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/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/ConstantRange.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/GetElementPtrTypeIterator.h"
50 #include "llvm/Support/InstVisitor.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
63 using namespace llvm::PatternMatch;
65 STATISTIC(NumCombined , "Number of insts combined");
66 STATISTIC(NumConstProp, "Number of constant folds");
67 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
68 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
69 STATISTIC(NumSunkInst , "Number of instructions sunk");
72 class VISIBILITY_HIDDEN InstCombiner
73 : public FunctionPass,
74 public InstVisitor<InstCombiner, Instruction*> {
75 // Worklist of all of the instructions that need to be simplified.
76 std::vector<Instruction*> Worklist;
77 DenseMap<Instruction*, unsigned> WorklistMap;
79 bool MustPreserveLCSSA;
81 static char ID; // Pass identification, replacement for typeid
82 InstCombiner() : FunctionPass((intptr_t)&ID) {}
84 /// AddToWorkList - Add the specified instruction to the worklist if it
85 /// isn't already in it.
86 void AddToWorkList(Instruction *I) {
87 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
88 Worklist.push_back(I);
91 // RemoveFromWorkList - remove I from the worklist if it exists.
92 void RemoveFromWorkList(Instruction *I) {
93 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
94 if (It == WorklistMap.end()) return; // Not in worklist.
96 // Don't bother moving everything down, just null out the slot.
97 Worklist[It->second] = 0;
99 WorklistMap.erase(It);
102 Instruction *RemoveOneFromWorkList() {
103 Instruction *I = Worklist.back();
105 WorklistMap.erase(I);
110 /// AddUsersToWorkList - When an instruction is simplified, add all users of
111 /// the instruction to the work lists because they might get more simplified
114 void AddUsersToWorkList(Value &I) {
115 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
117 AddToWorkList(cast<Instruction>(*UI));
120 /// AddUsesToWorkList - When an instruction is simplified, add operands to
121 /// the work lists because they might get more simplified now.
123 void AddUsesToWorkList(Instruction &I) {
124 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
125 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
129 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
130 /// dead. Add all of its operands to the worklist, turning them into
131 /// undef's to reduce the number of uses of those instructions.
133 /// Return the specified operand before it is turned into an undef.
135 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
136 Value *R = I.getOperand(op);
138 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
139 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
141 // Set the operand to undef to drop the use.
142 I.setOperand(i, UndefValue::get(Op->getType()));
149 virtual bool runOnFunction(Function &F);
151 bool DoOneIteration(Function &F, unsigned ItNum);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 AU.addRequired<TargetData>();
155 AU.addPreservedID(LCSSAID);
156 AU.setPreservesCFG();
159 TargetData &getTargetData() const { return *TD; }
161 // Visitation implementation - Implement instruction combining for different
162 // instruction types. The semantics are as follows:
164 // null - No change was made
165 // I - Change was made, I is still valid, I may be dead though
166 // otherwise - Change was made, replace I with returned instruction
168 Instruction *visitAdd(BinaryOperator &I);
169 Instruction *visitSub(BinaryOperator &I);
170 Instruction *visitMul(BinaryOperator &I);
171 Instruction *visitURem(BinaryOperator &I);
172 Instruction *visitSRem(BinaryOperator &I);
173 Instruction *visitFRem(BinaryOperator &I);
174 Instruction *commonRemTransforms(BinaryOperator &I);
175 Instruction *commonIRemTransforms(BinaryOperator &I);
176 Instruction *commonDivTransforms(BinaryOperator &I);
177 Instruction *commonIDivTransforms(BinaryOperator &I);
178 Instruction *visitUDiv(BinaryOperator &I);
179 Instruction *visitSDiv(BinaryOperator &I);
180 Instruction *visitFDiv(BinaryOperator &I);
181 Instruction *visitAnd(BinaryOperator &I);
182 Instruction *visitOr (BinaryOperator &I);
183 Instruction *visitXor(BinaryOperator &I);
184 Instruction *visitShl(BinaryOperator &I);
185 Instruction *visitAShr(BinaryOperator &I);
186 Instruction *visitLShr(BinaryOperator &I);
187 Instruction *commonShiftTransforms(BinaryOperator &I);
188 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
190 Instruction *visitFCmpInst(FCmpInst &I);
191 Instruction *visitICmpInst(ICmpInst &I);
192 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
193 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
196 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
197 ConstantInt *DivRHS);
199 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
200 ICmpInst::Predicate Cond, Instruction &I);
201 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
203 Instruction *commonCastTransforms(CastInst &CI);
204 Instruction *commonIntCastTransforms(CastInst &CI);
205 Instruction *commonPointerCastTransforms(CastInst &CI);
206 Instruction *visitTrunc(TruncInst &CI);
207 Instruction *visitZExt(ZExtInst &CI);
208 Instruction *visitSExt(SExtInst &CI);
209 Instruction *visitFPTrunc(FPTruncInst &CI);
210 Instruction *visitFPExt(CastInst &CI);
211 Instruction *visitFPToUI(FPToUIInst &FI);
212 Instruction *visitFPToSI(FPToSIInst &FI);
213 Instruction *visitUIToFP(CastInst &CI);
214 Instruction *visitSIToFP(CastInst &CI);
215 Instruction *visitPtrToInt(CastInst &CI);
216 Instruction *visitIntToPtr(IntToPtrInst &CI);
217 Instruction *visitBitCast(BitCastInst &CI);
218 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
220 Instruction *visitSelectInst(SelectInst &CI);
221 Instruction *visitCallInst(CallInst &CI);
222 Instruction *visitInvokeInst(InvokeInst &II);
223 Instruction *visitPHINode(PHINode &PN);
224 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
225 Instruction *visitAllocationInst(AllocationInst &AI);
226 Instruction *visitFreeInst(FreeInst &FI);
227 Instruction *visitLoadInst(LoadInst &LI);
228 Instruction *visitStoreInst(StoreInst &SI);
229 Instruction *visitBranchInst(BranchInst &BI);
230 Instruction *visitSwitchInst(SwitchInst &SI);
231 Instruction *visitInsertElementInst(InsertElementInst &IE);
232 Instruction *visitExtractElementInst(ExtractElementInst &EI);
233 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
235 // visitInstruction - Specify what to return for unhandled instructions...
236 Instruction *visitInstruction(Instruction &I) { return 0; }
239 Instruction *visitCallSite(CallSite CS);
240 bool transformConstExprCastCall(CallSite CS);
241 Instruction *transformCallThroughTrampoline(CallSite CS);
242 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
243 bool DoXform = true);
244 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
247 // InsertNewInstBefore - insert an instruction New before instruction Old
248 // in the program. Add the new instruction to the worklist.
250 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
251 assert(New && New->getParent() == 0 &&
252 "New instruction already inserted into a basic block!");
253 BasicBlock *BB = Old.getParent();
254 BB->getInstList().insert(&Old, New); // Insert inst
259 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
260 /// This also adds the cast to the worklist. Finally, this returns the
262 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
264 if (V->getType() == Ty) return V;
266 if (Constant *CV = dyn_cast<Constant>(V))
267 return ConstantExpr::getCast(opc, CV, Ty);
269 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
274 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
275 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
279 // ReplaceInstUsesWith - This method is to be used when an instruction is
280 // found to be dead, replacable with another preexisting expression. Here
281 // we add all uses of I to the worklist, replace all uses of I with the new
282 // value, then return I, so that the inst combiner will know that I was
285 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
286 AddUsersToWorkList(I); // Add all modified instrs to worklist
288 I.replaceAllUsesWith(V);
291 // If we are replacing the instruction with itself, this must be in a
292 // segment of unreachable code, so just clobber the instruction.
293 I.replaceAllUsesWith(UndefValue::get(I.getType()));
298 // UpdateValueUsesWith - This method is to be used when an value is
299 // found to be replacable with another preexisting expression or was
300 // updated. Here we add all uses of I to the worklist, replace all uses of
301 // I with the new value (unless the instruction was just updated), then
302 // return true, so that the inst combiner will know that I was modified.
304 bool UpdateValueUsesWith(Value *Old, Value *New) {
305 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
307 Old->replaceAllUsesWith(New);
308 if (Instruction *I = dyn_cast<Instruction>(Old))
310 if (Instruction *I = dyn_cast<Instruction>(New))
315 // EraseInstFromFunction - When dealing with an instruction that has side
316 // effects or produces a void value, we can't rely on DCE to delete the
317 // instruction. Instead, visit methods should return the value returned by
319 Instruction *EraseInstFromFunction(Instruction &I) {
320 assert(I.use_empty() && "Cannot erase instruction that is used!");
321 AddUsesToWorkList(I);
322 RemoveFromWorkList(&I);
324 return 0; // Don't do anything with FI
328 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
329 /// InsertBefore instruction. This is specialized a bit to avoid inserting
330 /// casts that are known to not do anything...
332 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
333 Value *V, const Type *DestTy,
334 Instruction *InsertBefore);
336 /// SimplifyCommutative - This performs a few simplifications for
337 /// commutative operators.
338 bool SimplifyCommutative(BinaryOperator &I);
340 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
341 /// most-complex to least-complex order.
342 bool SimplifyCompare(CmpInst &I);
344 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
345 /// on the demanded bits.
346 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
347 APInt& KnownZero, APInt& KnownOne,
350 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
351 uint64_t &UndefElts, unsigned Depth = 0);
353 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
354 // PHI node as operand #0, see if we can fold the instruction into the PHI
355 // (which is only possible if all operands to the PHI are constants).
356 Instruction *FoldOpIntoPhi(Instruction &I);
358 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
359 // operator and they all are only used by the PHI, PHI together their
360 // inputs, and do the operation once, to the result of the PHI.
361 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
362 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
365 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
366 ConstantInt *AndRHS, BinaryOperator &TheAnd);
368 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
369 bool isSub, Instruction &I);
370 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
371 bool isSigned, bool Inside, Instruction &IB);
372 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
373 Instruction *MatchBSwap(BinaryOperator &I);
374 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
375 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
376 Instruction *SimplifyMemSet(MemSetInst *MI);
379 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
381 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
382 APInt& KnownOne, unsigned Depth = 0) const;
383 bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0);
384 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const;
385 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
387 int &NumCastsRemoved);
388 unsigned GetOrEnforceKnownAlignment(Value *V,
389 unsigned PrefAlign = 0);
393 char InstCombiner::ID = 0;
394 static RegisterPass<InstCombiner>
395 X("instcombine", "Combine redundant instructions");
397 // getComplexity: Assign a complexity or rank value to LLVM Values...
398 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
399 static unsigned getComplexity(Value *V) {
400 if (isa<Instruction>(V)) {
401 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
405 if (isa<Argument>(V)) return 3;
406 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
409 // isOnlyUse - Return true if this instruction will be deleted if we stop using
411 static bool isOnlyUse(Value *V) {
412 return V->hasOneUse() || isa<Constant>(V);
415 // getPromotedType - Return the specified type promoted as it would be to pass
416 // though a va_arg area...
417 static const Type *getPromotedType(const Type *Ty) {
418 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
419 if (ITy->getBitWidth() < 32)
420 return Type::Int32Ty;
425 /// getBitCastOperand - If the specified operand is a CastInst or a constant
426 /// expression bitcast, return the operand value, otherwise return null.
427 static Value *getBitCastOperand(Value *V) {
428 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
429 return I->getOperand(0);
430 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
431 if (CE->getOpcode() == Instruction::BitCast)
432 return CE->getOperand(0);
436 /// This function is a wrapper around CastInst::isEliminableCastPair. It
437 /// simply extracts arguments and returns what that function returns.
438 static Instruction::CastOps
439 isEliminableCastPair(
440 const CastInst *CI, ///< The first cast instruction
441 unsigned opcode, ///< The opcode of the second cast instruction
442 const Type *DstTy, ///< The target type for the second cast instruction
443 TargetData *TD ///< The target data for pointer size
446 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
447 const Type *MidTy = CI->getType(); // B from above
449 // Get the opcodes of the two Cast instructions
450 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
451 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
453 return Instruction::CastOps(
454 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
455 DstTy, TD->getIntPtrType()));
458 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
459 /// in any code being generated. It does not require codegen if V is simple
460 /// enough or if the cast can be folded into other casts.
461 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
462 const Type *Ty, TargetData *TD) {
463 if (V->getType() == Ty || isa<Constant>(V)) return false;
465 // If this is another cast that can be eliminated, it isn't codegen either.
466 if (const CastInst *CI = dyn_cast<CastInst>(V))
467 if (isEliminableCastPair(CI, opcode, Ty, TD))
472 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
473 /// InsertBefore instruction. This is specialized a bit to avoid inserting
474 /// casts that are known to not do anything...
476 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
477 Value *V, const Type *DestTy,
478 Instruction *InsertBefore) {
479 if (V->getType() == DestTy) return V;
480 if (Constant *C = dyn_cast<Constant>(V))
481 return ConstantExpr::getCast(opcode, C, DestTy);
483 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
486 // SimplifyCommutative - This performs a few simplifications for commutative
489 // 1. Order operands such that they are listed from right (least complex) to
490 // left (most complex). This puts constants before unary operators before
493 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
494 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
496 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
497 bool Changed = false;
498 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
499 Changed = !I.swapOperands();
501 if (!I.isAssociative()) return Changed;
502 Instruction::BinaryOps Opcode = I.getOpcode();
503 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
504 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
505 if (isa<Constant>(I.getOperand(1))) {
506 Constant *Folded = ConstantExpr::get(I.getOpcode(),
507 cast<Constant>(I.getOperand(1)),
508 cast<Constant>(Op->getOperand(1)));
509 I.setOperand(0, Op->getOperand(0));
510 I.setOperand(1, Folded);
512 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
513 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
514 isOnlyUse(Op) && isOnlyUse(Op1)) {
515 Constant *C1 = cast<Constant>(Op->getOperand(1));
516 Constant *C2 = cast<Constant>(Op1->getOperand(1));
518 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
519 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
520 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
524 I.setOperand(0, New);
525 I.setOperand(1, Folded);
532 /// SimplifyCompare - For a CmpInst this function just orders the operands
533 /// so that theyare listed from right (least complex) to left (most complex).
534 /// This puts constants before unary operators before binary operators.
535 bool InstCombiner::SimplifyCompare(CmpInst &I) {
536 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
539 // Compare instructions are not associative so there's nothing else we can do.
543 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
544 // if the LHS is a constant zero (which is the 'negate' form).
546 static inline Value *dyn_castNegVal(Value *V) {
547 if (BinaryOperator::isNeg(V))
548 return BinaryOperator::getNegArgument(V);
550 // Constants can be considered to be negated values if they can be folded.
551 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
552 return ConstantExpr::getNeg(C);
554 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
555 if (C->getType()->getElementType()->isInteger())
556 return ConstantExpr::getNeg(C);
561 static inline Value *dyn_castNotVal(Value *V) {
562 if (BinaryOperator::isNot(V))
563 return BinaryOperator::getNotArgument(V);
565 // Constants can be considered to be not'ed values...
566 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
567 return ConstantInt::get(~C->getValue());
571 // dyn_castFoldableMul - If this value is a multiply that can be folded into
572 // other computations (because it has a constant operand), return the
573 // non-constant operand of the multiply, and set CST to point to the multiplier.
574 // Otherwise, return null.
576 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
577 if (V->hasOneUse() && V->getType()->isInteger())
578 if (Instruction *I = dyn_cast<Instruction>(V)) {
579 if (I->getOpcode() == Instruction::Mul)
580 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
581 return I->getOperand(0);
582 if (I->getOpcode() == Instruction::Shl)
583 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
584 // The multiplier is really 1 << CST.
585 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
586 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
587 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
588 return I->getOperand(0);
594 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
595 /// expression, return it.
596 static User *dyn_castGetElementPtr(Value *V) {
597 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
598 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
599 if (CE->getOpcode() == Instruction::GetElementPtr)
600 return cast<User>(V);
604 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
605 /// opcode value. Otherwise return UserOp1.
606 static unsigned getOpcode(Value *V) {
607 if (Instruction *I = dyn_cast<Instruction>(V))
608 return I->getOpcode();
609 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
610 return CE->getOpcode();
611 // Use UserOp1 to mean there's no opcode.
612 return Instruction::UserOp1;
615 /// AddOne - Add one to a ConstantInt
616 static ConstantInt *AddOne(ConstantInt *C) {
617 APInt Val(C->getValue());
618 return ConstantInt::get(++Val);
620 /// SubOne - Subtract one from a ConstantInt
621 static ConstantInt *SubOne(ConstantInt *C) {
622 APInt Val(C->getValue());
623 return ConstantInt::get(--Val);
625 /// Add - Add two ConstantInts together
626 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
627 return ConstantInt::get(C1->getValue() + C2->getValue());
629 /// And - Bitwise AND two ConstantInts together
630 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
631 return ConstantInt::get(C1->getValue() & C2->getValue());
633 /// Subtract - Subtract one ConstantInt from another
634 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
635 return ConstantInt::get(C1->getValue() - C2->getValue());
637 /// Multiply - Multiply two ConstantInts together
638 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
639 return ConstantInt::get(C1->getValue() * C2->getValue());
641 /// MultiplyOverflows - True if the multiply can not be expressed in an int
643 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
644 uint32_t W = C1->getBitWidth();
645 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
654 APInt MulExt = LHSExt * RHSExt;
657 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
658 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
659 return MulExt.slt(Min) || MulExt.sgt(Max);
661 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
664 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
665 /// known to be either zero or one and return them in the KnownZero/KnownOne
666 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
668 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
669 /// we cannot optimize based on the assumption that it is zero without changing
670 /// it to be an explicit zero. If we don't change it to zero, other code could
671 /// optimized based on the contradictory assumption that it is non-zero.
672 /// Because instcombine aggressively folds operations with undef args anyway,
673 /// this won't lose us code quality.
674 void InstCombiner::ComputeMaskedBits(Value *V, const APInt &Mask,
675 APInt& KnownZero, APInt& KnownOne,
676 unsigned Depth) const {
677 assert(V && "No Value?");
678 assert(Depth <= 6 && "Limit Search Depth");
679 uint32_t BitWidth = Mask.getBitWidth();
680 assert((V->getType()->isInteger() || isa<PointerType>(V->getType())) &&
681 "Not integer or pointer type!");
682 assert((!TD || TD->getTypeSizeInBits(V->getType()) == BitWidth) &&
683 (!isa<IntegerType>(V->getType()) ||
684 V->getType()->getPrimitiveSizeInBits() == BitWidth) &&
685 KnownZero.getBitWidth() == BitWidth &&
686 KnownOne.getBitWidth() == BitWidth &&
687 "V, Mask, KnownOne and KnownZero should have same BitWidth");
689 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
690 // We know all of the bits for a constant!
691 KnownOne = CI->getValue() & Mask;
692 KnownZero = ~KnownOne & Mask;
695 // Null is all-zeros.
696 if (isa<ConstantPointerNull>(V)) {
701 // The address of an aligned GlobalValue has trailing zeros.
702 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
703 unsigned Align = GV->getAlignment();
704 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
705 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
707 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
708 CountTrailingZeros_32(Align));
715 KnownZero.clear(); KnownOne.clear(); // Start out not knowing anything.
717 if (Depth == 6 || Mask == 0)
718 return; // Limit search depth.
720 User *I = dyn_cast<User>(V);
723 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
724 switch (getOpcode(I)) {
726 case Instruction::And: {
727 // If either the LHS or the RHS are Zero, the result is zero.
728 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
729 APInt Mask2(Mask & ~KnownZero);
730 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
731 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
732 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
734 // Output known-1 bits are only known if set in both the LHS & RHS.
735 KnownOne &= KnownOne2;
736 // Output known-0 are known to be clear if zero in either the LHS | RHS.
737 KnownZero |= KnownZero2;
740 case Instruction::Or: {
741 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
742 APInt Mask2(Mask & ~KnownOne);
743 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
745 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
747 // Output known-0 bits are only known if clear in both the LHS & RHS.
748 KnownZero &= KnownZero2;
749 // Output known-1 are known to be set if set in either the LHS | RHS.
750 KnownOne |= KnownOne2;
753 case Instruction::Xor: {
754 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
755 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
756 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
757 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
759 // Output known-0 bits are known if clear or set in both the LHS & RHS.
760 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
761 // Output known-1 are known to be set if set in only one of the LHS, RHS.
762 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
763 KnownZero = KnownZeroOut;
766 case Instruction::Mul: {
767 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
768 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
769 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
770 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
771 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
773 // If low bits are zero in either operand, output low known-0 bits.
774 // Also compute a conserative estimate for high known-0 bits.
775 // More trickiness is possible, but this is sufficient for the
776 // interesting case of alignment computation.
778 unsigned TrailZ = KnownZero.countTrailingOnes() +
779 KnownZero2.countTrailingOnes();
780 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
781 KnownZero2.countLeadingOnes(),
782 BitWidth) - BitWidth;
784 TrailZ = std::min(TrailZ, BitWidth);
785 LeadZ = std::min(LeadZ, BitWidth);
786 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
787 APInt::getHighBitsSet(BitWidth, LeadZ);
791 case Instruction::UDiv: {
792 // For the purposes of computing leading zeros we can conservatively
793 // treat a udiv as a logical right shift by the power of 2 known to
794 // be less than the denominator.
795 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
796 ComputeMaskedBits(I->getOperand(0),
797 AllOnes, KnownZero2, KnownOne2, Depth+1);
798 unsigned LeadZ = KnownZero2.countLeadingOnes();
802 ComputeMaskedBits(I->getOperand(1),
803 AllOnes, KnownZero2, KnownOne2, Depth+1);
804 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
805 if (RHSUnknownLeadingOnes != BitWidth)
806 LeadZ = std::min(BitWidth,
807 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
809 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
812 case Instruction::Select:
813 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
814 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
815 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
816 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
818 // Only known if known in both the LHS and RHS.
819 KnownOne &= KnownOne2;
820 KnownZero &= KnownZero2;
822 case Instruction::FPTrunc:
823 case Instruction::FPExt:
824 case Instruction::FPToUI:
825 case Instruction::FPToSI:
826 case Instruction::SIToFP:
827 case Instruction::UIToFP:
828 return; // Can't work with floating point.
829 case Instruction::PtrToInt:
830 case Instruction::IntToPtr:
831 // We can't handle these if we don't know the pointer size.
833 // FALL THROUGH and handle them the same as zext/trunc.
834 case Instruction::ZExt:
835 case Instruction::Trunc: {
836 // Note that we handle pointer operands here because of inttoptr/ptrtoint
837 // which fall through here.
838 const Type *SrcTy = I->getOperand(0)->getType();
839 uint32_t SrcBitWidth = TD ?
840 TD->getTypeSizeInBits(SrcTy) :
841 SrcTy->getPrimitiveSizeInBits();
843 MaskIn.zextOrTrunc(SrcBitWidth);
844 KnownZero.zextOrTrunc(SrcBitWidth);
845 KnownOne.zextOrTrunc(SrcBitWidth);
846 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
847 KnownZero.zextOrTrunc(BitWidth);
848 KnownOne.zextOrTrunc(BitWidth);
849 // Any top bits are known to be zero.
850 if (BitWidth > SrcBitWidth)
851 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
854 case Instruction::BitCast: {
855 const Type *SrcTy = I->getOperand(0)->getType();
856 if (SrcTy->isInteger() || isa<PointerType>(SrcTy)) {
857 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
862 case Instruction::SExt: {
863 // Compute the bits in the result that are not present in the input.
864 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
865 uint32_t SrcBitWidth = SrcTy->getBitWidth();
868 MaskIn.trunc(SrcBitWidth);
869 KnownZero.trunc(SrcBitWidth);
870 KnownOne.trunc(SrcBitWidth);
871 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
872 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
873 KnownZero.zext(BitWidth);
874 KnownOne.zext(BitWidth);
876 // If the sign bit of the input is known set or clear, then we know the
877 // top bits of the result.
878 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
879 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
880 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
881 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
884 case Instruction::Shl:
885 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
886 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
887 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
888 APInt Mask2(Mask.lshr(ShiftAmt));
889 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
890 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
891 KnownZero <<= ShiftAmt;
892 KnownOne <<= ShiftAmt;
893 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
897 case Instruction::LShr:
898 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
899 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
900 // Compute the new bits that are at the top now.
901 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
903 // Unsigned shift right.
904 APInt Mask2(Mask.shl(ShiftAmt));
905 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
906 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
907 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
908 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
909 // high bits known zero.
910 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
914 case Instruction::AShr:
915 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
916 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
917 // Compute the new bits that are at the top now.
918 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
920 // Signed shift right.
921 APInt Mask2(Mask.shl(ShiftAmt));
922 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
923 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
924 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
925 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
927 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
928 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
929 KnownZero |= HighBits;
930 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
931 KnownOne |= HighBits;
935 case Instruction::Sub: {
936 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
937 // We know that the top bits of C-X are clear if X contains less bits
938 // than C (i.e. no wrap-around can happen). For example, 20-X is
939 // positive if we can prove that X is >= 0 and < 16.
940 if (!CLHS->getValue().isNegative()) {
941 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
942 // NLZ can't be BitWidth with no sign bit
943 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
944 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
947 // If all of the MaskV bits are known to be zero, then we know the
948 // output top bits are zero, because we now know that the output is
950 if ((KnownZero2 & MaskV) == MaskV) {
951 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
952 // Top bits known zero.
953 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
959 case Instruction::Add: {
960 // Output known-0 bits are known if clear or set in both the low clear bits
961 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
963 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
964 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
965 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
966 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
968 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
969 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
970 KnownZeroOut = std::min(KnownZeroOut,
971 KnownZero2.countTrailingOnes());
973 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
976 case Instruction::SRem:
977 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
978 APInt RA = Rem->getValue();
979 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
980 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
981 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
982 ComputeMaskedBits(I->getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
984 // The sign of a remainder is equal to the sign of the first
985 // operand (zero being positive).
986 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
987 KnownZero2 |= ~LowBits;
988 else if (KnownOne2[BitWidth-1])
989 KnownOne2 |= ~LowBits;
991 KnownZero |= KnownZero2 & Mask;
992 KnownOne |= KnownOne2 & Mask;
994 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
998 case Instruction::URem: {
999 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1000 APInt RA = Rem->getValue();
1001 if (RA.isPowerOf2()) {
1002 APInt LowBits = (RA - 1);
1003 APInt Mask2 = LowBits & Mask;
1004 KnownZero |= ~LowBits & Mask;
1005 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
1006 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1011 // Since the result is less than or equal to either operand, any leading
1012 // zero bits in either operand must also exist in the result.
1013 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1014 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
1016 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
1019 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1020 KnownZero2.countLeadingOnes());
1022 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
1026 case Instruction::Alloca:
1027 case Instruction::Malloc: {
1028 AllocationInst *AI = cast<AllocationInst>(V);
1029 unsigned Align = AI->getAlignment();
1030 if (Align == 0 && TD) {
1031 if (isa<AllocaInst>(AI))
1032 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
1033 else if (isa<MallocInst>(AI)) {
1034 // Malloc returns maximally aligned memory.
1035 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1038 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
1041 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
1046 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
1047 CountTrailingZeros_32(Align));
1050 case Instruction::GetElementPtr: {
1051 // Analyze all of the subscripts of this getelementptr instruction
1052 // to determine if we can prove known low zero bits.
