
4785 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4786 ICmpInst::Predicate Cond,
4788 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4790 // Look through bitcasts.
4791 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
4792 RHS = BCI->getOperand(0);
4794 Value *PtrBase = GEPLHS->getOperand(0);
4795 if (PtrBase == RHS) {
4796 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4797 // This transformation (ignoring the base and scales) is valid because we
4798 // know pointers can't overflow. See if we can output an optimized form.
4799 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
4801 // If not, synthesize the offset the hard way.
4803 Offset = EmitGEPOffset(GEPLHS, I, *this);
4804 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4805 Constant::getNullValue(Offset->getType()));
4806 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4807 // If the base pointers are different, but the indices are the same, just
4808 // compare the base pointer.
4809 if (PtrBase != GEPRHS->getOperand(0)) {
4810 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4811 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4812 GEPRHS->getOperand(0)->getType();
4814 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4815 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4816 IndicesTheSame = false;
4820 // If all indices are the same, just compare the base pointers.
4822 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4823 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4825 // Otherwise, the base pointers are different and the indices are
4826 // different, bail out.
4830 // If one of the GEPs has all zero indices, recurse.
4831 bool AllZeros = true;
4832 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4833 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4834 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4839 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4840 ICmpInst::getSwappedPredicate(Cond), I);
4842 // If the other GEP has all zero indices, recurse.
4844 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4845 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4846 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4851 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4853 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4854 // If the GEPs only differ by one index, compare it.
4855 unsigned NumDifferences = 0; // Keep track of # differences.
4856 unsigned DiffOperand = 0; // The operand that differs.
4857 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4858 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4859 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4860 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4861 // Irreconcilable differences.
4865 if (NumDifferences++) break;
4870 if (NumDifferences == 0) // SAME GEP?
4871 return ReplaceInstUsesWith(I, // No comparison is needed here.
4872 ConstantInt::get(Type::Int1Ty,
4873 ICmpInst::isTrueWhenEqual(Cond)));
4875 else if (NumDifferences == 1) {
4876 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4877 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4878 // Make sure we do a signed comparison here.
4879 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4883 // Only lower this if the icmp is the only user of the GEP or if we expect
4884 // the result to fold to a constant!
4885 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4886 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4887 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4888 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4889 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4890 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4896 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
4898 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
4901 if (!isa<ConstantFP>(RHSC)) return 0;
4902 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
4904 // Get the width of the mantissa. We don't want to hack on conversions that
4905 // might lose information from the integer, e.g. "i64 -> float"
4906 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
4907 if (MantissaWidth == -1) return 0; // Unknown.
4909 // Check to see that the input is converted from an integer type that is small
4910 // enough that preserves all bits. TODO: check here for "known" sign bits.
4911 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
4912 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
4914 // If this is a uitofp instruction, we need an extra bit to hold the sign.
4915 if (isa<UIToFPInst>(LHSI))
4918 // If the conversion would lose info, don't hack on this.
4919 if ((int)InputSize > MantissaWidth)
4922 // Otherwise, we can potentially simplify the comparison. We know that it
4923 // will always come through as an integer value and we know the constant is
4924 // not a NAN (it would have been previously simplified).
4925 assert(!RHS.isNaN() && "NaN comparison not already folded!");
4927 ICmpInst::Predicate Pred;
4928 switch (I.getPredicate()) {
4929 default: assert(0 && "Unexpected predicate!");
4930 case FCmpInst::FCMP_UEQ:
4931 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
4932 case FCmpInst::FCMP_UGT:
4933 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
4934 case FCmpInst::FCMP_UGE:
4935 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
4936 case FCmpInst::FCMP_ULT:
4937 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
4938 case FCmpInst::FCMP_ULE:
4939 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
4940 case FCmpInst::FCMP_UNE:
4941 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
4942 case FCmpInst::FCMP_ORD:
4943 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4944 case FCmpInst::FCMP_UNO:
4945 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4948 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
4950 // Now we know that the APFloat is a normal number, zero or inf.
4952 // See if the FP constant is too large for the integer. For example,
4953 // comparing an i8 to 300.0.
4954 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
4956 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
4957 // and large values.
4958 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
4959 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
4960 APFloat::rmNearestTiesToEven);
4961 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
4962 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
4963 Pred == ICmpInst::ICMP_SLE)
4964 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4965 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4968 // See if the RHS value is < SignedMin.
4969 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
4970 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
4971 APFloat::rmNearestTiesToEven);
4972 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
4973 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
4974 Pred == ICmpInst::ICMP_SGE)
4975 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4976 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4979 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
4980 // it may still be fractional. See if it is fractional by casting the FP
4981 // value to the integer value and back, checking for equality. Don't do this
4982 // for zero, because -0.0 is not fractional.
4983 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
4984 if (!RHS.isZero() &&
4985 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
4986 // If we had a comparison against a fractional value, we have to adjust
4987 // the compare predicate and sometimes the value. RHSC is rounded towards
4988 // zero at this point.
4990 default: assert(0 && "Unexpected integer comparison!");
4991 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
4992 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4993 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
4994 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4995 case ICmpInst::ICMP_SLE:
4996 // (float)int <= 4.4 --> int <= 4
4997 // (float)int <= -4.4 --> int < -4
4998 if (RHS.isNegative())
4999 Pred = ICmpInst::ICMP_SLT;
5001 case ICmpInst::ICMP_SLT:
5002 // (float)int < -4.4 --> int < -4
5003 // (float)int < 4.4 --> int <= 4
5004 if (!RHS.isNegative())
5005 Pred = ICmpInst::ICMP_SLE;
5007 case ICmpInst::ICMP_SGT:
5008 // (float)int > 4.4 --> int > 4
5009 // (float)int > -4.4 --> int >= -4
5010 if (RHS.isNegative())
5011 Pred = ICmpInst::ICMP_SGE;
5013 case ICmpInst::ICMP_SGE:
5014 // (float)int >= -4.4 --> int >= -4
5015 // (float)int >= 4.4 --> int > 4
5016 if (!RHS.isNegative())
5017 Pred = ICmpInst::ICMP_SGT;
5022 // Lower this FP comparison into an appropriate integer version of the
5024 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5027 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5028 bool Changed = SimplifyCompare(I);
5029 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5031 // Fold trivial predicates.
5032 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5033 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5034 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5035 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5037 // Simplify 'fcmp pred X, X'
5039 switch (I.getPredicate()) {
5040 default: assert(0 && "Unknown predicate!");
5041 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5042 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5043 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5044 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5045 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5046 case FCmpInst::FCMP_OLT: // True if ordered and less than
5047 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5048 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5050 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5051 case FCmpInst::FCMP_ULT: // True if unordered or less than
5052 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5053 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5054 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5055 I.setPredicate(FCmpInst::FCMP_UNO);
5056 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5059 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5060 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5061 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5062 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5063 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5064 I.setPredicate(FCmpInst::FCMP_ORD);
5065 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5070 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5071 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5073 // Handle fcmp with constant RHS
5074 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5075 // If the constant is a nan, see if we can fold the comparison based on it.
5076 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5077 if (CFP->getValueAPF().isNaN()) {
5078 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5079 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5080 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5081 "Comparison must be either ordered or unordered!");
5082 // True if unordered.
5083 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5087 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5088 switch (LHSI->getOpcode()) {
5089 case Instruction::PHI:
5090 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5091 // block. If in the same block, we're encouraging jump threading. If
5092 // not, we are just pessimizing the code by making an i1 phi.
5093 if (LHSI->getParent() == I.getParent())
5094 if (Instruction *NV = FoldOpIntoPhi(I))
5097 case Instruction::SIToFP:
5098 case Instruction::UIToFP:
5099 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5102 case Instruction::Select:
5103 // If either operand of the select is a constant, we can fold the
5104 // comparison into the select arms, which will cause one to be
5105 // constant folded and the select turned into a bitwise or.
5106 Value *Op1 = 0, *Op2 = 0;
5107 if (LHSI->hasOneUse()) {
5108 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5109 // Fold the known value into the constant operand.
5110 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5111 // Insert a new FCmp of the other select operand.
5112 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5113 LHSI->getOperand(2), RHSC,
5115 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5116 // Fold the known value into the constant operand.
5117 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5118 // Insert a new FCmp of the other select operand.
5119 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5120 LHSI->getOperand(1), RHSC,
5126 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5131 return Changed ? &I : 0;
5134 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5135 bool Changed = SimplifyCompare(I);
5136 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5137 const Type *Ty = Op0->getType();
5141 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5142 I.isTrueWhenEqual()));
5144 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5145 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5147 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5148 // addresses never equal each other! We already know that Op0 != Op1.
5149 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5150 isa<ConstantPointerNull>(Op0)) &&
5151 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5152 isa<ConstantPointerNull>(Op1)))
5153 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5154 !I.isTrueWhenEqual()));
5156 // icmp's with boolean values can always be turned into bitwise operations
5157 if (Ty == Type::Int1Ty) {
5158 switch (I.getPredicate()) {
5159 default: assert(0 && "Invalid icmp instruction!");
5160 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5161 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5162 InsertNewInstBefore(Xor, I);
5163 return BinaryOperator::CreateNot(Xor);
5165 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5166 return BinaryOperator::CreateXor(Op0, Op1);
5168 case ICmpInst::ICMP_UGT:
5169 case ICmpInst::ICMP_SGT:
5170 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5172 case ICmpInst::ICMP_ULT:
5173 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5174 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5175 InsertNewInstBefore(Not, I);
5176 return BinaryOperator::CreateAnd(Not, Op1);
5178 case ICmpInst::ICMP_UGE:
5179 case ICmpInst::ICMP_SGE:
5180 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5182 case ICmpInst::ICMP_ULE:
5183 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5184 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5185 InsertNewInstBefore(Not, I);
5186 return BinaryOperator::CreateOr(Not, Op1);
5191 // See if we are doing a comparison between a constant and an instruction that
5192 // can be folded into the comparison.
5193 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5196 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5197 if (I.isEquality() && CI->isNullValue() &&
5198 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5199 // (icmp cond A B) if cond is equality
5200 return new ICmpInst(I.getPredicate(), A, B);
5203 switch (I.getPredicate()) {
5205 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5206 if (CI->isMinValue(false))
5207 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5208 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5209 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5210 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5211 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5212 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5213 if (CI->isMinValue(true))
5214 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5215 ConstantInt::getAllOnesValue(Op0->getType()));
5219 case ICmpInst::ICMP_SLT:
5220 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5221 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5222 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5223 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5224 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5225 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5228 case ICmpInst::ICMP_UGT:
5229 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5230 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5231 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5232 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5233 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5234 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5236 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5237 if (CI->isMaxValue(true))
5238 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5239 ConstantInt::getNullValue(Op0->getType()));
5242 case ICmpInst::ICMP_SGT:
5243 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5244 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5245 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5246 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5247 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5248 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5251 case ICmpInst::ICMP_ULE:
5252 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5253 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5254 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5255 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5256 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5257 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5260 case ICmpInst::ICMP_SLE:
5261 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5262 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5263 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5264 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5265 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5266 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5269 case ICmpInst::ICMP_UGE:
5270 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5271 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5272 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5273 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5274 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5275 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5278 case ICmpInst::ICMP_SGE:
5279 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5280 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5281 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5282 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5283 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5284 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5288 // If we still have a icmp le or icmp ge instruction, turn it into the
5289 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5290 // already been handled above, this requires little checking.
5292 switch (I.getPredicate()) {
5294 case ICmpInst::ICMP_ULE:
5295 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5296 case ICmpInst::ICMP_SLE:
5297 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5298 case ICmpInst::ICMP_UGE:
5299 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5300 case ICmpInst::ICMP_SGE:
5301 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5304 // See if we can fold the comparison based on bits known to be zero or one
5305 // in the input. If this comparison is a normal comparison, it demands all
5306 // bits, if it is a sign bit comparison, it only demands the sign bit.
5309 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5311 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5312 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5313 if (SimplifyDemandedBits(Op0,
5314 isSignBit ? APInt::getSignBit(BitWidth)
5315 : APInt::getAllOnesValue(BitWidth),
5316 KnownZero, KnownOne, 0))
5319 // Given the known and unknown bits, compute a range that the LHS could be
5321 if ((KnownOne | KnownZero) != 0) {
5322 // Compute the Min, Max and RHS values based on the known bits. For the
5323 // EQ and NE we use unsigned values.
5324 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5325 const APInt& RHSVal = CI->getValue();
5326 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5327 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5330 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5333 switch (I.getPredicate()) { // LE/GE have been folded already.
5334 default: assert(0 && "Unknown icmp opcode!");
5335 case ICmpInst::ICMP_EQ:
5336 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5337 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5339 case ICmpInst::ICMP_NE:
5340 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5341 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5343 case ICmpInst::ICMP_ULT:
5344 if (Max.ult(RHSVal))
5345 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5346 if (Min.uge(RHSVal))
5347 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5349 case ICmpInst::ICMP_UGT:
5350 if (Min.ugt(RHSVal))
5351 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5352 if (Max.ule(RHSVal))
5353 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5355 case ICmpInst::ICMP_SLT:
5356 if (Max.slt(RHSVal))
5357 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5358 if (Min.sgt(RHSVal))
5359 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5361 case ICmpInst::ICMP_SGT:
5362 if (Min.sgt(RHSVal))
5363 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5364 if (Max.sle(RHSVal))
5365 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5370 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5371 // instruction, see if that instruction also has constants so that the
5372 // instruction can be folded into the icmp
5373 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5374 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5378 // Handle icmp with constant (but not simple integer constant) RHS
5379 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5380 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5381 switch (LHSI->getOpcode()) {
5382 case Instruction::GetElementPtr:
5383 if (RHSC->isNullValue()) {
5384 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5385 bool isAllZeros = true;
5386 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5387 if (!isa<Constant>(LHSI->getOperand(i)) ||
5388 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5393 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5394 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5398 case Instruction::PHI:
5399 // Only fold icmp into the PHI if the phi and fcmp are in the same
5400 // block. If in the same block, we're encouraging jump threading. If
5401 // not, we are just pessimizing the code by making an i1 phi.
5402 if (LHSI->getParent() == I.getParent())
5403 if (Instruction *NV = FoldOpIntoPhi(I))
5406 case Instruction::Select: {
5407 // If either operand of the select is a constant, we can fold the
5408 // comparison into the select arms, which will cause one to be
5409 // constant folded and the select turned into a bitwise or.
5410 Value *Op1 = 0, *Op2 = 0;
5411 if (LHSI->hasOneUse()) {
5412 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5413 // Fold the known value into the constant operand.
5414 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5415 // Insert a new ICmp of the other select operand.
5416 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5417 LHSI->getOperand(2), RHSC,
5419 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5420 // Fold the known value into the constant operand.
5421 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5422 // Insert a new ICmp of the other select operand.
5423 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5424 LHSI->getOperand(1), RHSC,
5430 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5433 case Instruction::Malloc:
5434 // If we have (malloc != null), and if the malloc has a single use, we
5435 // can assume it is successful and remove the malloc.
5436 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5437 AddToWorkList(LHSI);
5438 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5439 !I.isTrueWhenEqual()));
5445 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5446 if (User *GEP = dyn_castGetElementPtr(Op0))
5447 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5449 if (User *GEP = dyn_castGetElementPtr(Op1))
5450 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5451 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5454 // Test to see if the operands of the icmp are casted versions of other
5455 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5457 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5458 if (isa<PointerType>(Op0->getType()) &&
5459 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5460 // We keep moving the cast from the left operand over to the right
5461 // operand, where it can often be eliminated completely.
5462 Op0 = CI->getOperand(0);
5464 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5465 // so eliminate it as well.
5466 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5467 Op1 = CI2->getOperand(0);
5469 // If Op1 is a constant, we can fold the cast into the constant.
5470 if (Op0->getType() != Op1->getType()) {
5471 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5472 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5474 // Otherwise, cast the RHS right before the icmp
5475 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5478 return new ICmpInst(I.getPredicate(), Op0, Op1);
5482 if (isa<CastInst>(Op0)) {
5483 // Handle the special case of: icmp (cast bool to X), <cst>
5484 // This comes up when you have code like
5487 // For generality, we handle any zero-extension of any operand comparison
5488 // with a constant or another cast from the same type.
5489 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5490 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5494 // ~x < ~y --> y < x
5496 if (match(Op0, m_Not(m_Value(A))) &&
5497 match(Op1, m_Not(m_Value(B))))
5498 return new ICmpInst(I.getPredicate(), B, A);
5501 if (I.isEquality()) {
5502 Value *A, *B, *C, *D;
5504 // -x == -y --> x == y
5505 if (match(Op0, m_Neg(m_Value(A))) &&
5506 match(Op1, m_Neg(m_Value(B))))
5507 return new ICmpInst(I.getPredicate(), A, B);
5509 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5510 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5511 Value *OtherVal = A == Op1 ? B : A;
5512 return new ICmpInst(I.getPredicate(), OtherVal,
5513 Constant::getNullValue(A->getType()));
5516 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5517 // A^c1 == C^c2 --> A == C^(c1^c2)
5518 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5519 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5520 if (Op1->hasOneUse()) {
5521 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5522 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5523 return new ICmpInst(I.getPredicate(), A,
5524 InsertNewInstBefore(Xor, I));
5527 // A^B == A^D -> B == D
5528 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5529 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5530 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5531 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5535 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5536 (A == Op0 || B == Op0)) {
5537 // A == (A^B) -> B == 0
5538 Value *OtherVal = A == Op0 ? B : A;
5539 return new ICmpInst(I.getPredicate(), OtherVal,
5540 Constant::getNullValue(A->getType()));
5542 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5543 // (A-B) == A -> B == 0
5544 return new ICmpInst(I.getPredicate(), B,
5545 Constant::getNullValue(B->getType()));
5547 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5548 // A == (A-B) -> B == 0
5549 return new ICmpInst(I.getPredicate(), B,
5550 Constant::getNullValue(B->getType()));
5553 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5554 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5555 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5556 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5557 Value *X = 0, *Y = 0, *Z = 0;
5560 X = B; Y = D; Z = A;
5561 } else if (A == D) {
5562 X = B; Y = C; Z = A;
5563 } else if (B == C) {
5564 X = A; Y = D; Z = B;
5565 } else if (B == D) {
5566 X = A; Y = C; Z = B;
5569 if (X) { // Build (X^Y) & Z
5570 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
5571 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
5572 I.setOperand(0, Op1);
5573 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5578 return Changed ? &I : 0;
5582 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5583 /// and CmpRHS are both known to be integer constants.
5584 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5585 ConstantInt *DivRHS) {
5586 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5587 const APInt &CmpRHSV = CmpRHS->getValue();
5589 // FIXME: If the operand types don't match the type of the divide
5590 // then don't attempt this transform. The code below doesn't have the
5591 // logic to deal with a signed divide and an unsigned compare (and
5592 // vice versa). This is because (x /s C1) <s C2 produces different
5593 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5594 // (x /u C1) <u C2. Simply casting the operands and result won't
5595 // work. :( The if statement below tests that condition and bails
5597 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5598 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5600 if (DivRHS->isZero())
5601 return 0; // The ProdOV computation fails on divide by zero.
5603 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5604 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5605 // C2 (CI). By solving for X we can turn this into a range check
5606 // instead of computing a divide.
5607 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5609 // Determine if the product overflows by seeing if the product is
5610 // not equal to the divide. Make sure we do the same kind of divide
5611 // as in the LHS instruction that we're folding.
5612 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5613 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5615 // Get the ICmp opcode
5616 ICmpInst::Predicate Pred = ICI.getPredicate();
5618 // Figure out the interval that is being checked. For example, a comparison
5619 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5620 // Compute this interval based on the constants involved and the signedness of
5621 // the compare/divide. This computes a half-open interval, keeping track of
5622 // whether either value in the interval overflows. After analysis each
5623 // overflow variable is set to 0 if it's corresponding bound variable is valid
5624 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5625 int LoOverflow = 0, HiOverflow = 0;
5626 ConstantInt *LoBound = 0, *HiBound = 0;
5629 if (!DivIsSigned) { // udiv
5630 // e.g. X/5 op 3 --> [15, 20)
5632 HiOverflow = LoOverflow = ProdOV;
5634 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5635 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5636 if (CmpRHSV == 0) { // (X / pos) op 0
5637 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5638 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5640 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5641 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5642 HiOverflow = LoOverflow = ProdOV;
5644 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5645 } else { // (X / pos) op neg
5646 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5647 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5648 LoOverflow = AddWithOverflow(LoBound, Prod,
5649 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5650 HiBound = AddOne(Prod);
5651 HiOverflow = ProdOV ? -1 : 0;
5653 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5654 if (CmpRHSV == 0) { // (X / neg) op 0
5655 // e.g. X/-5 op 0 --> [-4, 5)
5656 LoBound = AddOne(DivRHS);
5657 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5658 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5659 HiOverflow = 1; // [INTMIN+1, overflow)
5660 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5662 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5663 // e.g. X/-5 op 3 --> [-19, -14)
5664 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5666 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5667 HiBound = AddOne(Prod);
5668 } else { // (X / neg) op neg
5669 // e.g. X/-5 op -3 --> [15, 20)
5671 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5672 HiBound = Subtract(Prod, DivRHS);
5675 // Dividing by a negative swaps the condition. LT <-> GT
5676 Pred = ICmpInst::getSwappedPredicate(Pred);
5679 Value *X = DivI->getOperand(0);
5681 default: assert(0 && "Unhandled icmp opcode!");
5682 case ICmpInst::ICMP_EQ:
5683 if (LoOverflow && HiOverflow)
5684 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5685 else if (HiOverflow)
5686 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5687 ICmpInst::ICMP_UGE, X, LoBound);
5688 else if (LoOverflow)
5689 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5690 ICmpInst::ICMP_ULT, X, HiBound);
5692 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5693 case ICmpInst::ICMP_NE:
5694 if (LoOverflow && HiOverflow)
5695 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5696 else if (HiOverflow)
5697 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5698 ICmpInst::ICMP_ULT, X, LoBound);
5699 else if (LoOverflow)
5700 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5701 ICmpInst::ICMP_UGE, X, HiBound);
5703 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5704 case ICmpInst::ICMP_ULT:
5705 case ICmpInst::ICMP_SLT:
5706 if (LoOverflow == +1) // Low bound is greater than input range.