1053 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
1054 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1055 ComputeMaskedBits(I->getOperand(0), LocalMask,
1056 LocalKnownZero, LocalKnownOne, Depth+1);
1057 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1059 gep_type_iterator GTI = gep_type_begin(I);
1060 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1061 Value *Index = I->getOperand(i);
1062 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1063 // Handle struct member offset arithmetic.
1065 const StructLayout *SL = TD->getStructLayout(STy);
1066 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1067 uint64_t Offset = SL->getElementOffset(Idx);
1068 TrailZ = std::min(TrailZ,
1069 CountTrailingZeros_64(Offset));
1071 // Handle array index arithmetic.
1072 const Type *IndexedTy = GTI.getIndexedType();
1073 if (!IndexedTy->isSized()) return;
1074 unsigned GEPOpiBits = Index->getType()->getPrimitiveSizeInBits();
1075 uint64_t TypeSize = TD ? TD->getABITypeSize(IndexedTy) : 1;
1076 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
1077 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1078 ComputeMaskedBits(Index, LocalMask,
1079 LocalKnownZero, LocalKnownOne, Depth+1);
1080 TrailZ = std::min(TrailZ,
1081 CountTrailingZeros_64(TypeSize) +
1082 LocalKnownZero.countTrailingOnes());
1086 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
1089 case Instruction::PHI: {
1090 PHINode *P = cast<PHINode>(I);
1091 // Handle the case of a simple two-predecessor recurrence PHI.
1092 // There's a lot more that could theoretically be done here, but
1093 // this is sufficient to catch some interesting cases.
1094 if (P->getNumIncomingValues() == 2) {
1095 for (unsigned i = 0; i != 2; ++i) {
1096 Value *L = P->getIncomingValue(i);
1097 Value *R = P->getIncomingValue(!i);
1098 User *LU = dyn_cast<User>(L);
1099 unsigned Opcode = LU ? getOpcode(LU) : (unsigned)Instruction::UserOp1;
1100 // Check for operations that have the property that if
1101 // both their operands have low zero bits, the result
1102 // will have low zero bits.
1103 if (Opcode == Instruction::Add ||
1104 Opcode == Instruction::Sub ||
1105 Opcode == Instruction::And ||
1106 Opcode == Instruction::Or ||
1107 Opcode == Instruction::Mul) {
1108 Value *LL = LU->getOperand(0);
1109 Value *LR = LU->getOperand(1);
1110 // Find a recurrence.
1117 // Ok, we have a PHI of the form L op= R. Check for low
1119 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1120 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, Depth+1);
1121 Mask2 = APInt::getLowBitsSet(BitWidth,
1122 KnownZero2.countTrailingOnes());
1125 ComputeMaskedBits(L, Mask2, KnownZero2, KnownOne2, Depth+1);
1127 APInt::getLowBitsSet(BitWidth,
1128 KnownZero2.countTrailingOnes());
1135 case Instruction::Call:
1136 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1137 switch (II->getIntrinsicID()) {
1139 case Intrinsic::ctpop:
1140 case Intrinsic::ctlz:
1141 case Intrinsic::cttz: {
1142 unsigned LowBits = Log2_32(BitWidth)+1;
1143 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1152 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1153 /// this predicate to simplify operations downstream. Mask is known to be zero
1154 /// for bits that V cannot have.
1155 bool InstCombiner::MaskedValueIsZero(Value *V, const APInt& Mask,
1157 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1158 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
1159 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1160 return (KnownZero & Mask) == Mask;
1163 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1164 /// specified instruction is a constant integer. If so, check to see if there
1165 /// are any bits set in the constant that are not demanded. If so, shrink the
1166 /// constant and return true.
1167 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1169 assert(I && "No instruction?");
1170 assert(OpNo < I->getNumOperands() && "Operand index too large");
1172 // If the operand is not a constant integer, nothing to do.
1173 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1174 if (!OpC) return false;
1176 // If there are no bits set that aren't demanded, nothing to do.
1177 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
1178 if ((~Demanded & OpC->getValue()) == 0)
1181 // This instruction is producing bits that are not demanded. Shrink the RHS.
1182 Demanded &= OpC->getValue();
1183 I->setOperand(OpNo, ConstantInt::get(Demanded));
1187 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
1188 // set of known zero and one bits, compute the maximum and minimum values that
1189 // could have the specified known zero and known one bits, returning them in
1191 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
1192 const APInt& KnownZero,
1193 const APInt& KnownOne,
1194 APInt& Min, APInt& Max) {
1195 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
1196 assert(KnownZero.getBitWidth() == BitWidth &&
1197 KnownOne.getBitWidth() == BitWidth &&
1198 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
1199 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1200 APInt UnknownBits = ~(KnownZero|KnownOne);
1202 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
1203 // bit if it is unknown.
1205 Max = KnownOne|UnknownBits;
1207 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
1208 Min.set(BitWidth-1);
1209 Max.clear(BitWidth-1);
1213 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
1214 // a set of known zero and one bits, compute the maximum and minimum values that
1215 // could have the specified known zero and known one bits, returning them in
1217 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
1218 const APInt &KnownZero,
1219 const APInt &KnownOne,
1220 APInt &Min, APInt &Max) {
1221 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
1222 assert(KnownZero.getBitWidth() == BitWidth &&
1223 KnownOne.getBitWidth() == BitWidth &&
1224 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
1225 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1226 APInt UnknownBits = ~(KnownZero|KnownOne);
1228 // The minimum value is when the unknown bits are all zeros.
1230 // The maximum value is when the unknown bits are all ones.
1231 Max = KnownOne|UnknownBits;
1234 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
1235 /// value based on the demanded bits. When this function is called, it is known
1236 /// that only the bits set in DemandedMask of the result of V are ever used
1237 /// downstream. Consequently, depending on the mask and V, it may be possible
1238 /// to replace V with a constant or one of its operands. In such cases, this
1239 /// function does the replacement and returns true. In all other cases, it
1240 /// returns false after analyzing the expression and setting KnownOne and known
1241 /// to be one in the expression. KnownZero contains all the bits that are known
1242 /// to be zero in the expression. These are provided to potentially allow the
1243 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
1244 /// the expression. KnownOne and KnownZero always follow the invariant that
1245 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
1246 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
1247 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
1248 /// and KnownOne must all be the same.
1249 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
1250 APInt& KnownZero, APInt& KnownOne,
1252 assert(V != 0 && "Null pointer of Value???");
1253 assert(Depth <= 6 && "Limit Search Depth");
1254 uint32_t BitWidth = DemandedMask.getBitWidth();
1255 const IntegerType *VTy = cast<IntegerType>(V->getType());
1256 assert(VTy->getBitWidth() == BitWidth &&
1257 KnownZero.getBitWidth() == BitWidth &&
1258 KnownOne.getBitWidth() == BitWidth &&
1259 "Value *V, DemandedMask, KnownZero and KnownOne \
1260 must have same BitWidth");
1261 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1262 // We know all of the bits for a constant!
1263 KnownOne = CI->getValue() & DemandedMask;
1264 KnownZero = ~KnownOne & DemandedMask;
1270 if (!V->hasOneUse()) { // Other users may use these bits.
1271 if (Depth != 0) { // Not at the root.
1272 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1273 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1276 // If this is the root being simplified, allow it to have multiple uses,
1277 // just set the DemandedMask to all bits.
1278 DemandedMask = APInt::getAllOnesValue(BitWidth);
1279 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1280 if (V != UndefValue::get(VTy))
1281 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1283 } else if (Depth == 6) { // Limit search depth.
1287 Instruction *I = dyn_cast<Instruction>(V);
1288 if (!I) return false; // Only analyze instructions.
1290 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1291 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
1292 switch (I->getOpcode()) {
1294 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1296 case Instruction::And:
1297 // If either the LHS or the RHS are Zero, the result is zero.
1298 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1299 RHSKnownZero, RHSKnownOne, Depth+1))
1301 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1302 "Bits known to be one AND zero?");
1304 // If something is known zero on the RHS, the bits aren't demanded on the
1306 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
1307 LHSKnownZero, LHSKnownOne, Depth+1))
1309 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1310 "Bits known to be one AND zero?");
1312 // If all of the demanded bits are known 1 on one side, return the other.
1313 // These bits cannot contribute to the result of the 'and'.
1314 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1315 (DemandedMask & ~LHSKnownZero))
1316 return UpdateValueUsesWith(I, I->getOperand(0));
1317 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1318 (DemandedMask & ~RHSKnownZero))
1319 return UpdateValueUsesWith(I, I->getOperand(1));
1321 // If all of the demanded bits in the inputs are known zeros, return zero.
1322 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1323 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1325 // If the RHS is a constant, see if we can simplify it.
1326 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1327 return UpdateValueUsesWith(I, I);
1329 // Output known-1 bits are only known if set in both the LHS & RHS.
1330 RHSKnownOne &= LHSKnownOne;
1331 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1332 RHSKnownZero |= LHSKnownZero;
1334 case Instruction::Or:
1335 // If either the LHS or the RHS are One, the result is One.
1336 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1337 RHSKnownZero, RHSKnownOne, Depth+1))
1339 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1340 "Bits known to be one AND zero?");
1341 // If something is known one on the RHS, the bits aren't demanded on the
1343 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1344 LHSKnownZero, LHSKnownOne, Depth+1))
1346 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1347 "Bits known to be one AND zero?");
1349 // If all of the demanded bits are known zero on one side, return the other.
1350 // These bits cannot contribute to the result of the 'or'.
1351 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1352 (DemandedMask & ~LHSKnownOne))
1353 return UpdateValueUsesWith(I, I->getOperand(0));
1354 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1355 (DemandedMask & ~RHSKnownOne))
1356 return UpdateValueUsesWith(I, I->getOperand(1));
1358 // If all of the potentially set bits on one side are known to be set on
1359 // the other side, just use the 'other' side.
1360 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1361 (DemandedMask & (~RHSKnownZero)))
1362 return UpdateValueUsesWith(I, I->getOperand(0));
1363 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1364 (DemandedMask & (~LHSKnownZero)))
1365 return UpdateValueUsesWith(I, I->getOperand(1));
1367 // If the RHS is a constant, see if we can simplify it.
1368 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1369 return UpdateValueUsesWith(I, I);
1371 // Output known-0 bits are only known if clear in both the LHS & RHS.
1372 RHSKnownZero &= LHSKnownZero;
1373 // Output known-1 are known to be set if set in either the LHS | RHS.
1374 RHSKnownOne |= LHSKnownOne;
1376 case Instruction::Xor: {
1377 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1378 RHSKnownZero, RHSKnownOne, Depth+1))
1380 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1381 "Bits known to be one AND zero?");
1382 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1383 LHSKnownZero, LHSKnownOne, Depth+1))
1385 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1386 "Bits known to be one AND zero?");
1388 // If all of the demanded bits are known zero on one side, return the other.
1389 // These bits cannot contribute to the result of the 'xor'.
1390 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1391 return UpdateValueUsesWith(I, I->getOperand(0));
1392 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1393 return UpdateValueUsesWith(I, I->getOperand(1));
1395 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1396 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1397 (RHSKnownOne & LHSKnownOne);
1398 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1399 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1400 (RHSKnownOne & LHSKnownZero);
1402 // If all of the demanded bits are known to be zero on one side or the
1403 // other, turn this into an *inclusive* or.
1404 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1405 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1407 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1409 InsertNewInstBefore(Or, *I);
1410 return UpdateValueUsesWith(I, Or);
1413 // If all of the demanded bits on one side are known, and all of the set
1414 // bits on that side are also known to be set on the other side, turn this
1415 // into an AND, as we know the bits will be cleared.
1416 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1417 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1419 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1420 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1422 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1423 InsertNewInstBefore(And, *I);
1424 return UpdateValueUsesWith(I, And);
1428 // If the RHS is a constant, see if we can simplify it.
1429 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1430 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1431 return UpdateValueUsesWith(I, I);
1433 RHSKnownZero = KnownZeroOut;
1434 RHSKnownOne = KnownOneOut;
1437 case Instruction::Select:
1438 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1439 RHSKnownZero, RHSKnownOne, Depth+1))
1441 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1442 LHSKnownZero, LHSKnownOne, Depth+1))
1444 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1445 "Bits known to be one AND zero?");
1446 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1447 "Bits known to be one AND zero?");
1449 // If the operands are constants, see if we can simplify them.
1450 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1451 return UpdateValueUsesWith(I, I);
1452 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1453 return UpdateValueUsesWith(I, I);
1455 // Only known if known in both the LHS and RHS.
1456 RHSKnownOne &= LHSKnownOne;
1457 RHSKnownZero &= LHSKnownZero;
1459 case Instruction::Trunc: {
1461 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1462 DemandedMask.zext(truncBf);
1463 RHSKnownZero.zext(truncBf);
1464 RHSKnownOne.zext(truncBf);
1465 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1466 RHSKnownZero, RHSKnownOne, Depth+1))
1468 DemandedMask.trunc(BitWidth);
1469 RHSKnownZero.trunc(BitWidth);
1470 RHSKnownOne.trunc(BitWidth);
1471 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1472 "Bits known to be one AND zero?");
1475 case Instruction::BitCast:
1476 if (!I->getOperand(0)->getType()->isInteger())
1479 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1480 RHSKnownZero, RHSKnownOne, Depth+1))
1482 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1483 "Bits known to be one AND zero?");
1485 case Instruction::ZExt: {
1486 // Compute the bits in the result that are not present in the input.
1487 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1488 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1490 DemandedMask.trunc(SrcBitWidth);
1491 RHSKnownZero.trunc(SrcBitWidth);
1492 RHSKnownOne.trunc(SrcBitWidth);
1493 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1494 RHSKnownZero, RHSKnownOne, Depth+1))
1496 DemandedMask.zext(BitWidth);
1497 RHSKnownZero.zext(BitWidth);
1498 RHSKnownOne.zext(BitWidth);
1499 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1500 "Bits known to be one AND zero?");
1501 // The top bits are known to be zero.
1502 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1505 case Instruction::SExt: {
1506 // Compute the bits in the result that are not present in the input.
1507 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1508 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1510 APInt InputDemandedBits = DemandedMask &
1511 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1513 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1514 // If any of the sign extended bits are demanded, we know that the sign
1516 if ((NewBits & DemandedMask) != 0)
1517 InputDemandedBits.set(SrcBitWidth-1);
1519 InputDemandedBits.trunc(SrcBitWidth);
1520 RHSKnownZero.trunc(SrcBitWidth);
1521 RHSKnownOne.trunc(SrcBitWidth);
1522 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1523 RHSKnownZero, RHSKnownOne, Depth+1))
1525 InputDemandedBits.zext(BitWidth);
1526 RHSKnownZero.zext(BitWidth);
1527 RHSKnownOne.zext(BitWidth);
1528 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1529 "Bits known to be one AND zero?");
1531 // If the sign bit of the input is known set or clear, then we know the
1532 // top bits of the result.
1534 // If the input sign bit is known zero, or if the NewBits are not demanded
1535 // convert this into a zero extension.
1536 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1538 // Convert to ZExt cast
1539 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1540 return UpdateValueUsesWith(I, NewCast);
1541 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1542 RHSKnownOne |= NewBits;
1546 case Instruction::Add: {
1547 // Figure out what the input bits are. If the top bits of the and result
1548 // are not demanded, then the add doesn't demand them from its input
1550 uint32_t NLZ = DemandedMask.countLeadingZeros();
1552 // If there is a constant on the RHS, there are a variety of xformations
1554 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1555 // If null, this should be simplified elsewhere. Some of the xforms here
1556 // won't work if the RHS is zero.
1560 // If the top bit of the output is demanded, demand everything from the
1561 // input. Otherwise, we demand all the input bits except NLZ top bits.
1562 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1564 // Find information about known zero/one bits in the input.
1565 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1566 LHSKnownZero, LHSKnownOne, Depth+1))
1569 // If the RHS of the add has bits set that can't affect the input, reduce
1571 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1572 return UpdateValueUsesWith(I, I);
1574 // Avoid excess work.
1575 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1578 // Turn it into OR if input bits are zero.
1579 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1581 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1583 InsertNewInstBefore(Or, *I);
1584 return UpdateValueUsesWith(I, Or);
1587 // We can say something about the output known-zero and known-one bits,
1588 // depending on potential carries from the input constant and the
1589 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1590 // bits set and the RHS constant is 0x01001, then we know we have a known
1591 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1593 // To compute this, we first compute the potential carry bits. These are
1594 // the bits which may be modified. I'm not aware of a better way to do
1596 const APInt& RHSVal = RHS->getValue();
1597 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1599 // Now that we know which bits have carries, compute the known-1/0 sets.
1601 // Bits are known one if they are known zero in one operand and one in the
1602 // other, and there is no input carry.
1603 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1604 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1606 // Bits are known zero if they are known zero in both operands and there
1607 // is no input carry.
1608 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1610 // If the high-bits of this ADD are not demanded, then it does not demand
1611 // the high bits of its LHS or RHS.
1612 if (DemandedMask[BitWidth-1] == 0) {
1613 // Right fill the mask of bits for this ADD to demand the most
1614 // significant bit and all those below it.
1615 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1616 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1617 LHSKnownZero, LHSKnownOne, Depth+1))
1619 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1620 LHSKnownZero, LHSKnownOne, Depth+1))
1626 case Instruction::Sub:
1627 // If the high-bits of this SUB are not demanded, then it does not demand
1628 // the high bits of its LHS or RHS.
1629 if (DemandedMask[BitWidth-1] == 0) {
1630 // Right fill the mask of bits for this SUB to demand the most
1631 // significant bit and all those below it.
1632 uint32_t NLZ = DemandedMask.countLeadingZeros();
1633 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1634 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1635 LHSKnownZero, LHSKnownOne, Depth+1))
1637 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1638 LHSKnownZero, LHSKnownOne, Depth+1))
1641 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1642 // the known zeros and ones.
1643 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1645 case Instruction::Shl:
1646 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1647 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1648 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1649 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1650 RHSKnownZero, RHSKnownOne, Depth+1))
1652 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1653 "Bits known to be one AND zero?");
1654 RHSKnownZero <<= ShiftAmt;
1655 RHSKnownOne <<= ShiftAmt;
1656 // low bits known zero.
1658 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1661 case Instruction::LShr:
1662 // For a logical shift right
1663 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1664 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1666 // Unsigned shift right.
1667 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1668 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1669 RHSKnownZero, RHSKnownOne, Depth+1))
1671 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1672 "Bits known to be one AND zero?");
1673 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1674 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1676 // Compute the new bits that are at the top now.
1677 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1678 RHSKnownZero |= HighBits; // high bits known zero.
1682 case Instruction::AShr:
1683 // If this is an arithmetic shift right and only the low-bit is set, we can
1684 // always convert this into a logical shr, even if the shift amount is
1685 // variable. The low bit of the shift cannot be an input sign bit unless
1686 // the shift amount is >= the size of the datatype, which is undefined.
1687 if (DemandedMask == 1) {
1688 // Perform the logical shift right.
1689 Value *NewVal = BinaryOperator::CreateLShr(
1690 I->getOperand(0), I->getOperand(1), I->getName());
1691 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1692 return UpdateValueUsesWith(I, NewVal);
1695 // If the sign bit is the only bit demanded by this ashr, then there is no
1696 // need to do it, the shift doesn't change the high bit.
1697 if (DemandedMask.isSignBit())
1698 return UpdateValueUsesWith(I, I->getOperand(0));
1700 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1701 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1703 // Signed shift right.
1704 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1705 // If any of the "high bits" are demanded, we should set the sign bit as
1707 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1708 DemandedMaskIn.set(BitWidth-1);
1709 if (SimplifyDemandedBits(I->getOperand(0),
1711 RHSKnownZero, RHSKnownOne, Depth+1))
1713 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1714 "Bits known to be one AND zero?");
1715 // Compute the new bits that are at the top now.
1716 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1717 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1718 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1720 // Handle the sign bits.
1721 APInt SignBit(APInt::getSignBit(BitWidth));
1722 // Adjust to where it is now in the mask.
1723 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1725 // If the input sign bit is known to be zero, or if none of the top bits
1726 // are demanded, turn this into an unsigned shift right.
1727 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1728 (HighBits & ~DemandedMask) == HighBits) {
1729 // Perform the logical shift right.
1730 Value *NewVal = BinaryOperator::CreateLShr(
1731 I->getOperand(0), SA, I->getName());
1732 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1733 return UpdateValueUsesWith(I, NewVal);
1734 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1735 RHSKnownOne |= HighBits;
1739 case Instruction::SRem:
1740 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1741 APInt RA = Rem->getValue();
1742 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1743 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1744 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1745 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1746 LHSKnownZero, LHSKnownOne, Depth+1))
1749 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1750 LHSKnownZero |= ~LowBits;
1751 else if (LHSKnownOne[BitWidth-1])
1752 LHSKnownOne |= ~LowBits;
1754 KnownZero |= LHSKnownZero & DemandedMask;
1755 KnownOne |= LHSKnownOne & DemandedMask;
1757 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1761 case Instruction::URem: {
1762 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1763 APInt RA = Rem->getValue();
1764 if (RA.isPowerOf2()) {
1765 APInt LowBits = (RA - 1);
1766 APInt Mask2 = LowBits & DemandedMask;
1767 KnownZero |= ~LowBits & DemandedMask;
1768 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1769 KnownZero, KnownOne, Depth+1))
1772 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1777 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1778 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1779 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1780 KnownZero2, KnownOne2, Depth+1))
1783 uint32_t Leaders = KnownZero2.countLeadingOnes();
1784 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1785 KnownZero2, KnownOne2, Depth+1))
1788 Leaders = std::max(Leaders,
1789 KnownZero2.countLeadingOnes());
1790 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1795 // If the client is only demanding bits that we know, return the known
1797 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1798 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1803 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1804 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1805 /// actually used by the caller. This method analyzes which elements of the
1806 /// operand are undef and returns that information in UndefElts.
1808 /// If the information about demanded elements can be used to simplify the
1809 /// operation, the operation is simplified, then the resultant value is
1810 /// returned. This returns null if no change was made.
1811 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1812 uint64_t &UndefElts,
1814 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1815 assert(VWidth <= 64 && "Vector too wide to analyze!");
1816 uint64_t EltMask = ~0ULL >> (64-VWidth);
1817 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1818 "Invalid DemandedElts!");
1820 if (isa<UndefValue>(V)) {
1821 // If the entire vector is undefined, just return this info.
1822 UndefElts = EltMask;
1824 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1825 UndefElts = EltMask;
1826 return UndefValue::get(V->getType());
1830 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1831 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1832 Constant *Undef = UndefValue::get(EltTy);
1834 std::vector<Constant*> Elts;
1835 for (unsigned i = 0; i != VWidth; ++i)
1836 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1837 Elts.push_back(Undef);
1838 UndefElts |= (1ULL << i);
1839 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1840 Elts.push_back(Undef);
1841 UndefElts |= (1ULL << i);
1842 } else { // Otherwise, defined.
1843 Elts.push_back(CP->getOperand(i));
1846 // If we changed the constant, return it.
1847 Constant *NewCP = ConstantVector::get(Elts);
1848 return NewCP != CP ? NewCP : 0;
1849 } else if (isa<ConstantAggregateZero>(V)) {
1850 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1852 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1853 Constant *Zero = Constant::getNullValue(EltTy);
1854 Constant *Undef = UndefValue::get(EltTy);
1855 std::vector<Constant*> Elts;
1856 for (unsigned i = 0; i != VWidth; ++i)
1857 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1858 UndefElts = DemandedElts ^ EltMask;
1859 return ConstantVector::get(Elts);
1862 if (!V->hasOneUse()) { // Other users may use these bits.
1863 if (Depth != 0) { // Not at the root.
1864 // TODO: Just compute the UndefElts information recursively.
1868 } else if (Depth == 10) { // Limit search depth.
1872 Instruction *I = dyn_cast<Instruction>(V);
1873 if (!I) return false; // Only analyze instructions.
1875 bool MadeChange = false;
1876 uint64_t UndefElts2;
1878 switch (I->getOpcode()) {
1881 case Instruction::InsertElement: {
1882 // If this is a variable index, we don't know which element it overwrites.
1883 // demand exactly the same input as we produce.
1884 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1886 // Note that we can't propagate undef elt info, because we don't know
1887 // which elt is getting updated.
1888 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1889 UndefElts2, Depth+1);
1890 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1894 // If this is inserting an element that isn't demanded, remove this
1896 unsigned IdxNo = Idx->getZExtValue();
1897 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1898 return AddSoonDeadInstToWorklist(*I, 0);
1900 // Otherwise, the element inserted overwrites whatever was there, so the
1901 // input demanded set is simpler than the output set.
1902 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1903 DemandedElts & ~(1ULL << IdxNo),
1904 UndefElts, Depth+1);
1905 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1907 // The inserted element is defined.
1908 UndefElts |= 1ULL << IdxNo;
1911 case Instruction::BitCast: {
1912 // Vector->vector casts only.
1913 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1915 unsigned InVWidth = VTy->getNumElements();
1916 uint64_t InputDemandedElts = 0;
1919 if (VWidth == InVWidth) {
1920 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1921 // elements as are demanded of us.
1923 InputDemandedElts = DemandedElts;
1924 } else if (VWidth > InVWidth) {
1928 // If there are more elements in the result than there are in the source,
1929 // then an input element is live if any of the corresponding output
1930 // elements are live.
1931 Ratio = VWidth/InVWidth;
1932 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1933 if (DemandedElts & (1ULL << OutIdx))
1934 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1940 // If there are more elements in the source than there are in the result,
1941 // then an input element is live if the corresponding output element is
1943 Ratio = InVWidth/VWidth;
1944 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1945 if (DemandedElts & (1ULL << InIdx/Ratio))
1946 InputDemandedElts |= 1ULL << InIdx;
1949 // div/rem demand all inputs, because they don't want divide by zero.
1950 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1951 UndefElts2, Depth+1);
1953 I->setOperand(0, TmpV);
1957 UndefElts = UndefElts2;
1958 if (VWidth > InVWidth) {
1959 assert(0 && "Unimp");
1960 // If there are more elements in the result than there are in the source,
1961 // then an output element is undef if the corresponding input element is
1963 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1964 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1965 UndefElts |= 1ULL << OutIdx;
1966 } else if (VWidth < InVWidth) {
1967 assert(0 && "Unimp");
1968 // If there are more elements in the source than there are in the result,
1969 // then a result element is undef if all of the corresponding input
1970 // elements are undef.
1971 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1972 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1973 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1974 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1978 case Instruction::And:
1979 case Instruction::Or:
1980 case Instruction::Xor:
1981 case Instruction::Add:
1982 case Instruction::Sub:
1983 case Instruction::Mul:
1984 // div/rem demand all inputs, because they don't want divide by zero.
1985 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1986 UndefElts, Depth+1);
1987 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1988 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1989 UndefElts2, Depth+1);
1990 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1992 // Output elements are undefined if both are undefined. Consider things
1993 // like undef&0. The result is known zero, not undef.
1994 UndefElts &= UndefElts2;
1997 case Instruction::Call: {
1998 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
2000 switch (II->getIntrinsicID()) {
2003 // Binary vector operations that work column-wise. A dest element is a
2004 // function of the corresponding input elements from the two inputs.
2005 case Intrinsic::x86_sse_sub_ss:
2006 case Intrinsic::x86_sse_mul_ss:
2007 case Intrinsic::x86_sse_min_ss:
2008 case Intrinsic::x86_sse_max_ss:
2009 case Intrinsic::x86_sse2_sub_sd:
2010 case Intrinsic::x86_sse2_mul_sd:
2011 case Intrinsic::x86_sse2_min_sd:
2012 case Intrinsic::x86_sse2_max_sd:
2013 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
2014 UndefElts, Depth+1);
2015 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
2016 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
2017 UndefElts2, Depth+1);
2018 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
2020 // If only the low elt is demanded and this is a scalarizable intrinsic,
2021 // scalarize it now.
2022 if (DemandedElts == 1) {
2023 switch (II->getIntrinsicID()) {
2025 case Intrinsic::x86_sse_sub_ss:
2026 case Intrinsic::x86_sse_mul_ss:
2027 case Intrinsic::x86_sse2_sub_sd:
2028 case Intrinsic::x86_sse2_mul_sd:
2029 // TODO: Lower MIN/MAX/ABS/etc
2030 Value *LHS = II->getOperand(1);
2031 Value *RHS = II->getOperand(2);
2032 // Extract the element as scalars.
2033 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
2034 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
2036 switch (II->getIntrinsicID()) {
2037 default: assert(0 && "Case stmts out of sync!");
2038 case Intrinsic::x86_sse_sub_ss:
2039 case Intrinsic::x86_sse2_sub_sd:
2040 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
2041 II->getName()), *II);
2043 case Intrinsic::x86_sse_mul_ss:
2044 case Intrinsic::x86_sse2_mul_sd:
2045 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
2046 II->getName()), *II);
2051 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
2053 InsertNewInstBefore(New, *II);
2054 AddSoonDeadInstToWorklist(*II, 0);
2059 // Output elements are undefined if both are undefined. Consider things
2060 // like undef&0. The result is known zero, not undef.
2061 UndefElts &= UndefElts2;
2067 return MadeChange ? I : 0;
2070 /// ComputeNumSignBits - Return the number of times the sign bit of the
2071 /// register is replicated into the other bits. We know that at least 1 bit
2072 /// is always equal to the sign bit (itself), but other cases can give us
2073 /// information. For example, immediately after an "ashr X, 2", we know that
2074 /// the top 3 bits are all equal to each other, so we return 3.
2076 unsigned InstCombiner::ComputeNumSignBits(Value *V, unsigned Depth) const{
2077 const IntegerType *Ty = cast<IntegerType>(V->getType());
2078 unsigned TyBits = Ty->getBitWidth();
2080 unsigned FirstAnswer = 1;
2083 return 1; // Limit search depth.
2085 User *U = dyn_cast<User>(V);
2086 switch (getOpcode(V)) {
2088 case Instruction::SExt:
2089 Tmp = TyBits-cast<IntegerType>(U->getOperand(0)->getType())->getBitWidth();
2090 return ComputeNumSignBits(U->getOperand(0), Depth+1) + Tmp;
2092 case Instruction::AShr:
2093 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2094 // ashr X, C -> adds C sign bits.
2095 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
2096 Tmp += C->getZExtValue();
2097 if (Tmp > TyBits) Tmp = TyBits;
2100 case Instruction::Shl:
2101 if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
2102 // shl destroys sign bits.
2103 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2104 if (C->getZExtValue() >= TyBits || // Bad shift.
2105 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
2106 return Tmp - C->getZExtValue();
2109 case Instruction::And:
2110 case Instruction::Or:
2111 case Instruction::Xor: // NOT is handled here.
2112 // Logical binary ops preserve the number of sign bits at the worst.
2113 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2115 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth+1);
2116 FirstAnswer = std::min(Tmp, Tmp2);
2117 // We computed what we know about the sign bits as our first
2118 // answer. Now proceed to the generic code that uses
2119 // ComputeMaskedBits, and pick whichever answer is better.
2123 case Instruction::Select:
2124 Tmp = ComputeNumSignBits(U->getOperand(1), Depth+1);
2125 if (Tmp == 1) return 1; // Early out.
2126 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth+1);
2127 return std::min(Tmp, Tmp2);
2129 case Instruction::Add:
2130 // Add can have at most one carry bit. Thus we know that the output
2131 // is, at worst, one more bit than the inputs.
2132 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2133 if (Tmp == 1) return 1; // Early out.
2135 // Special case decrementing a value (ADD X, -1):
2136 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(0)))
2137 if (CRHS->isAllOnesValue()) {
2138 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2139 APInt Mask = APInt::getAllOnesValue(TyBits);
2140 ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
2142 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2144 if ((KnownZero | APInt(TyBits, 1)) == Mask)
2147 // If we are subtracting one from a positive number, there is no carry
2148 // out of the result.
2149 if (KnownZero.isNegative())
2153 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth+1);
2154 if (Tmp2 == 1) return 1;
2155 return std::min(Tmp, Tmp2)-1;
2158 case Instruction::Sub:
2159 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth+1);
2160 if (Tmp2 == 1) return 1;
2163 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
2164 if (CLHS->isNullValue()) {
2165 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2166 APInt Mask = APInt::getAllOnesValue(TyBits);
2167 ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
2168 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2170 if ((KnownZero | APInt(TyBits, 1)) == Mask)
2173 // If the input is known to be positive (the sign bit is known clear),
2174 // the output of the NEG has the same number of sign bits as the input.
2175 if (KnownZero.isNegative())
2178 // Otherwise, we treat this like a SUB.
2181 // Sub can have at most one carry bit. Thus we know that the output
2182 // is, at worst, one more bit than the inputs.
2183 Tmp = ComputeNumSignBits(U->getOperand(0), Depth+1);
2184 if (Tmp == 1) return 1; // Early out.
2185 return std::min(Tmp, Tmp2)-1;
2187 case Instruction::Trunc:
2188 // FIXME: it's tricky to do anything useful for this, but it is an important
2189 // case for targets like X86.
2193 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2194 // use this information.
2195 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
2196 APInt Mask = APInt::getAllOnesValue(TyBits);
2197 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
2199 if (KnownZero.isNegative()) { // sign bit is 0
2201 } else if (KnownOne.isNegative()) { // sign bit is 1;
2208 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
2209 // the number of identical bits in the top of the input value.
2211 Mask <<= Mask.getBitWidth()-TyBits;
2212 // Return # leading zeros. We use 'min' here in case Val was zero before
2213 // shifting. We don't want to return '64' as for an i32 "0".
2214 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
2218 /// AssociativeOpt - Perform an optimization on an associative operator. This
2219 /// function is designed to check a chain of associative operators for a
2220 /// potential to apply a certain optimization. Since the optimization may be
2221 /// applicable if the expression was reassociated, this checks the chain, then
2222 /// reassociates the expression as necessary to expose the optimization
2223 /// opportunity. This makes use of a special Functor, which must define
2224 /// 'shouldApply' and 'apply' methods.
2226 template<typename Functor>
2227 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
2228 unsigned Opcode = Root.getOpcode();
2229 Value *LHS = Root.getOperand(0);
2231 // Quick check, see if the immediate LHS matches...
2232 if (F.shouldApply(LHS))
2233 return F.apply(Root);
2235 // Otherwise, if the LHS is not of the same opcode as the root, return.
2236 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2237 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
2238 // Should we apply this transform to the RHS?
2239 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
2241 // If not to the RHS, check to see if we should apply to the LHS...
2242 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
2243 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
2247 // If the functor wants to apply the optimization to the RHS of LHSI,
2248 // reassociate the expression from ((? op A) op B) to (? op (A op B))
2250 BasicBlock *BB = Root.getParent();
2252 // Now all of the instructions are in the current basic block, go ahead
2253 // and perform the reassociation.
2254 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
2256 // First move the selected RHS to the LHS of the root...
2257 Root.setOperand(0, LHSI->getOperand(1));
2259 // Make what used to be the LHS of the root be the user of the root...
2260 Value *ExtraOperand = TmpLHSI->getOperand(1);
2261 if (&Root == TmpLHSI) {
2262 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
2265 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
2266 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
2267 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
2268 BasicBlock::iterator ARI = &Root; ++ARI;
2269 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
2272 // Now propagate the ExtraOperand down the chain of instructions until we
2274 while (TmpLHSI != LHSI) {
2275 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
2276 // Move the instruction to immediately before the chain we are
2277 // constructing to avoid breaking dominance properties.
2278 NextLHSI->getParent()->getInstList().remove(NextLHSI);
2279 BB->getInstList().insert(ARI, NextLHSI);
2282 Value *NextOp = NextLHSI->getOperand(1);
2283 NextLHSI->setOperand(1, ExtraOperand);
2285 ExtraOperand = NextOp;
2288 // Now that the instructions are reassociated, have the functor perform
2289 // the transformation...
2290 return F.apply(Root);
2293 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2300 // AddRHS - Implements: X + X --> X << 1
2303 AddRHS(Value *rhs) : RHS(rhs) {}
2304 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2305 Instruction *apply(BinaryOperator &Add) const {
2306 return BinaryOperator::CreateShl(Add.getOperand(0),
2307 ConstantInt::get(Add.getType(), 1));
2311 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2313 struct AddMaskingAnd {
2315 AddMaskingAnd(Constant *c) : C2(c) {}
2316 bool shouldApply(Value *LHS) const {
2318 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2319 ConstantExpr::getAnd(C1, C2)->isNullValue();
2321 Instruction *apply(BinaryOperator &Add) const {
2322 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
2328 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2330 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
2331 if (Constant *SOC = dyn_cast<Constant>(SO))
2332 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
2334 return IC->InsertNewInstBefore(CastInst::Create(
2335 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
2338 // Figure out if the constant is the left or the right argument.
2339 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2340 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2342 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2344 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2345 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2348 Value *Op0 = SO, *Op1 = ConstOperand;
2350 std::swap(Op0, Op1);
2352 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2353 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
2354 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2355 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
2356 SO->getName()+".cmp");
2358 assert(0 && "Unknown binary instruction type!");
2361 return IC->InsertNewInstBefore(New, I);
2364 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2365 // constant as the other operand, try to fold the binary operator into the
2366 // select arguments. This also works for Cast instructions, which obviously do
2367 // not have a second operand.
2368 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2370 // Don't modify shared select instructions
2371 if (!SI->hasOneUse()) return 0;
2372 Value *TV = SI->getOperand(1);
2373 Value *FV = SI->getOperand(2);
2375 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2376 // Bool selects with constant operands can be folded to logical ops.
2377 if (SI->getType() == Type::Int1Ty) return 0;
2379 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2380 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2382 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
2389 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
2390 /// node as operand #0, see if we can fold the instruction into the PHI (which
2391 /// is only possible if all operands to the PHI are constants).
2392 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
2393 PHINode *PN = cast<PHINode>(I.getOperand(0));
2394 unsigned NumPHIValues = PN->getNumIncomingValues();
2395 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
2397 // Check to see if all of the operands of the PHI are constants. If there is
2398 // one non-constant value, remember the BB it is. If there is more than one
2399 // or if *it* is a PHI, bail out.
2400 BasicBlock *NonConstBB = 0;
2401 for (unsigned i = 0; i != NumPHIValues; ++i)
2402 if (!isa<Constant>(PN->getIncomingValue(i))) {
2403 if (NonConstBB) return 0; // More than one non-const value.
2404 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2405 NonConstBB = PN->getIncomingBlock(i);
2407 // If the incoming non-constant value is in I's block, we have an infinite
2409 if (NonConstBB == I.getParent())
2413 // If there is exactly one non-constant value, we can insert a copy of the
2414 // operation in that block. However, if this is a critical edge, we would be
2415 // inserting the computation one some other paths (e.g. inside a loop). Only
2416 // do this if the pred block is unconditionally branching into the phi block.
2418 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2419 if (!BI || !BI->isUnconditional()) return 0;
2422 // Okay, we can do the transformation: create the new PHI node.
2423 PHINode *NewPN = PHINode::Create(I.getType(), "");
2424 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2425 InsertNewInstBefore(NewPN, *PN);
2426 NewPN->takeName(PN);
2428 // Next, add all of the operands to the PHI.
2429 if (I.getNumOperands() == 2) {
2430 Constant *C = cast<Constant>(I.getOperand(1));
2431 for (unsigned i = 0; i != NumPHIValues; ++i) {
2433 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2434 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2435 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2437 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2439 assert(PN->getIncomingBlock(i) == NonConstBB);
2440 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2441 InV = BinaryOperator::Create(BO->getOpcode(),
2442 PN->getIncomingValue(i), C, "phitmp",
2443 NonConstBB->getTerminator());
2444 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2445 InV = CmpInst::Create(CI->getOpcode(),
2447 PN->getIncomingValue(i), C, "phitmp",
2448 NonConstBB->getTerminator());
2450 assert(0 && "Unknown binop!");
2452 AddToWorkList(cast<Instruction>(InV));
2454 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2457 CastInst *CI = cast<CastInst>(&I);
2458 const Type *RetTy = CI->getType();
2459 for (unsigned i = 0; i != NumPHIValues; ++i) {
2461 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2462 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2464 assert(PN->getIncomingBlock(i) == NonConstBB);
2465 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2466 I.getType(), "phitmp",
2467 NonConstBB->getTerminator());
2468 AddToWorkList(cast<Instruction>(InV));
2470 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2473 return ReplaceInstUsesWith(I, NewPN);
2477 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
2478 /// value is never equal to -0.0.
2480 /// Note that this function will need to be revisited when we support nondefault
2483 static bool CannotBeNegativeZero(const Value *V) {
2484 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2485 return !CFP->getValueAPF().isNegZero();
2487 if (const Instruction *I = dyn_cast<Instruction>(V)) {
2488 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2489 if (I->getOpcode() == Instruction::Add &&
2490 isa<ConstantFP>(I->getOperand(1)) &&
2491 cast<ConstantFP>(I->getOperand(1))->isNullValue())
2494 // sitofp and uitofp turn into +0.0 for zero.
2495 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2498 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2499 if (II->getIntrinsicID() == Intrinsic::sqrt)
2500 return CannotBeNegativeZero(II->getOperand(1));
2502 if (const CallInst *CI = dyn_cast<CallInst>(I))
2503 if (const Function *F = CI->getCalledFunction()) {
2504 if (F->isDeclaration()) {
2505 switch (F->getNameLen()) {
2506 case 3: // abs(x) != -0.0
2507 if (!strcmp(F->getNameStart(), "abs")) return true;
2509 case 4: // abs[lf](x) != -0.0
2510 if (!strcmp(F->getNameStart(), "absf")) return true;
2511 if (!strcmp(F->getNameStart(), "absl")) return true;
2521 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2522 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2523 /// This basically requires proving that the add in the original type would not
2524 /// overflow to change the sign bit or have a carry out.
2525 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2526 // There are different heuristics we can use for this. Here are some simple
2529 // Add has the property that adding any two 2's complement numbers can only
2530 // have one carry bit which can change a sign. As such, if LHS and RHS each
2531 // have at least two sign bits, we know that the addition of the two values will
2532 // sign extend fine.
2533 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2537 // If one of the operands only has one non-zero bit, and if the other operand
2538 // has a known-zero bit in a more significant place than it (not including the
2539 // sign bit) the ripple may go up to and fill the zero, but won't change the
2540 // sign. For example, (X & ~4) + 1.
2548 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2549 bool Changed = SimplifyCommutative(I);
2550 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2552 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2553 // X + undef -> undef
2554 if (isa<UndefValue>(RHS))
2555 return ReplaceInstUsesWith(I, RHS);
2558 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2559 if (RHSC->isNullValue())
2560 return ReplaceInstUsesWith(I, LHS);
2561 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2562 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2563 (I.getType())->getValueAPF()))
2564 return ReplaceInstUsesWith(I, LHS);
2567 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2568 // X + (signbit) --> X ^ signbit
2569 const APInt& Val = CI->getValue();
2570 uint32_t BitWidth = Val.getBitWidth();
2571 if (Val == APInt::getSignBit(BitWidth))
2572 return BinaryOperator::CreateXor(LHS, RHS);
2574 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2575 // (X & 254)+1 -> (X&254)|1
2576 if (!isa<VectorType>(I.getType())) {
2577 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2578 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2579 KnownZero, KnownOne))
2584 if (isa<PHINode>(LHS))
2585 if (Instruction *NV = FoldOpIntoPhi(I))
2588 ConstantInt *XorRHS = 0;
2590 if (isa<ConstantInt>(RHSC) &&
2591 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2592 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2593 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2595 uint32_t Size = TySizeBits / 2;
2596 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2597 APInt CFF80Val(-C0080Val);
2599 if (TySizeBits > Size) {
2600 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2601 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2602 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2603 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2604 // This is a sign extend if the top bits are known zero.
2605 if (!MaskedValueIsZero(XorLHS,
2606 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2607 Size = 0; // Not a sign ext, but can't be any others either.
2612 C0080Val = APIntOps::lshr(C0080Val, Size);
2613 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2614 } while (Size >= 1);
2616 // FIXME: This shouldn't be necessary. When the backends can handle types
2617 // with funny bit widths then this switch statement should be removed. It
2618 // is just here to get the size of the "middle" type back up to something
2619 // that the back ends can handle.
2620 const Type *MiddleType = 0;
2623 case 32: MiddleType = Type::Int32Ty; break;
2624 case 16: MiddleType = Type::Int16Ty; break;
2625 case 8: MiddleType = Type::Int8Ty; break;
2628 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2629 InsertNewInstBefore(NewTrunc, I);
2630 return new SExtInst(NewTrunc, I.getType(), I.getName());
2636 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2637 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2639 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2640 if (RHSI->getOpcode() == Instruction::Sub)
2641 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2642 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2644 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2645 if (LHSI->getOpcode() == Instruction::Sub)
2646 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2647 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2652 // -A + -B --> -(A + B)
2653 if (Value *LHSV = dyn_castNegVal(LHS)) {
2654 if (LHS->getType()->isIntOrIntVector()) {
2655 if (Value *RHSV = dyn_castNegVal(RHS)) {
2656 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2657 InsertNewInstBefore(NewAdd, I);
2658 return BinaryOperator::CreateNeg(NewAdd);
2662 return BinaryOperator::CreateSub(RHS, LHSV);
2666 if (!isa<Constant>(RHS))
2667 if (Value *V = dyn_castNegVal(RHS))
2668 return BinaryOperator::CreateSub(LHS, V);
2672 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2673 if (X == RHS) // X*C + X --> X * (C+1)
2674 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2676 // X*C1 + X*C2 --> X * (C1+C2)
2678 if (X == dyn_castFoldableMul(RHS, C1))
2679 return BinaryOperator::CreateMul(X, Add(C1, C2));
2682 // X + X*C --> X * (C+1)
2683 if (dyn_castFoldableMul(RHS, C2) == LHS)
2684 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2686 // X + ~X --> -1 since ~X = -X-1
2687 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2688 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2691 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2692 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2693 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2696 // A+B --> A|B iff A and B have no bits set in common.
2697 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2698 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2699 APInt LHSKnownOne(IT->getBitWidth(), 0);
2700 APInt LHSKnownZero(IT->getBitWidth(), 0);
2701 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2702 if (LHSKnownZero != 0) {
2703 APInt RHSKnownOne(IT->getBitWidth(), 0);
2704 APInt RHSKnownZero(IT->getBitWidth(), 0);
2705 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2707 // No bits in common -> bitwise or.
2708 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2709 return BinaryOperator::CreateOr(LHS, RHS);
2713 // W*X + Y*Z --> W * (X+Z) iff W == Y
2714 if (I.getType()->isIntOrIntVector()) {
2715 Value *W, *X, *Y, *Z;
2716 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2717 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2721 } else if (Y == X) {
2723 } else if (X == Z) {
2730 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2731 LHS->getName()), I);
2732 return BinaryOperator::CreateMul(W, NewAdd);
2737 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2739 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2740 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2742 // (X & FF00) + xx00 -> (X+xx00) & FF00
2743 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2744 Constant *Anded = And(CRHS, C2);
2745 if (Anded == CRHS) {
2746 // See if all bits from the first bit set in the Add RHS up are included
2747 // in the mask. First, get the rightmost bit.
2748 const APInt& AddRHSV = CRHS->getValue();
2750 // Form a mask of all bits from the lowest bit added through the top.
2751 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2753 // See if the and mask includes all of these bits.
2754 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2756 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2757 // Okay, the xform is safe. Insert the new add pronto.
2758 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2759 LHS->getName()), I);
2760 return BinaryOperator::CreateAnd(NewAdd, C2);
2765 // Try to fold constant add into select arguments.
2766 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2767 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2771 // add (cast *A to intptrtype) B ->
2772 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2774 CastInst *CI = dyn_cast<CastInst>(LHS);
2777 CI = dyn_cast<CastInst>(RHS);
2780 if (CI && CI->getType()->isSized() &&
2781 (CI->getType()->getPrimitiveSizeInBits() ==
2782 TD->getIntPtrType()->getPrimitiveSizeInBits())
2783 && isa<PointerType>(CI->getOperand(0)->getType())) {
2785 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2786 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2787 PointerType::get(Type::Int8Ty, AS), I);
2788 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2789 return new PtrToIntInst(I2, CI->getType());
2793 // add (select X 0 (sub n A)) A --> select X A n
2795 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2798 SI = dyn_cast<SelectInst>(RHS);
2801 if (SI && SI->hasOneUse()) {
2802 Value *TV = SI->getTrueValue();
2803 Value *FV = SI->getFalseValue();
2806 // Can we fold the add into the argument of the select?
2807 // We check both true and false select arguments for a matching subtract.
2808 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2809 A == Other) // Fold the add into the true select value.
2810 return SelectInst::Create(SI->getCondition(), N, A);
2811 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2812 A == Other) // Fold the add into the false select value.
2813 return SelectInst::Create(SI->getCondition(), A, N);
2817 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2818 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2819 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2820 return ReplaceInstUsesWith(I, LHS);
2822 // Check for (add (sext x), y), see if we can merge this into an
2823 // integer add followed by a sext.
2824 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2825 // (add (sext x), cst) --> (sext (add x, cst'))
2826 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2828 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2829 if (LHSConv->hasOneUse() &&
2830 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2831 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2832 // Insert the new, smaller add.
2833 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2835 InsertNewInstBefore(NewAdd, I);
2836 return new SExtInst(NewAdd, I.getType());
2840 // (add (sext x), (sext y)) --> (sext (add int x, y))
2841 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2842 // Only do this if x/y have the same type, if at last one of them has a
2843 // single use (so we don't increase the number of sexts), and if the
2844 // integer add will not overflow.
2845 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2846 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2847 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2848 RHSConv->getOperand(0))) {
2849 // Insert the new integer add.
2850 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2851 RHSConv->getOperand(0),
2853 InsertNewInstBefore(NewAdd, I);
2854 return new SExtInst(NewAdd, I.getType());
2859 // Check for (add double (sitofp x), y), see if we can merge this into an
2860 // integer add followed by a promotion.
2861 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2862 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2863 // ... if the constant fits in the integer value. This is useful for things
2864 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2865 // requires a constant pool load, and generally allows the add to be better
2867 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2869 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2870 if (LHSConv->hasOneUse() &&
2871 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2872 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2873 // Insert the new integer add.
2874 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2876 InsertNewInstBefore(NewAdd, I);
2877 return new SIToFPInst(NewAdd, I.getType());
2881 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2882 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2883 // Only do this if x/y have the same type, if at last one of them has a
2884 // single use (so we don't increase the number of int->fp conversions),
2885 // and if the integer add will not overflow.
2886 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2887 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2888 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2889 RHSConv->getOperand(0))) {
2890 // Insert the new integer add.
2891 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2892 RHSConv->getOperand(0),
2894 InsertNewInstBefore(NewAdd, I);
2895 return new SIToFPInst(NewAdd, I.getType());
2900 return Changed ? &I : 0;
2903 // isSignBit - Return true if the value represented by the constant only has the
2904 // highest order bit set.
2905 static bool isSignBit(ConstantInt *CI) {
2906 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2907 return CI->getValue() == APInt::getSignBit(NumBits);
2910 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2911 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2913 if (Op0 == Op1) // sub X, X -> 0
2914 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2916 // If this is a 'B = x-(-A)', change to B = x+A...
2917 if (Value *V = dyn_castNegVal(Op1))
2918 return BinaryOperator::CreateAdd(Op0, V);
2920 if (isa<UndefValue>(Op0))
2921 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2922 if (isa<UndefValue>(Op1))
2923 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2925 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2926 // Replace (-1 - A) with (~A)...
2927 if (C->isAllOnesValue())
2928 return BinaryOperator::CreateNot(Op1);
2930 // C - ~X == X + (1+C)
2932 if (match(Op1, m_Not(m_Value(X))))
2933 return BinaryOperator::CreateAdd(X, AddOne(C));
2935 // -(X >>u 31) -> (X >>s 31)
2936 // -(X >>s 31) -> (X >>u 31)
2938 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2939 if (SI->getOpcode() == Instruction::LShr) {
2940 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2941 // Check to see if we are shifting out everything but the sign bit.
2942 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2943 SI->getType()->getPrimitiveSizeInBits()-1) {
2944 // Ok, the transformation is safe. Insert AShr.
2945 return BinaryOperator::Create(Instruction::AShr,
2946 SI->getOperand(0), CU, SI->getName());
2950 else if (SI->getOpcode() == Instruction::AShr) {
2951 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2952 // Check to see if we are shifting out everything but the sign bit.
2953 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2954 SI->getType()->getPrimitiveSizeInBits()-1) {
2955 // Ok, the transformation is safe. Insert LShr.
2956 return BinaryOperator::CreateLShr(
2957 SI->getOperand(0), CU, SI->getName());
2964 // Try to fold constant sub into select arguments.
2965 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2966 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2969 if (isa<PHINode>(Op0))
2970 if (Instruction *NV = FoldOpIntoPhi(I))
2974 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2975 if (Op1I->getOpcode() == Instruction::Add &&
2976 !Op0->getType()->isFPOrFPVector()) {
2977 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2978 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2979 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2980 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2981 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2982 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2983 // C1-(X+C2) --> (C1-C2)-X
2984 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2985 Op1I->getOperand(0));
2989 if (Op1I->hasOneUse()) {
2990 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2991 // is not used by anyone else...
2993 if (Op1I->getOpcode() == Instruction::Sub &&
2994 !Op1I->getType()->isFPOrFPVector()) {
2995 // Swap the two operands of the subexpr...
2996 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2997 Op1I->setOperand(0, IIOp1);
2998 Op1I->setOperand(1, IIOp0);
3000 // Create the new top level add instruction...