5707 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5708 if (LoOverflow == -1) // Low bound is less than input range.
5709 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5710 return new ICmpInst(Pred, X, LoBound);
5711 case ICmpInst::ICMP_UGT:
5712 case ICmpInst::ICMP_SGT:
5713 if (HiOverflow == +1) // High bound greater than input range.
5714 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5715 else if (HiOverflow == -1) // High bound less than input range.
5716 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5717 if (Pred == ICmpInst::ICMP_UGT)
5718 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5720 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5725 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5727 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5730 const APInt &RHSV = RHS->getValue();
5732 switch (LHSI->getOpcode()) {
5733 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5734 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5735 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5737 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5738 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5739 Value *CompareVal = LHSI->getOperand(0);
5741 // If the sign bit of the XorCST is not set, there is no change to
5742 // the operation, just stop using the Xor.
5743 if (!XorCST->getValue().isNegative()) {
5744 ICI.setOperand(0, CompareVal);
5745 AddToWorkList(LHSI);
5749 // Was the old condition true if the operand is positive?
5750 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5752 // If so, the new one isn't.
5753 isTrueIfPositive ^= true;
5755 if (isTrueIfPositive)
5756 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5758 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5762 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5763 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5764 LHSI->getOperand(0)->hasOneUse()) {
5765 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5767 // If the LHS is an AND of a truncating cast, we can widen the
5768 // and/compare to be the input width without changing the value
5769 // produced, eliminating a cast.
5770 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5771 // We can do this transformation if either the AND constant does not
5772 // have its sign bit set or if it is an equality comparison.
5773 // Extending a relational comparison when we're checking the sign
5774 // bit would not work.
5775 if (Cast->hasOneUse() &&
5776 (ICI.isEquality() ||
5777 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5779 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5780 APInt NewCST = AndCST->getValue();
5781 NewCST.zext(BitWidth);
5783 NewCI.zext(BitWidth);
5784 Instruction *NewAnd =
5785 BinaryOperator::CreateAnd(Cast->getOperand(0),
5786 ConstantInt::get(NewCST),LHSI->getName());
5787 InsertNewInstBefore(NewAnd, ICI);
5788 return new ICmpInst(ICI.getPredicate(), NewAnd,
5789 ConstantInt::get(NewCI));
5793 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5794 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5795 // happens a LOT in code produced by the C front-end, for bitfield
5797 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5798 if (Shift && !Shift->isShift())
5802 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5803 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5804 const Type *AndTy = AndCST->getType(); // Type of the and.
5806 // We can fold this as long as we can't shift unknown bits
5807 // into the mask. This can only happen with signed shift
5808 // rights, as they sign-extend.
5810 bool CanFold = Shift->isLogicalShift();
5812 // To test for the bad case of the signed shr, see if any
5813 // of the bits shifted in could be tested after the mask.
5814 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5815 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5817 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5818 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5819 AndCST->getValue()) == 0)
5825 if (Shift->getOpcode() == Instruction::Shl)
5826 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5828 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5830 // Check to see if we are shifting out any of the bits being
5832 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5833 // If we shifted bits out, the fold is not going to work out.
5834 // As a special case, check to see if this means that the
5835 // result is always true or false now.
5836 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5837 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5838 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5839 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5841 ICI.setOperand(1, NewCst);
5842 Constant *NewAndCST;
5843 if (Shift->getOpcode() == Instruction::Shl)
5844 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5846 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5847 LHSI->setOperand(1, NewAndCST);
5848 LHSI->setOperand(0, Shift->getOperand(0));
5849 AddToWorkList(Shift); // Shift is dead.
5850 AddUsesToWorkList(ICI);
5856 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5857 // preferable because it allows the C<<Y expression to be hoisted out
5858 // of a loop if Y is invariant and X is not.
5859 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5860 ICI.isEquality() && !Shift->isArithmeticShift() &&
5861 isa<Instruction>(Shift->getOperand(0))) {
5864 if (Shift->getOpcode() == Instruction::LShr) {
5865 NS = BinaryOperator::CreateShl(AndCST,
5866 Shift->getOperand(1), "tmp");
5868 // Insert a logical shift.
5869 NS = BinaryOperator::CreateLShr(AndCST,
5870 Shift->getOperand(1), "tmp");
5872 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5874 // Compute X & (C << Y).
5875 Instruction *NewAnd =
5876 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
5877 InsertNewInstBefore(NewAnd, ICI);
5879 ICI.setOperand(0, NewAnd);
5885 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5886 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5889 uint32_t TypeBits = RHSV.getBitWidth();
5891 // Check that the shift amount is in range. If not, don't perform
5892 // undefined shifts. When the shift is visited it will be
5894 if (ShAmt->uge(TypeBits))
5897 if (ICI.isEquality()) {
5898 // If we are comparing against bits always shifted out, the
5899 // comparison cannot succeed.
5901 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5902 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5903 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5904 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5905 return ReplaceInstUsesWith(ICI, Cst);
5908 if (LHSI->hasOneUse()) {
5909 // Otherwise strength reduce the shift into an and.
5910 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5912 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5915 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5916 Mask, LHSI->getName()+".mask");
5917 Value *And = InsertNewInstBefore(AndI, ICI);
5918 return new ICmpInst(ICI.getPredicate(), And,
5919 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5923 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5924 bool TrueIfSigned = false;
5925 if (LHSI->hasOneUse() &&
5926 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5927 // (X << 31) <s 0 --> (X&1) != 0
5928 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5929 (TypeBits-ShAmt->getZExtValue()-1));
5931 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5932 Mask, LHSI->getName()+".mask");
5933 Value *And = InsertNewInstBefore(AndI, ICI);
5935 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5936 And, Constant::getNullValue(And->getType()));
5941 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5942 case Instruction::AShr: {
5943 // Only handle equality comparisons of shift-by-constant.
5944 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5945 if (!ShAmt || !ICI.isEquality()) break;
5947 // Check that the shift amount is in range. If not, don't perform
5948 // undefined shifts. When the shift is visited it will be
5950 uint32_t TypeBits = RHSV.getBitWidth();
5951 if (ShAmt->uge(TypeBits))
5954 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5956 // If we are comparing against bits always shifted out, the
5957 // comparison cannot succeed.
5958 APInt Comp = RHSV << ShAmtVal;
5959 if (LHSI->getOpcode() == Instruction::LShr)
5960 Comp = Comp.lshr(ShAmtVal);
5962 Comp = Comp.ashr(ShAmtVal);
5964 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5965 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5966 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5967 return ReplaceInstUsesWith(ICI, Cst);
5970 // Otherwise, check to see if the bits shifted out are known to be zero.
5971 // If so, we can compare against the unshifted value:
5972 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
5973 if (LHSI->hasOneUse() &&
5974 MaskedValueIsZero(LHSI->getOperand(0),
5975 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
5976 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
5977 ConstantExpr::getShl(RHS, ShAmt));
5980 if (LHSI->hasOneUse()) {
5981 // Otherwise strength reduce the shift into an and.
5982 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5983 Constant *Mask = ConstantInt::get(Val);
5986 BinaryOperator::CreateAnd(LHSI->getOperand(0),
5987 Mask, LHSI->getName()+".mask");
5988 Value *And = InsertNewInstBefore(AndI, ICI);
5989 return new ICmpInst(ICI.getPredicate(), And,
5990 ConstantExpr::getShl(RHS, ShAmt));
5995 case Instruction::SDiv:
5996 case Instruction::UDiv:
5997 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5998 // Fold this div into the comparison, producing a range check.
5999 // Determine, based on the divide type, what the range is being
6000 // checked. If there is an overflow on the low or high side, remember
6001 // it, otherwise compute the range [low, hi) bounding the new value.
6002 // See: InsertRangeTest above for the kinds of replacements possible.
6003 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6004 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6009 case Instruction::Add:
6010 // Fold: icmp pred (add, X, C1), C2
6012 if (!ICI.isEquality()) {
6013 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6015 const APInt &LHSV = LHSC->getValue();
6017 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6020 if (ICI.isSignedPredicate()) {
6021 if (CR.getLower().isSignBit()) {
6022 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6023 ConstantInt::get(CR.getUpper()));
6024 } else if (CR.getUpper().isSignBit()) {
6025 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6026 ConstantInt::get(CR.getLower()));
6029 if (CR.getLower().isMinValue()) {
6030 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6031 ConstantInt::get(CR.getUpper()));
6032 } else if (CR.getUpper().isMinValue()) {
6033 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6034 ConstantInt::get(CR.getLower()));
6041 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6042 if (ICI.isEquality()) {
6043 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6045 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6046 // the second operand is a constant, simplify a bit.
6047 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6048 switch (BO->getOpcode()) {
6049 case Instruction::SRem:
6050 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6051 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6052 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6053 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6054 Instruction *NewRem =
6055 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6057 InsertNewInstBefore(NewRem, ICI);
6058 return new ICmpInst(ICI.getPredicate(), NewRem,
6059 Constant::getNullValue(BO->getType()));
6063 case Instruction::Add:
6064 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6065 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6066 if (BO->hasOneUse())
6067 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6068 Subtract(RHS, BOp1C));
6069 } else if (RHSV == 0) {
6070 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6071 // efficiently invertible, or if the add has just this one use.
6072 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6074 if (Value *NegVal = dyn_castNegVal(BOp1))
6075 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6076 else if (Value *NegVal = dyn_castNegVal(BOp0))
6077 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6078 else if (BO->hasOneUse()) {
6079 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6080 InsertNewInstBefore(Neg, ICI);
6082 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6086 case Instruction::Xor:
6087 // For the xor case, we can xor two constants together, eliminating
6088 // the explicit xor.
6089 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6090 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6091 ConstantExpr::getXor(RHS, BOC));
6094 case Instruction::Sub:
6095 // Replace (([sub|xor] A, B) != 0) with (A != B)
6097 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6101 case Instruction::Or:
6102 // If bits are being or'd in that are not present in the constant we
6103 // are comparing against, then the comparison could never succeed!
6104 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6105 Constant *NotCI = ConstantExpr::getNot(RHS);
6106 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6107 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6112 case Instruction::And:
6113 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6114 // If bits are being compared against that are and'd out, then the
6115 // comparison can never succeed!
6116 if ((RHSV & ~BOC->getValue()) != 0)
6117 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6120 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6121 if (RHS == BOC && RHSV.isPowerOf2())
6122 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6123 ICmpInst::ICMP_NE, LHSI,
6124 Constant::getNullValue(RHS->getType()));
6126 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6127 if (BOC->getValue().isSignBit()) {
6128 Value *X = BO->getOperand(0);
6129 Constant *Zero = Constant::getNullValue(X->getType());
6130 ICmpInst::Predicate pred = isICMP_NE ?
6131 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6132 return new ICmpInst(pred, X, Zero);
6135 // ((X & ~7) == 0) --> X < 8
6136 if (RHSV == 0 && isHighOnes(BOC)) {
6137 Value *X = BO->getOperand(0);
6138 Constant *NegX = ConstantExpr::getNeg(BOC);
6139 ICmpInst::Predicate pred = isICMP_NE ?
6140 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6141 return new ICmpInst(pred, X, NegX);
6146 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6147 // Handle icmp {eq|ne} <intrinsic>, intcst.
6148 if (II->getIntrinsicID() == Intrinsic::bswap) {
6150 ICI.setOperand(0, II->getOperand(1));
6151 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6155 } else { // Not a ICMP_EQ/ICMP_NE
6156 // If the LHS is a cast from an integral value of the same size,
6157 // then since we know the RHS is a constant, try to simlify.
6158 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6159 Value *CastOp = Cast->getOperand(0);
6160 const Type *SrcTy = CastOp->getType();
6161 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6162 if (SrcTy->isInteger() &&
6163 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6164 // If this is an unsigned comparison, try to make the comparison use
6165 // smaller constant values.
6166 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6167 // X u< 128 => X s> -1
6168 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6169 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6170 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6171 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6172 // X u> 127 => X s< 0
6173 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6174 Constant::getNullValue(SrcTy));
6182 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6183 /// We only handle extending casts so far.
6185 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6186 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6187 Value *LHSCIOp = LHSCI->getOperand(0);
6188 const Type *SrcTy = LHSCIOp->getType();
6189 const Type *DestTy = LHSCI->getType();
6192 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6193 // integer type is the same size as the pointer type.
6194 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6195 getTargetData().getPointerSizeInBits() ==
6196 cast<IntegerType>(DestTy)->getBitWidth()) {
6198 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6199 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6200 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6201 RHSOp = RHSC->getOperand(0);
6202 // If the pointer types don't match, insert a bitcast.
6203 if (LHSCIOp->getType() != RHSOp->getType())
6204 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6208 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6211 // The code below only handles extension cast instructions, so far.
6213 if (LHSCI->getOpcode() != Instruction::ZExt &&
6214 LHSCI->getOpcode() != Instruction::SExt)
6217 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6218 bool isSignedCmp = ICI.isSignedPredicate();
6220 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6221 // Not an extension from the same type?
6222 RHSCIOp = CI->getOperand(0);
6223 if (RHSCIOp->getType() != LHSCIOp->getType())
6226 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6227 // and the other is a zext), then we can't handle this.
6228 if (CI->getOpcode() != LHSCI->getOpcode())
6231 // Deal with equality cases early.
6232 if (ICI.isEquality())
6233 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6235 // A signed comparison of sign extended values simplifies into a
6236 // signed comparison.
6237 if (isSignedCmp && isSignedExt)
6238 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6240 // The other three cases all fold into an unsigned comparison.
6241 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6244 // If we aren't dealing with a constant on the RHS, exit early
6245 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6249 // Compute the constant that would happen if we truncated to SrcTy then
6250 // reextended to DestTy.
6251 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6252 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6254 // If the re-extended constant didn't change...
6256 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6257 // For example, we might have:
6258 // %A = sext short %X to uint
6259 // %B = icmp ugt uint %A, 1330
6260 // It is incorrect to transform this into
6261 // %B = icmp ugt short %X, 1330
6262 // because %A may have negative value.
6264 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6265 // OR operation is EQ/NE.
6266 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6267 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6272 // The re-extended constant changed so the constant cannot be represented
6273 // in the shorter type. Consequently, we cannot emit a simple comparison.
6275 // First, handle some easy cases. We know the result cannot be equal at this
6276 // point so handle the ICI.isEquality() cases
6277 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6278 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6279 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6280 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6282 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6283 // should have been folded away previously and not enter in here.
6286 // We're performing a signed comparison.
6287 if (cast<ConstantInt>(CI)->getValue().isNegative())
6288 Result = ConstantInt::getFalse(); // X < (small) --> false
6290 Result = ConstantInt::getTrue(); // X < (large) --> true
6292 // We're performing an unsigned comparison.
6294 // We're performing an unsigned comp with a sign extended value.
6295 // This is true if the input is >= 0. [aka >s -1]
6296 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6297 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6298 NegOne, ICI.getName()), ICI);
6300 // Unsigned extend & unsigned compare -> always true.
6301 Result = ConstantInt::getTrue();
6305 // Finally, return the value computed.
6306 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6307 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6308 return ReplaceInstUsesWith(ICI, Result);
6310 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6311 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6312 "ICmp should be folded!");
6313 if (Constant *CI = dyn_cast<Constant>(Result))
6314 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6316 return BinaryOperator::CreateNot(Result);
6320 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6321 return commonShiftTransforms(I);
6324 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6325 return commonShiftTransforms(I);
6328 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6329 if (Instruction *R = commonShiftTransforms(I))
6332 Value *Op0 = I.getOperand(0);
6334 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6335 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6336 if (CSI->isAllOnesValue())
6337 return ReplaceInstUsesWith(I, CSI);
6339 // See if we can turn a signed shr into an unsigned shr.
6340 if (MaskedValueIsZero(Op0,
6341 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6342 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6347 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6348 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6349 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6351 // shl X, 0 == X and shr X, 0 == X
6352 // shl 0, X == 0 and shr 0, X == 0
6353 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6354 Op0 == Constant::getNullValue(Op0->getType()))
6355 return ReplaceInstUsesWith(I, Op0);
6357 if (isa<UndefValue>(Op0)) {
6358 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6359 return ReplaceInstUsesWith(I, Op0);
6360 else // undef << X -> 0, undef >>u X -> 0
6361 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6363 if (isa<UndefValue>(Op1)) {
6364 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6365 return ReplaceInstUsesWith(I, Op0);
6366 else // X << undef, X >>u undef -> 0
6367 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6370 // Try to fold constant and into select arguments.
6371 if (isa<Constant>(Op0))
6372 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6373 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6376 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6377 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6382 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6383 BinaryOperator &I) {
6384 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6386 // See if we can simplify any instructions used by the instruction whose sole
6387 // purpose is to compute bits we don't care about.
6388 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6389 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6390 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6391 KnownZero, KnownOne))
6394 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6395 // of a signed value.
6397 if (Op1->uge(TypeBits)) {
6398 if (I.getOpcode() != Instruction::AShr)
6399 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6401 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6406 // ((X*C1) << C2) == (X * (C1 << C2))
6407 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6408 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6409 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6410 return BinaryOperator::CreateMul(BO->getOperand(0),
6411 ConstantExpr::getShl(BOOp, Op1));
6413 // Try to fold constant and into select arguments.
6414 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6415 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6417 if (isa<PHINode>(Op0))
6418 if (Instruction *NV = FoldOpIntoPhi(I))
6421 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6422 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6423 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6424 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6425 // place. Don't try to do this transformation in this case. Also, we
6426 // require that the input operand is a shift-by-constant so that we have
6427 // confidence that the shifts will get folded together. We could do this
6428 // xform in more cases, but it is unlikely to be profitable.
6429 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6430 isa<ConstantInt>(TrOp->getOperand(1))) {
6431 // Okay, we'll do this xform. Make the shift of shift.
6432 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6433 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6435 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6437 // For logical shifts, the truncation has the effect of making the high
6438 // part of the register be zeros. Emulate this by inserting an AND to
6439 // clear the top bits as needed. This 'and' will usually be zapped by
6440 // other xforms later if dead.
6441 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6442 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6443 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6445 // The mask we constructed says what the trunc would do if occurring
6446 // between the shifts. We want to know the effect *after* the second
6447 // shift. We know that it is a logical shift by a constant, so adjust the
6448 // mask as appropriate.
6449 if (I.getOpcode() == Instruction::Shl)
6450 MaskV <<= Op1->getZExtValue();
6452 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6453 MaskV = MaskV.lshr(Op1->getZExtValue());
6456 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6458 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6460 // Return the value truncated to the interesting size.
6461 return new TruncInst(And, I.getType());
6465 if (Op0->hasOneUse()) {
6466 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6467 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6470 switch (Op0BO->getOpcode()) {
6472 case Instruction::Add:
6473 case Instruction::And:
6474 case Instruction::Or:
6475 case Instruction::Xor: {
6476 // These operators commute.
6477 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6478 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6479 match(Op0BO->getOperand(1),
6480 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6481 Instruction *YS = BinaryOperator::CreateShl(
6482 Op0BO->getOperand(0), Op1,
6484 InsertNewInstBefore(YS, I); // (Y << C)
6486 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6487 Op0BO->getOperand(1)->getName());
6488 InsertNewInstBefore(X, I); // (X + (Y << C))
6489 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6490 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6491 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6494 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6495 Value *Op0BOOp1 = Op0BO->getOperand(1);
6496 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6498 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6499 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6501 Instruction *YS = BinaryOperator::CreateShl(
6502 Op0BO->getOperand(0), Op1,
6504 InsertNewInstBefore(YS, I); // (Y << C)
6506 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6507 V1->getName()+".mask");
6508 InsertNewInstBefore(XM, I); // X & (CC << C)
6510 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6515 case Instruction::Sub: {
6516 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6517 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6518 match(Op0BO->getOperand(0),
6519 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6520 Instruction *YS = BinaryOperator::CreateShl(
6521 Op0BO->getOperand(1), Op1,
6523 InsertNewInstBefore(YS, I); // (Y << C)
6525 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6526 Op0BO->getOperand(0)->getName());
6527 InsertNewInstBefore(X, I); // (X + (Y << C))
6528 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6529 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6530 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6533 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6534 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6535 match(Op0BO->getOperand(0),
6536 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6537 m_ConstantInt(CC))) && V2 == Op1 &&
6538 cast<BinaryOperator>(Op0BO->getOperand(0))
6539 ->getOperand(0)->hasOneUse()) {
6540 Instruction *YS = BinaryOperator::CreateShl(
6541 Op0BO->getOperand(1), Op1,
6543 InsertNewInstBefore(YS, I); // (Y << C)
6545 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6546 V1->getName()+".mask");
6547 InsertNewInstBefore(XM, I); // X & (CC << C)
6549 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
6557 // If the operand is an bitwise operator with a constant RHS, and the
6558 // shift is the only use, we can pull it out of the shift.
6559 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6560 bool isValid = true; // Valid only for And, Or, Xor
6561 bool highBitSet = false; // Transform if high bit of constant set?
6563 switch (Op0BO->getOpcode()) {
6564 default: isValid = false; break; // Do not perform transform!
6565 case Instruction::Add:
6566 isValid = isLeftShift;
6568 case Instruction::Or:
6569 case Instruction::Xor:
6572 case Instruction::And:
6577 // If this is a signed shift right, and the high bit is modified
6578 // by the logical operation, do not perform the transformation.
6579 // The highBitSet boolean indicates the value of the high bit of
6580 // the constant which would cause it to be modified for this
6583 if (isValid && I.getOpcode() == Instruction::AShr)
6584 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6587 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6589 Instruction *NewShift =
6590 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6591 InsertNewInstBefore(NewShift, I);
6592 NewShift->takeName(Op0BO);
6594 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
6601 // Find out if this is a shift of a shift by a constant.
6602 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6603 if (ShiftOp && !ShiftOp->isShift())
6606 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6607 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6608 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6609 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6610 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6611 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6612 Value *X = ShiftOp->getOperand(0);
6614 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6615 if (AmtSum > TypeBits)
6618 const IntegerType *Ty = cast<IntegerType>(I.getType());
6620 // Check for (X << c1) << c2 and (X >> c1) >> c2
6621 if (I.getOpcode() == ShiftOp->getOpcode()) {
6622 return BinaryOperator::Create(I.getOpcode(), X,
6623 ConstantInt::get(Ty, AmtSum));
6624 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6625 I.getOpcode() == Instruction::AShr) {
6626 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6627 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
6628 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6629 I.getOpcode() == Instruction::LShr) {
6630 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6631 Instruction *Shift =
6632 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
6633 InsertNewInstBefore(Shift, I);
6635 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6636 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6639 // Okay, if we get here, one shift must be left, and the other shift must be
6640 // right. See if the amounts are equal.
6641 if (ShiftAmt1 == ShiftAmt2) {
6642 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6643 if (I.getOpcode() == Instruction::Shl) {
6644 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6645 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6647 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6648 if (I.getOpcode() == Instruction::LShr) {
6649 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6650 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6652 // We can simplify ((X << C) >>s C) into a trunc + sext.
6653 // NOTE: we could do this for any C, but that would make 'unusual' integer
6654 // types. For now, just stick to ones well-supported by the code
6656 const Type *SExtType = 0;
6657 switch (Ty->getBitWidth() - ShiftAmt1) {
6664 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6669 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6670 InsertNewInstBefore(NewTrunc, I);
6671 return new SExtInst(NewTrunc, Ty);
6673 // Otherwise, we can't handle it yet.
6674 } else if (ShiftAmt1 < ShiftAmt2) {
6675 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6677 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6678 if (I.getOpcode() == Instruction::Shl) {
6679 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6680 ShiftOp->getOpcode() == Instruction::AShr);
6681 Instruction *Shift =
6682 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6683 InsertNewInstBefore(Shift, I);
6685 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6686 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6689 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6690 if (I.getOpcode() == Instruction::LShr) {
6691 assert(ShiftOp->getOpcode() == Instruction::Shl);
6692 Instruction *Shift =
6693 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
6694 InsertNewInstBefore(Shift, I);
6696 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6697 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6700 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6702 assert(ShiftAmt2 < ShiftAmt1);
6703 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6705 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6706 if (I.getOpcode() == Instruction::Shl) {
6707 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6708 ShiftOp->getOpcode() == Instruction::AShr);
6709 Instruction *Shift =
6710 BinaryOperator::Create(ShiftOp->getOpcode(), X,
6711 ConstantInt::get(Ty, ShiftDiff));
6712 InsertNewInstBefore(Shift, I);
6714 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6715 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6718 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6719 if (I.getOpcode() == Instruction::LShr) {
6720 assert(ShiftOp->getOpcode() == Instruction::Shl);
6721 Instruction *Shift =
6722 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6723 InsertNewInstBefore(Shift, I);
6725 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6726 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6729 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6736 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6737 /// expression. If so, decompose it, returning some value X, such that Val is
6740 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6742 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6743 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6744 Offset = CI->getZExtValue();
6746 return ConstantInt::get(Type::Int32Ty, 0);
6747 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6748 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6749 if (I->getOpcode() == Instruction::Shl) {
6750 // This is a value scaled by '1 << the shift amt'.
6751 Scale = 1U << RHS->getZExtValue();
6753 return I->getOperand(0);
6754 } else if (I->getOpcode() == Instruction::Mul) {
6755 // This value is scaled by 'RHS'.
6756 Scale = RHS->getZExtValue();
6758 return I->getOperand(0);
6759 } else if (I->getOpcode() == Instruction::Add) {
6760 // We have X+C. Check to see if we really have (X*C2)+C1,
6761 // where C1 is divisible by C2.
6764 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6765 Offset += RHS->getZExtValue();
6772 // Otherwise, we can't look past this.
6779 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6780 /// try to eliminate the cast by moving the type information into the alloc.
6781 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6782 AllocationInst &AI) {
6783 const PointerType *PTy = cast<PointerType>(CI.getType());
6785 // Remove any uses of AI that are dead.
6786 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6788 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6789 Instruction *User = cast<Instruction>(*UI++);
6790 if (isInstructionTriviallyDead(User)) {
6791 while (UI != E && *UI == User)
6792 ++UI; // If this instruction uses AI more than once, don't break UI.
6795 DOUT << "IC: DCE: " << *User;
6796 EraseInstFromFunction(*User);
6800 // Get the type really allocated and the type casted to.
6801 const Type *AllocElTy = AI.getAllocatedType();
6802 const Type *CastElTy = PTy->getElementType();
6803 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6805 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6806 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6807 if (CastElTyAlign < AllocElTyAlign) return 0;
6809 // If the allocation has multiple uses, only promote it if we are strictly
6810 // increasing the alignment of the resultant allocation. If we keep it the
6811 // same, we open the door to infinite loops of various kinds.
6812 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6814 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6815 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6816 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6818 // See if we can satisfy the modulus by pulling a scale out of the array
6820 unsigned ArraySizeScale;
6822 Value *NumElements = // See if the array size is a decomposable linear expr.
6823 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6825 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6827 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6828 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6830 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6835 // If the allocation size is constant, form a constant mul expression
6836 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6837 if (isa<ConstantInt>(NumElements))
6838 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6839 // otherwise multiply the amount and the number of elements
6840 else if (Scale != 1) {
6841 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
6842 Amt = InsertNewInstBefore(Tmp, AI);
6846 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6847 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6848 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
6849 Amt = InsertNewInstBefore(Tmp, AI);
6852 AllocationInst *New;
6853 if (isa<MallocInst>(AI))
6854 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6856 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6857 InsertNewInstBefore(New, AI);
6860 // If the allocation has multiple uses, insert a cast and change all things
6861 // that used it to use the new cast. This will also hack on CI, but it will
6863 if (!AI.hasOneUse()) {
6864 AddUsesToWorkList(AI);
6865 // New is the allocation instruction, pointer typed. AI is the original
6866 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6867 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6868 InsertNewInstBefore(NewCast, AI);
6869 AI.replaceAllUsesWith(NewCast);
6871 return ReplaceInstUsesWith(CI, New);
6874 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6875 /// and return it as type Ty without inserting any new casts and without
6876 /// changing the computed value. This is used by code that tries to decide
6877 /// whether promoting or shrinking integer operations to wider or smaller types
6878 /// will allow us to eliminate a truncate or extend.
6880 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6881 /// extension operation if Ty is larger.
6883 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
6884 /// should return true if trunc(V) can be computed by computing V in the smaller
6885 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
6886 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
6887 /// efficiently truncated.
6889 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
6890 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
6891 /// the final result.
6892 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6894 int &NumCastsRemoved) {
6895 // We can always evaluate constants in another type.
6896 if (isa<ConstantInt>(V))
6899 Instruction *I = dyn_cast<Instruction>(V);
6900 if (!I) return false;
6902 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6904 // If this is an extension or truncate, we can often eliminate it.
6905 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6906 // If this is a cast from the destination type, we can trivially eliminate
6907 // it, and this will remove a cast overall.
6908 if (I->getOperand(0)->getType() == Ty) {
6909 // If the first operand is itself a cast, and is eliminable, do not count
6910 // this as an eliminable cast. We would prefer to eliminate those two
6912 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
6918 // We can't extend or shrink something that has multiple uses: doing so would
6919 // require duplicating the instruction in general, which isn't profitable.
6920 if (!I->hasOneUse()) return false;
6922 switch (I->getOpcode()) {
6923 case Instruction::Add:
6924 case Instruction::Sub:
6925 case Instruction::And:
6926 case Instruction::Or:
6927 case Instruction::Xor:
6928 // These operators can all arbitrarily be extended or truncated.
6929 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6931 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6934 case Instruction::Mul:
6935 // A multiply can be truncated by truncating its operands.
6936 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6937 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6939 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6942 case Instruction::Shl:
6943 // If we are truncating the result of this SHL, and if it's a shift of a
6944 // constant amount, we can always perform a SHL in a smaller type.
6945 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6946 uint32_t BitWidth = Ty->getBitWidth();
6947 if (BitWidth < OrigTy->getBitWidth() &&
6948 CI->getLimitedValue(BitWidth) < BitWidth)
6949 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6953 case Instruction::LShr:
6954 // If this is a truncate of a logical shr, we can truncate it to a smaller
6955 // lshr iff we know that the bits we would otherwise be shifting in are
6957 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6958 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6959 uint32_t BitWidth = Ty->getBitWidth();
6960 if (BitWidth < OrigBitWidth &&
6961 MaskedValueIsZero(I->getOperand(0),
6962 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6963 CI->getLimitedValue(BitWidth) < BitWidth) {
6964 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6969 case Instruction::ZExt:
6970 case Instruction::SExt:
6971 case Instruction::Trunc:
6972 // If this is the same kind of case as our original (e.g. zext+zext), we
6973 // can safely replace it. Note that replacing it does not reduce the number
6974 // of casts in the input.
6975 if (I->getOpcode() == CastOpc)
6979 case Instruction::PHI: {
6980 // We can change a phi if we can change all operands.
6981 PHINode *PN = cast<PHINode>(I);
6982 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
6983 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
6989 // TODO: Can handle more cases here.
6996 /// EvaluateInDifferentType - Given an expression that
6997 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6998 /// evaluate the expression.
6999 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7001 if (Constant *C = dyn_cast<Constant>(V))
7002 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7004 // Otherwise, it must be an instruction.
7005 Instruction *I = cast<Instruction>(V);
7006 Instruction *Res = 0;
7007 switch (I->getOpcode()) {
7008 case Instruction::Add:
7009 case Instruction::Sub:
7010 case Instruction::Mul:
7011 case Instruction::And:
7012 case Instruction::Or:
7013 case Instruction::Xor:
7014 case Instruction::AShr:
7015 case Instruction::LShr:
7016 case Instruction::Shl: {
7017 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7018 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7019 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7023 case Instruction::Trunc:
7024 case Instruction::ZExt:
7025 case Instruction::SExt:
7026 // If the source type of the cast is the type we're trying for then we can
7027 // just return the source. There's no need to insert it because it is not
7029 if (I->getOperand(0)->getType() == Ty)
7030 return I->getOperand(0);
7032 // Otherwise, must be the same type of cast, so just reinsert a new one.
7033 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7036 case Instruction::PHI: {
7037 PHINode *OPN = cast<PHINode>(I);
7038 PHINode *NPN = PHINode::Create(Ty);
7039 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7040 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7041 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7047 // TODO: Can handle more cases here.
7048 assert(0 && "Unreachable!");
7053 return InsertNewInstBefore(Res, *I);
7056 /// @brief Implement the transforms common to all CastInst visitors.
7057 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7058 Value *Src = CI.getOperand(0);
7060 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7061 // eliminate it now.
7062 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7063 if (Instruction::CastOps opc =
7064 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7065 // The first cast (CSrc) is eliminable so we need to fix up or replace
7066 // the second cast (CI). CSrc will then have a good chance of being dead.
7067 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7071 // If we are casting a select then fold the cast into the select
7072 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7073 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7076 // If we are casting a PHI then fold the cast into the PHI
7077 if (isa<PHINode>(Src))
7078 if (Instruction *NV = FoldOpIntoPhi(CI))
7084 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7085 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7086 Value *Src = CI.getOperand(0);
7088 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7089 // If casting the result of a getelementptr instruction with no offset, turn
7090 // this into a cast of the original pointer!
7091 if (GEP->hasAllZeroIndices()) {
7092 // Changing the cast operand is usually not a good idea but it is safe
7093 // here because the pointer operand is being replaced with another
7094 // pointer operand so the opcode doesn't need to change.
7096 CI.setOperand(0, GEP->getOperand(0));
7100 // If the GEP has a single use, and the base pointer is a bitcast, and the
7101 // GEP computes a constant offset, see if we can convert these three
7102 // instructions into fewer. This typically happens with unions and other
7103 // non-type-safe code.
7104 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7105 if (GEP->hasAllConstantIndices()) {
7106 // We are guaranteed to get a constant from EmitGEPOffset.
7107 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7108 int64_t Offset = OffsetV->getSExtValue();
7110 // Get the base pointer input of the bitcast, and the type it points to.
7111 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7112 const Type *GEPIdxTy =
7113 cast<PointerType>(OrigBase->getType())->getElementType();
7114 if (GEPIdxTy->isSized()) {
7115 SmallVector<Value*, 8> NewIndices;
7117 // Start with the index over the outer type. Note that the type size
7118 // might be zero (even if the offset isn't zero) if the indexed type
7119 // is something like [0 x {int, int}]
7120 const Type *IntPtrTy = TD->getIntPtrType();
7121 int64_t FirstIdx = 0;
7122 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7123 FirstIdx = Offset/TySize;
7126 // Handle silly modulus not returning values values [0..TySize).
7130 assert(Offset >= 0);
7132 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7135 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7137 // Index into the types. If we fail, set OrigBase to null.
7139 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7140 const StructLayout *SL = TD->getStructLayout(STy);
7141 if (Offset < (int64_t)SL->getSizeInBytes()) {
7142 unsigned Elt = SL->getElementContainingOffset(Offset);
7143 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7145 Offset -= SL->getElementOffset(Elt);
7146 GEPIdxTy = STy->getElementType(Elt);
7148 // Otherwise, we can't index into this, bail out.
7152 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7153 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7154 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7155 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7158 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7160 GEPIdxTy = STy->getElementType();
7162 // Otherwise, we can't index into this, bail out.
7168 // If we were able to index down into an element, create the GEP
7169 // and bitcast the result. This eliminates one bitcast, potentially
7171 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7173 NewIndices.end(), "");
7174 InsertNewInstBefore(NGEP, CI);
7175 NGEP->takeName(GEP);
7177 if (isa<BitCastInst>(CI))
7178 return new BitCastInst(NGEP, CI.getType());
7179 assert(isa<PtrToIntInst>(CI));
7180 return new PtrToIntInst(NGEP, CI.getType());
7187 return commonCastTransforms(CI);
7192 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7193 /// integer types. This function implements the common transforms for all those
7195 /// @brief Implement the transforms common to CastInst with integer operands
7196 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7197 if (Instruction *Result = commonCastTransforms(CI))
7200 Value *Src = CI.getOperand(0);
7201 const Type *SrcTy = Src->getType();
7202 const Type *DestTy = CI.getType();
7203 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7204 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7206 // See if we can simplify any instructions used by the LHS whose sole
7207 // purpose is to compute bits we don't care about.
7208 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7209 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7210 KnownZero, KnownOne))
7213 // If the source isn't an instruction or has more than one use then we
7214 // can't do anything more.
7215 Instruction *SrcI = dyn_cast<Instruction>(Src);
7216 if (!SrcI || !Src->hasOneUse())
7219 // Attempt to propagate the cast into the instruction for int->int casts.
7220 int NumCastsRemoved = 0;
7221 if (!isa<BitCastInst>(CI) &&
7222 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7223 CI.getOpcode(), NumCastsRemoved)) {
7224 // If this cast is a truncate, evaluting in a different type always
7225 // eliminates the cast, so it is always a win. If this is a zero-extension,
7226 // we need to do an AND to maintain the clear top-part of the computation,
7227 // so we require that the input have eliminated at least one cast. If this
7228 // is a sign extension, we insert two new casts (to do the extension) so we
7229 // require that two casts have been eliminated.
7231 switch (CI.getOpcode()) {
7233 // All the others use floating point so we shouldn't actually
7234 // get here because of the check above.
7235 assert(0 && "Unknown cast type");
7236 case Instruction::Trunc:
7239 case Instruction::ZExt:
7240 DoXForm = NumCastsRemoved >= 1;
7242 case Instruction::SExt:
7243 DoXForm = NumCastsRemoved >= 2;
7248 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7249 CI.getOpcode() == Instruction::SExt);
7250 assert(Res->getType() == DestTy);
7251 switch (CI.getOpcode()) {
7252 default: assert(0 && "Unknown cast type!");
7253 case Instruction::Trunc:
7254 case Instruction::BitCast:
7255 // Just replace this cast with the result.
7256 return ReplaceInstUsesWith(CI, Res);
7257 case Instruction::ZExt: {
7258 // We need to emit an AND to clear the high bits.
7259 assert(SrcBitSize < DestBitSize && "Not a zext?");
7260 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7262 return BinaryOperator::CreateAnd(Res, C);
7264 case Instruction::SExt:
7265 // We need to emit a cast to truncate, then a cast to sext.
7266 return CastInst::Create(Instruction::SExt,
7267 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7273 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7274 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7276 switch (SrcI->getOpcode()) {
7277 case Instruction::Add:
7278 case Instruction::Mul:
7279 case Instruction::And:
7280 case Instruction::Or:
7281 case Instruction::Xor:
7282 // If we are discarding information, rewrite.
7283 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7284 // Don't insert two casts if they cannot be eliminated. We allow
7285 // two casts to be inserted if the sizes are the same. This could
7286 // only be converting signedness, which is a noop.
7287 if (DestBitSize == SrcBitSize ||
7288 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7289 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7290 Instruction::CastOps opcode = CI.getOpcode();
7291 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7292 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7293 return BinaryOperator::Create(
7294 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7298 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7299 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7300 SrcI->getOpcode() == Instruction::Xor &&
7301 Op1 == ConstantInt::getTrue() &&
7302 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7303 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7304 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7307 case Instruction::SDiv:
7308 case Instruction::UDiv:
7309 case Instruction::SRem:
7310 case Instruction::URem:
7311 // If we are just changing the sign, rewrite.
7312 if (DestBitSize == SrcBitSize) {
7313 // Don't insert two casts if they cannot be eliminated. We allow
7314 // two casts to be inserted if the sizes are the same. This could
7315 // only be converting signedness, which is a noop.
7316 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7317 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7318 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7320 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7322 return BinaryOperator::Create(
7323 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7328 case Instruction::Shl:
7329 // Allow changing the sign of the source operand. Do not allow
7330 // changing the size of the shift, UNLESS the shift amount is a
7331 // constant. We must not change variable sized shifts to a smaller
7332 // size, because it is undefined to shift more bits out than exist
7334 if (DestBitSize == SrcBitSize ||
7335 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7336 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7337 Instruction::BitCast : Instruction::Trunc);
7338 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7339 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7340 return BinaryOperator::CreateShl(Op0c, Op1c);
7343 case Instruction::AShr:
7344 // If this is a signed shr, and if all bits shifted in are about to be
7345 // truncated off, turn it into an unsigned shr to allow greater
7347 if (DestBitSize < SrcBitSize &&
7348 isa<ConstantInt>(Op1)) {
7349 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7350 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7351 // Insert the new logical shift right.
7352 return BinaryOperator::CreateLShr(Op0, Op1);
7360 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7361 if (Instruction *Result = commonIntCastTransforms(CI))
7364 Value *Src = CI.getOperand(0);
7365 const Type *Ty = CI.getType();
7366 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7367 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7369 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7370 switch (SrcI->getOpcode()) {
7372 case Instruction::LShr:
7373 // We can shrink lshr to something smaller if we know the bits shifted in
7374 // are already zeros.
7375 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7376 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7378 // Get a mask for the bits shifting in.