3001 return BinaryOperator::CreateAdd(Op0, Op1);
3004 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
3006 if (Op1I->getOpcode() == Instruction::And &&
3007 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
3008 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
3011 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
3012 return BinaryOperator::CreateAnd(Op0, NewNot);
3015 // 0 - (X sdiv C) -> (X sdiv -C)
3016 if (Op1I->getOpcode() == Instruction::SDiv)
3017 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
3019 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
3020 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
3021 ConstantExpr::getNeg(DivRHS));
3023 // X - X*C --> X * (1-C)
3024 ConstantInt *C2 = 0;
3025 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
3026 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
3027 return BinaryOperator::CreateMul(Op0, CP1);
3030 // X - ((X / Y) * Y) --> X % Y
3031 if (Op1I->getOpcode() == Instruction::Mul)
3032 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
3033 if (Op0 == I->getOperand(0) &&
3034 Op1I->getOperand(1) == I->getOperand(1)) {
3035 if (I->getOpcode() == Instruction::SDiv)
3036 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
3037 if (I->getOpcode() == Instruction::UDiv)
3038 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
3043 if (!Op0->getType()->isFPOrFPVector())
3044 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3045 if (Op0I->getOpcode() == Instruction::Add) {
3046 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
3047 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3048 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
3049 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3050 } else if (Op0I->getOpcode() == Instruction::Sub) {
3051 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
3052 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
3057 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
3058 if (X == Op1) // X*C - X --> X * (C-1)
3059 return BinaryOperator::CreateMul(Op1, SubOne(C1));
3061 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
3062 if (X == dyn_castFoldableMul(Op1, C2))
3063 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
3068 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
3069 /// comparison only checks the sign bit. If it only checks the sign bit, set
3070 /// TrueIfSigned if the result of the comparison is true when the input value is
3072 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
3073 bool &TrueIfSigned) {
3075 case ICmpInst::ICMP_SLT: // True if LHS s< 0
3076 TrueIfSigned = true;
3077 return RHS->isZero();
3078 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
3079 TrueIfSigned = true;
3080 return RHS->isAllOnesValue();
3081 case ICmpInst::ICMP_SGT: // True if LHS s> -1
3082 TrueIfSigned = false;
3083 return RHS->isAllOnesValue();
3084 case ICmpInst::ICMP_UGT:
3085 // True if LHS u> RHS and RHS == high-bit-mask - 1
3086 TrueIfSigned = true;
3087 return RHS->getValue() ==
3088 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
3089 case ICmpInst::ICMP_UGE:
3090 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
3091 TrueIfSigned = true;
3092 return RHS->getValue() ==
3093 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
3099 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
3100 bool Changed = SimplifyCommutative(I);
3101 Value *Op0 = I.getOperand(0);
3103 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
3104 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3106 // Simplify mul instructions with a constant RHS...
3107 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
3108 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
3110 // ((X << C1)*C2) == (X * (C2 << C1))
3111 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
3112 if (SI->getOpcode() == Instruction::Shl)
3113 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
3114 return BinaryOperator::CreateMul(SI->getOperand(0),
3115 ConstantExpr::getShl(CI, ShOp));
3118 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
3119 if (CI->equalsInt(1)) // X * 1 == X
3120 return ReplaceInstUsesWith(I, Op0);
3121 if (CI->isAllOnesValue()) // X * -1 == 0 - X
3122 return BinaryOperator::CreateNeg(Op0, I.getName());
3124 const APInt& Val = cast<ConstantInt>(CI)->getValue();
3125 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
3126 return BinaryOperator::CreateShl(Op0,
3127 ConstantInt::get(Op0->getType(), Val.logBase2()));
3129 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
3130 if (Op1F->isNullValue())
3131 return ReplaceInstUsesWith(I, Op1);
3133 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
3134 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
3135 // We need a better interface for long double here.
3136 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
3137 if (Op1F->isExactlyValue(1.0))
3138 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
3141 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3142 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
3143 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
3144 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
3145 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
3147 InsertNewInstBefore(Add, I);
3148 Value *C1C2 = ConstantExpr::getMul(Op1,
3149 cast<Constant>(Op0I->getOperand(1)));
3150 return BinaryOperator::CreateAdd(Add, C1C2);
3154 // Try to fold constant mul into select arguments.
3155 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3156 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3159 if (isa<PHINode>(Op0))
3160 if (Instruction *NV = FoldOpIntoPhi(I))
3164 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
3165 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
3166 return BinaryOperator::CreateMul(Op0v, Op1v);
3168 // If one of the operands of the multiply is a cast from a boolean value, then
3169 // we know the bool is either zero or one, so this is a 'masking' multiply.
3170 // See if we can simplify things based on how the boolean was originally
3172 CastInst *BoolCast = 0;
3173 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
3174 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3177 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
3178 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3181 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
3182 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
3183 const Type *SCOpTy = SCIOp0->getType();
3186 // If the icmp is true iff the sign bit of X is set, then convert this
3187 // multiply into a shift/and combination.
3188 if (isa<ConstantInt>(SCIOp1) &&
3189 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
3191 // Shift the X value right to turn it into "all signbits".
3192 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
3193 SCOpTy->getPrimitiveSizeInBits()-1);
3195 InsertNewInstBefore(
3196 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
3197 BoolCast->getOperand(0)->getName()+
3200 // If the multiply type is not the same as the source type, sign extend
3201 // or truncate to the multiply type.
3202 if (I.getType() != V->getType()) {
3203 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
3204 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
3205 Instruction::CastOps opcode =
3206 (SrcBits == DstBits ? Instruction::BitCast :
3207 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
3208 V = InsertCastBefore(opcode, V, I.getType(), I);
3211 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
3212 return BinaryOperator::CreateAnd(V, OtherOp);
3217 return Changed ? &I : 0;
3220 /// This function implements the transforms on div instructions that work
3221 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
3222 /// used by the visitors to those instructions.
3223 /// @brief Transforms common to all three div instructions
3224 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
3225 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3227 // undef / X -> 0 for integer.
3228 // undef / X -> undef for FP (the undef could be a snan).
3229 if (isa<UndefValue>(Op0)) {
3230 if (Op0->getType()->isFPOrFPVector())
3231 return ReplaceInstUsesWith(I, Op0);
3232 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3235 // X / undef -> undef
3236 if (isa<UndefValue>(Op1))
3237 return ReplaceInstUsesWith(I, Op1);
3239 // Handle cases involving: [su]div X, (select Cond, Y, Z)
3240 // This does not apply for fdiv.
3241 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3242 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
3243 // the same basic block, then we replace the select with Y, and the
3244 // condition of the select with false (if the cond value is in the same BB).
3245 // If the select has uses other than the div, this allows them to be
3246 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
3247 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
3248 if (ST->isNullValue()) {
3249 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3250 if (CondI && CondI->getParent() == I.getParent())
3251 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3252 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3253 I.setOperand(1, SI->getOperand(2));
3255 UpdateValueUsesWith(SI, SI->getOperand(2));
3259 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
3260 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
3261 if (ST->isNullValue()) {
3262 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3263 if (CondI && CondI->getParent() == I.getParent())
3264 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3265 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3266 I.setOperand(1, SI->getOperand(1));
3268 UpdateValueUsesWith(SI, SI->getOperand(1));
3276 /// This function implements the transforms common to both integer division
3277 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3278 /// division instructions.
3279 /// @brief Common integer divide transforms
3280 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3281 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3283 // (sdiv X, X) --> 1 (udiv X, X) --> 1
3285 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
3286 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
3287 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
3288 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
3291 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
3292 return ReplaceInstUsesWith(I, CI);
3295 if (Instruction *Common = commonDivTransforms(I))
3298 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3300 if (RHS->equalsInt(1))
3301 return ReplaceInstUsesWith(I, Op0);
3303 // (X / C1) / C2 -> X / (C1*C2)
3304 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3305 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3306 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3307 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
3308 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3310 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3311 Multiply(RHS, LHSRHS));
3314 if (!RHS->isZero()) { // avoid X udiv 0
3315 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3316 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3318 if (isa<PHINode>(Op0))
3319 if (Instruction *NV = FoldOpIntoPhi(I))
3324 // 0 / X == 0, we don't need to preserve faults!
3325 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3326 if (LHS->equalsInt(0))
3327 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3332 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3333 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3335 // Handle the integer div common cases
3336 if (Instruction *Common = commonIDivTransforms(I))
3339 // X udiv C^2 -> X >> C
3340 // Check to see if this is an unsigned division with an exact power of 2,
3341 // if so, convert to a right shift.
3342 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3343 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3344 return BinaryOperator::CreateLShr(Op0,
3345 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3348 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3349 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3350 if (RHSI->getOpcode() == Instruction::Shl &&
3351 isa<ConstantInt>(RHSI->getOperand(0))) {
3352 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3353 if (C1.isPowerOf2()) {
3354 Value *N = RHSI->getOperand(1);
3355 const Type *NTy = N->getType();
3356 if (uint32_t C2 = C1.logBase2()) {
3357 Constant *C2V = ConstantInt::get(NTy, C2);
3358 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3360 return BinaryOperator::CreateLShr(Op0, N);
3365 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3366 // where C1&C2 are powers of two.
3367 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3368 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3369 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3370 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3371 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3372 // Compute the shift amounts
3373 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3374 // Construct the "on true" case of the select
3375 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3376 Instruction *TSI = BinaryOperator::CreateLShr(
3377 Op0, TC, SI->getName()+".t");
3378 TSI = InsertNewInstBefore(TSI, I);
3380 // Construct the "on false" case of the select
3381 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3382 Instruction *FSI = BinaryOperator::CreateLShr(
3383 Op0, FC, SI->getName()+".f");
3384 FSI = InsertNewInstBefore(FSI, I);
3386 // construct the select instruction and return it.
3387 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3393 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3394 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3396 // Handle the integer div common cases
3397 if (Instruction *Common = commonIDivTransforms(I))
3400 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3402 if (RHS->isAllOnesValue())
3403 return BinaryOperator::CreateNeg(Op0);
3406 if (Value *LHSNeg = dyn_castNegVal(Op0))
3407 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
3410 // If the sign bits of both operands are zero (i.e. we can prove they are
3411 // unsigned inputs), turn this into a udiv.
3412 if (I.getType()->isInteger()) {
3413 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3414 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3415 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3416 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3423 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3424 return commonDivTransforms(I);
3427 /// This function implements the transforms on rem instructions that work
3428 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3429 /// is used by the visitors to those instructions.
3430 /// @brief Transforms common to all three rem instructions
3431 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3432 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3434 // 0 % X == 0 for integer, we don't need to preserve faults!
3435 if (Constant *LHS = dyn_cast<Constant>(Op0))
3436 if (LHS->isNullValue())
3437 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3439 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3440 if (I.getType()->isFPOrFPVector())
3441 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3442 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3444 if (isa<UndefValue>(Op1))
3445 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3447 // Handle cases involving: rem X, (select Cond, Y, Z)
3448 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3449 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
3450 // the same basic block, then we replace the select with Y, and the
3451 // condition of the select with false (if the cond value is in the same
3452 // BB). If the select has uses other than the div, this allows them to be
3454 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3455 if (ST->isNullValue()) {
3456 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3457 if (CondI && CondI->getParent() == I.getParent())
3458 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3459 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3460 I.setOperand(1, SI->getOperand(2));
3462 UpdateValueUsesWith(SI, SI->getOperand(2));
3465 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
3466 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3467 if (ST->isNullValue()) {
3468 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3469 if (CondI && CondI->getParent() == I.getParent())
3470 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3471 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3472 I.setOperand(1, SI->getOperand(1));
3474 UpdateValueUsesWith(SI, SI->getOperand(1));
3482 /// This function implements the transforms common to both integer remainder
3483 /// instructions (urem and srem). It is called by the visitors to those integer
3484 /// remainder instructions.
3485 /// @brief Common integer remainder transforms
3486 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3487 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3489 if (Instruction *common = commonRemTransforms(I))
3492 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3493 // X % 0 == undef, we don't need to preserve faults!
3494 if (RHS->equalsInt(0))
3495 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3497 if (RHS->equalsInt(1)) // X % 1 == 0
3498 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3500 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3501 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3502 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3504 } else if (isa<PHINode>(Op0I)) {
3505 if (Instruction *NV = FoldOpIntoPhi(I))
3509 // See if we can fold away this rem instruction.
3510 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3511 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3512 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3513 KnownZero, KnownOne))
3521 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3522 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3524 if (Instruction *common = commonIRemTransforms(I))
3527 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3528 // X urem C^2 -> X and C
3529 // Check to see if this is an unsigned remainder with an exact power of 2,
3530 // if so, convert to a bitwise and.
3531 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3532 if (C->getValue().isPowerOf2())
3533 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3536 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3537 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3538 if (RHSI->getOpcode() == Instruction::Shl &&
3539 isa<ConstantInt>(RHSI->getOperand(0))) {
3540 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3541 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3542 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3544 return BinaryOperator::CreateAnd(Op0, Add);
3549 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3550 // where C1&C2 are powers of two.
3551 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3552 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3553 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3554 // STO == 0 and SFO == 0 handled above.
3555 if ((STO->getValue().isPowerOf2()) &&
3556 (SFO->getValue().isPowerOf2())) {
3557 Value *TrueAnd = InsertNewInstBefore(
3558 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3559 Value *FalseAnd = InsertNewInstBefore(
3560 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3561 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3569 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3570 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3572 // Handle the integer rem common cases
3573 if (Instruction *common = commonIRemTransforms(I))
3576 if (Value *RHSNeg = dyn_castNegVal(Op1))
3577 if (!isa<ConstantInt>(RHSNeg) ||
3578 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
3580 AddUsesToWorkList(I);
3581 I.setOperand(1, RHSNeg);
3585 // If the sign bits of both operands are zero (i.e. we can prove they are
3586 // unsigned inputs), turn this into a urem.
3587 if (I.getType()->isInteger()) {
3588 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3589 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3590 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3591 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3598 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3599 return commonRemTransforms(I);
3602 // isMaxValueMinusOne - return true if this is Max-1
3603 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
3604 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3606 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
3607 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
3610 // isMinValuePlusOne - return true if this is Min+1
3611 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3613 return C->getValue() == 1; // unsigned
3615 // Calculate 1111111111000000000000
3616 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3617 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
3620 // isOneBitSet - Return true if there is exactly one bit set in the specified
3622 static bool isOneBitSet(const ConstantInt *CI) {
3623 return CI->getValue().isPowerOf2();
3626 // isHighOnes - Return true if the constant is of the form 1+0+.
3627 // This is the same as lowones(~X).
3628 static bool isHighOnes(const ConstantInt *CI) {
3629 return (~CI->getValue() + 1).isPowerOf2();
3632 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3633 /// are carefully arranged to allow folding of expressions such as:
3635 /// (A < B) | (A > B) --> (A != B)
3637 /// Note that this is only valid if the first and second predicates have the
3638 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3640 /// Three bits are used to represent the condition, as follows:
3645 /// <=> Value Definition
3646 /// 000 0 Always false
3653 /// 111 7 Always true
3655 static unsigned getICmpCode(const ICmpInst *ICI) {
3656 switch (ICI->getPredicate()) {
3658 case ICmpInst::ICMP_UGT: return 1; // 001
3659 case ICmpInst::ICMP_SGT: return 1; // 001
3660 case ICmpInst::ICMP_EQ: return 2; // 010
3661 case ICmpInst::ICMP_UGE: return 3; // 011
3662 case ICmpInst::ICMP_SGE: return 3; // 011
3663 case ICmpInst::ICMP_ULT: return 4; // 100
3664 case ICmpInst::ICMP_SLT: return 4; // 100
3665 case ICmpInst::ICMP_NE: return 5; // 101
3666 case ICmpInst::ICMP_ULE: return 6; // 110
3667 case ICmpInst::ICMP_SLE: return 6; // 110
3670 assert(0 && "Invalid ICmp predicate!");
3675 /// getICmpValue - This is the complement of getICmpCode, which turns an
3676 /// opcode and two operands into either a constant true or false, or a brand
3677 /// new ICmp instruction. The sign is passed in to determine which kind
3678 /// of predicate to use in new icmp instructions.
3679 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3681 default: assert(0 && "Illegal ICmp code!");
3682 case 0: return ConstantInt::getFalse();
3685 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3687 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3688 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3691 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3693 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3696 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3698 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3699 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3702 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3704 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3705 case 7: return ConstantInt::getTrue();
3709 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3710 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3711 (ICmpInst::isSignedPredicate(p1) &&
3712 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3713 (ICmpInst::isSignedPredicate(p2) &&
3714 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3718 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3719 struct FoldICmpLogical {
3722 ICmpInst::Predicate pred;
3723 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3724 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3725 pred(ICI->getPredicate()) {}
3726 bool shouldApply(Value *V) const {
3727 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3728 if (PredicatesFoldable(pred, ICI->getPredicate()))
3729 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3730 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3733 Instruction *apply(Instruction &Log) const {
3734 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3735 if (ICI->getOperand(0) != LHS) {
3736 assert(ICI->getOperand(1) == LHS);
3737 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3740 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3741 unsigned LHSCode = getICmpCode(ICI);
3742 unsigned RHSCode = getICmpCode(RHSICI);
3744 switch (Log.getOpcode()) {
3745 case Instruction::And: Code = LHSCode & RHSCode; break;
3746 case Instruction::Or: Code = LHSCode | RHSCode; break;
3747 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3748 default: assert(0 && "Illegal logical opcode!"); return 0;
3751 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3752 ICmpInst::isSignedPredicate(ICI->getPredicate());
3754 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3755 if (Instruction *I = dyn_cast<Instruction>(RV))
3757 // Otherwise, it's a constant boolean value...
3758 return IC.ReplaceInstUsesWith(Log, RV);
3761 } // end anonymous namespace
3763 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3764 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3765 // guaranteed to be a binary operator.
3766 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3768 ConstantInt *AndRHS,
3769 BinaryOperator &TheAnd) {
3770 Value *X = Op->getOperand(0);
3771 Constant *Together = 0;
3773 Together = And(AndRHS, OpRHS);
3775 switch (Op->getOpcode()) {
3776 case Instruction::Xor:
3777 if (Op->hasOneUse()) {
3778 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3779 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3780 InsertNewInstBefore(And, TheAnd);
3782 return BinaryOperator::CreateXor(And, Together);
3785 case Instruction::Or:
3786 if (Together == AndRHS) // (X | C) & C --> C
3787 return ReplaceInstUsesWith(TheAnd, AndRHS);
3789 if (Op->hasOneUse() && Together != OpRHS) {
3790 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3791 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3792 InsertNewInstBefore(Or, TheAnd);
3794 return BinaryOperator::CreateAnd(Or, AndRHS);
3797 case Instruction::Add:
3798 if (Op->hasOneUse()) {
3799 // Adding a one to a single bit bit-field should be turned into an XOR
3800 // of the bit. First thing to check is to see if this AND is with a
3801 // single bit constant.
3802 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3804 // If there is only one bit set...
3805 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3806 // Ok, at this point, we know that we are masking the result of the
3807 // ADD down to exactly one bit. If the constant we are adding has
3808 // no bits set below this bit, then we can eliminate the ADD.
3809 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3811 // Check to see if any bits below the one bit set in AndRHSV are set.
3812 if ((AddRHS & (AndRHSV-1)) == 0) {
3813 // If not, the only thing that can effect the output of the AND is
3814 // the bit specified by AndRHSV. If that bit is set, the effect of
3815 // the XOR is to toggle the bit. If it is clear, then the ADD has
3817 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3818 TheAnd.setOperand(0, X);
3821 // Pull the XOR out of the AND.
3822 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3823 InsertNewInstBefore(NewAnd, TheAnd);
3824 NewAnd->takeName(Op);
3825 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3832 case Instruction::Shl: {
3833 // We know that the AND will not produce any of the bits shifted in, so if
3834 // the anded constant includes them, clear them now!
3836 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3837 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3838 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3839 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3841 if (CI->getValue() == ShlMask) {
3842 // Masking out bits that the shift already masks
3843 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3844 } else if (CI != AndRHS) { // Reducing bits set in and.
3845 TheAnd.setOperand(1, CI);
3850 case Instruction::LShr:
3852 // We know that the AND will not produce any of the bits shifted in, so if
3853 // the anded constant includes them, clear them now! This only applies to
3854 // unsigned shifts, because a signed shr may bring in set bits!
3856 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3857 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3858 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3859 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3861 if (CI->getValue() == ShrMask) {
3862 // Masking out bits that the shift already masks.
3863 return ReplaceInstUsesWith(TheAnd, Op);
3864 } else if (CI != AndRHS) {
3865 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3870 case Instruction::AShr:
3872 // See if this is shifting in some sign extension, then masking it out
3874 if (Op->hasOneUse()) {
3875 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3876 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3877 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3878 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3879 if (C == AndRHS) { // Masking out bits shifted in.
3880 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3881 // Make the argument unsigned.
3882 Value *ShVal = Op->getOperand(0);
3883 ShVal = InsertNewInstBefore(
3884 BinaryOperator::CreateLShr(ShVal, OpRHS,
3885 Op->getName()), TheAnd);
3886 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3895 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3896 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3897 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3898 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3899 /// insert new instructions.
3900 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3901 bool isSigned, bool Inside,
3903 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3904 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3905 "Lo is not <= Hi in range emission code!");
3908 if (Lo == Hi) // Trivially false.
3909 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3911 // V >= Min && V < Hi --> V < Hi
3912 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3913 ICmpInst::Predicate pred = (isSigned ?
3914 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3915 return new ICmpInst(pred, V, Hi);
3918 // Emit V-Lo <u Hi-Lo
3919 Constant *NegLo = ConstantExpr::getNeg(Lo);
3920 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3921 InsertNewInstBefore(Add, IB);
3922 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3923 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3926 if (Lo == Hi) // Trivially true.
3927 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3929 // V < Min || V >= Hi -> V > Hi-1
3930 Hi = SubOne(cast<ConstantInt>(Hi));
3931 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3932 ICmpInst::Predicate pred = (isSigned ?
3933 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3934 return new ICmpInst(pred, V, Hi);
3937 // Emit V-Lo >u Hi-1-Lo
3938 // Note that Hi has already had one subtracted from it, above.
3939 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3940 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3941 InsertNewInstBefore(Add, IB);
3942 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3943 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3946 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3947 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3948 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3949 // not, since all 1s are not contiguous.
3950 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3951 const APInt& V = Val->getValue();
3952 uint32_t BitWidth = Val->getType()->getBitWidth();
3953 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3955 // look for the first zero bit after the run of ones
3956 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3957 // look for the first non-zero bit
3958 ME = V.getActiveBits();
3962 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3963 /// where isSub determines whether the operator is a sub. If we can fold one of
3964 /// the following xforms:
3966 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3967 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3968 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3970 /// return (A +/- B).
3972 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3973 ConstantInt *Mask, bool isSub,
3975 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3976 if (!LHSI || LHSI->getNumOperands() != 2 ||
3977 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3979 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3981 switch (LHSI->getOpcode()) {
3983 case Instruction::And:
3984 if (And(N, Mask) == Mask) {
3985 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3986 if ((Mask->getValue().countLeadingZeros() +
3987 Mask->getValue().countPopulation()) ==
3988 Mask->getValue().getBitWidth())
3991 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3992 // part, we don't need any explicit masks to take them out of A. If that
3993 // is all N is, ignore it.
3994 uint32_t MB = 0, ME = 0;
3995 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3996 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3997 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3998 if (MaskedValueIsZero(RHS, Mask))
4003 case Instruction::Or:
4004 case Instruction::Xor:
4005 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
4006 if ((Mask->getValue().countLeadingZeros() +
4007 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
4008 && And(N, Mask)->isZero())
4015 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
4017 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
4018 return InsertNewInstBefore(New, I);
4021 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4022 bool Changed = SimplifyCommutative(I);
4023 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4025 if (isa<UndefValue>(Op1)) // X & undef -> 0
4026 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4030 return ReplaceInstUsesWith(I, Op1);
4032 // See if we can simplify any instructions used by the instruction whose sole
4033 // purpose is to compute bits we don't care about.
4034 if (!isa<VectorType>(I.getType())) {
4035 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4036 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4037 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4038 KnownZero, KnownOne))
4041 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4042 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4043 return ReplaceInstUsesWith(I, I.getOperand(0));
4044 } else if (isa<ConstantAggregateZero>(Op1)) {
4045 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4049 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4050 const APInt& AndRHSMask = AndRHS->getValue();
4051 APInt NotAndRHS(~AndRHSMask);
4053 // Optimize a variety of ((val OP C1) & C2) combinations...
4054 if (isa<BinaryOperator>(Op0)) {
4055 Instruction *Op0I = cast<Instruction>(Op0);
4056 Value *Op0LHS = Op0I->getOperand(0);
4057 Value *Op0RHS = Op0I->getOperand(1);
4058 switch (Op0I->getOpcode()) {
4059 case Instruction::Xor:
4060 case Instruction::Or:
4061 // If the mask is only needed on one incoming arm, push it up.
4062 if (Op0I->hasOneUse()) {
4063 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4064 // Not masking anything out for the LHS, move to RHS.
4065 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4066 Op0RHS->getName()+".masked");
4067 InsertNewInstBefore(NewRHS, I);
4068 return BinaryOperator::Create(
4069 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4071 if (!isa<Constant>(Op0RHS) &&
4072 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4073 // Not masking anything out for the RHS, move to LHS.
4074 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4075 Op0LHS->getName()+".masked");
4076 InsertNewInstBefore(NewLHS, I);
4077 return BinaryOperator::Create(
4078 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4083 case Instruction::Add:
4084 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4085 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4086 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4087 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4088 return BinaryOperator::CreateAnd(V, AndRHS);
4089 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4090 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4093 case Instruction::Sub:
4094 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4095 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4096 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4097 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4098 return BinaryOperator::CreateAnd(V, AndRHS);
4102 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4103 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4105 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4106 // If this is an integer truncation or change from signed-to-unsigned, and
4107 // if the source is an and/or with immediate, transform it. This
4108 // frequently occurs for bitfield accesses.
4109 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4110 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4111 CastOp->getNumOperands() == 2)
4112 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4113 if (CastOp->getOpcode() == Instruction::And) {
4114 // Change: and (cast (and X, C1) to T), C2
4115 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4116 // This will fold the two constants together, which may allow
4117 // other simplifications.
4118 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4119 CastOp->getOperand(0), I.getType(),
4120 CastOp->getName()+".shrunk");
4121 NewCast = InsertNewInstBefore(NewCast, I);
4122 // trunc_or_bitcast(C1)&C2
4123 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4124 C3 = ConstantExpr::getAnd(C3, AndRHS);
4125 return BinaryOperator::CreateAnd(NewCast, C3);
4126 } else if (CastOp->getOpcode() == Instruction::Or) {
4127 // Change: and (cast (or X, C1) to T), C2
4128 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4129 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4130 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
4131 return ReplaceInstUsesWith(I, AndRHS);
4137 // Try to fold constant and into select arguments.