7379 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7380 Value* SrcIOp0 = SrcI->getOperand(0);
7381 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7382 if (ShAmt >= DestBitWidth) // All zeros.
7383 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7385 // Okay, we can shrink this. Truncate the input, then return a new
7387 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7388 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7390 return BinaryOperator::CreateLShr(V1, V2);
7392 } else { // This is a variable shr.
7394 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7395 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7396 // loop-invariant and CSE'd.
7397 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7398 Value *One = ConstantInt::get(SrcI->getType(), 1);
7400 Value *V = InsertNewInstBefore(
7401 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7403 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7404 SrcI->getOperand(0),
7406 Value *Zero = Constant::getNullValue(V->getType());
7407 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7417 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7418 /// in order to eliminate the icmp.
7419 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7421 // If we are just checking for a icmp eq of a single bit and zext'ing it
7422 // to an integer, then shift the bit to the appropriate place and then
7423 // cast to integer to avoid the comparison.
7424 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7425 const APInt &Op1CV = Op1C->getValue();
7427 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7428 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7429 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7430 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7431 if (!DoXform) return ICI;
7433 Value *In = ICI->getOperand(0);
7434 Value *Sh = ConstantInt::get(In->getType(),
7435 In->getType()->getPrimitiveSizeInBits()-1);
7436 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7437 In->getName()+".lobit"),
7439 if (In->getType() != CI.getType())
7440 In = CastInst::CreateIntegerCast(In, CI.getType(),
7441 false/*ZExt*/, "tmp", &CI);
7443 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7444 Constant *One = ConstantInt::get(In->getType(), 1);
7445 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7446 In->getName()+".not"),
7450 return ReplaceInstUsesWith(CI, In);
7455 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7456 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7457 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7458 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7459 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7460 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7461 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7462 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7463 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7464 // This only works for EQ and NE
7465 ICI->isEquality()) {
7466 // If Op1C some other power of two, convert:
7467 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7468 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7469 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7470 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7472 APInt KnownZeroMask(~KnownZero);
7473 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7474 if (!DoXform) return ICI;
7476 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7477 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7478 // (X&4) == 2 --> false
7479 // (X&4) != 2 --> true
7480 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7481 Res = ConstantExpr::getZExt(Res, CI.getType());
7482 return ReplaceInstUsesWith(CI, Res);
7485 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7486 Value *In = ICI->getOperand(0);
7488 // Perform a logical shr by shiftamt.
7489 // Insert the shift to put the result in the low bit.
7490 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7491 ConstantInt::get(In->getType(), ShiftAmt),
7492 In->getName()+".lobit"), CI);
7495 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7496 Constant *One = ConstantInt::get(In->getType(), 1);
7497 In = BinaryOperator::CreateXor(In, One, "tmp");
7498 InsertNewInstBefore(cast<Instruction>(In), CI);
7501 if (CI.getType() == In->getType())
7502 return ReplaceInstUsesWith(CI, In);
7504 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7512 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7513 // If one of the common conversion will work ..
7514 if (Instruction *Result = commonIntCastTransforms(CI))
7517 Value *Src = CI.getOperand(0);
7519 // If this is a cast of a cast
7520 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7521 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7522 // types and if the sizes are just right we can convert this into a logical
7523 // 'and' which will be much cheaper than the pair of casts.
7524 if (isa<TruncInst>(CSrc)) {
7525 // Get the sizes of the types involved
7526 Value *A = CSrc->getOperand(0);
7527 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7528 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7529 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7530 // If we're actually extending zero bits and the trunc is a no-op
7531 if (MidSize < DstSize && SrcSize == DstSize) {
7532 // Replace both of the casts with an And of the type mask.
7533 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7534 Constant *AndConst = ConstantInt::get(AndValue);
7536 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
7537 // Unfortunately, if the type changed, we need to cast it back.
7538 if (And->getType() != CI.getType()) {
7539 And->setName(CSrc->getName()+".mask");
7540 InsertNewInstBefore(And, CI);
7541 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
7548 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7549 return transformZExtICmp(ICI, CI);
7551 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7552 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7553 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7554 // of the (zext icmp) will be transformed.
7555 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7556 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7557 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7558 (transformZExtICmp(LHS, CI, false) ||
7559 transformZExtICmp(RHS, CI, false))) {
7560 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7561 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7562 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
7569 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7570 if (Instruction *I = commonIntCastTransforms(CI))
7573 Value *Src = CI.getOperand(0);
7575 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7576 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7577 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7578 // If we are just checking for a icmp eq of a single bit and zext'ing it
7579 // to an integer, then shift the bit to the appropriate place and then
7580 // cast to integer to avoid the comparison.
7581 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7582 const APInt &Op1CV = Op1C->getValue();
7584 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7585 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7586 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7587 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7588 Value *In = ICI->getOperand(0);
7589 Value *Sh = ConstantInt::get(In->getType(),
7590 In->getType()->getPrimitiveSizeInBits()-1);
7591 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
7592 In->getName()+".lobit"),
7594 if (In->getType() != CI.getType())
7595 In = CastInst::CreateIntegerCast(In, CI.getType(),
7596 true/*SExt*/, "tmp", &CI);
7598 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7599 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
7600 In->getName()+".not"), CI);
7602 return ReplaceInstUsesWith(CI, In);
7607 // See if the value being truncated is already sign extended. If so, just
7608 // eliminate the trunc/sext pair.
7609 if (getOpcode(Src) == Instruction::Trunc) {
7610 Value *Op = cast<User>(Src)->getOperand(0);
7611 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
7612 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
7613 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
7614 unsigned NumSignBits = ComputeNumSignBits(Op);
7616 if (OpBits == DestBits) {
7617 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7618 // bits, it is already ready.
7619 if (NumSignBits > DestBits-MidBits)
7620 return ReplaceInstUsesWith(CI, Op);
7621 } else if (OpBits < DestBits) {
7622 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7623 // bits, just sext from i32.
7624 if (NumSignBits > OpBits-MidBits)
7625 return new SExtInst(Op, CI.getType(), "tmp");
7627 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7628 // bits, just truncate to i32.
7629 if (NumSignBits > OpBits-MidBits)
7630 return new TruncInst(Op, CI.getType(), "tmp");
7637 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7638 /// in the specified FP type without changing its value.
7639 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7640 APFloat F = CFP->getValueAPF();
7641 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7642 return ConstantFP::get(F);
7646 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7647 /// through it until we get the source value.
7648 static Value *LookThroughFPExtensions(Value *V) {
7649 if (Instruction *I = dyn_cast<Instruction>(V))
7650 if (I->getOpcode() == Instruction::FPExt)
7651 return LookThroughFPExtensions(I->getOperand(0));
7653 // If this value is a constant, return the constant in the smallest FP type
7654 // that can accurately represent it. This allows us to turn
7655 // (float)((double)X+2.0) into x+2.0f.
7656 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7657 if (CFP->getType() == Type::PPC_FP128Ty)
7658 return V; // No constant folding of this.
7659 // See if the value can be truncated to float and then reextended.
7660 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7662 if (CFP->getType() == Type::DoubleTy)
7663 return V; // Won't shrink.
7664 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7666 // Don't try to shrink to various long double types.
7672 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7673 if (Instruction *I = commonCastTransforms(CI))
7676 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7677 // smaller than the destination type, we can eliminate the truncate by doing
7678 // the add as the smaller type. This applies to add/sub/mul/div as well as
7679 // many builtins (sqrt, etc).
7680 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7681 if (OpI && OpI->hasOneUse()) {
7682 switch (OpI->getOpcode()) {
7684 case Instruction::Add:
7685 case Instruction::Sub:
7686 case Instruction::Mul:
7687 case Instruction::FDiv:
7688 case Instruction::FRem:
7689 const Type *SrcTy = OpI->getType();
7690 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7691 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7692 if (LHSTrunc->getType() != SrcTy &&
7693 RHSTrunc->getType() != SrcTy) {
7694 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7695 // If the source types were both smaller than the destination type of
7696 // the cast, do this xform.
7697 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7698 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7699 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7701 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7703 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7712 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7713 return commonCastTransforms(CI);
7716 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
7717 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
7718 // mantissa to accurately represent all values of X. For example, do not
7719 // do this with i64->float->i64.
7720 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
7721 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7722 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
7723 SrcI->getType()->getFPMantissaWidth())
7724 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7726 return commonCastTransforms(FI);
7729 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
7730 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
7731 // mantissa to accurately represent all values of X. For example, do not
7732 // do this with i64->float->i64.
7733 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
7734 if (SrcI->getOperand(0)->getType() == FI.getType() &&
7735 (int)FI.getType()->getPrimitiveSizeInBits() <=
7736 SrcI->getType()->getFPMantissaWidth())
7737 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
7739 return commonCastTransforms(FI);
7742 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7743 return commonCastTransforms(CI);
7746 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7747 return commonCastTransforms(CI);
7750 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7751 return commonPointerCastTransforms(CI);
7754 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7755 if (Instruction *I = commonCastTransforms(CI))
7758 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7759 if (!DestPointee->isSized()) return 0;
7761 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7764 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7765 m_ConstantInt(Cst)))) {
7766 // If the source and destination operands have the same type, see if this
7767 // is a single-index GEP.
7768 if (X->getType() == CI.getType()) {
7769 // Get the size of the pointee type.
7770 uint64_t Size = TD->getABITypeSize(DestPointee);
7772 // Convert the constant to intptr type.
7773 APInt Offset = Cst->getValue();
7774 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7776 // If Offset is evenly divisible by Size, we can do this xform.
7777 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7778 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7779 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7782 // TODO: Could handle other cases, e.g. where add is indexing into field of
7784 } else if (CI.getOperand(0)->hasOneUse() &&
7785 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7786 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7787 // "inttoptr+GEP" instead of "add+intptr".
7789 // Get the size of the pointee type.
7790 uint64_t Size = TD->getABITypeSize(DestPointee);
7792 // Convert the constant to intptr type.
7793 APInt Offset = Cst->getValue();
7794 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7796 // If Offset is evenly divisible by Size, we can do this xform.
7797 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7798 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7800 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7802 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7808 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7809 // If the operands are integer typed then apply the integer transforms,
7810 // otherwise just apply the common ones.
7811 Value *Src = CI.getOperand(0);
7812 const Type *SrcTy = Src->getType();
7813 const Type *DestTy = CI.getType();
7815 if (SrcTy->isInteger() && DestTy->isInteger()) {
7816 if (Instruction *Result = commonIntCastTransforms(CI))
7818 } else if (isa<PointerType>(SrcTy)) {
7819 if (Instruction *I = commonPointerCastTransforms(CI))
7822 if (Instruction *Result = commonCastTransforms(CI))
7827 // Get rid of casts from one type to the same type. These are useless and can
7828 // be replaced by the operand.
7829 if (DestTy == Src->getType())
7830 return ReplaceInstUsesWith(CI, Src);
7832 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7833 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7834 const Type *DstElTy = DstPTy->getElementType();
7835 const Type *SrcElTy = SrcPTy->getElementType();
7837 // If the address spaces don't match, don't eliminate the bitcast, which is
7838 // required for changing types.
7839 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7842 // If we are casting a malloc or alloca to a pointer to a type of the same
7843 // size, rewrite the allocation instruction to allocate the "right" type.
7844 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7845 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7848 // If the source and destination are pointers, and this cast is equivalent
7849 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7850 // This can enhance SROA and other transforms that want type-safe pointers.
7851 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7852 unsigned NumZeros = 0;
7853 while (SrcElTy != DstElTy &&
7854 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7855 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7856 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7860 // If we found a path from the src to dest, create the getelementptr now.
7861 if (SrcElTy == DstElTy) {
7862 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7863 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7864 ((Instruction*) NULL));
7868 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7869 if (SVI->hasOneUse()) {
7870 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7871 // a bitconvert to a vector with the same # elts.
7872 if (isa<VectorType>(DestTy) &&
7873 cast<VectorType>(DestTy)->getNumElements() ==
7874 SVI->getType()->getNumElements()) {
7876 // If either of the operands is a cast from CI.getType(), then
7877 // evaluating the shuffle in the casted destination's type will allow
7878 // us to eliminate at least one cast.
7879 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7880 Tmp->getOperand(0)->getType() == DestTy) ||
7881 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7882 Tmp->getOperand(0)->getType() == DestTy)) {
7883 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7884 SVI->getOperand(0), DestTy, &CI);
7885 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7886 SVI->getOperand(1), DestTy, &CI);
7887 // Return a new shuffle vector. Use the same element ID's, as we
7888 // know the vector types match #elts.
7889 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7897 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7899 /// %D = select %cond, %C, %A
7901 /// %C = select %cond, %B, 0
7904 /// Assuming that the specified instruction is an operand to the select, return
7905 /// a bitmask indicating which operands of this instruction are foldable if they
7906 /// equal the other incoming value of the select.
7908 static unsigned GetSelectFoldableOperands(Instruction *I) {
7909 switch (I->getOpcode()) {
7910 case Instruction::Add:
7911 case Instruction::Mul:
7912 case Instruction::And:
7913 case Instruction::Or:
7914 case Instruction::Xor:
7915 return 3; // Can fold through either operand.
7916 case Instruction::Sub: // Can only fold on the amount subtracted.
7917 case Instruction::Shl: // Can only fold on the shift amount.
7918 case Instruction::LShr:
7919 case Instruction::AShr:
7922 return 0; // Cannot fold
7926 /// GetSelectFoldableConstant - For the same transformation as the previous
7927 /// function, return the identity constant that goes into the select.
7928 static Constant *GetSelectFoldableConstant(Instruction *I) {
7929 switch (I->getOpcode()) {
7930 default: assert(0 && "This cannot happen!"); abort();
7931 case Instruction::Add:
7932 case Instruction::Sub:
7933 case Instruction::Or:
7934 case Instruction::Xor:
7935 case Instruction::Shl:
7936 case Instruction::LShr:
7937 case Instruction::AShr:
7938 return Constant::getNullValue(I->getType());
7939 case Instruction::And:
7940 return Constant::getAllOnesValue(I->getType());
7941 case Instruction::Mul:
7942 return ConstantInt::get(I->getType(), 1);
7946 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7947 /// have the same opcode and only one use each. Try to simplify this.
7948 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7950 if (TI->getNumOperands() == 1) {
7951 // If this is a non-volatile load or a cast from the same type,
7954 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7957 return 0; // unknown unary op.
7960 // Fold this by inserting a select from the input values.
7961 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
7962 FI->getOperand(0), SI.getName()+".v");
7963 InsertNewInstBefore(NewSI, SI);
7964 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
7968 // Only handle binary operators here.
7969 if (!isa<BinaryOperator>(TI))
7972 // Figure out if the operations have any operands in common.
7973 Value *MatchOp, *OtherOpT, *OtherOpF;
7975 if (TI->getOperand(0) == FI->getOperand(0)) {
7976 MatchOp = TI->getOperand(0);
7977 OtherOpT = TI->getOperand(1);
7978 OtherOpF = FI->getOperand(1);
7979 MatchIsOpZero = true;
7980 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7981 MatchOp = TI->getOperand(1);
7982 OtherOpT = TI->getOperand(0);
7983 OtherOpF = FI->getOperand(0);
7984 MatchIsOpZero = false;
7985 } else if (!TI->isCommutative()) {
7987 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7988 MatchOp = TI->getOperand(0);
7989 OtherOpT = TI->getOperand(1);
7990 OtherOpF = FI->getOperand(0);
7991 MatchIsOpZero = true;
7992 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7993 MatchOp = TI->getOperand(1);
7994 OtherOpT = TI->getOperand(0);
7995 OtherOpF = FI->getOperand(1);
7996 MatchIsOpZero = true;
8001 // If we reach here, they do have operations in common.
8002 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8003 OtherOpF, SI.getName()+".v");
8004 InsertNewInstBefore(NewSI, SI);
8006 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8008 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8010 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8012 assert(0 && "Shouldn't get here");
8016 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8017 Value *CondVal = SI.getCondition();
8018 Value *TrueVal = SI.getTrueValue();
8019 Value *FalseVal = SI.getFalseValue();
8021 // select true, X, Y -> X
8022 // select false, X, Y -> Y
8023 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8024 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8026 // select C, X, X -> X
8027 if (TrueVal == FalseVal)
8028 return ReplaceInstUsesWith(SI, TrueVal);
8030 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8031 return ReplaceInstUsesWith(SI, FalseVal);
8032 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8033 return ReplaceInstUsesWith(SI, TrueVal);
8034 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8035 if (isa<Constant>(TrueVal))
8036 return ReplaceInstUsesWith(SI, TrueVal);
8038 return ReplaceInstUsesWith(SI, FalseVal);
8041 if (SI.getType() == Type::Int1Ty) {
8042 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8043 if (C->getZExtValue()) {
8044 // Change: A = select B, true, C --> A = or B, C
8045 return BinaryOperator::CreateOr(CondVal, FalseVal);
8047 // Change: A = select B, false, C --> A = and !B, C
8049 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8050 "not."+CondVal->getName()), SI);
8051 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8053 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8054 if (C->getZExtValue() == false) {
8055 // Change: A = select B, C, false --> A = and B, C
8056 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8058 // Change: A = select B, C, true --> A = or !B, C
8060 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8061 "not."+CondVal->getName()), SI);
8062 return BinaryOperator::CreateOr(NotCond, TrueVal);
8066 // select a, b, a -> a&b
8067 // select a, a, b -> a|b
8068 if (CondVal == TrueVal)
8069 return BinaryOperator::CreateOr(CondVal, FalseVal);
8070 else if (CondVal == FalseVal)
8071 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8074 // Selecting between two integer constants?
8075 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8076 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8077 // select C, 1, 0 -> zext C to int
8078 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8079 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8080 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8081 // select C, 0, 1 -> zext !C to int
8083 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8084 "not."+CondVal->getName()), SI);
8085 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8088 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8090 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8092 // (x <s 0) ? -1 : 0 -> ashr x, 31
8093 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8094 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8095 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8096 // The comparison constant and the result are not neccessarily the
8097 // same width. Make an all-ones value by inserting a AShr.
8098 Value *X = IC->getOperand(0);
8099 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8100 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8101 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8103 InsertNewInstBefore(SRA, SI);
8105 // Finally, convert to the type of the select RHS. We figure out
8106 // if this requires a SExt, Trunc or BitCast based on the sizes.
8107 Instruction::CastOps opc = Instruction::BitCast;
8108 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8109 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8110 if (SRASize < SISize)
8111 opc = Instruction::SExt;
8112 else if (SRASize > SISize)
8113 opc = Instruction::Trunc;
8114 return CastInst::Create(opc, SRA, SI.getType());
8119 // If one of the constants is zero (we know they can't both be) and we
8120 // have an icmp instruction with zero, and we have an 'and' with the
8121 // non-constant value, eliminate this whole mess. This corresponds to
8122 // cases like this: ((X & 27) ? 27 : 0)
8123 if (TrueValC->isZero() || FalseValC->isZero())
8124 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8125 cast<Constant>(IC->getOperand(1))->isNullValue())
8126 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8127 if (ICA->getOpcode() == Instruction::And &&
8128 isa<ConstantInt>(ICA->getOperand(1)) &&
8129 (ICA->getOperand(1) == TrueValC ||
8130 ICA->getOperand(1) == FalseValC) &&
8131 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8132 // Okay, now we know that everything is set up, we just don't
8133 // know whether we have a icmp_ne or icmp_eq and whether the
8134 // true or false val is the zero.
8135 bool ShouldNotVal = !TrueValC->isZero();
8136 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8139 V = InsertNewInstBefore(BinaryOperator::Create(
8140 Instruction::Xor, V, ICA->getOperand(1)), SI);
8141 return ReplaceInstUsesWith(SI, V);
8146 // See if we are selecting two values based on a comparison of the two values.
8147 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8148 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8149 // Transform (X == Y) ? X : Y -> Y
8150 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8151 // This is not safe in general for floating point:
8152 // consider X== -0, Y== +0.
8153 // It becomes safe if either operand is a nonzero constant.
8154 ConstantFP *CFPt, *CFPf;
8155 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8156 !CFPt->getValueAPF().isZero()) ||
8157 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8158 !CFPf->getValueAPF().isZero()))
8159 return ReplaceInstUsesWith(SI, FalseVal);
8161 // Transform (X != Y) ? X : Y -> X
8162 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8163 return ReplaceInstUsesWith(SI, TrueVal);
8164 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8166 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8167 // Transform (X == Y) ? Y : X -> X
8168 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8169 // This is not safe in general for floating point:
8170 // consider X== -0, Y== +0.
8171 // It becomes safe if either operand is a nonzero constant.
8172 ConstantFP *CFPt, *CFPf;
8173 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8174 !CFPt->getValueAPF().isZero()) ||
8175 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8176 !CFPf->getValueAPF().isZero()))
8177 return ReplaceInstUsesWith(SI, FalseVal);
8179 // Transform (X != Y) ? Y : X -> Y
8180 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8181 return ReplaceInstUsesWith(SI, TrueVal);
8182 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8186 // See if we are selecting two values based on a comparison of the two values.
8187 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8188 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8189 // Transform (X == Y) ? X : Y -> Y
8190 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8191 return ReplaceInstUsesWith(SI, FalseVal);
8192 // Transform (X != Y) ? X : Y -> X
8193 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8194 return ReplaceInstUsesWith(SI, TrueVal);
8195 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8197 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8198 // Transform (X == Y) ? Y : X -> X
8199 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8200 return ReplaceInstUsesWith(SI, FalseVal);
8201 // Transform (X != Y) ? Y : X -> Y
8202 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8203 return ReplaceInstUsesWith(SI, TrueVal);
8204 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8208 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8209 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8210 if (TI->hasOneUse() && FI->hasOneUse()) {
8211 Instruction *AddOp = 0, *SubOp = 0;
8213 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8214 if (TI->getOpcode() == FI->getOpcode())
8215 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8218 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8219 // even legal for FP.
8220 if (TI->getOpcode() == Instruction::Sub &&
8221 FI->getOpcode() == Instruction::Add) {
8222 AddOp = FI; SubOp = TI;
8223 } else if (FI->getOpcode() == Instruction::Sub &&
8224 TI->getOpcode() == Instruction::Add) {
8225 AddOp = TI; SubOp = FI;
8229 Value *OtherAddOp = 0;
8230 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8231 OtherAddOp = AddOp->getOperand(1);
8232 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8233 OtherAddOp = AddOp->getOperand(0);
8237 // So at this point we know we have (Y -> OtherAddOp):
8238 // select C, (add X, Y), (sub X, Z)
8239 Value *NegVal; // Compute -Z
8240 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8241 NegVal = ConstantExpr::getNeg(C);
8243 NegVal = InsertNewInstBefore(
8244 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8247 Value *NewTrueOp = OtherAddOp;
8248 Value *NewFalseOp = NegVal;
8250 std::swap(NewTrueOp, NewFalseOp);
8251 Instruction *NewSel =
8252 SelectInst::Create(CondVal, NewTrueOp,
8253 NewFalseOp, SI.getName() + ".p");
8255 NewSel = InsertNewInstBefore(NewSel, SI);
8256 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8261 // See if we can fold the select into one of our operands.
8262 if (SI.getType()->isInteger()) {
8263 // See the comment above GetSelectFoldableOperands for a description of the
8264 // transformation we are doing here.
8265 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8266 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8267 !isa<Constant>(FalseVal))
8268 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8269 unsigned OpToFold = 0;
8270 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8272 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8277 Constant *C = GetSelectFoldableConstant(TVI);
8278 Instruction *NewSel =
8279 SelectInst::Create(SI.getCondition(),
8280 TVI->getOperand(2-OpToFold), C);
8281 InsertNewInstBefore(NewSel, SI);
8282 NewSel->takeName(TVI);
8283 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8284 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8286 assert(0 && "Unknown instruction!!");
8291 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8292 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8293 !isa<Constant>(TrueVal))
8294 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8295 unsigned OpToFold = 0;
8296 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8298 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8303 Constant *C = GetSelectFoldableConstant(FVI);
8304 Instruction *NewSel =
8305 SelectInst::Create(SI.getCondition(), C,
8306 FVI->getOperand(2-OpToFold));
8307 InsertNewInstBefore(NewSel, SI);
8308 NewSel->takeName(FVI);
8309 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8310 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8312 assert(0 && "Unknown instruction!!");
8317 if (BinaryOperator::isNot(CondVal)) {
8318 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8319 SI.setOperand(1, FalseVal);
8320 SI.setOperand(2, TrueVal);
8327 /// EnforceKnownAlignment - If the specified pointer points to an object that
8328 /// we control, modify the object's alignment to PrefAlign. This isn't
8329 /// often possible though. If alignment is important, a more reliable approach
8330 /// is to simply align all global variables and allocation instructions to
8331 /// their preferred alignment from the beginning.
8333 static unsigned EnforceKnownAlignment(Value *V,
8334 unsigned Align, unsigned PrefAlign) {
8336 User *U = dyn_cast<User>(V);
8337 if (!U) return Align;
8339 switch (getOpcode(U)) {
8341 case Instruction::BitCast:
8342 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8343 case Instruction::GetElementPtr: {
8344 // If all indexes are zero, it is just the alignment of the base pointer.
8345 bool AllZeroOperands = true;
8346 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8347 if (!isa<Constant>(*i) ||
8348 !cast<Constant>(*i)->isNullValue()) {
8349 AllZeroOperands = false;
8353 if (AllZeroOperands) {
8354 // Treat this like a bitcast.
8355 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8361 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8362 // If there is a large requested alignment and we can, bump up the alignment
8364 if (!GV->isDeclaration()) {
8365 GV->setAlignment(PrefAlign);
8368 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8369 // If there is a requested alignment and if this is an alloca, round up. We
8370 // don't do this for malloc, because some systems can't respect the request.
8371 if (isa<AllocaInst>(AI)) {
8372 AI->setAlignment(PrefAlign);
8380 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8381 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8382 /// and it is more than the alignment of the ultimate object, see if we can
8383 /// increase the alignment of the ultimate object, making this check succeed.
8384 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8385 unsigned PrefAlign) {
8386 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8387 sizeof(PrefAlign) * CHAR_BIT;
8388 APInt Mask = APInt::getAllOnesValue(BitWidth);
8389 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8390 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8391 unsigned TrailZ = KnownZero.countTrailingOnes();
8392 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8394 if (PrefAlign > Align)
8395 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8397 // We don't need to make any adjustment.
8401 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8402 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8403 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8404 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8405 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8407 if (CopyAlign < MinAlign) {
8408 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8412 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8414 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8415 if (MemOpLength == 0) return 0;
8417 // Source and destination pointer types are always "i8*" for intrinsic. See
8418 // if the size is something we can handle with a single primitive load/store.
8419 // A single load+store correctly handles overlapping memory in the memmove
8421 unsigned Size = MemOpLength->getZExtValue();
8422 if (Size == 0) return MI; // Delete this mem transfer.
8424 if (Size > 8 || (Size&(Size-1)))
8425 return 0; // If not 1/2/4/8 bytes, exit.
8427 // Use an integer load+store unless we can find something better.
8428 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8430 // Memcpy forces the use of i8* for the source and destination. That means
8431 // that if you're using memcpy to move one double around, you'll get a cast
8432 // from double* to i8*. We'd much rather use a double load+store rather than
8433 // an i64 load+store, here because this improves the odds that the source or
8434 // dest address will be promotable. See if we can find a better type than the
8435 // integer datatype.
8436 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8437 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8438 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8439 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8440 // down through these levels if so.
8441 while (!SrcETy->isSingleValueType()) {
8442 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8443 if (STy->getNumElements() == 1)
8444 SrcETy = STy->getElementType(0);
8447 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8448 if (ATy->getNumElements() == 1)
8449 SrcETy = ATy->getElementType();
8456 if (SrcETy->isSingleValueType())
8457 NewPtrTy = PointerType::getUnqual(SrcETy);
8462 // If the memcpy/memmove provides better alignment info than we can
8464 SrcAlign = std::max(SrcAlign, CopyAlign);
8465 DstAlign = std::max(DstAlign, CopyAlign);
8467 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8468 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8469 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8470 InsertNewInstBefore(L, *MI);
8471 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8473 // Set the size of the copy to 0, it will be deleted on the next iteration.
8474 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8478 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8479 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8480 if (MI->getAlignment()->getZExtValue() < Alignment) {
8481 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8485 // Extract the length and alignment and fill if they are constant.
8486 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8487 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8488 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8490 uint64_t Len = LenC->getZExtValue();
8491 Alignment = MI->getAlignment()->getZExtValue();
8493 // If the length is zero, this is a no-op
8494 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8496 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8497 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8498 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8500 Value *Dest = MI->getDest();
8501 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8503 // Alignment 0 is identity for alignment 1 for memset, but not store.
8504 if (Alignment == 0) Alignment = 1;
8506 // Extract the fill value and store.
8507 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8508 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8511 // Set the size of the copy to 0, it will be deleted on the next iteration.
8512 MI->setLength(Constant::getNullValue(LenC->getType()));
8520 /// visitCallInst - CallInst simplification. This mostly only handles folding
8521 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8522 /// the heavy lifting.
8524 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8525 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8526 if (!II) return visitCallSite(&CI);
8528 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8530 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8531 bool Changed = false;
8533 // memmove/cpy/set of zero bytes is a noop.
8534 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8535 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8537 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8538 if (CI->getZExtValue() == 1) {
8539 // Replace the instruction with just byte operations. We would
8540 // transform other cases to loads/stores, but we don't know if
8541 // alignment is sufficient.
8545 // If we have a memmove and the source operation is a constant global,
8546 // then the source and dest pointers can't alias, so we can change this
8547 // into a call to memcpy.
8548 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8549 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8550 if (GVSrc->isConstant()) {
8551 Module *M = CI.getParent()->getParent()->getParent();
8552 Intrinsic::ID MemCpyID;
8553 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8554 MemCpyID = Intrinsic::memcpy_i32;
8556 MemCpyID = Intrinsic::memcpy_i64;
8557 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8561 // memmove(x,x,size) -> noop.
8562 if (MMI->getSource() == MMI->getDest())
8563 return EraseInstFromFunction(CI);
8566 // If we can determine a pointer alignment that is bigger than currently
8567 // set, update the alignment.
8568 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8569 if (Instruction *I = SimplifyMemTransfer(MI))
8571 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8572 if (Instruction *I = SimplifyMemSet(MSI))
8576 if (Changed) return II;
8579 switch (II->getIntrinsicID()) {
8581 case Intrinsic::bswap:
8582 // bswap(bswap(x)) -> x
8583 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
8584 if (Operand->getIntrinsicID() == Intrinsic::bswap)
8585 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
8587 case Intrinsic::ppc_altivec_lvx:
8588 case Intrinsic::ppc_altivec_lvxl:
8589 case Intrinsic::x86_sse_loadu_ps:
8590 case Intrinsic::x86_sse2_loadu_pd:
8591 case Intrinsic::x86_sse2_loadu_dq:
8592 // Turn PPC lvx -> load if the pointer is known aligned.
8593 // Turn X86 loadups -> load if the pointer is known aligned.
8594 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8595 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8596 PointerType::getUnqual(II->getType()),
8598 return new LoadInst(Ptr);
8601 case Intrinsic::ppc_altivec_stvx:
8602 case Intrinsic::ppc_altivec_stvxl:
8603 // Turn stvx -> store if the pointer is known aligned.
8604 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8605 const Type *OpPtrTy =
8606 PointerType::getUnqual(II->getOperand(1)->getType());
8607 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8608 return new StoreInst(II->getOperand(1), Ptr);
8611 case Intrinsic::x86_sse_storeu_ps:
8612 case Intrinsic::x86_sse2_storeu_pd:
8613 case Intrinsic::x86_sse2_storeu_dq:
8614 case Intrinsic::x86_sse2_storel_dq:
8615 // Turn X86 storeu -> store if the pointer is known aligned.
8616 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8617 const Type *OpPtrTy =
8618 PointerType::getUnqual(II->getOperand(2)->getType());
8619 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8620 return new StoreInst(II->getOperand(2), Ptr);
8624 case Intrinsic::x86_sse_cvttss2si: {
8625 // These intrinsics only demands the 0th element of its input vector. If
8626 // we can simplify the input based on that, do so now.
8628 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8630 II->setOperand(1, V);
8636 case Intrinsic::ppc_altivec_vperm:
8637 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8638 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8639 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8641 // Check that all of the elements are integer constants or undefs.
8642 bool AllEltsOk = true;
8643 for (unsigned i = 0; i != 16; ++i) {
8644 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8645 !isa<UndefValue>(Mask->getOperand(i))) {
8652 // Cast the input vectors to byte vectors.
8653 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8654 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8655 Value *Result = UndefValue::get(Op0->getType());
8657 // Only extract each element once.
8658 Value *ExtractedElts[32];
8659 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8661 for (unsigned i = 0; i != 16; ++i) {
8662 if (isa<UndefValue>(Mask->getOperand(i)))
8664 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8665 Idx &= 31; // Match the hardware behavior.
8667 if (ExtractedElts[Idx] == 0) {
8669 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8670 InsertNewInstBefore(Elt, CI);
8671 ExtractedElts[Idx] = Elt;
8674 // Insert this value into the result vector.
8675 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
8677 InsertNewInstBefore(cast<Instruction>(Result), CI);
8679 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
8684 case Intrinsic::stackrestore: {
8685 // If the save is right next to the restore, remove the restore. This can
8686 // happen when variable allocas are DCE'd.
8687 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8688 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8689 BasicBlock::iterator BI = SS;
8691 return EraseInstFromFunction(CI);
8695 // Scan down this block to see if there is another stack restore in the
8696 // same block without an intervening call/alloca.
8697 BasicBlock::iterator BI = II;
8698 TerminatorInst *TI = II->getParent()->getTerminator();
8699 bool CannotRemove = false;
8700 for (++BI; &*BI != TI; ++BI) {
8701 if (isa<AllocaInst>(BI)) {
8702 CannotRemove = true;
8705 if (isa<CallInst>(BI)) {
8706 if (!isa<IntrinsicInst>(BI)) {
8707 CannotRemove = true;
8710 // If there is a stackrestore below this one, remove this one.
8711 return EraseInstFromFunction(CI);
8715 // If the stack restore is in a return/unwind block and if there are no
8716 // allocas or calls between the restore and the return, nuke the restore.
8717 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8718 return EraseInstFromFunction(CI);
8723 return visitCallSite(II);
8726 // InvokeInst simplification
8728 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8729 return visitCallSite(&II);
8732 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8733 /// passed through the varargs area, we can eliminate the use of the cast.
8734 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8735 const CastInst * const CI,
8736 const TargetData * const TD,
8738 if (!CI->isLosslessCast())
8741 // The size of ByVal arguments is derived from the type, so we
8742 // can't change to a type with a different size. If the size were
8743 // passed explicitly we could avoid this check.
8744 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
8748 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
8749 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
8750 if (!SrcTy->isSized() || !DstTy->isSized())
8752 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
8757 // visitCallSite - Improvements for call and invoke instructions.
8759 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8760 bool Changed = false;
8762 // If the callee is a constexpr cast of a function, attempt to move the cast
8763 // to the arguments of the call/invoke.
8764 if (transformConstExprCastCall(CS)) return 0;
8766 Value *Callee = CS.getCalledValue();
8768 if (Function *CalleeF = dyn_cast<Function>(Callee))
8769 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8770 Instruction *OldCall = CS.getInstruction();
8771 // If the call and callee calling conventions don't match, this call must
8772 // be unreachable, as the call is undefined.
8773 new StoreInst(ConstantInt::getTrue(),
8774 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8776 if (!OldCall->use_empty())
8777 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8778 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8779 return EraseInstFromFunction(*OldCall);
8783 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8784 // This instruction is not reachable, just remove it. We insert a store to
8785 // undef so that we know that this code is not reachable, despite the fact
8786 // that we can't modify the CFG here.
8787 new StoreInst(ConstantInt::getTrue(),
8788 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8789 CS.getInstruction());
8791 if (!CS.getInstruction()->use_empty())
8792 CS.getInstruction()->
8793 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8795 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8796 // Don't break the CFG, insert a dummy cond branch.
8797 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8798 ConstantInt::getTrue(), II);
8800 return EraseInstFromFunction(*CS.getInstruction());
8803 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8804 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8805 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8806 return transformCallThroughTrampoline(CS);
8808 const PointerType *PTy = cast<PointerType>(Callee->getType());
8809 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8810 if (FTy->isVarArg()) {
8811 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
8812 // See if we can optimize any arguments passed through the varargs area of
8814 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8815 E = CS.arg_end(); I != E; ++I, ++ix) {
8816 CastInst *CI = dyn_cast<CastInst>(*I);
8817 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
8818 *I = CI->getOperand(0);
8824 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8825 // Inline asm calls cannot throw - mark them 'nounwind'.
8826 CS.setDoesNotThrow();
8830 return Changed ? CS.getInstruction() : 0;
8833 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8834 // attempt to move the cast to the arguments of the call/invoke.
8836 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8837 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8838 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8839 if (CE->getOpcode() != Instruction::BitCast ||
8840 !isa<Function>(CE->getOperand(0)))
8842 Function *Callee = cast<Function>(CE->getOperand(0));
8843 Instruction *Caller = CS.getInstruction();
8844 const PAListPtr &CallerPAL = CS.getParamAttrs();
8846 // Okay, this is a cast from a function to a different type. Unless doing so
8847 // would cause a type conversion of one of our arguments, change this call to
8848 // be a direct call with arguments casted to the appropriate types.
8850 const FunctionType *FT = Callee->getFunctionType();
8851 const Type *OldRetTy = Caller->getType();
8852 const Type *NewRetTy = FT->getReturnType();
8854 if (isa<StructType>(NewRetTy))
8855 return false; // TODO: Handle multiple return values.
8857 // Check to see if we are changing the return type...
8858 if (OldRetTy != NewRetTy) {
8859 if (Callee->isDeclaration() &&
8860 // Conversion is ok if changing from one pointer type to another or from
8861 // a pointer to an integer of the same size.
8862 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
8863 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
8864 return false; // Cannot transform this return value.
8866 if (!Caller->use_empty() &&
8867 // void -> non-void is handled specially
8868 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
8869 return false; // Cannot transform this return value.
8871 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8872 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8873 if (RAttrs & ParamAttr::typeIncompatible(NewRetTy))
8874 return false; // Attribute not compatible with transformed value.
8877 // If the callsite is an invoke instruction, and the return value is used by
8878 // a PHI node in a successor, we cannot change the return type of the call
8879 // because there is no place to put the cast instruction (without breaking
8880 // the critical edge). Bail out in this case.
8881 if (!Caller->use_empty())
8882 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8883 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8885 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8886 if (PN->getParent() == II->getNormalDest() ||
8887 PN->getParent() == II->getUnwindDest())
8891 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8892 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8894 CallSite::arg_iterator AI = CS.arg_begin();
8895 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8896 const Type *ParamTy = FT->getParamType(i);
8897 const Type *ActTy = (*AI)->getType();
8899 if (!CastInst::isCastable(ActTy, ParamTy))
8900 return false; // Cannot transform this parameter value.
8902 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8903 return false; // Attribute not compatible with transformed value.
8905 // Converting from one pointer type to another or between a pointer and an
8906 // integer of the same size is safe even if we do not have a body.
8907 bool isConvertible = ActTy == ParamTy ||
8908 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
8909 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
8910 if (Callee->isDeclaration() && !isConvertible) return false;
8913 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8914 Callee->isDeclaration())
8915 return false; // Do not delete arguments unless we have a function body.
8917 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
8918 !CallerPAL.isEmpty())
8919 // In this case we have more arguments than the new function type, but we
8920 // won't be dropping them. Check that these extra arguments have attributes
8921 // that are compatible with being a vararg call argument.
8922 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
8923 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
8925 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
8926 if (PAttrs & ParamAttr::VarArgsIncompatible)
8930 // Okay, we decided that this is a safe thing to do: go ahead and start
8931 // inserting cast instructions as necessary...
8932 std::vector<Value*> Args;
8933 Args.reserve(NumActualArgs);
8934 SmallVector<ParamAttrsWithIndex, 8> attrVec;
8935 attrVec.reserve(NumCommonArgs);
8937 // Get any return attributes.
8938 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8940 // If the return value is not being used, the type may not be compatible
8941 // with the existing attributes. Wipe out any problematic attributes.
8942 RAttrs &= ~ParamAttr::typeIncompatible(NewRetTy);
8944 // Add the new return attributes.
8946 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8948 AI = CS.arg_begin();
8949 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8950 const Type *ParamTy = FT->getParamType(i);
8951 if ((*AI)->getType() == ParamTy) {
8952 Args.push_back(*AI);
8954 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8955 false, ParamTy, false);
8956 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
8957 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8960 // Add any parameter attributes.
8961 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8962 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8965 // If the function takes more arguments than the call was taking, add them
8967 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8968 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8970 // If we are removing arguments to the function, emit an obnoxious warning...
8971 if (FT->getNumParams() < NumActualArgs) {
8972 if (!FT->isVarArg()) {
8973 cerr << "WARNING: While resolving call to function '"
8974 << Callee->getName() << "' arguments were dropped!\n";
8976 // Add all of the arguments in their promoted form to the arg list...
8977 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8978 const Type *PTy = getPromotedType((*AI)->getType());
8979 if (PTy != (*AI)->getType()) {
8980 // Must promote to pass through va_arg area!
8981 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8983 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
8984 InsertNewInstBefore(Cast, *Caller);
8985 Args.push_back(Cast);
8987 Args.push_back(*AI);
8990 // Add any parameter attributes.
8991 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8992 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8997 if (NewRetTy == Type::VoidTy)
8998 Caller->setName(""); // Void type should not have a name.
9000 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9003 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9004 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9005 Args.begin(), Args.end(),
9006 Caller->getName(), Caller);
9007 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9008 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9010 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9011 Caller->getName(), Caller);
9012 CallInst *CI = cast<CallInst>(Caller);
9013 if (CI->isTailCall())
9014 cast<CallInst>(NC)->setTailCall();
9015 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9016 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9019 // Insert a cast of the return type as necessary.