4138 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4139 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4141 if (isa<PHINode>(Op0))
4142 if (Instruction *NV = FoldOpIntoPhi(I))
4146 Value *Op0NotVal = dyn_castNotVal(Op0);
4147 Value *Op1NotVal = dyn_castNotVal(Op1);
4149 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4150 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4152 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4153 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4154 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4155 I.getName()+".demorgan");
4156 InsertNewInstBefore(Or, I);
4157 return BinaryOperator::CreateNot(Or);
4161 Value *A = 0, *B = 0, *C = 0, *D = 0;
4162 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4163 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4164 return ReplaceInstUsesWith(I, Op1);
4166 // (A|B) & ~(A&B) -> A^B
4167 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4168 if ((A == C && B == D) || (A == D && B == C))
4169 return BinaryOperator::CreateXor(A, B);
4173 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4174 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4175 return ReplaceInstUsesWith(I, Op0);
4177 // ~(A&B) & (A|B) -> A^B
4178 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4179 if ((A == C && B == D) || (A == D && B == C))
4180 return BinaryOperator::CreateXor(A, B);
4184 if (Op0->hasOneUse() &&
4185 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4186 if (A == Op1) { // (A^B)&A -> A&(A^B)
4187 I.swapOperands(); // Simplify below
4188 std::swap(Op0, Op1);
4189 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4190 cast<BinaryOperator>(Op0)->swapOperands();
4191 I.swapOperands(); // Simplify below
4192 std::swap(Op0, Op1);
4195 if (Op1->hasOneUse() &&
4196 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4197 if (B == Op0) { // B&(A^B) -> B&(B^A)
4198 cast<BinaryOperator>(Op1)->swapOperands();
4201 if (A == Op0) { // A&(A^B) -> A & ~B
4202 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4203 InsertNewInstBefore(NotB, I);
4204 return BinaryOperator::CreateAnd(A, NotB);
4209 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4210 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4211 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4214 Value *LHSVal, *RHSVal;
4215 ConstantInt *LHSCst, *RHSCst;
4216 ICmpInst::Predicate LHSCC, RHSCC;
4217 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4218 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4219 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
4220 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
4221 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4222 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4223 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4224 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4226 // Don't try to fold ICMP_SLT + ICMP_ULT.
4227 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
4228 ICmpInst::isSignedPredicate(LHSCC) ==
4229 ICmpInst::isSignedPredicate(RHSCC))) {
4230 // Ensure that the larger constant is on the RHS.
4231 ICmpInst::Predicate GT;
4232 if (ICmpInst::isSignedPredicate(LHSCC) ||
4233 (ICmpInst::isEquality(LHSCC) &&
4234 ICmpInst::isSignedPredicate(RHSCC)))
4235 GT = ICmpInst::ICMP_SGT;
4237 GT = ICmpInst::ICMP_UGT;
4239 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
4240 ICmpInst *LHS = cast<ICmpInst>(Op0);
4241 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
4242 std::swap(LHS, RHS);
4243 std::swap(LHSCst, RHSCst);
4244 std::swap(LHSCC, RHSCC);
4247 // At this point, we know we have have two icmp instructions
4248 // comparing a value against two constants and and'ing the result
4249 // together. Because of the above check, we know that we only have
4250 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
4251 // (from the FoldICmpLogical check above), that the two constants
4252 // are not equal and that the larger constant is on the RHS
4253 assert(LHSCst != RHSCst && "Compares not folded above?");
4256 default: assert(0 && "Unknown integer condition code!");
4257 case ICmpInst::ICMP_EQ:
4259 default: assert(0 && "Unknown integer condition code!");
4260 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
4261 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
4262 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
4263 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4264 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
4265 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
4266 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
4267 return ReplaceInstUsesWith(I, LHS);
4269 case ICmpInst::ICMP_NE:
4271 default: assert(0 && "Unknown integer condition code!");
4272 case ICmpInst::ICMP_ULT:
4273 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
4274 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
4275 break; // (X != 13 & X u< 15) -> no change
4276 case ICmpInst::ICMP_SLT:
4277 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
4278 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
4279 break; // (X != 13 & X s< 15) -> no change
4280 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
4281 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
4282 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
4283 return ReplaceInstUsesWith(I, RHS);
4284 case ICmpInst::ICMP_NE:
4285 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
4286 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4287 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4288 LHSVal->getName()+".off");
4289 InsertNewInstBefore(Add, I);
4290 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
4291 ConstantInt::get(Add->getType(), 1));
4293 break; // (X != 13 & X != 15) -> no change
4296 case ICmpInst::ICMP_ULT:
4298 default: assert(0 && "Unknown integer condition code!");
4299 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
4300 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
4301 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4302 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
4304 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
4305 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
4306 return ReplaceInstUsesWith(I, LHS);
4307 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
4311 case ICmpInst::ICMP_SLT:
4313 default: assert(0 && "Unknown integer condition code!");
4314 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
4315 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
4316 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4317 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
4319 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
4320 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
4321 return ReplaceInstUsesWith(I, LHS);
4322 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
4326 case ICmpInst::ICMP_UGT:
4328 default: assert(0 && "Unknown integer condition code!");
4329 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
4330 return ReplaceInstUsesWith(I, LHS);
4331 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
4332 return ReplaceInstUsesWith(I, RHS);
4333 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
4335 case ICmpInst::ICMP_NE:
4336 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
4337 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4338 break; // (X u> 13 & X != 15) -> no change
4339 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
4340 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
4342 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4346 case ICmpInst::ICMP_SGT:
4348 default: assert(0 && "Unknown integer condition code!");
4349 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
4350 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4351 return ReplaceInstUsesWith(I, RHS);
4352 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4354 case ICmpInst::ICMP_NE:
4355 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4356 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4357 break; // (X s> 13 & X != 15) -> no change
4358 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
4359 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
4361 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4369 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4370 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4371 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4372 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4373 const Type *SrcTy = Op0C->getOperand(0)->getType();
4374 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4375 // Only do this if the casts both really cause code to be generated.
4376 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4378 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4380 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4381 Op1C->getOperand(0),
4383 InsertNewInstBefore(NewOp, I);
4384 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4388 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4389 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4390 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4391 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4392 SI0->getOperand(1) == SI1->getOperand(1) &&
4393 (SI0->hasOneUse() || SI1->hasOneUse())) {
4394 Instruction *NewOp =
4395 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4397 SI0->getName()), I);
4398 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4399 SI1->getOperand(1));
4403 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4404 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4405 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4406 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4407 RHS->getPredicate() == FCmpInst::FCMP_ORD)
4408 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4409 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4410 // If either of the constants are nans, then the whole thing returns
4412 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4413 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4414 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4415 RHS->getOperand(0));
4420 return Changed ? &I : 0;
4423 /// CollectBSwapParts - Look to see if the specified value defines a single byte
4424 /// in the result. If it does, and if the specified byte hasn't been filled in
4425 /// yet, fill it in and return false.
4426 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
4427 Instruction *I = dyn_cast<Instruction>(V);
4428 if (I == 0) return true;
4430 // If this is an or instruction, it is an inner node of the bswap.
4431 if (I->getOpcode() == Instruction::Or)
4432 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
4433 CollectBSwapParts(I->getOperand(1), ByteValues);
4435 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
4436 // If this is a shift by a constant int, and it is "24", then its operand
4437 // defines a byte. We only handle unsigned types here.
4438 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
4439 // Not shifting the entire input by N-1 bytes?
4440 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
4441 8*(ByteValues.size()-1))
4445 if (I->getOpcode() == Instruction::Shl) {
4446 // X << 24 defines the top byte with the lowest of the input bytes.
4447 DestNo = ByteValues.size()-1;
4449 // X >>u 24 defines the low byte with the highest of the input bytes.
4453 // If the destination byte value is already defined, the values are or'd
4454 // together, which isn't a bswap (unless it's an or of the same bits).
4455 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
4457 ByteValues[DestNo] = I->getOperand(0);
4461 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
4463 Value *Shift = 0, *ShiftLHS = 0;
4464 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
4465 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
4466 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
4468 Instruction *SI = cast<Instruction>(Shift);
4470 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
4471 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
4472 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
4475 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
4477 if (AndAmt->getValue().getActiveBits() > 64)
4479 uint64_t AndAmtVal = AndAmt->getZExtValue();
4480 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
4481 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
4483 // Unknown mask for bswap.
4484 if (DestByte == ByteValues.size()) return true;
4486 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
4488 if (SI->getOpcode() == Instruction::Shl)
4489 SrcByte = DestByte - ShiftBytes;
4491 SrcByte = DestByte + ShiftBytes;
4493 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
4494 if (SrcByte != ByteValues.size()-DestByte-1)
4497 // If the destination byte value is already defined, the values are or'd
4498 // together, which isn't a bswap (unless it's an or of the same bits).
4499 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
4501 ByteValues[DestByte] = SI->getOperand(0);
4505 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4506 /// If so, insert the new bswap intrinsic and return it.
4507 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4508 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4509 if (!ITy || ITy->getBitWidth() % 16)
4510 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4512 /// ByteValues - For each byte of the result, we keep track of which value
4513 /// defines each byte.
4514 SmallVector<Value*, 8> ByteValues;
4515 ByteValues.resize(ITy->getBitWidth()/8);
4517 // Try to find all the pieces corresponding to the bswap.
4518 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
4519 CollectBSwapParts(I.getOperand(1), ByteValues))
4522 // Check to see if all of the bytes come from the same value.
4523 Value *V = ByteValues[0];
4524 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4526 // Check to make sure that all of the bytes come from the same value.
4527 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4528 if (ByteValues[i] != V)
4530 const Type *Tys[] = { ITy };
4531 Module *M = I.getParent()->getParent()->getParent();
4532 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4533 return CallInst::Create(F, V);
4537 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4538 bool Changed = SimplifyCommutative(I);
4539 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4541 if (isa<UndefValue>(Op1)) // X | undef -> -1
4542 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4546 return ReplaceInstUsesWith(I, Op0);
4548 // See if we can simplify any instructions used by the instruction whose sole
4549 // purpose is to compute bits we don't care about.
4550 if (!isa<VectorType>(I.getType())) {
4551 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4552 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4553 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4554 KnownZero, KnownOne))
4556 } else if (isa<ConstantAggregateZero>(Op1)) {
4557 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4558 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4559 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4560 return ReplaceInstUsesWith(I, I.getOperand(1));
4566 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4567 ConstantInt *C1 = 0; Value *X = 0;
4568 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4569 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4570 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4571 InsertNewInstBefore(Or, I);
4573 return BinaryOperator::CreateAnd(Or,
4574 ConstantInt::get(RHS->getValue() | C1->getValue()));
4577 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4578 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4579 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4580 InsertNewInstBefore(Or, I);
4582 return BinaryOperator::CreateXor(Or,
4583 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4586 // Try to fold constant and into select arguments.
4587 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4588 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4590 if (isa<PHINode>(Op0))
4591 if (Instruction *NV = FoldOpIntoPhi(I))
4595 Value *A = 0, *B = 0;
4596 ConstantInt *C1 = 0, *C2 = 0;
4598 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4599 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4600 return ReplaceInstUsesWith(I, Op1);
4601 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4602 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4603 return ReplaceInstUsesWith(I, Op0);
4605 // (A | B) | C and A | (B | C) -> bswap if possible.
4606 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4607 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4608 match(Op1, m_Or(m_Value(), m_Value())) ||
4609 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4610 match(Op1, m_Shift(m_Value(), m_Value())))) {
4611 if (Instruction *BSwap = MatchBSwap(I))
4615 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4616 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4617 MaskedValueIsZero(Op1, C1->getValue())) {
4618 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4619 InsertNewInstBefore(NOr, I);
4621 return BinaryOperator::CreateXor(NOr, C1);
4624 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4625 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4626 MaskedValueIsZero(Op0, C1->getValue())) {
4627 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4628 InsertNewInstBefore(NOr, I);
4630 return BinaryOperator::CreateXor(NOr, C1);
4634 Value *C = 0, *D = 0;
4635 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4636 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4637 Value *V1 = 0, *V2 = 0, *V3 = 0;
4638 C1 = dyn_cast<ConstantInt>(C);
4639 C2 = dyn_cast<ConstantInt>(D);
4640 if (C1 && C2) { // (A & C1)|(B & C2)
4641 // If we have: ((V + N) & C1) | (V & C2)
4642 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4643 // replace with V+N.
4644 if (C1->getValue() == ~C2->getValue()) {
4645 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4646 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4647 // Add commutes, try both ways.
4648 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4649 return ReplaceInstUsesWith(I, A);
4650 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4651 return ReplaceInstUsesWith(I, A);
4653 // Or commutes, try both ways.
4654 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4655 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4656 // Add commutes, try both ways.
4657 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4658 return ReplaceInstUsesWith(I, B);
4659 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4660 return ReplaceInstUsesWith(I, B);
4663 V1 = 0; V2 = 0; V3 = 0;
4666 // Check to see if we have any common things being and'ed. If so, find the
4667 // terms for V1 & (V2|V3).
4668 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4669 if (A == B) // (A & C)|(A & D) == A & (C|D)
4670 V1 = A, V2 = C, V3 = D;
4671 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4672 V1 = A, V2 = B, V3 = C;
4673 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4674 V1 = C, V2 = A, V3 = D;
4675 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4676 V1 = C, V2 = A, V3 = B;
4680 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4681 return BinaryOperator::CreateAnd(V1, Or);
4686 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4687 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4688 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4689 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4690 SI0->getOperand(1) == SI1->getOperand(1) &&
4691 (SI0->hasOneUse() || SI1->hasOneUse())) {
4692 Instruction *NewOp =
4693 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4695 SI0->getName()), I);
4696 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4697 SI1->getOperand(1));
4701 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4702 if (A == Op1) // ~A | A == -1
4703 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4707 // Note, A is still live here!
4708 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4710 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4712 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4713 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4714 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4715 I.getName()+".demorgan"), I);
4716 return BinaryOperator::CreateNot(And);
4720 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4721 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4722 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4725 Value *LHSVal, *RHSVal;
4726 ConstantInt *LHSCst, *RHSCst;
4727 ICmpInst::Predicate LHSCC, RHSCC;
4728 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4729 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4730 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4731 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4732 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4733 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4734 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4735 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4736 // We can't fold (ugt x, C) | (sgt x, C2).
4737 PredicatesFoldable(LHSCC, RHSCC)) {
4738 // Ensure that the larger constant is on the RHS.
4739 ICmpInst *LHS = cast<ICmpInst>(Op0);
4741 if (ICmpInst::isSignedPredicate(LHSCC))
4742 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4744 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4747 std::swap(LHS, RHS);
4748 std::swap(LHSCst, RHSCst);
4749 std::swap(LHSCC, RHSCC);
4752 // At this point, we know we have have two icmp instructions
4753 // comparing a value against two constants and or'ing the result
4754 // together. Because of the above check, we know that we only have
4755 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4756 // FoldICmpLogical check above), that the two constants are not
4758 assert(LHSCst != RHSCst && "Compares not folded above?");
4761 default: assert(0 && "Unknown integer condition code!");
4762 case ICmpInst::ICMP_EQ:
4764 default: assert(0 && "Unknown integer condition code!");
4765 case ICmpInst::ICMP_EQ:
4766 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4767 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4768 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4769 LHSVal->getName()+".off");
4770 InsertNewInstBefore(Add, I);
4771 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4772 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4774 break; // (X == 13 | X == 15) -> no change
4775 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4776 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4778 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4779 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4780 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4781 return ReplaceInstUsesWith(I, RHS);
4784 case ICmpInst::ICMP_NE:
4786 default: assert(0 && "Unknown integer condition code!");
4787 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4788 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4789 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4790 return ReplaceInstUsesWith(I, LHS);
4791 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4792 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4793 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4794 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4797 case ICmpInst::ICMP_ULT:
4799 default: assert(0 && "Unknown integer condition code!");
4800 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4802 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4803 // If RHSCst is [us]MAXINT, it is always false. Not handling
4804 // this can cause overflow.
4805 if (RHSCst->isMaxValue(false))
4806 return ReplaceInstUsesWith(I, LHS);
4807 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4809 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4811 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4812 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4813 return ReplaceInstUsesWith(I, RHS);
4814 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4818 case ICmpInst::ICMP_SLT:
4820 default: assert(0 && "Unknown integer condition code!");
4821 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4823 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4824 // If RHSCst is [us]MAXINT, it is always false. Not handling
4825 // this can cause overflow.
4826 if (RHSCst->isMaxValue(true))
4827 return ReplaceInstUsesWith(I, LHS);
4828 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4830 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4832 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4833 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4834 return ReplaceInstUsesWith(I, RHS);
4835 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4839 case ICmpInst::ICMP_UGT:
4841 default: assert(0 && "Unknown integer condition code!");
4842 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4843 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4844 return ReplaceInstUsesWith(I, LHS);
4845 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4847 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4848 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4849 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4850 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4854 case ICmpInst::ICMP_SGT:
4856 default: assert(0 && "Unknown integer condition code!");
4857 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4858 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4859 return ReplaceInstUsesWith(I, LHS);
4860 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4862 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4863 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4864 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4865 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4873 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4874 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4875 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4876 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4877 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4878 !isa<ICmpInst>(Op1C->getOperand(0))) {
4879 const Type *SrcTy = Op0C->getOperand(0)->getType();
4880 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4881 // Only do this if the casts both really cause code to be
4883 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4885 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4887 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4888 Op1C->getOperand(0),
4890 InsertNewInstBefore(NewOp, I);
4891 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4898 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4899 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4900 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4901 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4902 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4903 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4904 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4905 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4906 // If either of the constants are nans, then the whole thing returns
4908 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4909 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4911 // Otherwise, no need to compare the two constants, compare the
4913 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4914 RHS->getOperand(0));
4919 return Changed ? &I : 0;
4924 // XorSelf - Implements: X ^ X --> 0
4927 XorSelf(Value *rhs) : RHS(rhs) {}
4928 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4929 Instruction *apply(BinaryOperator &Xor) const {
4936 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4937 bool Changed = SimplifyCommutative(I);
4938 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4940 if (isa<UndefValue>(Op1)) {
4941 if (isa<UndefValue>(Op0))
4942 // Handle undef ^ undef -> 0 special case. This is a common
4944 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4945 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4948 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4949 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4950 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4951 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4954 // See if we can simplify any instructions used by the instruction whose sole
4955 // purpose is to compute bits we don't care about.
4956 if (!isa<VectorType>(I.getType())) {
4957 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4958 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4959 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4960 KnownZero, KnownOne))
4962 } else if (isa<ConstantAggregateZero>(Op1)) {
4963 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4966 // Is this a ~ operation?
4967 if (Value *NotOp = dyn_castNotVal(&I)) {
4968 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4969 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4970 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4971 if (Op0I->getOpcode() == Instruction::And ||
4972 Op0I->getOpcode() == Instruction::Or) {
4973 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4974 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4976 BinaryOperator::CreateNot(Op0I->getOperand(1),
4977 Op0I->getOperand(1)->getName()+".not");
4978 InsertNewInstBefore(NotY, I);
4979 if (Op0I->getOpcode() == Instruction::And)
4980 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4982 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4989 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4990 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4991 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4992 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4993 return new ICmpInst(ICI->getInversePredicate(),
4994 ICI->getOperand(0), ICI->getOperand(1));
4996 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4997 return new FCmpInst(FCI->getInversePredicate(),
4998 FCI->getOperand(0), FCI->getOperand(1));
5001 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5002 // ~(c-X) == X-c-1 == X+(-c-1)
5003 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5004 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5005 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5006 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5007 ConstantInt::get(I.getType(), 1));
5008 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5011 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5012 if (Op0I->getOpcode() == Instruction::Add) {
5013 // ~(X-c) --> (-c-1)-X
5014 if (RHS->isAllOnesValue()) {
5015 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5016 return BinaryOperator::CreateSub(
5017 ConstantExpr::getSub(NegOp0CI,
5018 ConstantInt::get(I.getType(), 1)),
5019 Op0I->getOperand(0));
5020 } else if (RHS->getValue().isSignBit()) {
5021 // (X + C) ^ signbit -> (X + C + signbit)
5022 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
5023 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5026 } else if (Op0I->getOpcode() == Instruction::Or) {
5027 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5028 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5029 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5030 // Anything in both C1 and C2 is known to be zero, remove it from
5032 Constant *CommonBits = And(Op0CI, RHS);
5033 NewRHS = ConstantExpr::getAnd(NewRHS,
5034 ConstantExpr::getNot(CommonBits));
5035 AddToWorkList(Op0I);
5036 I.setOperand(0, Op0I->getOperand(0));
5037 I.setOperand(1, NewRHS);
5044 // Try to fold constant and into select arguments.
5045 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5046 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5048 if (isa<PHINode>(Op0))
5049 if (Instruction *NV = FoldOpIntoPhi(I))
5053 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5055 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5057 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5059 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5062 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5065 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5066 if (A == Op0) { // B^(B|A) == (A|B)^B
5067 Op1I->swapOperands();
5069 std::swap(Op0, Op1);
5070 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5071 I.swapOperands(); // Simplified below.
5072 std::swap(Op0, Op1);
5074 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
5075 if (Op0 == A) // A^(A^B) == B
5076 return ReplaceInstUsesWith(I, B);
5077 else if (Op0 == B) // A^(B^A) == B
5078 return ReplaceInstUsesWith(I, A);
5079 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
5080 if (A == Op0) { // A^(A&B) -> A^(B&A)
5081 Op1I->swapOperands();
5084 if (B == Op0) { // A^(B&A) -> (B&A)^A
5085 I.swapOperands(); // Simplified below.
5086 std::swap(Op0, Op1);
5091 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5094 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
5095 if (A == Op1) // (B|A)^B == (A|B)^B
5097 if (B == Op1) { // (A|B)^B == A & ~B
5099 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5100 return BinaryOperator::CreateAnd(A, NotB);
5102 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
5103 if (Op1 == A) // (A^B)^A == B
5104 return ReplaceInstUsesWith(I, B);
5105 else if (Op1 == B) // (B^A)^A == B
5106 return ReplaceInstUsesWith(I, A);
5107 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5108 if (A == Op1) // (A&B)^A -> (B&A)^A
5110 if (B == Op1 && // (B&A)^A == ~B & A
5111 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5113 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5114 return BinaryOperator::CreateAnd(N, Op1);
5119 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5120 if (Op0I && Op1I && Op0I->isShift() &&
5121 Op0I->getOpcode() == Op1I->getOpcode() &&
5122 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5123 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5124 Instruction *NewOp =
5125 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5126 Op1I->getOperand(0),
5127 Op0I->getName()), I);
5128 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5129 Op1I->getOperand(1));
5133 Value *A, *B, *C, *D;
5134 // (A & B)^(A | B) -> A ^ B
5135 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5136 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5137 if ((A == C && B == D) || (A == D && B == C))
5138 return BinaryOperator::CreateXor(A, B);
5140 // (A | B)^(A & B) -> A ^ B
5141 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5142 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5143 if ((A == C && B == D) || (A == D && B == C))
5144 return BinaryOperator::CreateXor(A, B);
5148 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5149 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5150 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5151 // (X & Y)^(X & Y) -> (Y^Z) & X
5152 Value *X = 0, *Y = 0, *Z = 0;
5154 X = A, Y = B, Z = D;
5156 X = A, Y = B, Z = C;
5158 X = B, Y = A, Z = D;
5160 X = B, Y = A, Z = C;
5163 Instruction *NewOp =
5164 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5165 return BinaryOperator::CreateAnd(NewOp, X);
5170 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5171 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5172 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5175 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5176 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5177 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5178 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5179 const Type *SrcTy = Op0C->getOperand(0)->getType();
5180 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5181 // Only do this if the casts both really cause code to be generated.
5182 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5184 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5186 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5187 Op1C->getOperand(0),
5189 InsertNewInstBefore(NewOp, I);
5190 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5194 return Changed ? &I : 0;
5197 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5198 /// overflowed for this type.
5199 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5200 ConstantInt *In2, bool IsSigned = false) {
5201 Result = cast<ConstantInt>(Add(In1, In2));
5204 if (In2->getValue().isNegative())
5205 return Result->getValue().sgt(In1->getValue());
5207 return Result->getValue().slt(In1->getValue());
5209 return Result->getValue().ult(In1->getValue());
5212 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5213 /// code necessary to compute the offset from the base pointer (without adding
5214 /// in the base pointer). Return the result as a signed integer of intptr size.
5215 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5216 TargetData &TD = IC.getTargetData();
5217 gep_type_iterator GTI = gep_type_begin(GEP);
5218 const Type *IntPtrTy = TD.getIntPtrType();
5219 Value *Result = Constant::getNullValue(IntPtrTy);
5221 // Build a mask for high order bits.
5222 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5223 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5225 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
5226 Value *Op = GEP->getOperand(i);
5227 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5228 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5229 if (OpC->isZero()) continue;
5231 // Handle a struct index, which adds its field offset to the pointer.
5232 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5233 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5235 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5236 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5238 Result = IC.InsertNewInstBefore(
5239 BinaryOperator::CreateAdd(Result,
5240 ConstantInt::get(IntPtrTy, Size),
5241 GEP->getName()+".offs"), I);
5245 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5246 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5247 Scale = ConstantExpr::getMul(OC, Scale);
5248 if (Constant *RC = dyn_cast<Constant>(Result))
5249 Result = ConstantExpr::getAdd(RC, Scale);
5251 // Emit an add instruction.
5252 Result = IC.InsertNewInstBefore(
5253 BinaryOperator::CreateAdd(Result, Scale,
5254 GEP->getName()+".offs"), I);
5258 // Convert to correct type.
5259 if (Op->getType() != IntPtrTy) {
5260 if (Constant *OpC = dyn_cast<Constant>(Op))
5261 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5263 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5264 Op->getName()+".c"), I);
5267 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5268 if (Constant *OpC = dyn_cast<Constant>(Op))
5269 Op = ConstantExpr::getMul(OpC, Scale);
5270 else // We'll let instcombine(mul) convert this to a shl if possible.
5271 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5272 GEP->getName()+".idx"), I);
5275 // Emit an add instruction.
5276 if (isa<Constant>(Op) && isa<Constant>(Result))
5277 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5278 cast<Constant>(Result));
5280 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5281 GEP->getName()+".offs"), I);
5287 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5288 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5289 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5290 /// complex, and scales are involved. The above expression would also be legal
5291 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5292 /// later form is less amenable to optimization though, and we are allowed to
5293 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5295 /// If we can't emit an optimized form for this expression, this returns null.
5297 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5299 TargetData &TD = IC.getTargetData();
5300 gep_type_iterator GTI = gep_type_begin(GEP);
5302 // Check to see if this gep only has a single variable index. If so, and if
5303 // any constant indices are a multiple of its scale, then we can compute this
5304 // in terms of the scale of the variable index. For example, if the GEP
5305 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5306 // because the expression will cross zero at the same point.
5307 unsigned i, e = GEP->getNumOperands();
5309 for (i = 1; i != e; ++i, ++GTI) {
5310 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5311 // Compute the aggregate offset of constant indices.
5312 if (CI->isZero()) continue;
5314 // Handle a struct index, which adds its field offset to the pointer.
5315 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5316 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5318 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5319 Offset += Size*CI->getSExtValue();
5322 // Found our variable index.
5327 // If there are no variable indices, we must have a constant offset, just
5328 // evaluate it the general way.
5329 if (i == e) return 0;
5331 Value *VariableIdx = GEP->getOperand(i);
5332 // Determine the scale factor of the variable element. For example, this is
5333 // 4 if the variable index is into an array of i32.
5334 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5336 // Verify that there are no other variable indices. If so, emit the hard way.
5337 for (++i, ++GTI; i != e; ++i, ++GTI) {
5338 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5341 // Compute the aggregate offset of constant indices.
5342 if (CI->isZero()) continue;
5344 // Handle a struct index, which adds its field offset to the pointer.
5345 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5346 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5348 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5349 Offset += Size*CI->getSExtValue();
5353 // Okay, we know we have a single variable index, which must be a
5354 // pointer/array/vector index. If there is no offset, life is simple, return
5356 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5358 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5359 // we don't need to bother extending: the extension won't affect where the
5360 // computation crosses zero.
5361 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5362 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5363 VariableIdx->getNameStart(), &I);
5367 // Otherwise, there is an index. The computation we will do will be modulo
5368 // the pointer size, so get it.