9021 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9022 if (NV->getType() != Type::VoidTy) {
9023 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9025 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9027 // If this is an invoke instruction, we should insert it after the first
9028 // non-phi, instruction in the normal successor block.
9029 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9030 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9031 InsertNewInstBefore(NC, *I);
9033 // Otherwise, it's a call, just insert cast right after the call instr
9034 InsertNewInstBefore(NC, *Caller);
9036 AddUsersToWorkList(*Caller);
9038 NV = UndefValue::get(Caller->getType());
9042 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9043 Caller->replaceAllUsesWith(NV);
9044 Caller->eraseFromParent();
9045 RemoveFromWorkList(Caller);
9049 // transformCallThroughTrampoline - Turn a call to a function created by the
9050 // init_trampoline intrinsic into a direct call to the underlying function.
9052 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9053 Value *Callee = CS.getCalledValue();
9054 const PointerType *PTy = cast<PointerType>(Callee->getType());
9055 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9056 const PAListPtr &Attrs = CS.getParamAttrs();
9058 // If the call already has the 'nest' attribute somewhere then give up -
9059 // otherwise 'nest' would occur twice after splicing in the chain.
9060 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9063 IntrinsicInst *Tramp =
9064 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9066 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9067 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9068 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9070 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9071 if (!NestAttrs.isEmpty()) {
9072 unsigned NestIdx = 1;
9073 const Type *NestTy = 0;
9074 ParameterAttributes NestAttr = ParamAttr::None;
9076 // Look for a parameter marked with the 'nest' attribute.
9077 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9078 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9079 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9080 // Record the parameter type and any other attributes.
9082 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9087 Instruction *Caller = CS.getInstruction();
9088 std::vector<Value*> NewArgs;
9089 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9091 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9092 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9094 // Insert the nest argument into the call argument list, which may
9095 // mean appending it. Likewise for attributes.
9097 // Add any function result attributes.
9098 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9099 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9103 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9105 if (Idx == NestIdx) {
9106 // Add the chain argument and attributes.
9107 Value *NestVal = Tramp->getOperand(3);
9108 if (NestVal->getType() != NestTy)
9109 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9110 NewArgs.push_back(NestVal);
9111 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9117 // Add the original argument and attributes.
9118 NewArgs.push_back(*I);
9119 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9121 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9127 // The trampoline may have been bitcast to a bogus type (FTy).
9128 // Handle this by synthesizing a new function type, equal to FTy
9129 // with the chain parameter inserted.
9131 std::vector<const Type*> NewTypes;
9132 NewTypes.reserve(FTy->getNumParams()+1);
9134 // Insert the chain's type into the list of parameter types, which may
9135 // mean appending it.
9138 FunctionType::param_iterator I = FTy->param_begin(),
9139 E = FTy->param_end();
9143 // Add the chain's type.
9144 NewTypes.push_back(NestTy);
9149 // Add the original type.
9150 NewTypes.push_back(*I);
9156 // Replace the trampoline call with a direct call. Let the generic
9157 // code sort out any function type mismatches.
9158 FunctionType *NewFTy =
9159 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9160 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9161 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9162 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9164 Instruction *NewCaller;
9165 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9166 NewCaller = InvokeInst::Create(NewCallee,
9167 II->getNormalDest(), II->getUnwindDest(),
9168 NewArgs.begin(), NewArgs.end(),
9169 Caller->getName(), Caller);
9170 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9171 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9173 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9174 Caller->getName(), Caller);
9175 if (cast<CallInst>(Caller)->isTailCall())
9176 cast<CallInst>(NewCaller)->setTailCall();
9177 cast<CallInst>(NewCaller)->
9178 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9179 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9181 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9182 Caller->replaceAllUsesWith(NewCaller);
9183 Caller->eraseFromParent();
9184 RemoveFromWorkList(Caller);
9189 // Replace the trampoline call with a direct call. Since there is no 'nest'
9190 // parameter, there is no need to adjust the argument list. Let the generic
9191 // code sort out any function type mismatches.
9192 Constant *NewCallee =
9193 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9194 CS.setCalledFunction(NewCallee);
9195 return CS.getInstruction();
9198 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9199 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9200 /// and a single binop.
9201 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9202 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9203 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9204 isa<CmpInst>(FirstInst));
9205 unsigned Opc = FirstInst->getOpcode();
9206 Value *LHSVal = FirstInst->getOperand(0);
9207 Value *RHSVal = FirstInst->getOperand(1);
9209 const Type *LHSType = LHSVal->getType();
9210 const Type *RHSType = RHSVal->getType();
9212 // Scan to see if all operands are the same opcode, all have one use, and all
9213 // kill their operands (i.e. the operands have one use).
9214 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9215 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9216 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9217 // Verify type of the LHS matches so we don't fold cmp's of different
9218 // types or GEP's with different index types.
9219 I->getOperand(0)->getType() != LHSType ||
9220 I->getOperand(1)->getType() != RHSType)
9223 // If they are CmpInst instructions, check their predicates
9224 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9225 if (cast<CmpInst>(I)->getPredicate() !=
9226 cast<CmpInst>(FirstInst)->getPredicate())
9229 // Keep track of which operand needs a phi node.
9230 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9231 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9234 // Otherwise, this is safe to transform, determine if it is profitable.
9236 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9237 // Indexes are often folded into load/store instructions, so we don't want to
9238 // hide them behind a phi.
9239 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9242 Value *InLHS = FirstInst->getOperand(0);
9243 Value *InRHS = FirstInst->getOperand(1);
9244 PHINode *NewLHS = 0, *NewRHS = 0;
9246 NewLHS = PHINode::Create(LHSType,
9247 FirstInst->getOperand(0)->getName() + ".pn");
9248 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9249 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9250 InsertNewInstBefore(NewLHS, PN);
9255 NewRHS = PHINode::Create(RHSType,
9256 FirstInst->getOperand(1)->getName() + ".pn");
9257 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9258 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9259 InsertNewInstBefore(NewRHS, PN);
9263 // Add all operands to the new PHIs.
9264 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9266 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9267 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9270 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9271 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9275 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9276 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9277 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9278 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9281 assert(isa<GetElementPtrInst>(FirstInst));
9282 return GetElementPtrInst::Create(LHSVal, RHSVal);
9286 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9287 /// of the block that defines it. This means that it must be obvious the value
9288 /// of the load is not changed from the point of the load to the end of the
9291 /// Finally, it is safe, but not profitable, to sink a load targetting a
9292 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9294 static bool isSafeToSinkLoad(LoadInst *L) {
9295 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9297 for (++BBI; BBI != E; ++BBI)
9298 if (BBI->mayWriteToMemory())
9301 // Check for non-address taken alloca. If not address-taken already, it isn't
9302 // profitable to do this xform.
9303 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9304 bool isAddressTaken = false;
9305 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9307 if (isa<LoadInst>(UI)) continue;
9308 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9309 // If storing TO the alloca, then the address isn't taken.
9310 if (SI->getOperand(1) == AI) continue;
9312 isAddressTaken = true;
9316 if (!isAddressTaken)
9324 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9325 // operator and they all are only used by the PHI, PHI together their
9326 // inputs, and do the operation once, to the result of the PHI.
9327 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9328 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9330 // Scan the instruction, looking for input operations that can be folded away.
9331 // If all input operands to the phi are the same instruction (e.g. a cast from
9332 // the same type or "+42") we can pull the operation through the PHI, reducing
9333 // code size and simplifying code.
9334 Constant *ConstantOp = 0;
9335 const Type *CastSrcTy = 0;
9336 bool isVolatile = false;
9337 if (isa<CastInst>(FirstInst)) {
9338 CastSrcTy = FirstInst->getOperand(0)->getType();
9339 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9340 // Can fold binop, compare or shift here if the RHS is a constant,
9341 // otherwise call FoldPHIArgBinOpIntoPHI.
9342 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9343 if (ConstantOp == 0)
9344 return FoldPHIArgBinOpIntoPHI(PN);
9345 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9346 isVolatile = LI->isVolatile();
9347 // We can't sink the load if the loaded value could be modified between the
9348 // load and the PHI.
9349 if (LI->getParent() != PN.getIncomingBlock(0) ||
9350 !isSafeToSinkLoad(LI))
9352 } else if (isa<GetElementPtrInst>(FirstInst)) {
9353 if (FirstInst->getNumOperands() == 2)
9354 return FoldPHIArgBinOpIntoPHI(PN);
9355 // Can't handle general GEPs yet.
9358 return 0; // Cannot fold this operation.
9361 // Check to see if all arguments are the same operation.
9362 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9363 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9364 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9365 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9368 if (I->getOperand(0)->getType() != CastSrcTy)
9369 return 0; // Cast operation must match.
9370 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9371 // We can't sink the load if the loaded value could be modified between
9372 // the load and the PHI.
9373 if (LI->isVolatile() != isVolatile ||
9374 LI->getParent() != PN.getIncomingBlock(i) ||
9375 !isSafeToSinkLoad(LI))
9378 // If the PHI is volatile and its block has multiple successors, sinking
9379 // it would remove a load of the volatile value from the path through the
9382 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9386 } else if (I->getOperand(1) != ConstantOp) {
9391 // Okay, they are all the same operation. Create a new PHI node of the
9392 // correct type, and PHI together all of the LHS's of the instructions.
9393 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9394 PN.getName()+".in");
9395 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9397 Value *InVal = FirstInst->getOperand(0);
9398 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9400 // Add all operands to the new PHI.
9401 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9402 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9403 if (NewInVal != InVal)
9405 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9410 // The new PHI unions all of the same values together. This is really
9411 // common, so we handle it intelligently here for compile-time speed.
9415 InsertNewInstBefore(NewPN, PN);
9419 // Insert and return the new operation.
9420 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9421 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9422 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9423 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9424 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9425 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9426 PhiVal, ConstantOp);
9427 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9429 // If this was a volatile load that we are merging, make sure to loop through
9430 // and mark all the input loads as non-volatile. If we don't do this, we will
9431 // insert a new volatile load and the old ones will not be deletable.
9433 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9434 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9436 return new LoadInst(PhiVal, "", isVolatile);
9439 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9441 static bool DeadPHICycle(PHINode *PN,
9442 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9443 if (PN->use_empty()) return true;
9444 if (!PN->hasOneUse()) return false;
9446 // Remember this node, and if we find the cycle, return.
9447 if (!PotentiallyDeadPHIs.insert(PN))
9450 // Don't scan crazily complex things.
9451 if (PotentiallyDeadPHIs.size() == 16)
9454 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9455 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9460 /// PHIsEqualValue - Return true if this phi node is always equal to
9461 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9462 /// z = some value; x = phi (y, z); y = phi (x, z)
9463 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9464 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9465 // See if we already saw this PHI node.
9466 if (!ValueEqualPHIs.insert(PN))
9469 // Don't scan crazily complex things.
9470 if (ValueEqualPHIs.size() == 16)
9473 // Scan the operands to see if they are either phi nodes or are equal to
9475 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9476 Value *Op = PN->getIncomingValue(i);
9477 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9478 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9480 } else if (Op != NonPhiInVal)
9488 // PHINode simplification
9490 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9491 // If LCSSA is around, don't mess with Phi nodes
9492 if (MustPreserveLCSSA) return 0;
9494 if (Value *V = PN.hasConstantValue())
9495 return ReplaceInstUsesWith(PN, V);
9497 // If all PHI operands are the same operation, pull them through the PHI,
9498 // reducing code size.
9499 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9500 PN.getIncomingValue(0)->hasOneUse())
9501 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9504 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9505 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9506 // PHI)... break the cycle.
9507 if (PN.hasOneUse()) {
9508 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9509 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9510 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9511 PotentiallyDeadPHIs.insert(&PN);
9512 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9513 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9516 // If this phi has a single use, and if that use just computes a value for
9517 // the next iteration of a loop, delete the phi. This occurs with unused
9518 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9519 // common case here is good because the only other things that catch this
9520 // are induction variable analysis (sometimes) and ADCE, which is only run
9522 if (PHIUser->hasOneUse() &&
9523 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9524 PHIUser->use_back() == &PN) {
9525 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9529 // We sometimes end up with phi cycles that non-obviously end up being the
9530 // same value, for example:
9531 // z = some value; x = phi (y, z); y = phi (x, z)
9532 // where the phi nodes don't necessarily need to be in the same block. Do a
9533 // quick check to see if the PHI node only contains a single non-phi value, if
9534 // so, scan to see if the phi cycle is actually equal to that value.
9536 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9537 // Scan for the first non-phi operand.
9538 while (InValNo != NumOperandVals &&
9539 isa<PHINode>(PN.getIncomingValue(InValNo)))
9542 if (InValNo != NumOperandVals) {
9543 Value *NonPhiInVal = PN.getOperand(InValNo);
9545 // Scan the rest of the operands to see if there are any conflicts, if so
9546 // there is no need to recursively scan other phis.
9547 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9548 Value *OpVal = PN.getIncomingValue(InValNo);
9549 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9553 // If we scanned over all operands, then we have one unique value plus
9554 // phi values. Scan PHI nodes to see if they all merge in each other or
9556 if (InValNo == NumOperandVals) {
9557 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9558 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9559 return ReplaceInstUsesWith(PN, NonPhiInVal);
9566 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9567 Instruction *InsertPoint,
9569 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9570 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9571 // We must cast correctly to the pointer type. Ensure that we
9572 // sign extend the integer value if it is smaller as this is
9573 // used for address computation.
9574 Instruction::CastOps opcode =
9575 (VTySize < PtrSize ? Instruction::SExt :
9576 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9577 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9581 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9582 Value *PtrOp = GEP.getOperand(0);
9583 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9584 // If so, eliminate the noop.
9585 if (GEP.getNumOperands() == 1)
9586 return ReplaceInstUsesWith(GEP, PtrOp);
9588 if (isa<UndefValue>(GEP.getOperand(0)))
9589 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9591 bool HasZeroPointerIndex = false;
9592 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9593 HasZeroPointerIndex = C->isNullValue();
9595 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9596 return ReplaceInstUsesWith(GEP, PtrOp);
9598 // Eliminate unneeded casts for indices.
9599 bool MadeChange = false;
9601 gep_type_iterator GTI = gep_type_begin(GEP);
9602 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
9603 i != e; ++i, ++GTI) {
9604 if (isa<SequentialType>(*GTI)) {
9605 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
9606 if (CI->getOpcode() == Instruction::ZExt ||
9607 CI->getOpcode() == Instruction::SExt) {
9608 const Type *SrcTy = CI->getOperand(0)->getType();
9609 // We can eliminate a cast from i32 to i64 iff the target
9610 // is a 32-bit pointer target.
9611 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9613 *i = CI->getOperand(0);
9617 // If we are using a wider index than needed for this platform, shrink it
9618 // to what we need. If the incoming value needs a cast instruction,
9619 // insert it. This explicit cast can make subsequent optimizations more
9622 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9623 if (Constant *C = dyn_cast<Constant>(Op)) {
9624 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
9627 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9635 if (MadeChange) return &GEP;
9637 // If this GEP instruction doesn't move the pointer, and if the input operand
9638 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9639 // real input to the dest type.
9640 if (GEP.hasAllZeroIndices()) {
9641 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9642 // If the bitcast is of an allocation, and the allocation will be
9643 // converted to match the type of the cast, don't touch this.
9644 if (isa<AllocationInst>(BCI->getOperand(0))) {
9645 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9646 if (Instruction *I = visitBitCast(*BCI)) {
9649 BCI->getParent()->getInstList().insert(BCI, I);
9650 ReplaceInstUsesWith(*BCI, I);
9655 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9659 // Combine Indices - If the source pointer to this getelementptr instruction
9660 // is a getelementptr instruction, combine the indices of the two
9661 // getelementptr instructions into a single instruction.
9663 SmallVector<Value*, 8> SrcGEPOperands;
9664 if (User *Src = dyn_castGetElementPtr(PtrOp))
9665 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9667 if (!SrcGEPOperands.empty()) {
9668 // Note that if our source is a gep chain itself that we wait for that
9669 // chain to be resolved before we perform this transformation. This
9670 // avoids us creating a TON of code in some cases.
9672 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9673 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9674 return 0; // Wait until our source is folded to completion.
9676 SmallVector<Value*, 8> Indices;
9678 // Find out whether the last index in the source GEP is a sequential idx.
9679 bool EndsWithSequential = false;
9680 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9681 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9682 EndsWithSequential = !isa<StructType>(*I);
9684 // Can we combine the two pointer arithmetics offsets?
9685 if (EndsWithSequential) {
9686 // Replace: gep (gep %P, long B), long A, ...
9687 // With: T = long A+B; gep %P, T, ...
9689 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9690 if (SO1 == Constant::getNullValue(SO1->getType())) {
9692 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9695 // If they aren't the same type, convert both to an integer of the
9696 // target's pointer size.
9697 if (SO1->getType() != GO1->getType()) {
9698 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9699 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9700 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9701 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9703 unsigned PS = TD->getPointerSizeInBits();
9704 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9705 // Convert GO1 to SO1's type.
9706 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9708 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9709 // Convert SO1 to GO1's type.
9710 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9712 const Type *PT = TD->getIntPtrType();
9713 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9714 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9718 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9719 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9721 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
9722 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9726 // Recycle the GEP we already have if possible.
9727 if (SrcGEPOperands.size() == 2) {
9728 GEP.setOperand(0, SrcGEPOperands[0]);
9729 GEP.setOperand(1, Sum);
9732 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9733 SrcGEPOperands.end()-1);
9734 Indices.push_back(Sum);
9735 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9737 } else if (isa<Constant>(*GEP.idx_begin()) &&
9738 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9739 SrcGEPOperands.size() != 1) {
9740 // Otherwise we can do the fold if the first index of the GEP is a zero
9741 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9742 SrcGEPOperands.end());
9743 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9746 if (!Indices.empty())
9747 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9748 Indices.end(), GEP.getName());
9750 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9751 // GEP of global variable. If all of the indices for this GEP are
9752 // constants, we can promote this to a constexpr instead of an instruction.
9754 // Scan for nonconstants...
9755 SmallVector<Constant*, 8> Indices;
9756 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9757 for (; I != E && isa<Constant>(*I); ++I)
9758 Indices.push_back(cast<Constant>(*I));
9760 if (I == E) { // If they are all constants...
9761 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9762 &Indices[0],Indices.size());
9764 // Replace all uses of the GEP with the new constexpr...
9765 return ReplaceInstUsesWith(GEP, CE);
9767 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9768 if (!isa<PointerType>(X->getType())) {
9769 // Not interesting. Source pointer must be a cast from pointer.
9770 } else if (HasZeroPointerIndex) {
9771 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9772 // into : GEP [10 x i8]* X, i32 0, ...
9774 // This occurs when the program declares an array extern like "int X[];"
9776 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9777 const PointerType *XTy = cast<PointerType>(X->getType());
9778 if (const ArrayType *XATy =
9779 dyn_cast<ArrayType>(XTy->getElementType()))
9780 if (const ArrayType *CATy =
9781 dyn_cast<ArrayType>(CPTy->getElementType()))
9782 if (CATy->getElementType() == XATy->getElementType()) {
9783 // At this point, we know that the cast source type is a pointer
9784 // to an array of the same type as the destination pointer
9785 // array. Because the array type is never stepped over (there
9786 // is a leading zero) we can fold the cast into this GEP.
9787 GEP.setOperand(0, X);
9790 } else if (GEP.getNumOperands() == 2) {
9791 // Transform things like:
9792 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9793 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9794 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9795 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9796 if (isa<ArrayType>(SrcElTy) &&
9797 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9798 TD->getABITypeSize(ResElTy)) {
9800 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9801 Idx[1] = GEP.getOperand(1);
9802 Value *V = InsertNewInstBefore(
9803 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9804 // V and GEP are both pointer types --> BitCast
9805 return new BitCastInst(V, GEP.getType());
9808 // Transform things like:
9809 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9810 // (where tmp = 8*tmp2) into:
9811 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9813 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9814 uint64_t ArrayEltSize =
9815 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9817 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9818 // allow either a mul, shift, or constant here.
9820 ConstantInt *Scale = 0;
9821 if (ArrayEltSize == 1) {
9822 NewIdx = GEP.getOperand(1);
9823 Scale = ConstantInt::get(NewIdx->getType(), 1);
9824 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9825 NewIdx = ConstantInt::get(CI->getType(), 1);
9827 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9828 if (Inst->getOpcode() == Instruction::Shl &&
9829 isa<ConstantInt>(Inst->getOperand(1))) {
9830 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9831 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9832 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9833 NewIdx = Inst->getOperand(0);
9834 } else if (Inst->getOpcode() == Instruction::Mul &&
9835 isa<ConstantInt>(Inst->getOperand(1))) {
9836 Scale = cast<ConstantInt>(Inst->getOperand(1));
9837 NewIdx = Inst->getOperand(0);
9841 // If the index will be to exactly the right offset with the scale taken
9842 // out, perform the transformation. Note, we don't know whether Scale is
9843 // signed or not. We'll use unsigned version of division/modulo
9844 // operation after making sure Scale doesn't have the sign bit set.
9845 if (Scale && Scale->getSExtValue() >= 0LL &&
9846 Scale->getZExtValue() % ArrayEltSize == 0) {
9847 Scale = ConstantInt::get(Scale->getType(),
9848 Scale->getZExtValue() / ArrayEltSize);
9849 if (Scale->getZExtValue() != 1) {
9850 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9852 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
9853 NewIdx = InsertNewInstBefore(Sc, GEP);
9856 // Insert the new GEP instruction.
9858 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9860 Instruction *NewGEP =
9861 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9862 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9863 // The NewGEP must be pointer typed, so must the old one -> BitCast
9864 return new BitCastInst(NewGEP, GEP.getType());
9873 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9874 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9875 if (AI.isArrayAllocation()) { // Check C != 1
9876 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9878 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9879 AllocationInst *New = 0;
9881 // Create and insert the replacement instruction...
9882 if (isa<MallocInst>(AI))
9883 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9885 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9886 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9889 InsertNewInstBefore(New, AI);
9891 // Scan to the end of the allocation instructions, to skip over a block of
9892 // allocas if possible...
9894 BasicBlock::iterator It = New;
9895 while (isa<AllocationInst>(*It)) ++It;
9897 // Now that I is pointing to the first non-allocation-inst in the block,
9898 // insert our getelementptr instruction...
9900 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9904 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9905 New->getName()+".sub", It);
9907 // Now make everything use the getelementptr instead of the original
9909 return ReplaceInstUsesWith(AI, V);
9910 } else if (isa<UndefValue>(AI.getArraySize())) {
9911 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9915 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9916 // Note that we only do this for alloca's, because malloc should allocate and
9917 // return a unique pointer, even for a zero byte allocation.
9918 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9919 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9920 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9925 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9926 Value *Op = FI.getOperand(0);
9928 // free undef -> unreachable.
9929 if (isa<UndefValue>(Op)) {
9930 // Insert a new store to null because we cannot modify the CFG here.
9931 new StoreInst(ConstantInt::getTrue(),
9932 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9933 return EraseInstFromFunction(FI);
9936 // If we have 'free null' delete the instruction. This can happen in stl code
9937 // when lots of inlining happens.
9938 if (isa<ConstantPointerNull>(Op))
9939 return EraseInstFromFunction(FI);
9941 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9942 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9943 FI.setOperand(0, CI->getOperand(0));
9947 // Change free (gep X, 0,0,0,0) into free(X)
9948 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9949 if (GEPI->hasAllZeroIndices()) {
9950 AddToWorkList(GEPI);
9951 FI.setOperand(0, GEPI->getOperand(0));
9956 // Change free(malloc) into nothing, if the malloc has a single use.
9957 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9958 if (MI->hasOneUse()) {
9959 EraseInstFromFunction(FI);
9960 return EraseInstFromFunction(*MI);
9967 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9968 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9969 const TargetData *TD) {
9970 User *CI = cast<User>(LI.getOperand(0));
9971 Value *CastOp = CI->getOperand(0);
9973 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9974 // Instead of loading constant c string, use corresponding integer value
9975 // directly if string length is small enough.
9976 const std::string &Str = CE->getOperand(0)->getStringValue();
9978 unsigned len = Str.length();
9979 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9980 unsigned numBits = Ty->getPrimitiveSizeInBits();
9981 // Replace LI with immediate integer store.
9982 if ((numBits >> 3) == len + 1) {
9983 APInt StrVal(numBits, 0);
9984 APInt SingleChar(numBits, 0);
9985 if (TD->isLittleEndian()) {
9986 for (signed i = len-1; i >= 0; i--) {
9987 SingleChar = (uint64_t) Str[i];
9988 StrVal = (StrVal << 8) | SingleChar;
9991 for (unsigned i = 0; i < len; i++) {
9992 SingleChar = (uint64_t) Str[i];
9993 StrVal = (StrVal << 8) | SingleChar;
9995 // Append NULL at the end.
9997 StrVal = (StrVal << 8) | SingleChar;
9999 Value *NL = ConstantInt::get(StrVal);
10000 return IC.ReplaceInstUsesWith(LI, NL);
10005 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10006 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10007 const Type *SrcPTy = SrcTy->getElementType();
10009 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10010 isa<VectorType>(DestPTy)) {
10011 // If the source is an array, the code below will not succeed. Check to
10012 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10014 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10015 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10016 if (ASrcTy->getNumElements() != 0) {
10018 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10019 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10020 SrcTy = cast<PointerType>(CastOp->getType());
10021 SrcPTy = SrcTy->getElementType();
10024 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10025 isa<VectorType>(SrcPTy)) &&
10026 // Do not allow turning this into a load of an integer, which is then
10027 // casted to a pointer, this pessimizes pointer analysis a lot.
10028 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10029 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10030 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10032 // Okay, we are casting from one integer or pointer type to another of
10033 // the same size. Instead of casting the pointer before the load, cast
10034 // the result of the loaded value.
10035 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10037 LI.isVolatile()),LI);
10038 // Now cast the result of the load.
10039 return new BitCastInst(NewLoad, LI.getType());
10046 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10047 /// from this value cannot trap. If it is not obviously safe to load from the
10048 /// specified pointer, we do a quick local scan of the basic block containing
10049 /// ScanFrom, to determine if the address is already accessed.
10050 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10051 // If it is an alloca it is always safe to load from.
10052 if (isa<AllocaInst>(V)) return true;
10054 // If it is a global variable it is mostly safe to load from.
10055 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10056 // Don't try to evaluate aliases. External weak GV can be null.
10057 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10059 // Otherwise, be a little bit agressive by scanning the local block where we
10060 // want to check to see if the pointer is already being loaded or stored
10061 // from/to. If so, the previous load or store would have already trapped,
10062 // so there is no harm doing an extra load (also, CSE will later eliminate
10063 // the load entirely).
10064 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10069 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10070 if (LI->getOperand(0) == V) return true;
10071 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10072 if (SI->getOperand(1) == V) return true;
10078 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10079 /// until we find the underlying object a pointer is referring to or something
10080 /// we don't understand. Note that the returned pointer may be offset from the
10081 /// input, because we ignore GEP indices.
10082 static Value *GetUnderlyingObject(Value *Ptr) {
10084 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10085 if (CE->getOpcode() == Instruction::BitCast ||
10086 CE->getOpcode() == Instruction::GetElementPtr)
10087 Ptr = CE->getOperand(0);
10090 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10091 Ptr = BCI->getOperand(0);
10092 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10093 Ptr = GEP->getOperand(0);
10100 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10101 Value *Op = LI.getOperand(0);
10103 // Attempt to improve the alignment.
10104 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10106 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10107 LI.getAlignment()))
10108 LI.setAlignment(KnownAlign);
10110 // load (cast X) --> cast (load X) iff safe
10111 if (isa<CastInst>(Op))
10112 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10115 // None of the following transforms are legal for volatile loads.
10116 if (LI.isVolatile()) return 0;
10118 if (&LI.getParent()->front() != &LI) {
10119 BasicBlock::iterator BBI = &LI; --BBI;
10120 // If the instruction immediately before this is a store to the same
10121 // address, do a simple form of store->load forwarding.
10122 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10123 if (SI->getOperand(1) == LI.getOperand(0))
10124 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10125 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10126 if (LIB->getOperand(0) == LI.getOperand(0))
10127 return ReplaceInstUsesWith(LI, LIB);
10130 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10131 const Value *GEPI0 = GEPI->getOperand(0);
10132 // TODO: Consider a target hook for valid address spaces for this xform.
10133 if (isa<ConstantPointerNull>(GEPI0) &&
10134 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10135 // Insert a new store to null instruction before the load to indicate
10136 // that this code is not reachable. We do this instead of inserting
10137 // an unreachable instruction directly because we cannot modify the
10139 new StoreInst(UndefValue::get(LI.getType()),
10140 Constant::getNullValue(Op->getType()), &LI);
10141 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10145 if (Constant *C = dyn_cast<Constant>(Op)) {
10146 // load null/undef -> undef
10147 // TODO: Consider a target hook for valid address spaces for this xform.
10148 if (isa<UndefValue>(C) || (C->isNullValue() &&
10149 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10150 // Insert a new store to null instruction before the load to indicate that
10151 // this code is not reachable. We do this instead of inserting an
10152 // unreachable instruction directly because we cannot modify the CFG.
10153 new StoreInst(UndefValue::get(LI.getType()),
10154 Constant::getNullValue(Op->getType()), &LI);
10155 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10158 // Instcombine load (constant global) into the value loaded.
10159 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10160 if (GV->isConstant() && !GV->isDeclaration())
10161 return ReplaceInstUsesWith(LI, GV->getInitializer());
10163 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10164 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10165 if (CE->getOpcode() == Instruction::GetElementPtr) {
10166 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10167 if (GV->isConstant() && !GV->isDeclaration())
10169 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10170 return ReplaceInstUsesWith(LI, V);
10171 if (CE->getOperand(0)->isNullValue()) {
10172 // Insert a new store to null instruction before the load to indicate
10173 // that this code is not reachable. We do this instead of inserting
10174 // an unreachable instruction directly because we cannot modify the
10176 new StoreInst(UndefValue::get(LI.getType()),
10177 Constant::getNullValue(Op->getType()), &LI);
10178 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10181 } else if (CE->isCast()) {
10182 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10188 // If this load comes from anywhere in a constant global, and if the global
10189 // is all undef or zero, we know what it loads.
10190 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10191 if (GV->isConstant() && GV->hasInitializer()) {
10192 if (GV->getInitializer()->isNullValue())
10193 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10194 else if (isa<UndefValue>(GV->getInitializer()))
10195 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10199 if (Op->hasOneUse()) {
10200 // Change select and PHI nodes to select values instead of addresses: this
10201 // helps alias analysis out a lot, allows many others simplifications, and
10202 // exposes redundancy in the code.
10204 // Note that we cannot do the transformation unless we know that the
10205 // introduced loads cannot trap! Something like this is valid as long as
10206 // the condition is always false: load (select bool %C, int* null, int* %G),
10207 // but it would not be valid if we transformed it to load from null
10208 // unconditionally.
10210 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10211 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10212 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10213 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10214 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10215 SI->getOperand(1)->getName()+".val"), LI);
10216 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10217 SI->getOperand(2)->getName()+".val"), LI);
10218 return SelectInst::Create(SI->getCondition(), V1, V2);
10221 // load (select (cond, null, P)) -> load P
10222 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10223 if (C->isNullValue()) {
10224 LI.setOperand(0, SI->getOperand(2));
10228 // load (select (cond, P, null)) -> load P
10229 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10230 if (C->isNullValue()) {
10231 LI.setOperand(0, SI->getOperand(1));
10239 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10241 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10242 User *CI = cast<User>(SI.getOperand(1));
10243 Value *CastOp = CI->getOperand(0);
10245 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10246 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10247 const Type *SrcPTy = SrcTy->getElementType();
10249 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10250 // If the source is an array, the code below will not succeed. Check to
10251 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10253 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10254 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10255 if (ASrcTy->getNumElements() != 0) {
10257 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10258 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10259 SrcTy = cast<PointerType>(CastOp->getType());
10260 SrcPTy = SrcTy->getElementType();
10263 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10264 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10265 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10267 // Okay, we are casting from one integer or pointer type to another of
10268 // the same size. Instead of casting the pointer before
10269 // the store, cast the value to be stored.
10271 Value *SIOp0 = SI.getOperand(0);
10272 Instruction::CastOps opcode = Instruction::BitCast;
10273 const Type* CastSrcTy = SIOp0->getType();
10274 const Type* CastDstTy = SrcPTy;
10275 if (isa<PointerType>(CastDstTy)) {
10276 if (CastSrcTy->isInteger())
10277 opcode = Instruction::IntToPtr;
10278 } else if (isa<IntegerType>(CastDstTy)) {
10279 if (isa<PointerType>(SIOp0->getType()))
10280 opcode = Instruction::PtrToInt;
10282 if (Constant *C = dyn_cast<Constant>(SIOp0))
10283 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10285 NewCast = IC.InsertNewInstBefore(
10286 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10288 return new StoreInst(NewCast, CastOp);
10295 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10296 Value *Val = SI.getOperand(0);
10297 Value *Ptr = SI.getOperand(1);
10299 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10300 EraseInstFromFunction(SI);
10305 // If the RHS is an alloca with a single use, zapify the store, making the
10307 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10308 if (isa<AllocaInst>(Ptr)) {
10309 EraseInstFromFunction(SI);
10314 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10315 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10316 GEP->getOperand(0)->hasOneUse()) {
10317 EraseInstFromFunction(SI);
10323 // Attempt to improve the alignment.
10324 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10326 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10327 SI.getAlignment()))
10328 SI.setAlignment(KnownAlign);
10330 // Do really simple DSE, to catch cases where there are several consequtive
10331 // stores to the same location, separated by a few arithmetic operations. This
10332 // situation often occurs with bitfield accesses.
10333 BasicBlock::iterator BBI = &SI;
10334 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10338 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10339 // Prev store isn't volatile, and stores to the same location?
10340 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10343 EraseInstFromFunction(*PrevSI);
10349 // If this is a load, we have to stop. However, if the loaded value is from
10350 // the pointer we're loading and is producing the pointer we're storing,
10351 // then *this* store is dead (X = load P; store X -> P).
10352 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10353 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10354 EraseInstFromFunction(SI);
10358 // Otherwise, this is a load from some other location. Stores before it
10359 // may not be dead.
10363 // Don't skip over loads or things that can modify memory.
10364 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10369 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10371 // store X, null -> turns into 'unreachable' in SimplifyCFG
10372 if (isa<ConstantPointerNull>(Ptr)) {
10373 if (!isa<UndefValue>(Val)) {
10374 SI.setOperand(0, UndefValue::get(Val->getType()));
10375 if (Instruction *U = dyn_cast<Instruction>(Val))
10376 AddToWorkList(U); // Dropped a use.
10379 return 0; // Do not modify these!
10382 // store undef, Ptr -> noop
10383 if (isa<UndefValue>(Val)) {
10384 EraseInstFromFunction(SI);
10389 // If the pointer destination is a cast, see if we can fold the cast into the
10391 if (isa<CastInst>(Ptr))
10392 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10394 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10396 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10400 // If this store is the last instruction in the basic block, and if the block
10401 // ends with an unconditional branch, try to move it to the successor block.
10403 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10404 if (BI->isUnconditional())
10405 if (SimplifyStoreAtEndOfBlock(SI))
10406 return 0; // xform done!
10411 /// SimplifyStoreAtEndOfBlock - Turn things like:
10412 /// if () { *P = v1; } else { *P = v2 }
10413 /// into a phi node with a store in the successor.
10415 /// Simplify things like:
10416 /// *P = v1; if () { *P = v2; }
10417 /// into a phi node with a store in the successor.
10419 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10420 BasicBlock *StoreBB = SI.getParent();
10422 // Check to see if the successor block has exactly two incoming edges. If
10423 // so, see if the other predecessor contains a store to the same location.
10424 // if so, insert a PHI node (if needed) and move the stores down.
10425 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10427 // Determine whether Dest has exactly two predecessors and, if so, compute
10428 // the other predecessor.
10429 pred_iterator PI = pred_begin(DestBB);
10430 BasicBlock *OtherBB = 0;
10431 if (*PI != StoreBB)
10434 if (PI == pred_end(DestBB))
10437 if (*PI != StoreBB) {
10442 if (++PI != pred_end(DestBB))
10445 // Bail out if all the relevant blocks aren't distinct (this can happen,
10446 // for example, if SI is in an infinite loop)
10447 if (StoreBB == DestBB || OtherBB == DestBB)
10450 // Verify that the other block ends in a branch and is not otherwise empty.
10451 BasicBlock::iterator BBI = OtherBB->getTerminator();
10452 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10453 if (!OtherBr || BBI == OtherBB->begin())
10456 // If the other block ends in an unconditional branch, check for the 'if then
10457 // else' case. there is an instruction before the branch.
10458 StoreInst *OtherStore = 0;
10459 if (OtherBr->isUnconditional()) {
10460 // If this isn't a store, or isn't a store to the same location, bail out.
10462 OtherStore = dyn_cast<StoreInst>(BBI);
10463 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10466 // Otherwise, the other block ended with a conditional branch. If one of the
10467 // destinations is StoreBB, then we have the if/then case.
10468 if (OtherBr->getSuccessor(0) != StoreBB &&
10469 OtherBr->getSuccessor(1) != StoreBB)
10472 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10473 // if/then triangle. See if there is a store to the same ptr as SI that
10474 // lives in OtherBB.
10476 // Check to see if we find the matching store.
10477 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10478 if (OtherStore->getOperand(1) != SI.getOperand(1))
10482 // If we find something that may be using or overwriting the stored
10483 // value, or if we run out of instructions, we can't do the xform.
10484 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
10485 BBI == OtherBB->begin())
10489 // In order to eliminate the store in OtherBr, we have to
10490 // make sure nothing reads or overwrites the stored value in
10492 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10493 // FIXME: This should really be AA driven.
10494 if (I->mayReadFromMemory() || I->mayWriteToMemory())
10499 // Insert a PHI node now if we need it.
10500 Value *MergedVal = OtherStore->getOperand(0);
10501 if (MergedVal != SI.getOperand(0)) {
10502 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10503 PN->reserveOperandSpace(2);
10504 PN->addIncoming(SI.getOperand(0), SI.getParent());
10505 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10506 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10509 // Advance to a place where it is safe to insert the new store and
10511 BBI = DestBB->getFirstNonPHI();
10512 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10513 OtherStore->isVolatile()), *BBI);
10515 // Nuke the old stores.
10516 EraseInstFromFunction(SI);
10517 EraseInstFromFunction(*OtherStore);
10523 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10524 // Change br (not X), label True, label False to: br X, label False, True
10526 BasicBlock *TrueDest;
10527 BasicBlock *FalseDest;
10528 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10529 !isa<Constant>(X)) {
10530 // Swap Destinations and condition...
10531 BI.setCondition(X);
10532 BI.setSuccessor(0, FalseDest);
10533 BI.setSuccessor(1, TrueDest);
10537 // Cannonicalize fcmp_one -> fcmp_oeq
10538 FCmpInst::Predicate FPred; Value *Y;
10539 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10540 TrueDest, FalseDest)))
10541 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10542 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10543 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10544 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10545 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10546 NewSCC->takeName(I);
10547 // Swap Destinations and condition...
10548 BI.setCondition(NewSCC);
10549 BI.setSuccessor(0, FalseDest);
10550 BI.setSuccessor(1, TrueDest);
10551 RemoveFromWorkList(I);
10552 I->eraseFromParent();
10553 AddToWorkList(NewSCC);
10557 // Cannonicalize icmp_ne -> icmp_eq
10558 ICmpInst::Predicate IPred;
10559 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10560 TrueDest, FalseDest)))
10561 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10562 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10563 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10564 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10565 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10566 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10567 NewSCC->takeName(I);
10568 // Swap Destinations and condition...
10569 BI.setCondition(NewSCC);
10570 BI.setSuccessor(0, FalseDest);
10571 BI.setSuccessor(1, TrueDest);
10572 RemoveFromWorkList(I);
10573 I->eraseFromParent();;
10574 AddToWorkList(NewSCC);
10581 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10582 Value *Cond = SI.getCondition();
10583 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10584 if (I->getOpcode() == Instruction::Add)
10585 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10586 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10587 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10588 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10590 SI.setOperand(0, I->getOperand(0));
10598 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
10599 // See if we are trying to extract a known value. If so, use that instead.
10600 if (Value *Elt = FindInsertedValue(EV.getOperand(0), EV.idx_begin(),
10601 EV.idx_end(), &EV))
10602 return ReplaceInstUsesWith(EV, Elt);
10608 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10609 /// is to leave as a vector operation.
10610 static bool CheapToScalarize(Value *V, bool isConstant) {
10611 if (isa<ConstantAggregateZero>(V))
10613 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10614 if (isConstant) return true;
10615 // If all elts are the same, we can extract.
10616 Constant *Op0 = C->getOperand(0);
10617 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10618 if (C->getOperand(i) != Op0)
10622 Instruction *I = dyn_cast<Instruction>(V);
10623 if (!I) return false;
10625 // Insert element gets simplified to the inserted element or is deleted if
10626 // this is constant idx extract element and its a constant idx insertelt.
10627 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10628 isa<ConstantInt>(I->getOperand(2)))
10630 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10632 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10633 if (BO->hasOneUse() &&
10634 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10635 CheapToScalarize(BO->getOperand(1), isConstant)))
10637 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10638 if (CI->hasOneUse() &&
10639 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10640 CheapToScalarize(CI->getOperand(1), isConstant)))
10646 /// Read and decode a shufflevector mask.
10648 /// It turns undef elements into values that are larger than the number of
10649 /// elements in the input.
10650 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10651 unsigned NElts = SVI->getType()->getNumElements();
10652 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10653 return std::vector<unsigned>(NElts, 0);
10654 if (isa<UndefValue>(SVI->getOperand(2)))
10655 return std::vector<unsigned>(NElts, 2*NElts);
10657 std::vector<unsigned> Result;
10658 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10659 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
10660 if (isa<UndefValue>(*i))
10661 Result.push_back(NElts*2); // undef -> 8
10663 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
10667 /// FindScalarElement - Given a vector and an element number, see if the scalar
10668 /// value is already around as a register, for example if it were inserted then
10669 /// extracted from the vector.