5369 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5371 Offset &= PtrSizeMask;
5372 VariableScale &= PtrSizeMask;
5374 // To do this transformation, any constant index must be a multiple of the
5375 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5376 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5377 // multiple of the variable scale.
5378 int64_t NewOffs = Offset / (int64_t)VariableScale;
5379 if (Offset != NewOffs*(int64_t)VariableScale)
5382 // Okay, we can do this evaluation. Start by converting the index to intptr.
5383 const Type *IntPtrTy = TD.getIntPtrType();
5384 if (VariableIdx->getType() != IntPtrTy)
5385 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5387 VariableIdx->getNameStart(), &I);
5388 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5389 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5393 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5394 /// else. At this point we know that the GEP is on the LHS of the comparison.
5395 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5396 ICmpInst::Predicate Cond,
5398 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5400 // Look through bitcasts.
5401 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5402 RHS = BCI->getOperand(0);
5404 Value *PtrBase = GEPLHS->getOperand(0);
5405 if (PtrBase == RHS) {
5406 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5407 // This transformation (ignoring the base and scales) is valid because we
5408 // know pointers can't overflow. See if we can output an optimized form.
5409 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5411 // If not, synthesize the offset the hard way.
5413 Offset = EmitGEPOffset(GEPLHS, I, *this);
5414 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5415 Constant::getNullValue(Offset->getType()));
5416 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5417 // If the base pointers are different, but the indices are the same, just
5418 // compare the base pointer.
5419 if (PtrBase != GEPRHS->getOperand(0)) {
5420 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5421 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5422 GEPRHS->getOperand(0)->getType();
5424 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5425 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5426 IndicesTheSame = false;
5430 // If all indices are the same, just compare the base pointers.
5432 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5433 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5435 // Otherwise, the base pointers are different and the indices are
5436 // different, bail out.
5440 // If one of the GEPs has all zero indices, recurse.
5441 bool AllZeros = true;
5442 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5443 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5444 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5449 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5450 ICmpInst::getSwappedPredicate(Cond), I);
5452 // If the other GEP has all zero indices, recurse.
5454 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5455 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5456 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5461 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5463 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5464 // If the GEPs only differ by one index, compare it.
5465 unsigned NumDifferences = 0; // Keep track of # differences.
5466 unsigned DiffOperand = 0; // The operand that differs.
5467 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5468 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5469 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5470 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5471 // Irreconcilable differences.
5475 if (NumDifferences++) break;
5480 if (NumDifferences == 0) // SAME GEP?
5481 return ReplaceInstUsesWith(I, // No comparison is needed here.
5482 ConstantInt::get(Type::Int1Ty,
5483 ICmpInst::isTrueWhenEqual(Cond)));
5485 else if (NumDifferences == 1) {
5486 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5487 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5488 // Make sure we do a signed comparison here.
5489 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5493 // Only lower this if the icmp is the only user of the GEP or if we expect
5494 // the result to fold to a constant!
5495 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5496 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5497 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5498 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5499 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5500 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5506 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5508 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5511 if (!isa<ConstantFP>(RHSC)) return 0;
5512 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5514 // Get the width of the mantissa. We don't want to hack on conversions that
5515 // might lose information from the integer, e.g. "i64 -> float"
5516 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5517 if (MantissaWidth == -1) return 0; // Unknown.
5519 // Check to see that the input is converted from an integer type that is small
5520 // enough that preserves all bits. TODO: check here for "known" sign bits.
5521 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5522 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5524 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5525 if (isa<UIToFPInst>(LHSI))
5528 // If the conversion would lose info, don't hack on this.
5529 if ((int)InputSize > MantissaWidth)
5532 // Otherwise, we can potentially simplify the comparison. We know that it
5533 // will always come through as an integer value and we know the constant is
5534 // not a NAN (it would have been previously simplified).
5535 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5537 ICmpInst::Predicate Pred;
5538 switch (I.getPredicate()) {
5539 default: assert(0 && "Unexpected predicate!");
5540 case FCmpInst::FCMP_UEQ:
5541 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
5542 case FCmpInst::FCMP_UGT:
5543 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
5544 case FCmpInst::FCMP_UGE:
5545 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
5546 case FCmpInst::FCMP_ULT:
5547 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
5548 case FCmpInst::FCMP_ULE:
5549 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
5550 case FCmpInst::FCMP_UNE:
5551 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
5552 case FCmpInst::FCMP_ORD:
5553 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5554 case FCmpInst::FCMP_UNO:
5555 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5558 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5560 // Now we know that the APFloat is a normal number, zero or inf.
5562 // See if the FP constant is too large for the integer. For example,
5563 // comparing an i8 to 300.0.
5564 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5566 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5567 // and large values.
5568 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5569 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5570 APFloat::rmNearestTiesToEven);
5571 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5572 if (ICmpInst::ICMP_NE || ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
5573 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5574 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5577 // See if the RHS value is < SignedMin.
5578 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5579 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5580 APFloat::rmNearestTiesToEven);
5581 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5582 if (ICmpInst::ICMP_NE || ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
5583 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5584 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5587 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
5588 // it may still be fractional. See if it is fractional by casting the FP
5589 // value to the integer value and back, checking for equality. Don't do this
5590 // for zero, because -0.0 is not fractional.
5591 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5592 if (!RHS.isZero() &&
5593 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5594 // If we had a comparison against a fractional value, we have to adjust
5595 // the compare predicate and sometimes the value. RHSC is rounded towards
5596 // zero at this point.
5598 default: assert(0 && "Unexpected integer comparison!");
5599 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5600 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5601 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5602 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5603 case ICmpInst::ICMP_SLE:
5604 // (float)int <= 4.4 --> int <= 4
5605 // (float)int <= -4.4 --> int < -4
5606 if (RHS.isNegative())
5607 Pred = ICmpInst::ICMP_SLT;
5609 case ICmpInst::ICMP_SLT:
5610 // (float)int < -4.4 --> int < -4
5611 // (float)int < 4.4 --> int <= 4
5612 if (!RHS.isNegative())
5613 Pred = ICmpInst::ICMP_SLE;
5615 case ICmpInst::ICMP_SGT:
5616 // (float)int > 4.4 --> int > 4
5617 // (float)int > -4.4 --> int >= -4
5618 if (RHS.isNegative())
5619 Pred = ICmpInst::ICMP_SGE;
5621 case ICmpInst::ICMP_SGE:
5622 // (float)int >= -4.4 --> int >= -4
5623 // (float)int >= 4.4 --> int > 4
5624 if (!RHS.isNegative())
5625 Pred = ICmpInst::ICMP_SGT;
5630 // Lower this FP comparison into an appropriate integer version of the
5632 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5635 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5636 bool Changed = SimplifyCompare(I);
5637 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5639 // Fold trivial predicates.
5640 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5641 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5642 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5643 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5645 // Simplify 'fcmp pred X, X'
5647 switch (I.getPredicate()) {
5648 default: assert(0 && "Unknown predicate!");
5649 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5650 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5651 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5652 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5653 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5654 case FCmpInst::FCMP_OLT: // True if ordered and less than
5655 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5656 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5658 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5659 case FCmpInst::FCMP_ULT: // True if unordered or less than
5660 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5661 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5662 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5663 I.setPredicate(FCmpInst::FCMP_UNO);
5664 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5667 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5668 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5669 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5670 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5671 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5672 I.setPredicate(FCmpInst::FCMP_ORD);
5673 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5678 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5679 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5681 // Handle fcmp with constant RHS
5682 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5683 // If the constant is a nan, see if we can fold the comparison based on it.
5684 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5685 if (CFP->getValueAPF().isNaN()) {
5686 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5687 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5688 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5689 "Comparison must be either ordered or unordered!");
5690 // True if unordered.
5691 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5695 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5696 switch (LHSI->getOpcode()) {
5697 case Instruction::PHI:
5698 if (Instruction *NV = FoldOpIntoPhi(I))
5701 case Instruction::SIToFP:
5702 case Instruction::UIToFP:
5703 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5706 case Instruction::Select:
5707 // If either operand of the select is a constant, we can fold the
5708 // comparison into the select arms, which will cause one to be
5709 // constant folded and the select turned into a bitwise or.
5710 Value *Op1 = 0, *Op2 = 0;
5711 if (LHSI->hasOneUse()) {
5712 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5713 // Fold the known value into the constant operand.
5714 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5715 // Insert a new FCmp of the other select operand.
5716 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5717 LHSI->getOperand(2), RHSC,
5719 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5720 // Fold the known value into the constant operand.
5721 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5722 // Insert a new FCmp of the other select operand.
5723 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5724 LHSI->getOperand(1), RHSC,
5730 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5735 return Changed ? &I : 0;
5738 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5739 bool Changed = SimplifyCompare(I);
5740 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5741 const Type *Ty = Op0->getType();
5745 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5746 I.isTrueWhenEqual()));
5748 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5749 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5751 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5752 // addresses never equal each other! We already know that Op0 != Op1.
5753 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5754 isa<ConstantPointerNull>(Op0)) &&
5755 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5756 isa<ConstantPointerNull>(Op1)))
5757 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5758 !I.isTrueWhenEqual()));
5760 // icmp's with boolean values can always be turned into bitwise operations
5761 if (Ty == Type::Int1Ty) {
5762 switch (I.getPredicate()) {
5763 default: assert(0 && "Invalid icmp instruction!");
5764 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5765 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5766 InsertNewInstBefore(Xor, I);
5767 return BinaryOperator::CreateNot(Xor);
5769 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5770 return BinaryOperator::CreateXor(Op0, Op1);
5772 case ICmpInst::ICMP_UGT:
5773 case ICmpInst::ICMP_SGT:
5774 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5776 case ICmpInst::ICMP_ULT:
5777 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5778 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5779 InsertNewInstBefore(Not, I);
5780 return BinaryOperator::CreateAnd(Not, Op1);
5782 case ICmpInst::ICMP_UGE:
5783 case ICmpInst::ICMP_SGE:
5784 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5786 case ICmpInst::ICMP_ULE:
5787 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5788 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5789 InsertNewInstBefore(Not, I);
5790 return BinaryOperator::CreateOr(Not, Op1);
5795 // See if we are doing a comparison between a constant and an instruction that
5796 // can be folded into the comparison.
5797 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5800 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5801 if (I.isEquality() && CI->isNullValue() &&
5802 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5803 // (icmp cond A B) if cond is equality
5804 return new ICmpInst(I.getPredicate(), A, B);
5807 switch (I.getPredicate()) {
5809 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5810 if (CI->isMinValue(false))
5811 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5812 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5813 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5814 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5815 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5816 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5817 if (CI->isMinValue(true))
5818 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5819 ConstantInt::getAllOnesValue(Op0->getType()));
5823 case ICmpInst::ICMP_SLT:
5824 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5825 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5826 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5827 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5828 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5829 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5832 case ICmpInst::ICMP_UGT:
5833 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5834 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5835 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5836 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5837 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5838 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5840 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5841 if (CI->isMaxValue(true))
5842 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5843 ConstantInt::getNullValue(Op0->getType()));
5846 case ICmpInst::ICMP_SGT:
5847 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5848 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5849 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5850 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5851 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5852 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5855 case ICmpInst::ICMP_ULE:
5856 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5857 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5858 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5859 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5860 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5861 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5864 case ICmpInst::ICMP_SLE:
5865 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5866 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5867 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5868 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5869 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5870 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5873 case ICmpInst::ICMP_UGE:
5874 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5875 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5876 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5877 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5878 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5879 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5882 case ICmpInst::ICMP_SGE:
5883 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5884 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5885 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5886 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5887 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5888 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5892 // If we still have a icmp le or icmp ge instruction, turn it into the
5893 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5894 // already been handled above, this requires little checking.
5896 switch (I.getPredicate()) {
5898 case ICmpInst::ICMP_ULE:
5899 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5900 case ICmpInst::ICMP_SLE:
5901 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5902 case ICmpInst::ICMP_UGE:
5903 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5904 case ICmpInst::ICMP_SGE:
5905 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5908 // See if we can fold the comparison based on bits known to be zero or one
5909 // in the input. If this comparison is a normal comparison, it demands all
5910 // bits, if it is a sign bit comparison, it only demands the sign bit.
5913 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5915 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5916 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5917 if (SimplifyDemandedBits(Op0,
5918 isSignBit ? APInt::getSignBit(BitWidth)
5919 : APInt::getAllOnesValue(BitWidth),
5920 KnownZero, KnownOne, 0))
5923 // Given the known and unknown bits, compute a range that the LHS could be
5925 if ((KnownOne | KnownZero) != 0) {
5926 // Compute the Min, Max and RHS values based on the known bits. For the
5927 // EQ and NE we use unsigned values.
5928 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5929 const APInt& RHSVal = CI->getValue();
5930 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5931 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5934 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5937 switch (I.getPredicate()) { // LE/GE have been folded already.
5938 default: assert(0 && "Unknown icmp opcode!");
5939 case ICmpInst::ICMP_EQ:
5940 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5941 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5943 case ICmpInst::ICMP_NE:
5944 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5945 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5947 case ICmpInst::ICMP_ULT:
5948 if (Max.ult(RHSVal))
5949 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5950 if (Min.uge(RHSVal))
5951 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5953 case ICmpInst::ICMP_UGT:
5954 if (Min.ugt(RHSVal))
5955 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5956 if (Max.ule(RHSVal))
5957 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5959 case ICmpInst::ICMP_SLT:
5960 if (Max.slt(RHSVal))
5961 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5962 if (Min.sgt(RHSVal))
5963 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5965 case ICmpInst::ICMP_SGT:
5966 if (Min.sgt(RHSVal))
5967 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5968 if (Max.sle(RHSVal))
5969 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5974 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5975 // instruction, see if that instruction also has constants so that the
5976 // instruction can be folded into the icmp
5977 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5978 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5982 // Handle icmp with constant (but not simple integer constant) RHS
5983 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5984 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5985 switch (LHSI->getOpcode()) {
5986 case Instruction::GetElementPtr:
5987 if (RHSC->isNullValue()) {
5988 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5989 bool isAllZeros = true;
5990 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5991 if (!isa<Constant>(LHSI->getOperand(i)) ||
5992 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5997 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5998 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6002 case Instruction::PHI:
6003 if (Instruction *NV = FoldOpIntoPhi(I))
6006 case Instruction::Select: {
6007 // If either operand of the select is a constant, we can fold the
6008 // comparison into the select arms, which will cause one to be
6009 // constant folded and the select turned into a bitwise or.
6010 Value *Op1 = 0, *Op2 = 0;
6011 if (LHSI->hasOneUse()) {
6012 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6013 // Fold the known value into the constant operand.
6014 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6015 // Insert a new ICmp of the other select operand.
6016 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6017 LHSI->getOperand(2), RHSC,
6019 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6020 // Fold the known value into the constant operand.
6021 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6022 // Insert a new ICmp of the other select operand.
6023 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6024 LHSI->getOperand(1), RHSC,
6030 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6033 case Instruction::Malloc:
6034 // If we have (malloc != null), and if the malloc has a single use, we
6035 // can assume it is successful and remove the malloc.
6036 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6037 AddToWorkList(LHSI);
6038 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6039 !I.isTrueWhenEqual()));
6045 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6046 if (User *GEP = dyn_castGetElementPtr(Op0))
6047 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6049 if (User *GEP = dyn_castGetElementPtr(Op1))
6050 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6051 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6054 // Test to see if the operands of the icmp are casted versions of other
6055 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6057 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6058 if (isa<PointerType>(Op0->getType()) &&
6059 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6060 // We keep moving the cast from the left operand over to the right
6061 // operand, where it can often be eliminated completely.
6062 Op0 = CI->getOperand(0);
6064 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6065 // so eliminate it as well.
6066 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6067 Op1 = CI2->getOperand(0);
6069 // If Op1 is a constant, we can fold the cast into the constant.
6070 if (Op0->getType() != Op1->getType()) {
6071 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6072 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6074 // Otherwise, cast the RHS right before the icmp
6075 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6078 return new ICmpInst(I.getPredicate(), Op0, Op1);
6082 if (isa<CastInst>(Op0)) {
6083 // Handle the special case of: icmp (cast bool to X), <cst>
6084 // This comes up when you have code like
6087 // For generality, we handle any zero-extension of any operand comparison
6088 // with a constant or another cast from the same type.
6089 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6090 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6094 // ~x < ~y --> y < x
6096 if (match(Op0, m_Not(m_Value(A))) &&
6097 match(Op1, m_Not(m_Value(B))))
6098 return new ICmpInst(I.getPredicate(), B, A);
6101 if (I.isEquality()) {
6102 Value *A, *B, *C, *D;
6104 // -x == -y --> x == y
6105 if (match(Op0, m_Neg(m_Value(A))) &&
6106 match(Op1, m_Neg(m_Value(B))))
6107 return new ICmpInst(I.getPredicate(), A, B);
6109 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6110 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6111 Value *OtherVal = A == Op1 ? B : A;
6112 return new ICmpInst(I.getPredicate(), OtherVal,
6113 Constant::getNullValue(A->getType()));
6116 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6117 // A^c1 == C^c2 --> A == C^(c1^c2)
6118 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
6119 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
6120 if (Op1->hasOneUse()) {
6121 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6122 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6123 return new ICmpInst(I.getPredicate(), A,
6124 InsertNewInstBefore(Xor, I));
6127 // A^B == A^D -> B == D
6128 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6129 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6130 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6131 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6135 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6136 (A == Op0 || B == Op0)) {
6137 // A == (A^B) -> B == 0
6138 Value *OtherVal = A == Op0 ? B : A;
6139 return new ICmpInst(I.getPredicate(), OtherVal,
6140 Constant::getNullValue(A->getType()));
6142 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
6143 // (A-B) == A -> B == 0
6144 return new ICmpInst(I.getPredicate(), B,
6145 Constant::getNullValue(B->getType()));
6147 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
6148 // A == (A-B) -> B == 0
6149 return new ICmpInst(I.getPredicate(), B,
6150 Constant::getNullValue(B->getType()));
6153 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6154 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6155 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6156 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6157 Value *X = 0, *Y = 0, *Z = 0;
6160 X = B; Y = D; Z = A;
6161 } else if (A == D) {
6162 X = B; Y = C; Z = A;
6163 } else if (B == C) {
6164 X = A; Y = D; Z = B;
6165 } else if (B == D) {
6166 X = A; Y = C; Z = B;
6169 if (X) { // Build (X^Y) & Z
6170 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6171 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6172 I.setOperand(0, Op1);
6173 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6178 return Changed ? &I : 0;
6182 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6183 /// and CmpRHS are both known to be integer constants.
6184 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6185 ConstantInt *DivRHS) {
6186 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6187 const APInt &CmpRHSV = CmpRHS->getValue();
6189 // FIXME: If the operand types don't match the type of the divide
6190 // then don't attempt this transform. The code below doesn't have the
6191 // logic to deal with a signed divide and an unsigned compare (and
6192 // vice versa). This is because (x /s C1) <s C2 produces different
6193 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6194 // (x /u C1) <u C2. Simply casting the operands and result won't
6195 // work. :( The if statement below tests that condition and bails
6197 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6198 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6200 if (DivRHS->isZero())
6201 return 0; // The ProdOV computation fails on divide by zero.
6203 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6204 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6205 // C2 (CI). By solving for X we can turn this into a range check
6206 // instead of computing a divide.
6207 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6209 // Determine if the product overflows by seeing if the product is
6210 // not equal to the divide. Make sure we do the same kind of divide
6211 // as in the LHS instruction that we're folding.
6212 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6213 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6215 // Get the ICmp opcode
6216 ICmpInst::Predicate Pred = ICI.getPredicate();
6218 // Figure out the interval that is being checked. For example, a comparison
6219 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6220 // Compute this interval based on the constants involved and the signedness of
6221 // the compare/divide. This computes a half-open interval, keeping track of
6222 // whether either value in the interval overflows. After analysis each
6223 // overflow variable is set to 0 if it's corresponding bound variable is valid
6224 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6225 int LoOverflow = 0, HiOverflow = 0;
6226 ConstantInt *LoBound = 0, *HiBound = 0;
6229 if (!DivIsSigned) { // udiv
6230 // e.g. X/5 op 3 --> [15, 20)
6232 HiOverflow = LoOverflow = ProdOV;
6234 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6235 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6236 if (CmpRHSV == 0) { // (X / pos) op 0
6237 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6238 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6240 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6241 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6242 HiOverflow = LoOverflow = ProdOV;
6244 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6245 } else { // (X / pos) op neg
6246 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6247 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
6248 LoOverflow = AddWithOverflow(LoBound, Prod,
6249 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
6250 HiBound = AddOne(Prod);
6251 HiOverflow = ProdOV ? -1 : 0;
6253 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6254 if (CmpRHSV == 0) { // (X / neg) op 0
6255 // e.g. X/-5 op 0 --> [-4, 5)
6256 LoBound = AddOne(DivRHS);
6257 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6258 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6259 HiOverflow = 1; // [INTMIN+1, overflow)
6260 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6262 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6263 // e.g. X/-5 op 3 --> [-19, -14)
6264 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6266 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
6267 HiBound = AddOne(Prod);
6268 } else { // (X / neg) op neg
6269 // e.g. X/-5 op -3 --> [15, 20)
6271 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
6272 HiBound = Subtract(Prod, DivRHS);
6275 // Dividing by a negative swaps the condition. LT <-> GT
6276 Pred = ICmpInst::getSwappedPredicate(Pred);
6279 Value *X = DivI->getOperand(0);
6281 default: assert(0 && "Unhandled icmp opcode!");
6282 case ICmpInst::ICMP_EQ:
6283 if (LoOverflow && HiOverflow)
6284 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6285 else if (HiOverflow)
6286 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6287 ICmpInst::ICMP_UGE, X, LoBound);
6288 else if (LoOverflow)
6289 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6290 ICmpInst::ICMP_ULT, X, HiBound);
6292 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6293 case ICmpInst::ICMP_NE:
6294 if (LoOverflow && HiOverflow)
6295 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6296 else if (HiOverflow)
6297 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6298 ICmpInst::ICMP_ULT, X, LoBound);
6299 else if (LoOverflow)
6300 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6301 ICmpInst::ICMP_UGE, X, HiBound);
6303 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6304 case ICmpInst::ICMP_ULT:
6305 case ICmpInst::ICMP_SLT:
6306 if (LoOverflow == +1) // Low bound is greater than input range.
6307 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6308 if (LoOverflow == -1) // Low bound is less than input range.
6309 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6310 return new ICmpInst(Pred, X, LoBound);
6311 case ICmpInst::ICMP_UGT:
6312 case ICmpInst::ICMP_SGT:
6313 if (HiOverflow == +1) // High bound greater than input range.
6314 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6315 else if (HiOverflow == -1) // High bound less than input range.
6316 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6317 if (Pred == ICmpInst::ICMP_UGT)
6318 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6320 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6325 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6327 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6330 const APInt &RHSV = RHS->getValue();
6332 switch (LHSI->getOpcode()) {
6333 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6334 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6335 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6337 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6338 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6339 Value *CompareVal = LHSI->getOperand(0);
6341 // If the sign bit of the XorCST is not set, there is no change to
6342 // the operation, just stop using the Xor.
6343 if (!XorCST->getValue().isNegative()) {
6344 ICI.setOperand(0, CompareVal);
6345 AddToWorkList(LHSI);
6349 // Was the old condition true if the operand is positive?
6350 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6352 // If so, the new one isn't.
6353 isTrueIfPositive ^= true;
6355 if (isTrueIfPositive)
6356 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6358 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6362 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6363 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6364 LHSI->getOperand(0)->hasOneUse()) {
6365 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6367 // If the LHS is an AND of a truncating cast, we can widen the
6368 // and/compare to be the input width without changing the value
6369 // produced, eliminating a cast.
6370 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6371 // We can do this transformation if either the AND constant does not
6372 // have its sign bit set or if it is an equality comparison.
6373 // Extending a relational comparison when we're checking the sign
6374 // bit would not work.
6375 if (Cast->hasOneUse() &&
6376 (ICI.isEquality() ||
6377 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6379 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6380 APInt NewCST = AndCST->getValue();
6381 NewCST.zext(BitWidth);
6383 NewCI.zext(BitWidth);
6384 Instruction *NewAnd =
6385 BinaryOperator::CreateAnd(Cast->getOperand(0),
6386 ConstantInt::get(NewCST),LHSI->getName());
6387 InsertNewInstBefore(NewAnd, ICI);
6388 return new ICmpInst(ICI.getPredicate(), NewAnd,
6389 ConstantInt::get(NewCI));
6393 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6394 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6395 // happens a LOT in code produced by the C front-end, for bitfield
6397 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6398 if (Shift && !Shift->isShift())
6402 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6403 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6404 const Type *AndTy = AndCST->getType(); // Type of the and.
6406 // We can fold this as long as we can't shift unknown bits
6407 // into the mask. This can only happen with signed shift
6408 // rights, as they sign-extend.
6410 bool CanFold = Shift->isLogicalShift();
6412 // To test for the bad case of the signed shr, see if any
6413 // of the bits shifted in could be tested after the mask.
6414 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6415 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6417 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6418 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6419 AndCST->getValue()) == 0)
6425 if (Shift->getOpcode() == Instruction::Shl)
6426 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6428 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6430 // Check to see if we are shifting out any of the bits being
6432 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6433 // If we shifted bits out, the fold is not going to work out.
6434 // As a special case, check to see if this means that the
6435 // result is always true or false now.
6436 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6437 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6438 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6439 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6441 ICI.setOperand(1, NewCst);
6442 Constant *NewAndCST;
6443 if (Shift->getOpcode() == Instruction::Shl)
6444 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6446 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6447 LHSI->setOperand(1, NewAndCST);
6448 LHSI->setOperand(0, Shift->getOperand(0));
6449 AddToWorkList(Shift); // Shift is dead.
6450 AddUsesToWorkList(ICI);
6456 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6457 // preferable because it allows the C<<Y expression to be hoisted out
6458 // of a loop if Y is invariant and X is not.
6459 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6460 ICI.isEquality() && !Shift->isArithmeticShift() &&
6461 isa<Instruction>(Shift->getOperand(0))) {
6464 if (Shift->getOpcode() == Instruction::LShr) {
6465 NS = BinaryOperator::CreateShl(AndCST,
6466 Shift->getOperand(1), "tmp");
6468 // Insert a logical shift.
6469 NS = BinaryOperator::CreateLShr(AndCST,
6470 Shift->getOperand(1), "tmp");
6472 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6474 // Compute X & (C << Y).
6475 Instruction *NewAnd =
6476 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6477 InsertNewInstBefore(NewAnd, ICI);
6479 ICI.setOperand(0, NewAnd);
6485 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6486 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6489 uint32_t TypeBits = RHSV.getBitWidth();
6491 // Check that the shift amount is in range. If not, don't perform
6492 // undefined shifts. When the shift is visited it will be
6494 if (ShAmt->uge(TypeBits))
6497 if (ICI.isEquality()) {
6498 // If we are comparing against bits always shifted out, the
6499 // comparison cannot succeed.
6501 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6502 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6503 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6504 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6505 return ReplaceInstUsesWith(ICI, Cst);
6508 if (LHSI->hasOneUse()) {
6509 // Otherwise strength reduce the shift into an and.
6510 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6512 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6515 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6516 Mask, LHSI->getName()+".mask");
6517 Value *And = InsertNewInstBefore(AndI, ICI);
6518 return new ICmpInst(ICI.getPredicate(), And,
6519 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6523 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6524 bool TrueIfSigned = false;
6525 if (LHSI->hasOneUse() &&
6526 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6527 // (X << 31) <s 0 --> (X&1) != 0
6528 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6529 (TypeBits-ShAmt->getZExtValue()-1));
6531 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6532 Mask, LHSI->getName()+".mask");
6533 Value *And = InsertNewInstBefore(AndI, ICI);
6535 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6536 And, Constant::getNullValue(And->getType()));
6541 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6542 case Instruction::AShr: {
6543 // Only handle equality comparisons of shift-by-constant.
6544 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6545 if (!ShAmt || !ICI.isEquality()) break;
6547 // Check that the shift amount is in range. If not, don't perform
6548 // undefined shifts. When the shift is visited it will be
6550 uint32_t TypeBits = RHSV.getBitWidth();
6551 if (ShAmt->uge(TypeBits))
6554 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6556 // If we are comparing against bits always shifted out, the
6557 // comparison cannot succeed.
6558 APInt Comp = RHSV << ShAmtVal;
6559 if (LHSI->getOpcode() == Instruction::LShr)
6560 Comp = Comp.lshr(ShAmtVal);
6562 Comp = Comp.ashr(ShAmtVal);
6564 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6565 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6566 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6567 return ReplaceInstUsesWith(ICI, Cst);
6570 // Otherwise, check to see if the bits shifted out are known to be zero.
6571 // If so, we can compare against the unshifted value:
6572 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6573 if (LHSI->hasOneUse() &&
6574 MaskedValueIsZero(LHSI->getOperand(0),
6575 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6576 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6577 ConstantExpr::getShl(RHS, ShAmt));
6580 if (LHSI->hasOneUse()) {
6581 // Otherwise strength reduce the shift into an and.
6582 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6583 Constant *Mask = ConstantInt::get(Val);
6586 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6587 Mask, LHSI->getName()+".mask");
6588 Value *And = InsertNewInstBefore(AndI, ICI);
6589 return new ICmpInst(ICI.getPredicate(), And,
6590 ConstantExpr::getShl(RHS, ShAmt));
6595 case Instruction::SDiv:
6596 case Instruction::UDiv:
6597 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6598 // Fold this div into the comparison, producing a range check.
6599 // Determine, based on the divide type, what the range is being
6600 // checked. If there is an overflow on the low or high side, remember
6601 // it, otherwise compute the range [low, hi) bounding the new value.
6602 // See: InsertRangeTest above for the kinds of replacements possible.
6603 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6604 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6609 case Instruction::Add:
6610 // Fold: icmp pred (add, X, C1), C2
6612 if (!ICI.isEquality()) {
6613 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6615 const APInt &LHSV = LHSC->getValue();
6617 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6620 if (ICI.isSignedPredicate()) {
6621 if (CR.getLower().isSignBit()) {
6622 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6623 ConstantInt::get(CR.getUpper()));
6624 } else if (CR.getUpper().isSignBit()) {
6625 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6626 ConstantInt::get(CR.getLower()));
6629 if (CR.getLower().isMinValue()) {
6630 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6631 ConstantInt::get(CR.getUpper()));
6632 } else if (CR.getUpper().isMinValue()) {
6633 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6634 ConstantInt::get(CR.getLower()));
6641 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6642 if (ICI.isEquality()) {
6643 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6645 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6646 // the second operand is a constant, simplify a bit.
6647 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6648 switch (BO->getOpcode()) {
6649 case Instruction::SRem:
6650 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6651 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6652 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6653 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6654 Instruction *NewRem =
6655 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6657 InsertNewInstBefore(NewRem, ICI);
6658 return new ICmpInst(ICI.getPredicate(), NewRem,
6659 Constant::getNullValue(BO->getType()));
6663 case Instruction::Add:
6664 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6665 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6666 if (BO->hasOneUse())
6667 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6668 Subtract(RHS, BOp1C));
6669 } else if (RHSV == 0) {
6670 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6671 // efficiently invertible, or if the add has just this one use.
6672 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6674 if (Value *NegVal = dyn_castNegVal(BOp1))
6675 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6676 else if (Value *NegVal = dyn_castNegVal(BOp0))
6677 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6678 else if (BO->hasOneUse()) {
6679 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6680 InsertNewInstBefore(Neg, ICI);
6682 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6686 case Instruction::Xor:
6687 // For the xor case, we can xor two constants together, eliminating
6688 // the explicit xor.
6689 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6690 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6691 ConstantExpr::getXor(RHS, BOC));
6694 case Instruction::Sub:
6695 // Replace (([sub|xor] A, B) != 0) with (A != B)
6697 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6701 case Instruction::Or:
6702 // If bits are being or'd in that are not present in the constant we
6703 // are comparing against, then the comparison could never succeed!
6704 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6705 Constant *NotCI = ConstantExpr::getNot(RHS);
6706 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6707 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6712 case Instruction::And:
6713 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6714 // If bits are being compared against that are and'd out, then the
6715 // comparison can never succeed!
6716 if ((RHSV & ~BOC->getValue()) != 0)
6717 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6720 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6721 if (RHS == BOC && RHSV.isPowerOf2())
6722 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6723 ICmpInst::ICMP_NE, LHSI,
6724 Constant::getNullValue(RHS->getType()));
6726 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6727 if (isSignBit(BOC)) {
6728 Value *X = BO->getOperand(0);
6729 Constant *Zero = Constant::getNullValue(X->getType());
6730 ICmpInst::Predicate pred = isICMP_NE ?
6731 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6732 return new ICmpInst(pred, X, Zero);
6735 // ((X & ~7) == 0) --> X < 8
6736 if (RHSV == 0 && isHighOnes(BOC)) {
6737 Value *X = BO->getOperand(0);
6738 Constant *NegX = ConstantExpr::getNeg(BOC);
6739 ICmpInst::Predicate pred = isICMP_NE ?
6740 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6741 return new ICmpInst(pred, X, NegX);
6746 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6747 // Handle icmp {eq|ne} <intrinsic>, intcst.
6748 if (II->getIntrinsicID() == Intrinsic::bswap) {
6750 ICI.setOperand(0, II->getOperand(1));
6751 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6755 } else { // Not a ICMP_EQ/ICMP_NE
6756 // If the LHS is a cast from an integral value of the same size,
6757 // then since we know the RHS is a constant, try to simlify.
6758 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6759 Value *CastOp = Cast->getOperand(0);
6760 const Type *SrcTy = CastOp->getType();
6761 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6762 if (SrcTy->isInteger() &&
6763 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6764 // If this is an unsigned comparison, try to make the comparison use
6765 // smaller constant values.
6766 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6767 // X u< 128 => X s> -1
6768 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6769 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6770 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6771 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6772 // X u> 127 => X s< 0
6773 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6774 Constant::getNullValue(SrcTy));
6782 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6783 /// We only handle extending casts so far.
6785 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6786 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6787 Value *LHSCIOp = LHSCI->getOperand(0);
6788 const Type *SrcTy = LHSCIOp->getType();
6789 const Type *DestTy = LHSCI->getType();
6792 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6793 // integer type is the same size as the pointer type.
6794 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6795 getTargetData().getPointerSizeInBits() ==
6796 cast<IntegerType>(DestTy)->getBitWidth()) {
6798 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6799 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6800 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6801 RHSOp = RHSC->getOperand(0);
6802 // If the pointer types don't match, insert a bitcast.
6803 if (LHSCIOp->getType() != RHSOp->getType())
6804 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6808 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6811 // The code below only handles extension cast instructions, so far.
6813 if (LHSCI->getOpcode() != Instruction::ZExt &&
6814 LHSCI->getOpcode() != Instruction::SExt)
6817 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6818 bool isSignedCmp = ICI.isSignedPredicate();
6820 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6821 // Not an extension from the same type?
6822 RHSCIOp = CI->getOperand(0);
6823 if (RHSCIOp->getType() != LHSCIOp->getType())
6826 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6827 // and the other is a zext), then we can't handle this.
6828 if (CI->getOpcode() != LHSCI->getOpcode())
6831 // Deal with equality cases early.
6832 if (ICI.isEquality())
6833 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6835 // A signed comparison of sign extended values simplifies into a
6836 // signed comparison.
6837 if (isSignedCmp && isSignedExt)
6838 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6840 // The other three cases all fold into an unsigned comparison.
6841 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6844 // If we aren't dealing with a constant on the RHS, exit early
6845 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6849 // Compute the constant that would happen if we truncated to SrcTy then
6850 // reextended to DestTy.
6851 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6852 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6854 // If the re-extended constant didn't change...
6856 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6857 // For example, we might have:
6858 // %A = sext short %X to uint
6859 // %B = icmp ugt uint %A, 1330
6860 // It is incorrect to transform this into
6861 // %B = icmp ugt short %X, 1330
6862 // because %A may have negative value.
6864 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6865 // OR operation is EQ/NE.
6866 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6867 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6872 // The re-extended constant changed so the constant cannot be represented
6873 // in the shorter type. Consequently, we cannot emit a simple comparison.
6875 // First, handle some easy cases. We know the result cannot be equal at this
6876 // point so handle the ICI.isEquality() cases
6877 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6878 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6879 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6880 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6882 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6883 // should have been folded away previously and not enter in here.
6886 // We're performing a signed comparison.
6887 if (cast<ConstantInt>(CI)->getValue().isNegative())
6888 Result = ConstantInt::getFalse(); // X < (small) --> false
6890 Result = ConstantInt::getTrue(); // X < (large) --> true
6892 // We're performing an unsigned comparison.
6894 // We're performing an unsigned comp with a sign extended value.
6895 // This is true if the input is >= 0. [aka >s -1]
6896 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6897 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6898 NegOne, ICI.getName()), ICI);
6900 // Unsigned extend & unsigned compare -> always true.
6901 Result = ConstantInt::getTrue();
6905 // Finally, return the value computed.
6906 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6907 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6908 return ReplaceInstUsesWith(ICI, Result);
6910 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6911 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6912 "ICmp should be folded!");
6913 if (Constant *CI = dyn_cast<Constant>(Result))
6914 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6916 return BinaryOperator::CreateNot(Result);
6920 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6921 return commonShiftTransforms(I);
6924 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6925 return commonShiftTransforms(I);
6928 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6929 if (Instruction *R = commonShiftTransforms(I))
6932 Value *Op0 = I.getOperand(0);
6934 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6935 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6936 if (CSI->isAllOnesValue())
6937 return ReplaceInstUsesWith(I, CSI);
6939 // See if we can turn a signed shr into an unsigned shr.
6940 if (MaskedValueIsZero(Op0,
6941 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6942 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6947 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6948 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6949 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6951 // shl X, 0 == X and shr X, 0 == X
6952 // shl 0, X == 0 and shr 0, X == 0
6953 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6954 Op0 == Constant::getNullValue(Op0->getType()))
6955 return ReplaceInstUsesWith(I, Op0);
6957 if (isa<UndefValue>(Op0)) {
6958 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6959 return ReplaceInstUsesWith(I, Op0);
6960 else // undef << X -> 0, undef >>u X -> 0
6961 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6963 if (isa<UndefValue>(Op1)) {
6964 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6965 return ReplaceInstUsesWith(I, Op0);
6966 else // X << undef, X >>u undef -> 0
6967 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6970 // Try to fold constant and into select arguments.
6971 if (isa<Constant>(Op0))
6972 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6973 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6976 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6977 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6982 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6983 BinaryOperator &I) {
6984 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6986 // See if we can simplify any instructions used by the instruction whose sole
6987 // purpose is to compute bits we don't care about.
6988 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6989 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6990 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6991 KnownZero, KnownOne))
6994 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6995 // of a signed value.
6997 if (Op1->uge(TypeBits)) {
6998 if (I.getOpcode() != Instruction::AShr)
6999 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7001 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7006 // ((X*C1) << C2) == (X * (C1 << C2))
7007 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7008 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7009 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7010 return BinaryOperator::CreateMul(BO->getOperand(0),
7011 ConstantExpr::getShl(BOOp, Op1));
7013 // Try to fold constant and into select arguments.
7014 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7015 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7017 if (isa<PHINode>(Op0))
7018 if (Instruction *NV = FoldOpIntoPhi(I))
7021 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7022 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7023 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7024 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7025 // place. Don't try to do this transformation in this case. Also, we
7026 // require that the input operand is a shift-by-constant so that we have
7027 // confidence that the shifts will get folded together. We could do this
7028 // xform in more cases, but it is unlikely to be profitable.
7029 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7030 isa<ConstantInt>(TrOp->getOperand(1))) {
7031 // Okay, we'll do this xform. Make the shift of shift.
7032 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7033 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7035 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7037 // For logical shifts, the truncation has the effect of making the high
7038 // part of the register be zeros. Emulate this by inserting an AND to
7039 // clear the top bits as needed. This 'and' will usually be zapped by
7040 // other xforms later if dead.
7041 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7042 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7043 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7045 // The mask we constructed says what the trunc would do if occurring
7046 // between the shifts. We want to know the effect *after* the second
7047 // shift. We know that it is a logical shift by a constant, so adjust the
7048 // mask as appropriate.
7049 if (I.getOpcode() == Instruction::Shl)
7050 MaskV <<= Op1->getZExtValue();
7052 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7053 MaskV = MaskV.lshr(Op1->getZExtValue());
7056 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7058 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7060 // Return the value truncated to the interesting size.
7061 return new TruncInst(And, I.getType());
7065 if (Op0->hasOneUse()) {
7066 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7067 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7070 switch (Op0BO->getOpcode()) {
7072 case Instruction::Add:
7073 case Instruction::And:
7074 case Instruction::Or:
7075 case Instruction::Xor: {
7076 // These operators commute.
7077 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7078 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7079 match(Op0BO->getOperand(1),
7080 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
7081 Instruction *YS = BinaryOperator::CreateShl(
7082 Op0BO->getOperand(0), Op1,
7084 InsertNewInstBefore(YS, I); // (Y << C)
7086 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7087 Op0BO->getOperand(1)->getName());
7088 InsertNewInstBefore(X, I); // (X + (Y << C))
7089 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7090 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7091 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7094 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7095 Value *Op0BOOp1 = Op0BO->getOperand(1);
7096 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7098 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
7099 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
7101 Instruction *YS = BinaryOperator::CreateShl(
7102 Op0BO->getOperand(0), Op1,
7104 InsertNewInstBefore(YS, I); // (Y << C)
7106 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7107 V1->getName()+".mask");
7108 InsertNewInstBefore(XM, I); // X & (CC << C)
7110 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7115 case Instruction::Sub: {
7116 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7117 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7118 match(Op0BO->getOperand(0),
7119 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
7120 Instruction *YS = BinaryOperator::CreateShl(
7121 Op0BO->getOperand(1), Op1,
7123 InsertNewInstBefore(YS, I); // (Y << C)
7125 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7126 Op0BO->getOperand(0)->getName());
7127 InsertNewInstBefore(X, I); // (X + (Y << C))
7128 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7129 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7130 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7133 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7134 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7135 match(Op0BO->getOperand(0),
7136 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7137 m_ConstantInt(CC))) && V2 == Op1 &&
7138 cast<BinaryOperator>(Op0BO->getOperand(0))
7139 ->getOperand(0)->hasOneUse()) {
7140 Instruction *YS = BinaryOperator::CreateShl(
7141 Op0BO->getOperand(1), Op1,
7143 InsertNewInstBefore(YS, I); // (Y << C)
7145 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7146 V1->getName()+".mask");
7147 InsertNewInstBefore(XM, I); // X & (CC << C)
7149 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7157 // If the operand is an bitwise operator with a constant RHS, and the
7158 // shift is the only use, we can pull it out of the shift.
7159 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7160 bool isValid = true; // Valid only for And, Or, Xor
7161 bool highBitSet = false; // Transform if high bit of constant set?
7163 switch (Op0BO->getOpcode()) {
7164 default: isValid = false; break; // Do not perform transform!
7165 case Instruction::Add:
7166 isValid = isLeftShift;
7168 case Instruction::Or:
7169 case Instruction::Xor:
7172 case Instruction::And:
7177 // If this is a signed shift right, and the high bit is modified
7178 // by the logical operation, do not perform the transformation.
7179 // The highBitSet boolean indicates the value of the high bit of
7180 // the constant which would cause it to be modified for this
7183 if (isValid && I.getOpcode() == Instruction::AShr)
7184 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7187 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7189 Instruction *NewShift =
7190 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7191 InsertNewInstBefore(NewShift, I);
7192 NewShift->takeName(Op0BO);
7194 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7201 // Find out if this is a shift of a shift by a constant.
7202 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7203 if (ShiftOp && !ShiftOp->isShift())
7206 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7207 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7208 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7209 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7210 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7211 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7212 Value *X = ShiftOp->getOperand(0);
7214 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7215 if (AmtSum > TypeBits)
7218 const IntegerType *Ty = cast<IntegerType>(I.getType());
7220 // Check for (X << c1) << c2 and (X >> c1) >> c2
7221 if (I.getOpcode() == ShiftOp->getOpcode()) {
7222 return BinaryOperator::Create(I.getOpcode(), X,
7223 ConstantInt::get(Ty, AmtSum));
7224 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7225 I.getOpcode() == Instruction::AShr) {
7226 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7227 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7228 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7229 I.getOpcode() == Instruction::LShr) {
7230 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7231 Instruction *Shift =
7232 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7233 InsertNewInstBefore(Shift, I);
7235 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7236 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7239 // Okay, if we get here, one shift must be left, and the other shift must be
7240 // right. See if the amounts are equal.
7241 if (ShiftAmt1 == ShiftAmt2) {
7242 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7243 if (I.getOpcode() == Instruction::Shl) {
7244 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7245 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7247 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7248 if (I.getOpcode() == Instruction::LShr) {
7249 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7250 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7252 // We can simplify ((X << C) >>s C) into a trunc + sext.
7253 // NOTE: we could do this for any C, but that would make 'unusual' integer
7254 // types. For now, just stick to ones well-supported by the code
7256 const Type *SExtType = 0;
7257 switch (Ty->getBitWidth() - ShiftAmt1) {
7264 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7269 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7270 InsertNewInstBefore(NewTrunc, I);
7271 return new SExtInst(NewTrunc, Ty);
7273 // Otherwise, we can't handle it yet.
7274 } else if (ShiftAmt1 < ShiftAmt2) {
7275 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7277 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7278 if (I.getOpcode() == Instruction::Shl) {
7279 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7280 ShiftOp->getOpcode() == Instruction::AShr);
7281 Instruction *Shift =
7282 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7283 InsertNewInstBefore(Shift, I);
7285 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7286 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7289 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7290 if (I.getOpcode() == Instruction::LShr) {
7291 assert(ShiftOp->getOpcode() == Instruction::Shl);
7292 Instruction *Shift =
7293 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7294 InsertNewInstBefore(Shift, I);
7296 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7297 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7300 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7302 assert(ShiftAmt2 < ShiftAmt1);
7303 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7305 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7306 if (I.getOpcode() == Instruction::Shl) {
7307 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7308 ShiftOp->getOpcode() == Instruction::AShr);
7309 Instruction *Shift =
7310 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7311 ConstantInt::get(Ty, ShiftDiff));
7312 InsertNewInstBefore(Shift, I);
7314 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7315 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7318 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7319 if (I.getOpcode() == Instruction::LShr) {
7320 assert(ShiftOp->getOpcode() == Instruction::Shl);
7321 Instruction *Shift =
7322 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7323 InsertNewInstBefore(Shift, I);
7325 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7326 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7329 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7336 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7337 /// expression. If so, decompose it, returning some value X, such that Val is
7340 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7342 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7343 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7344 Offset = CI->getZExtValue();
7346 return ConstantInt::get(Type::Int32Ty, 0);
7347 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7348 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7349 if (I->getOpcode() == Instruction::Shl) {
7350 // This is a value scaled by '1 << the shift amt'.
7351 Scale = 1U << RHS->getZExtValue();
7353 return I->getOperand(0);
7354 } else if (I->getOpcode() == Instruction::Mul) {
7355 // This value is scaled by 'RHS'.
7356 Scale = RHS->getZExtValue();
7358 return I->getOperand(0);
7359 } else if (I->getOpcode() == Instruction::Add) {
7360 // We have X+C. Check to see if we really have (X*C2)+C1,
7361 // where C1 is divisible by C2.
7364 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7365 Offset += RHS->getZExtValue();
7372 // Otherwise, we can't look past this.
7379 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7380 /// try to eliminate the cast by moving the type information into the alloc.
7381 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7382 AllocationInst &AI) {
7383 const PointerType *PTy = cast<PointerType>(CI.getType());
7385 // Remove any uses of AI that are dead.
7386 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7388 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7389 Instruction *User = cast<Instruction>(*UI++);
7390 if (isInstructionTriviallyDead(User)) {
7391 while (UI != E && *UI == User)
7392 ++UI; // If this instruction uses AI more than once, don't break UI.
7395 DOUT << "IC: DCE: " << *User;
7396 EraseInstFromFunction(*User);
7400 // Get the type really allocated and the type casted to.
7401 const Type *AllocElTy = AI.getAllocatedType();
7402 const Type *CastElTy = PTy->getElementType();
7403 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7405 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7406 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7407 if (CastElTyAlign < AllocElTyAlign) return 0;
7409 // If the allocation has multiple uses, only promote it if we are strictly
7410 // increasing the alignment of the resultant allocation. If we keep it the
7411 // same, we open the door to infinite loops of various kinds.
7412 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7414 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7415 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7416 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7418 // See if we can satisfy the modulus by pulling a scale out of the array
7420 unsigned ArraySizeScale;
7422 Value *NumElements = // See if the array size is a decomposable linear expr.
7423 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7425 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7427 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7428 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7430 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7435 // If the allocation size is constant, form a constant mul expression
7436 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7437 if (isa<ConstantInt>(NumElements))
7438 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7439 // otherwise multiply the amount and the number of elements
7440 else if (Scale != 1) {
7441 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7442 Amt = InsertNewInstBefore(Tmp, AI);
7446 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7447 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7448 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7449 Amt = InsertNewInstBefore(Tmp, AI);
7452 AllocationInst *New;
7453 if (isa<MallocInst>(AI))
7454 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7456 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7457 InsertNewInstBefore(New, AI);
7460 // If the allocation has multiple uses, insert a cast and change all things
7461 // that used it to use the new cast. This will also hack on CI, but it will
7463 if (!AI.hasOneUse()) {
7464 AddUsesToWorkList(AI);
7465 // New is the allocation instruction, pointer typed. AI is the original
7466 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7467 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7468 InsertNewInstBefore(NewCast, AI);
7469 AI.replaceAllUsesWith(NewCast);
7471 return ReplaceInstUsesWith(CI, New);
7474 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7475 /// and return it as type Ty without inserting any new casts and without
7476 /// changing the computed value. This is used by code that tries to decide
7477 /// whether promoting or shrinking integer operations to wider or smaller types
7478 /// will allow us to eliminate a truncate or extend.
7480 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7481 /// extension operation if Ty is larger.
7482 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7484 int &NumCastsRemoved) {
7485 // We can always evaluate constants in another type.
7486 if (isa<ConstantInt>(V))
7489 Instruction *I = dyn_cast<Instruction>(V);
7490 if (!I) return false;
7492 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7494 // If this is an extension or truncate, we can often eliminate it.
7495 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7496 // If this is a cast from the destination type, we can trivially eliminate
7497 // it, and this will remove a cast overall.
7498 if (I->getOperand(0)->getType() == Ty) {
7499 // If the first operand is itself a cast, and is eliminable, do not count
7500 // this as an eliminable cast. We would prefer to eliminate those two
7502 if (!isa<CastInst>(I->getOperand(0)))
7508 // We can't extend or shrink something that has multiple uses: doing so would
7509 // require duplicating the instruction in general, which isn't profitable.
7510 if (!I->hasOneUse()) return false;
7512 switch (I->getOpcode()) {
7513 case Instruction::Add:
7514 case Instruction::Sub:
7515 case Instruction::And:
7516 case Instruction::Or:
7517 case Instruction::Xor:
7518 // These operators can all arbitrarily be extended or truncated.
7519 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7521 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7524 case Instruction::Mul:
7525 // A multiply can be truncated by truncating its operands.
7526 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
7527 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7529 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7532 case Instruction::Shl:
7533 // If we are truncating the result of this SHL, and if it's a shift of a
7534 // constant amount, we can always perform a SHL in a smaller type.
7535 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7536 uint32_t BitWidth = Ty->getBitWidth();
7537 if (BitWidth < OrigTy->getBitWidth() &&
7538 CI->getLimitedValue(BitWidth) < BitWidth)
7539 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7543 case Instruction::LShr:
7544 // If this is a truncate of a logical shr, we can truncate it to a smaller
7545 // lshr iff we know that the bits we would otherwise be shifting in are
7547 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7548 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7549 uint32_t BitWidth = Ty->getBitWidth();
7550 if (BitWidth < OrigBitWidth &&
7551 MaskedValueIsZero(I->getOperand(0),
7552 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7553 CI->getLimitedValue(BitWidth) < BitWidth) {
7554 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7559 case Instruction::ZExt:
7560 case Instruction::SExt:
7561 case Instruction::Trunc:
7562 // If this is the same kind of case as our original (e.g. zext+zext), we
7563 // can safely replace it. Note that replacing it does not reduce the number
7564 // of casts in the input.
7565 if (I->getOpcode() == CastOpc)
7570 // TODO: Can handle more cases here.
7577 /// EvaluateInDifferentType - Given an expression that
7578 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7579 /// evaluate the expression.
7580 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7582 if (Constant *C = dyn_cast<Constant>(V))
7583 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7585 // Otherwise, it must be an instruction.
7586 Instruction *I = cast<Instruction>(V);
7587 Instruction *Res = 0;
7588 switch (I->getOpcode()) {
7589 case Instruction::Add:
7590 case Instruction::Sub:
7591 case Instruction::Mul:
7592 case Instruction::And:
7593 case Instruction::Or:
7594 case Instruction::Xor:
7595 case Instruction::AShr:
7596 case Instruction::LShr:
7597 case Instruction::Shl: {
7598 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7599 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7600 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7601 LHS, RHS, I->getName());
7604 case Instruction::Trunc:
7605 case Instruction::ZExt:
7606 case Instruction::SExt:
7607 // If the source type of the cast is the type we're trying for then we can
7608 // just return the source. There's no need to insert it because it is not
7610 if (I->getOperand(0)->getType() == Ty)
7611 return I->getOperand(0);
7613 // Otherwise, must be the same type of case, so just reinsert a new one.
7614 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7618 // TODO: Can handle more cases here.
7619 assert(0 && "Unreachable!");
7623 return InsertNewInstBefore(Res, *I);
7626 /// @brief Implement the transforms common to all CastInst visitors.
7627 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7628 Value *Src = CI.getOperand(0);
7630 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7631 // eliminate it now.
7632 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7633 if (Instruction::CastOps opc =
7634 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7635 // The first cast (CSrc) is eliminable so we need to fix up or replace
7636 // the second cast (CI). CSrc will then have a good chance of being dead.