10670 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10671 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10672 const VectorType *PTy = cast<VectorType>(V->getType());
10673 unsigned Width = PTy->getNumElements();
10674 if (EltNo >= Width) // Out of range access.
10675 return UndefValue::get(PTy->getElementType());
10677 if (isa<UndefValue>(V))
10678 return UndefValue::get(PTy->getElementType());
10679 else if (isa<ConstantAggregateZero>(V))
10680 return Constant::getNullValue(PTy->getElementType());
10681 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10682 return CP->getOperand(EltNo);
10683 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10684 // If this is an insert to a variable element, we don't know what it is.
10685 if (!isa<ConstantInt>(III->getOperand(2)))
10687 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10689 // If this is an insert to the element we are looking for, return the
10691 if (EltNo == IIElt)
10692 return III->getOperand(1);
10694 // Otherwise, the insertelement doesn't modify the value, recurse on its
10696 return FindScalarElement(III->getOperand(0), EltNo);
10697 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10698 unsigned InEl = getShuffleMask(SVI)[EltNo];
10700 return FindScalarElement(SVI->getOperand(0), InEl);
10701 else if (InEl < Width*2)
10702 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10704 return UndefValue::get(PTy->getElementType());
10707 // Otherwise, we don't know.
10711 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10712 // If vector val is undef, replace extract with scalar undef.
10713 if (isa<UndefValue>(EI.getOperand(0)))
10714 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10716 // If vector val is constant 0, replace extract with scalar 0.
10717 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10718 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10720 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10721 // If vector val is constant with all elements the same, replace EI with
10722 // that element. When the elements are not identical, we cannot replace yet
10723 // (we do that below, but only when the index is constant).
10724 Constant *op0 = C->getOperand(0);
10725 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10726 if (C->getOperand(i) != op0) {
10731 return ReplaceInstUsesWith(EI, op0);
10734 // If extracting a specified index from the vector, see if we can recursively
10735 // find a previously computed scalar that was inserted into the vector.
10736 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10737 unsigned IndexVal = IdxC->getZExtValue();
10738 unsigned VectorWidth =
10739 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10741 // If this is extracting an invalid index, turn this into undef, to avoid
10742 // crashing the code below.
10743 if (IndexVal >= VectorWidth)
10744 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10746 // This instruction only demands the single element from the input vector.
10747 // If the input vector has a single use, simplify it based on this use
10749 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10750 uint64_t UndefElts;
10751 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10754 EI.setOperand(0, V);
10759 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10760 return ReplaceInstUsesWith(EI, Elt);
10762 // If the this extractelement is directly using a bitcast from a vector of
10763 // the same number of elements, see if we can find the source element from
10764 // it. In this case, we will end up needing to bitcast the scalars.
10765 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10766 if (const VectorType *VT =
10767 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10768 if (VT->getNumElements() == VectorWidth)
10769 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10770 return new BitCastInst(Elt, EI.getType());
10774 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10775 if (I->hasOneUse()) {
10776 // Push extractelement into predecessor operation if legal and
10777 // profitable to do so
10778 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10779 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10780 if (CheapToScalarize(BO, isConstantElt)) {
10781 ExtractElementInst *newEI0 =
10782 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10783 EI.getName()+".lhs");
10784 ExtractElementInst *newEI1 =
10785 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10786 EI.getName()+".rhs");
10787 InsertNewInstBefore(newEI0, EI);
10788 InsertNewInstBefore(newEI1, EI);
10789 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
10791 } else if (isa<LoadInst>(I)) {
10793 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10794 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10795 PointerType::get(EI.getType(), AS),EI);
10796 GetElementPtrInst *GEP =
10797 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
10798 InsertNewInstBefore(GEP, EI);
10799 return new LoadInst(GEP);
10802 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10803 // Extracting the inserted element?
10804 if (IE->getOperand(2) == EI.getOperand(1))
10805 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10806 // If the inserted and extracted elements are constants, they must not
10807 // be the same value, extract from the pre-inserted value instead.
10808 if (isa<Constant>(IE->getOperand(2)) &&
10809 isa<Constant>(EI.getOperand(1))) {
10810 AddUsesToWorkList(EI);
10811 EI.setOperand(0, IE->getOperand(0));
10814 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10815 // If this is extracting an element from a shufflevector, figure out where
10816 // it came from and extract from the appropriate input element instead.
10817 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10818 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10820 if (SrcIdx < SVI->getType()->getNumElements())
10821 Src = SVI->getOperand(0);
10822 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10823 SrcIdx -= SVI->getType()->getNumElements();
10824 Src = SVI->getOperand(1);
10826 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10828 return new ExtractElementInst(Src, SrcIdx);
10835 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10836 /// elements from either LHS or RHS, return the shuffle mask and true.
10837 /// Otherwise, return false.
10838 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10839 std::vector<Constant*> &Mask) {
10840 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10841 "Invalid CollectSingleShuffleElements");
10842 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10844 if (isa<UndefValue>(V)) {
10845 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10847 } else if (V == LHS) {
10848 for (unsigned i = 0; i != NumElts; ++i)
10849 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10851 } else if (V == RHS) {
10852 for (unsigned i = 0; i != NumElts; ++i)
10853 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10855 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10856 // If this is an insert of an extract from some other vector, include it.
10857 Value *VecOp = IEI->getOperand(0);
10858 Value *ScalarOp = IEI->getOperand(1);
10859 Value *IdxOp = IEI->getOperand(2);
10861 if (!isa<ConstantInt>(IdxOp))
10863 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10865 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10866 // Okay, we can handle this if the vector we are insertinting into is
10867 // transitively ok.
10868 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10869 // If so, update the mask to reflect the inserted undef.
10870 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10873 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10874 if (isa<ConstantInt>(EI->getOperand(1)) &&
10875 EI->getOperand(0)->getType() == V->getType()) {
10876 unsigned ExtractedIdx =
10877 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10879 // This must be extracting from either LHS or RHS.
10880 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10881 // Okay, we can handle this if the vector we are insertinting into is
10882 // transitively ok.
10883 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10884 // If so, update the mask to reflect the inserted value.
10885 if (EI->getOperand(0) == LHS) {
10886 Mask[InsertedIdx & (NumElts-1)] =
10887 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10889 assert(EI->getOperand(0) == RHS);
10890 Mask[InsertedIdx & (NumElts-1)] =
10891 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10900 // TODO: Handle shufflevector here!
10905 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10906 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10907 /// that computes V and the LHS value of the shuffle.
10908 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10910 assert(isa<VectorType>(V->getType()) &&
10911 (RHS == 0 || V->getType() == RHS->getType()) &&
10912 "Invalid shuffle!");
10913 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10915 if (isa<UndefValue>(V)) {
10916 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10918 } else if (isa<ConstantAggregateZero>(V)) {
10919 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10921 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10922 // If this is an insert of an extract from some other vector, include it.
10923 Value *VecOp = IEI->getOperand(0);
10924 Value *ScalarOp = IEI->getOperand(1);
10925 Value *IdxOp = IEI->getOperand(2);
10927 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10928 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10929 EI->getOperand(0)->getType() == V->getType()) {
10930 unsigned ExtractedIdx =
10931 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10932 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10934 // Either the extracted from or inserted into vector must be RHSVec,
10935 // otherwise we'd end up with a shuffle of three inputs.
10936 if (EI->getOperand(0) == RHS || RHS == 0) {
10937 RHS = EI->getOperand(0);
10938 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10939 Mask[InsertedIdx & (NumElts-1)] =
10940 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10944 if (VecOp == RHS) {
10945 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10946 // Everything but the extracted element is replaced with the RHS.
10947 for (unsigned i = 0; i != NumElts; ++i) {
10948 if (i != InsertedIdx)
10949 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10954 // If this insertelement is a chain that comes from exactly these two
10955 // vectors, return the vector and the effective shuffle.
10956 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10957 return EI->getOperand(0);
10962 // TODO: Handle shufflevector here!
10964 // Otherwise, can't do anything fancy. Return an identity vector.
10965 for (unsigned i = 0; i != NumElts; ++i)
10966 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10970 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10971 Value *VecOp = IE.getOperand(0);
10972 Value *ScalarOp = IE.getOperand(1);
10973 Value *IdxOp = IE.getOperand(2);
10975 // Inserting an undef or into an undefined place, remove this.
10976 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10977 ReplaceInstUsesWith(IE, VecOp);
10979 // If the inserted element was extracted from some other vector, and if the
10980 // indexes are constant, try to turn this into a shufflevector operation.
10981 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10982 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10983 EI->getOperand(0)->getType() == IE.getType()) {
10984 unsigned NumVectorElts = IE.getType()->getNumElements();
10985 unsigned ExtractedIdx =
10986 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10987 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10989 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10990 return ReplaceInstUsesWith(IE, VecOp);
10992 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10993 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10995 // If we are extracting a value from a vector, then inserting it right
10996 // back into the same place, just use the input vector.
10997 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10998 return ReplaceInstUsesWith(IE, VecOp);
11000 // We could theoretically do this for ANY input. However, doing so could
11001 // turn chains of insertelement instructions into a chain of shufflevector
11002 // instructions, and right now we do not merge shufflevectors. As such,
11003 // only do this in a situation where it is clear that there is benefit.
11004 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11005 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11006 // the values of VecOp, except then one read from EIOp0.
11007 // Build a new shuffle mask.
11008 std::vector<Constant*> Mask;
11009 if (isa<UndefValue>(VecOp))
11010 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11012 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11013 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11016 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11017 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11018 ConstantVector::get(Mask));
11021 // If this insertelement isn't used by some other insertelement, turn it
11022 // (and any insertelements it points to), into one big shuffle.
11023 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11024 std::vector<Constant*> Mask;
11026 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11027 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11028 // We now have a shuffle of LHS, RHS, Mask.
11029 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11038 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11039 Value *LHS = SVI.getOperand(0);
11040 Value *RHS = SVI.getOperand(1);
11041 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11043 bool MadeChange = false;
11045 // Undefined shuffle mask -> undefined value.
11046 if (isa<UndefValue>(SVI.getOperand(2)))
11047 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11049 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11050 // the undef, change them to undefs.
11051 if (isa<UndefValue>(SVI.getOperand(1))) {
11052 // Scan to see if there are any references to the RHS. If so, replace them
11053 // with undef element refs and set MadeChange to true.
11054 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11055 if (Mask[i] >= e && Mask[i] != 2*e) {
11062 // Remap any references to RHS to use LHS.
11063 std::vector<Constant*> Elts;
11064 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11065 if (Mask[i] == 2*e)
11066 Elts.push_back(UndefValue::get(Type::Int32Ty));
11068 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11070 SVI.setOperand(2, ConstantVector::get(Elts));
11074 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11075 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11076 if (LHS == RHS || isa<UndefValue>(LHS)) {
11077 if (isa<UndefValue>(LHS) && LHS == RHS) {
11078 // shuffle(undef,undef,mask) -> undef.
11079 return ReplaceInstUsesWith(SVI, LHS);
11082 // Remap any references to RHS to use LHS.
11083 std::vector<Constant*> Elts;
11084 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11085 if (Mask[i] >= 2*e)
11086 Elts.push_back(UndefValue::get(Type::Int32Ty));
11088 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11089 (Mask[i] < e && isa<UndefValue>(LHS)))
11090 Mask[i] = 2*e; // Turn into undef.
11092 Mask[i] &= (e-1); // Force to LHS.
11093 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11096 SVI.setOperand(0, SVI.getOperand(1));
11097 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11098 SVI.setOperand(2, ConstantVector::get(Elts));
11099 LHS = SVI.getOperand(0);
11100 RHS = SVI.getOperand(1);
11104 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11105 bool isLHSID = true, isRHSID = true;
11107 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11108 if (Mask[i] >= e*2) continue; // Ignore undef values.
11109 // Is this an identity shuffle of the LHS value?
11110 isLHSID &= (Mask[i] == i);
11112 // Is this an identity shuffle of the RHS value?
11113 isRHSID &= (Mask[i]-e == i);
11116 // Eliminate identity shuffles.
11117 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11118 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11120 // If the LHS is a shufflevector itself, see if we can combine it with this
11121 // one without producing an unusual shuffle. Here we are really conservative:
11122 // we are absolutely afraid of producing a shuffle mask not in the input
11123 // program, because the code gen may not be smart enough to turn a merged
11124 // shuffle into two specific shuffles: it may produce worse code. As such,
11125 // we only merge two shuffles if the result is one of the two input shuffle
11126 // masks. In this case, merging the shuffles just removes one instruction,
11127 // which we know is safe. This is good for things like turning:
11128 // (splat(splat)) -> splat.
11129 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11130 if (isa<UndefValue>(RHS)) {
11131 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11133 std::vector<unsigned> NewMask;
11134 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11135 if (Mask[i] >= 2*e)
11136 NewMask.push_back(2*e);
11138 NewMask.push_back(LHSMask[Mask[i]]);
11140 // If the result mask is equal to the src shuffle or this shuffle mask, do
11141 // the replacement.
11142 if (NewMask == LHSMask || NewMask == Mask) {
11143 std::vector<Constant*> Elts;
11144 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11145 if (NewMask[i] >= e*2) {
11146 Elts.push_back(UndefValue::get(Type::Int32Ty));
11148 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11151 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11152 LHSSVI->getOperand(1),
11153 ConstantVector::get(Elts));
11158 return MadeChange ? &SVI : 0;
11164 /// TryToSinkInstruction - Try to move the specified instruction from its
11165 /// current block into the beginning of DestBlock, which can only happen if it's
11166 /// safe to move the instruction past all of the instructions between it and the
11167 /// end of its block.
11168 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11169 assert(I->hasOneUse() && "Invariants didn't hold!");
11171 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11172 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11175 // Do not sink alloca instructions out of the entry block.
11176 if (isa<AllocaInst>(I) && I->getParent() ==
11177 &DestBlock->getParent()->getEntryBlock())
11180 // We can only sink load instructions if there is nothing between the load and
11181 // the end of block that could change the value.
11182 if (I->mayReadFromMemory()) {
11183 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11185 if (Scan->mayWriteToMemory())
11189 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11191 I->moveBefore(InsertPos);
11197 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11198 /// all reachable code to the worklist.
11200 /// This has a couple of tricks to make the code faster and more powerful. In
11201 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11202 /// them to the worklist (this significantly speeds up instcombine on code where
11203 /// many instructions are dead or constant). Additionally, if we find a branch
11204 /// whose condition is a known constant, we only visit the reachable successors.
11206 static void AddReachableCodeToWorklist(BasicBlock *BB,
11207 SmallPtrSet<BasicBlock*, 64> &Visited,
11209 const TargetData *TD) {
11210 std::vector<BasicBlock*> Worklist;
11211 Worklist.push_back(BB);
11213 while (!Worklist.empty()) {
11214 BB = Worklist.back();
11215 Worklist.pop_back();
11217 // We have now visited this block! If we've already been here, ignore it.
11218 if (!Visited.insert(BB)) continue;
11220 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11221 Instruction *Inst = BBI++;
11223 // DCE instruction if trivially dead.
11224 if (isInstructionTriviallyDead(Inst)) {
11226 DOUT << "IC: DCE: " << *Inst;
11227 Inst->eraseFromParent();
11231 // ConstantProp instruction if trivially constant.
11232 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11233 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11234 Inst->replaceAllUsesWith(C);
11236 Inst->eraseFromParent();
11240 IC.AddToWorkList(Inst);
11243 // Recursively visit successors. If this is a branch or switch on a
11244 // constant, only visit the reachable successor.
11245 TerminatorInst *TI = BB->getTerminator();
11246 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11247 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11248 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11249 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11250 Worklist.push_back(ReachableBB);
11253 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11254 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11255 // See if this is an explicit destination.
11256 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11257 if (SI->getCaseValue(i) == Cond) {
11258 BasicBlock *ReachableBB = SI->getSuccessor(i);
11259 Worklist.push_back(ReachableBB);
11263 // Otherwise it is the default destination.
11264 Worklist.push_back(SI->getSuccessor(0));
11269 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11270 Worklist.push_back(TI->getSuccessor(i));
11274 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11275 bool Changed = false;
11276 TD = &getAnalysis<TargetData>();
11278 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11279 << F.getNameStr() << "\n");
11282 // Do a depth-first traversal of the function, populate the worklist with
11283 // the reachable instructions. Ignore blocks that are not reachable. Keep
11284 // track of which blocks we visit.
11285 SmallPtrSet<BasicBlock*, 64> Visited;
11286 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11288 // Do a quick scan over the function. If we find any blocks that are
11289 // unreachable, remove any instructions inside of them. This prevents
11290 // the instcombine code from having to deal with some bad special cases.
11291 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11292 if (!Visited.count(BB)) {
11293 Instruction *Term = BB->getTerminator();
11294 while (Term != BB->begin()) { // Remove instrs bottom-up
11295 BasicBlock::iterator I = Term; --I;
11297 DOUT << "IC: DCE: " << *I;
11300 if (!I->use_empty())
11301 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11302 I->eraseFromParent();
11307 while (!Worklist.empty()) {
11308 Instruction *I = RemoveOneFromWorkList();
11309 if (I == 0) continue; // skip null values.
11311 // Check to see if we can DCE the instruction.
11312 if (isInstructionTriviallyDead(I)) {
11313 // Add operands to the worklist.
11314 if (I->getNumOperands() < 4)
11315 AddUsesToWorkList(*I);
11318 DOUT << "IC: DCE: " << *I;
11320 I->eraseFromParent();
11321 RemoveFromWorkList(I);
11325 // Instruction isn't dead, see if we can constant propagate it.
11326 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11327 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11329 // Add operands to the worklist.
11330 AddUsesToWorkList(*I);
11331 ReplaceInstUsesWith(*I, C);
11334 I->eraseFromParent();
11335 RemoveFromWorkList(I);
11339 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
11340 // See if we can constant fold its operands.
11341 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
11342 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
11343 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
11349 // See if we can trivially sink this instruction to a successor basic block.
11350 // FIXME: Remove GetResultInst test when first class support for aggregates
11352 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11353 BasicBlock *BB = I->getParent();
11354 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11355 if (UserParent != BB) {
11356 bool UserIsSuccessor = false;
11357 // See if the user is one of our successors.
11358 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11359 if (*SI == UserParent) {
11360 UserIsSuccessor = true;
11364 // If the user is one of our immediate successors, and if that successor
11365 // only has us as a predecessors (we'd have to split the critical edge
11366 // otherwise), we can keep going.
11367 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11368 next(pred_begin(UserParent)) == pred_end(UserParent))
11369 // Okay, the CFG is simple enough, try to sink this instruction.
11370 Changed |= TryToSinkInstruction(I, UserParent);
11374 // Now that we have an instruction, try combining it to simplify it...
11378 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11379 if (Instruction *Result = visit(*I)) {
11381 // Should we replace the old instruction with a new one?
11383 DOUT << "IC: Old = " << *I
11384 << " New = " << *Result;
11386 // Everything uses the new instruction now.
11387 I->replaceAllUsesWith(Result);
11389 // Push the new instruction and any users onto the worklist.
11390 AddToWorkList(Result);
11391 AddUsersToWorkList(*Result);
11393 // Move the name to the new instruction first.
11394 Result->takeName(I);
11396 // Insert the new instruction into the basic block...
11397 BasicBlock *InstParent = I->getParent();
11398 BasicBlock::iterator InsertPos = I;
11400 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11401 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11404 InstParent->getInstList().insert(InsertPos, Result);
11406 // Make sure that we reprocess all operands now that we reduced their
11408 AddUsesToWorkList(*I);
11410 // Instructions can end up on the worklist more than once. Make sure
11411 // we do not process an instruction that has been deleted.
11412 RemoveFromWorkList(I);
11414 // Erase the old instruction.
11415 InstParent->getInstList().erase(I);
11418 DOUT << "IC: Mod = " << OrigI
11419 << " New = " << *I;
11422 // If the instruction was modified, it's possible that it is now dead.
11423 // if so, remove it.
11424 if (isInstructionTriviallyDead(I)) {
11425 // Make sure we process all operands now that we are reducing their
11427 AddUsesToWorkList(*I);
11429 // Instructions may end up in the worklist more than once. Erase all
11430 // occurrences of this instruction.
11431 RemoveFromWorkList(I);
11432 I->eraseFromParent();
11435 AddUsersToWorkList(*I);
11442 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11444 // Do an explicit clear, this shrinks the map if needed.
11445 WorklistMap.clear();
11450 bool InstCombiner::runOnFunction(Function &F) {
11451 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11453 bool EverMadeChange = false;
11455 // Iterate while there is work to do.
11456 unsigned Iteration = 0;
11457 while (DoOneIteration(F, Iteration++))
11458 EverMadeChange = true;
11459 return EverMadeChange;
11462 FunctionPass *llvm::createInstructionCombiningPass() {
11463 return new InstCombiner();