7637 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7641 // If we are casting a select then fold the cast into the select
7642 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7643 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7646 // If we are casting a PHI then fold the cast into the PHI
7647 if (isa<PHINode>(Src))
7648 if (Instruction *NV = FoldOpIntoPhi(CI))
7654 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7655 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7656 Value *Src = CI.getOperand(0);
7658 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7659 // If casting the result of a getelementptr instruction with no offset, turn
7660 // this into a cast of the original pointer!
7661 if (GEP->hasAllZeroIndices()) {
7662 // Changing the cast operand is usually not a good idea but it is safe
7663 // here because the pointer operand is being replaced with another
7664 // pointer operand so the opcode doesn't need to change.
7666 CI.setOperand(0, GEP->getOperand(0));
7670 // If the GEP has a single use, and the base pointer is a bitcast, and the
7671 // GEP computes a constant offset, see if we can convert these three
7672 // instructions into fewer. This typically happens with unions and other
7673 // non-type-safe code.
7674 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7675 if (GEP->hasAllConstantIndices()) {
7676 // We are guaranteed to get a constant from EmitGEPOffset.
7677 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7678 int64_t Offset = OffsetV->getSExtValue();
7680 // Get the base pointer input of the bitcast, and the type it points to.
7681 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7682 const Type *GEPIdxTy =
7683 cast<PointerType>(OrigBase->getType())->getElementType();
7684 if (GEPIdxTy->isSized()) {
7685 SmallVector<Value*, 8> NewIndices;
7687 // Start with the index over the outer type. Note that the type size
7688 // might be zero (even if the offset isn't zero) if the indexed type
7689 // is something like [0 x {int, int}]
7690 const Type *IntPtrTy = TD->getIntPtrType();
7691 int64_t FirstIdx = 0;
7692 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7693 FirstIdx = Offset/TySize;
7696 // Handle silly modulus not returning values values [0..TySize).
7700 assert(Offset >= 0);
7702 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7705 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7707 // Index into the types. If we fail, set OrigBase to null.
7709 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7710 const StructLayout *SL = TD->getStructLayout(STy);
7711 if (Offset < (int64_t)SL->getSizeInBytes()) {
7712 unsigned Elt = SL->getElementContainingOffset(Offset);
7713 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7715 Offset -= SL->getElementOffset(Elt);
7716 GEPIdxTy = STy->getElementType(Elt);
7718 // Otherwise, we can't index into this, bail out.
7722 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7723 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7724 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7725 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7728 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7730 GEPIdxTy = STy->getElementType();
7732 // Otherwise, we can't index into this, bail out.
7738 // If we were able to index down into an element, create the GEP
7739 // and bitcast the result. This eliminates one bitcast, potentially
7741 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7743 NewIndices.end(), "");
7744 InsertNewInstBefore(NGEP, CI);
7745 NGEP->takeName(GEP);
7747 if (isa<BitCastInst>(CI))
7748 return new BitCastInst(NGEP, CI.getType());
7749 assert(isa<PtrToIntInst>(CI));
7750 return new PtrToIntInst(NGEP, CI.getType());
7757 return commonCastTransforms(CI);
7762 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7763 /// integer types. This function implements the common transforms for all those
7765 /// @brief Implement the transforms common to CastInst with integer operands
7766 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7767 if (Instruction *Result = commonCastTransforms(CI))
7770 Value *Src = CI.getOperand(0);
7771 const Type *SrcTy = Src->getType();
7772 const Type *DestTy = CI.getType();
7773 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7774 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7776 // See if we can simplify any instructions used by the LHS whose sole
7777 // purpose is to compute bits we don't care about.
7778 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7779 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7780 KnownZero, KnownOne))
7783 // If the source isn't an instruction or has more than one use then we
7784 // can't do anything more.
7785 Instruction *SrcI = dyn_cast<Instruction>(Src);
7786 if (!SrcI || !Src->hasOneUse())
7789 // Attempt to propagate the cast into the instruction for int->int casts.
7790 int NumCastsRemoved = 0;
7791 if (!isa<BitCastInst>(CI) &&
7792 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7793 CI.getOpcode(), NumCastsRemoved)) {
7794 // If this cast is a truncate, evaluting in a different type always
7795 // eliminates the cast, so it is always a win. If this is a zero-extension,
7796 // we need to do an AND to maintain the clear top-part of the computation,
7797 // so we require that the input have eliminated at least one cast. If this
7798 // is a sign extension, we insert two new casts (to do the extension) so we
7799 // require that two casts have been eliminated.
7801 switch (CI.getOpcode()) {
7803 // All the others use floating point so we shouldn't actually
7804 // get here because of the check above.
7805 assert(0 && "Unknown cast type");
7806 case Instruction::Trunc:
7809 case Instruction::ZExt:
7810 DoXForm = NumCastsRemoved >= 1;
7812 case Instruction::SExt:
7813 DoXForm = NumCastsRemoved >= 2;
7818 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7819 CI.getOpcode() == Instruction::SExt);
7820 assert(Res->getType() == DestTy);
7821 switch (CI.getOpcode()) {
7822 default: assert(0 && "Unknown cast type!");
7823 case Instruction::Trunc:
7824 case Instruction::BitCast:
7825 // Just replace this cast with the result.
7826 return ReplaceInstUsesWith(CI, Res);
7827 case Instruction::ZExt: {
7828 // We need to emit an AND to clear the high bits.
7829 assert(SrcBitSize < DestBitSize && "Not a zext?");
7830 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7832 return BinaryOperator::CreateAnd(Res, C);
7834 case Instruction::SExt:
7835 // We need to emit a cast to truncate, then a cast to sext.
7836 return CastInst::Create(Instruction::SExt,
7837 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7843 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7844 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7846 switch (SrcI->getOpcode()) {
7847 case Instruction::Add:
7848 case Instruction::Mul:
7849 case Instruction::And:
7850 case Instruction::Or:
7851 case Instruction::Xor:
7852 // If we are discarding information, rewrite.
7853 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7854 // Don't insert two casts if they cannot be eliminated. We allow
7855 // two casts to be inserted if the sizes are the same. This could
7856 // only be converting signedness, which is a noop.
7857 if (DestBitSize == SrcBitSize ||
7858 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7859 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7860 Instruction::CastOps opcode = CI.getOpcode();
7861 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7862 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7863 return BinaryOperator::Create(
7864 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7868 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7869 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7870 SrcI->getOpcode() == Instruction::Xor &&
7871 Op1 == ConstantInt::getTrue() &&
7872 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7873 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7874 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7877 case Instruction::SDiv:
7878 case Instruction::UDiv:
7879 case Instruction::SRem:
7880 case Instruction::URem:
7881 // If we are just changing the sign, rewrite.
7882 if (DestBitSize == SrcBitSize) {
7883 // Don't insert two casts if they cannot be eliminated. We allow
7884 // two casts to be inserted if the sizes are the same. This could
7885 // only be converting signedness, which is a noop.
7886 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7887 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7888 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7890 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7892 return BinaryOperator::Create(
7893 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7898 case Instruction::Shl:
7899 // Allow changing the sign of the source operand. Do not allow
7900 // changing the size of the shift, UNLESS the shift amount is a
7901 // constant. We must not change variable sized shifts to a smaller
7902 // size, because it is undefined to shift more bits out than exist
7904 if (DestBitSize == SrcBitSize ||
7905 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7906 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7907 Instruction::BitCast : Instruction::Trunc);
7908 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7909 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7910 return BinaryOperator::CreateShl(Op0c, Op1c);
7913 case Instruction::AShr:
7914 // If this is a signed shr, and if all bits shifted in are about to be
7915 // truncated off, turn it into an unsigned shr to allow greater
7917 if (DestBitSize < SrcBitSize &&
7918 isa<ConstantInt>(Op1)) {
7919 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7920 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7921 // Insert the new logical shift right.
7922 return BinaryOperator::CreateLShr(Op0, Op1);
7930 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7931 if (Instruction *Result = commonIntCastTransforms(CI))
7934 Value *Src = CI.getOperand(0);
7935 const Type *Ty = CI.getType();
7936 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7937 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7939 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7940 switch (SrcI->getOpcode()) {
7942 case Instruction::LShr:
7943 // We can shrink lshr to something smaller if we know the bits shifted in
7944 // are already zeros.
7945 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7946 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7948 // Get a mask for the bits shifting in.
7949 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7950 Value* SrcIOp0 = SrcI->getOperand(0);
7951 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7952 if (ShAmt >= DestBitWidth) // All zeros.
7953 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7955 // Okay, we can shrink this. Truncate the input, then return a new
7957 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7958 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7960 return BinaryOperator::CreateLShr(V1, V2);
7962 } else { // This is a variable shr.
7964 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7965 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7966 // loop-invariant and CSE'd.
7967 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7968 Value *One = ConstantInt::get(SrcI->getType(), 1);
7970 Value *V = InsertNewInstBefore(
7971 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7973 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7974 SrcI->getOperand(0),
7976 Value *Zero = Constant::getNullValue(V->getType());
7977 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7987 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7988 /// in order to eliminate the icmp.
7989 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7991 // If we are just checking for a icmp eq of a single bit and zext'ing it
7992 // to an integer, then shift the bit to the appropriate place and then
7993 // cast to integer to avoid the comparison.
7994 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7995 const APInt &Op1CV = Op1C->getValue();
7997 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7998 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7999 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8000 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8001 if (!DoXform) return ICI;
8003 Value *In = ICI->getOperand(0);
8004 Value *Sh = ConstantInt::get(In->getType(),
8005 In->getType()->getPrimitiveSizeInBits()-1);
8006 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8007 In->getName()+".lobit"),
8009 if (In->getType() != CI.getType())
8010 In = CastInst::CreateIntegerCast(In, CI.getType(),
8011 false/*ZExt*/, "tmp", &CI);
8013 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8014 Constant *One = ConstantInt::get(In->getType(), 1);
8015 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8016 In->getName()+".not"),
8020 return ReplaceInstUsesWith(CI, In);
8025 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8026 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8027 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8028 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8029 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8030 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8031 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8032 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8033 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8034 // This only works for EQ and NE
8035 ICI->isEquality()) {
8036 // If Op1C some other power of two, convert:
8037 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8038 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8039 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8040 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8042 APInt KnownZeroMask(~KnownZero);
8043 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8044 if (!DoXform) return ICI;
8046 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8047 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8048 // (X&4) == 2 --> false
8049 // (X&4) != 2 --> true
8050 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8051 Res = ConstantExpr::getZExt(Res, CI.getType());
8052 return ReplaceInstUsesWith(CI, Res);
8055 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8056 Value *In = ICI->getOperand(0);
8058 // Perform a logical shr by shiftamt.
8059 // Insert the shift to put the result in the low bit.
8060 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8061 ConstantInt::get(In->getType(), ShiftAmt),
8062 In->getName()+".lobit"), CI);
8065 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8066 Constant *One = ConstantInt::get(In->getType(), 1);
8067 In = BinaryOperator::CreateXor(In, One, "tmp");
8068 InsertNewInstBefore(cast<Instruction>(In), CI);
8071 if (CI.getType() == In->getType())
8072 return ReplaceInstUsesWith(CI, In);
8074 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8082 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8083 // If one of the common conversion will work ..
8084 if (Instruction *Result = commonIntCastTransforms(CI))
8087 Value *Src = CI.getOperand(0);
8089 // If this is a cast of a cast
8090 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8091 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8092 // types and if the sizes are just right we can convert this into a logical
8093 // 'and' which will be much cheaper than the pair of casts.
8094 if (isa<TruncInst>(CSrc)) {
8095 // Get the sizes of the types involved
8096 Value *A = CSrc->getOperand(0);
8097 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8098 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8099 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8100 // If we're actually extending zero bits and the trunc is a no-op
8101 if (MidSize < DstSize && SrcSize == DstSize) {
8102 // Replace both of the casts with an And of the type mask.
8103 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8104 Constant *AndConst = ConstantInt::get(AndValue);
8106 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8107 // Unfortunately, if the type changed, we need to cast it back.
8108 if (And->getType() != CI.getType()) {
8109 And->setName(CSrc->getName()+".mask");
8110 InsertNewInstBefore(And, CI);
8111 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8118 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8119 return transformZExtICmp(ICI, CI);
8121 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8122 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8123 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8124 // of the (zext icmp) will be transformed.
8125 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8126 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8127 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8128 (transformZExtICmp(LHS, CI, false) ||
8129 transformZExtICmp(RHS, CI, false))) {
8130 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8131 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8132 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8139 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8140 if (Instruction *I = commonIntCastTransforms(CI))
8143 Value *Src = CI.getOperand(0);
8145 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
8146 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
8147 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
8148 // If we are just checking for a icmp eq of a single bit and zext'ing it
8149 // to an integer, then shift the bit to the appropriate place and then
8150 // cast to integer to avoid the comparison.
8151 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8152 const APInt &Op1CV = Op1C->getValue();
8154 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8155 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8156 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8157 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
8158 Value *In = ICI->getOperand(0);
8159 Value *Sh = ConstantInt::get(In->getType(),
8160 In->getType()->getPrimitiveSizeInBits()-1);
8161 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8162 In->getName()+".lobit"),
8164 if (In->getType() != CI.getType())
8165 In = CastInst::CreateIntegerCast(In, CI.getType(),
8166 true/*SExt*/, "tmp", &CI);
8168 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
8169 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8170 In->getName()+".not"), CI);
8172 return ReplaceInstUsesWith(CI, In);
8177 // See if the value being truncated is already sign extended. If so, just
8178 // eliminate the trunc/sext pair.
8179 if (getOpcode(Src) == Instruction::Trunc) {
8180 Value *Op = cast<User>(Src)->getOperand(0);
8181 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8182 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8183 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8184 unsigned NumSignBits = ComputeNumSignBits(Op);
8186 if (OpBits == DestBits) {
8187 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8188 // bits, it is already ready.
8189 if (NumSignBits > DestBits-MidBits)
8190 return ReplaceInstUsesWith(CI, Op);
8191 } else if (OpBits < DestBits) {
8192 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8193 // bits, just sext from i32.
8194 if (NumSignBits > OpBits-MidBits)
8195 return new SExtInst(Op, CI.getType(), "tmp");
8197 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8198 // bits, just truncate to i32.
8199 if (NumSignBits > OpBits-MidBits)
8200 return new TruncInst(Op, CI.getType(), "tmp");
8207 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8208 /// in the specified FP type without changing its value.
8209 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8210 APFloat F = CFP->getValueAPF();
8211 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
8212 return ConstantFP::get(F);
8216 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8217 /// through it until we get the source value.
8218 static Value *LookThroughFPExtensions(Value *V) {
8219 if (Instruction *I = dyn_cast<Instruction>(V))
8220 if (I->getOpcode() == Instruction::FPExt)
8221 return LookThroughFPExtensions(I->getOperand(0));
8223 // If this value is a constant, return the constant in the smallest FP type
8224 // that can accurately represent it. This allows us to turn
8225 // (float)((double)X+2.0) into x+2.0f.
8226 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8227 if (CFP->getType() == Type::PPC_FP128Ty)
8228 return V; // No constant folding of this.
8229 // See if the value can be truncated to float and then reextended.
8230 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8232 if (CFP->getType() == Type::DoubleTy)
8233 return V; // Won't shrink.
8234 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8236 // Don't try to shrink to various long double types.
8242 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8243 if (Instruction *I = commonCastTransforms(CI))
8246 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8247 // smaller than the destination type, we can eliminate the truncate by doing
8248 // the add as the smaller type. This applies to add/sub/mul/div as well as
8249 // many builtins (sqrt, etc).
8250 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8251 if (OpI && OpI->hasOneUse()) {
8252 switch (OpI->getOpcode()) {
8254 case Instruction::Add:
8255 case Instruction::Sub:
8256 case Instruction::Mul:
8257 case Instruction::FDiv:
8258 case Instruction::FRem:
8259 const Type *SrcTy = OpI->getType();
8260 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8261 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8262 if (LHSTrunc->getType() != SrcTy &&
8263 RHSTrunc->getType() != SrcTy) {
8264 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8265 // If the source types were both smaller than the destination type of
8266 // the cast, do this xform.
8267 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8268 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8269 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8271 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8273 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8282 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8283 return commonCastTransforms(CI);
8286 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8287 // fptoui(uitofp(X)) --> X if the intermediate type has enough bits in its
8288 // mantissa to accurately represent all values of X. For example, do not
8289 // do this with i64->float->i64.
8290 if (UIToFPInst *SrcI = dyn_cast<UIToFPInst>(FI.getOperand(0)))
8291 if (SrcI->getOperand(0)->getType() == FI.getType() &&
8292 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8293 SrcI->getType()->getFPMantissaWidth())
8294 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
8296 return commonCastTransforms(FI);
8299 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8300 // fptosi(sitofp(X)) --> X if the intermediate type has enough bits in its
8301 // mantissa to accurately represent all values of X. For example, do not
8302 // do this with i64->float->i64.
8303 if (SIToFPInst *SrcI = dyn_cast<SIToFPInst>(FI.getOperand(0)))
8304 if (SrcI->getOperand(0)->getType() == FI.getType() &&
8305 (int)FI.getType()->getPrimitiveSizeInBits() <=
8306 SrcI->getType()->getFPMantissaWidth())
8307 return ReplaceInstUsesWith(FI, SrcI->getOperand(0));
8309 return commonCastTransforms(FI);
8312 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8313 return commonCastTransforms(CI);
8316 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8317 return commonCastTransforms(CI);
8320 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8321 return commonPointerCastTransforms(CI);
8324 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8325 if (Instruction *I = commonCastTransforms(CI))
8328 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8329 if (!DestPointee->isSized()) return 0;
8331 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8334 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8335 m_ConstantInt(Cst)))) {
8336 // If the source and destination operands have the same type, see if this
8337 // is a single-index GEP.
8338 if (X->getType() == CI.getType()) {
8339 // Get the size of the pointee type.
8340 uint64_t Size = TD->getABITypeSize(DestPointee);
8342 // Convert the constant to intptr type.
8343 APInt Offset = Cst->getValue();
8344 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8346 // If Offset is evenly divisible by Size, we can do this xform.
8347 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8348 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8349 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8352 // TODO: Could handle other cases, e.g. where add is indexing into field of
8354 } else if (CI.getOperand(0)->hasOneUse() &&
8355 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8356 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8357 // "inttoptr+GEP" instead of "add+intptr".
8359 // Get the size of the pointee type.
8360 uint64_t Size = TD->getABITypeSize(DestPointee);
8362 // Convert the constant to intptr type.
8363 APInt Offset = Cst->getValue();
8364 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8366 // If Offset is evenly divisible by Size, we can do this xform.
8367 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8368 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8370 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8372 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8378 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8379 // If the operands are integer typed then apply the integer transforms,
8380 // otherwise just apply the common ones.
8381 Value *Src = CI.getOperand(0);
8382 const Type *SrcTy = Src->getType();
8383 const Type *DestTy = CI.getType();
8385 if (SrcTy->isInteger() && DestTy->isInteger()) {
8386 if (Instruction *Result = commonIntCastTransforms(CI))
8388 } else if (isa<PointerType>(SrcTy)) {
8389 if (Instruction *I = commonPointerCastTransforms(CI))
8392 if (Instruction *Result = commonCastTransforms(CI))
8397 // Get rid of casts from one type to the same type. These are useless and can
8398 // be replaced by the operand.
8399 if (DestTy == Src->getType())
8400 return ReplaceInstUsesWith(CI, Src);
8402 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8403 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8404 const Type *DstElTy = DstPTy->getElementType();
8405 const Type *SrcElTy = SrcPTy->getElementType();
8407 // If the address spaces don't match, don't eliminate the bitcast, which is
8408 // required for changing types.
8409 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8412 // If we are casting a malloc or alloca to a pointer to a type of the same
8413 // size, rewrite the allocation instruction to allocate the "right" type.
8414 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8415 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8418 // If the source and destination are pointers, and this cast is equivalent
8419 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8420 // This can enhance SROA and other transforms that want type-safe pointers.
8421 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8422 unsigned NumZeros = 0;
8423 while (SrcElTy != DstElTy &&
8424 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8425 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8426 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8430 // If we found a path from the src to dest, create the getelementptr now.
8431 if (SrcElTy == DstElTy) {
8432 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8433 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8434 ((Instruction*) NULL));
8438 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8439 if (SVI->hasOneUse()) {
8440 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8441 // a bitconvert to a vector with the same # elts.
8442 if (isa<VectorType>(DestTy) &&
8443 cast<VectorType>(DestTy)->getNumElements() ==
8444 SVI->getType()->getNumElements()) {
8446 // If either of the operands is a cast from CI.getType(), then
8447 // evaluating the shuffle in the casted destination's type will allow
8448 // us to eliminate at least one cast.
8449 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8450 Tmp->getOperand(0)->getType() == DestTy) ||
8451 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8452 Tmp->getOperand(0)->getType() == DestTy)) {
8453 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8454 SVI->getOperand(0), DestTy, &CI);
8455 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8456 SVI->getOperand(1), DestTy, &CI);
8457 // Return a new shuffle vector. Use the same element ID's, as we
8458 // know the vector types match #elts.
8459 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8467 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8469 /// %D = select %cond, %C, %A
8471 /// %C = select %cond, %B, 0
8474 /// Assuming that the specified instruction is an operand to the select, return
8475 /// a bitmask indicating which operands of this instruction are foldable if they
8476 /// equal the other incoming value of the select.
8478 static unsigned GetSelectFoldableOperands(Instruction *I) {
8479 switch (I->getOpcode()) {
8480 case Instruction::Add:
8481 case Instruction::Mul:
8482 case Instruction::And:
8483 case Instruction::Or:
8484 case Instruction::Xor:
8485 return 3; // Can fold through either operand.
8486 case Instruction::Sub: // Can only fold on the amount subtracted.
8487 case Instruction::Shl: // Can only fold on the shift amount.
8488 case Instruction::LShr:
8489 case Instruction::AShr:
8492 return 0; // Cannot fold
8496 /// GetSelectFoldableConstant - For the same transformation as the previous
8497 /// function, return the identity constant that goes into the select.
8498 static Constant *GetSelectFoldableConstant(Instruction *I) {
8499 switch (I->getOpcode()) {
8500 default: assert(0 && "This cannot happen!"); abort();
8501 case Instruction::Add:
8502 case Instruction::Sub:
8503 case Instruction::Or:
8504 case Instruction::Xor:
8505 case Instruction::Shl:
8506 case Instruction::LShr:
8507 case Instruction::AShr:
8508 return Constant::getNullValue(I->getType());
8509 case Instruction::And:
8510 return Constant::getAllOnesValue(I->getType());
8511 case Instruction::Mul:
8512 return ConstantInt::get(I->getType(), 1);
8516 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8517 /// have the same opcode and only one use each. Try to simplify this.
8518 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8520 if (TI->getNumOperands() == 1) {
8521 // If this is a non-volatile load or a cast from the same type,
8524 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8527 return 0; // unknown unary op.
8530 // Fold this by inserting a select from the input values.
8531 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8532 FI->getOperand(0), SI.getName()+".v");
8533 InsertNewInstBefore(NewSI, SI);
8534 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8538 // Only handle binary operators here.
8539 if (!isa<BinaryOperator>(TI))
8542 // Figure out if the operations have any operands in common.
8543 Value *MatchOp, *OtherOpT, *OtherOpF;
8545 if (TI->getOperand(0) == FI->getOperand(0)) {
8546 MatchOp = TI->getOperand(0);
8547 OtherOpT = TI->getOperand(1);
8548 OtherOpF = FI->getOperand(1);
8549 MatchIsOpZero = true;
8550 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8551 MatchOp = TI->getOperand(1);
8552 OtherOpT = TI->getOperand(0);
8553 OtherOpF = FI->getOperand(0);
8554 MatchIsOpZero = false;
8555 } else if (!TI->isCommutative()) {
8557 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8558 MatchOp = TI->getOperand(0);
8559 OtherOpT = TI->getOperand(1);
8560 OtherOpF = FI->getOperand(0);
8561 MatchIsOpZero = true;
8562 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8563 MatchOp = TI->getOperand(1);
8564 OtherOpT = TI->getOperand(0);
8565 OtherOpF = FI->getOperand(1);
8566 MatchIsOpZero = true;
8571 // If we reach here, they do have operations in common.
8572 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8573 OtherOpF, SI.getName()+".v");
8574 InsertNewInstBefore(NewSI, SI);
8576 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8578 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8580 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8582 assert(0 && "Shouldn't get here");
8586 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8587 Value *CondVal = SI.getCondition();
8588 Value *TrueVal = SI.getTrueValue();
8589 Value *FalseVal = SI.getFalseValue();
8591 // select true, X, Y -> X
8592 // select false, X, Y -> Y
8593 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8594 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8596 // select C, X, X -> X
8597 if (TrueVal == FalseVal)
8598 return ReplaceInstUsesWith(SI, TrueVal);
8600 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8601 return ReplaceInstUsesWith(SI, FalseVal);
8602 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8603 return ReplaceInstUsesWith(SI, TrueVal);
8604 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8605 if (isa<Constant>(TrueVal))
8606 return ReplaceInstUsesWith(SI, TrueVal);
8608 return ReplaceInstUsesWith(SI, FalseVal);
8611 if (SI.getType() == Type::Int1Ty) {
8612 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8613 if (C->getZExtValue()) {
8614 // Change: A = select B, true, C --> A = or B, C
8615 return BinaryOperator::CreateOr(CondVal, FalseVal);
8617 // Change: A = select B, false, C --> A = and !B, C
8619 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8620 "not."+CondVal->getName()), SI);
8621 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8623 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8624 if (C->getZExtValue() == false) {
8625 // Change: A = select B, C, false --> A = and B, C
8626 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8628 // Change: A = select B, C, true --> A = or !B, C
8630 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8631 "not."+CondVal->getName()), SI);
8632 return BinaryOperator::CreateOr(NotCond, TrueVal);
8636 // select a, b, a -> a&b
8637 // select a, a, b -> a|b
8638 if (CondVal == TrueVal)
8639 return BinaryOperator::CreateOr(CondVal, FalseVal);
8640 else if (CondVal == FalseVal)
8641 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8644 // Selecting between two integer constants?
8645 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8646 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8647 // select C, 1, 0 -> zext C to int
8648 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8649 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8650 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8651 // select C, 0, 1 -> zext !C to int
8653 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8654 "not."+CondVal->getName()), SI);
8655 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8658 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8660 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8662 // (x <s 0) ? -1 : 0 -> ashr x, 31
8663 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8664 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8665 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8666 // The comparison constant and the result are not neccessarily the
8667 // same width. Make an all-ones value by inserting a AShr.
8668 Value *X = IC->getOperand(0);
8669 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8670 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8671 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8673 InsertNewInstBefore(SRA, SI);
8675 // Finally, convert to the type of the select RHS. We figure out
8676 // if this requires a SExt, Trunc or BitCast based on the sizes.
8677 Instruction::CastOps opc = Instruction::BitCast;
8678 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8679 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8680 if (SRASize < SISize)
8681 opc = Instruction::SExt;
8682 else if (SRASize > SISize)
8683 opc = Instruction::Trunc;
8684 return CastInst::Create(opc, SRA, SI.getType());
8689 // If one of the constants is zero (we know they can't both be) and we
8690 // have an icmp instruction with zero, and we have an 'and' with the
8691 // non-constant value, eliminate this whole mess. This corresponds to
8692 // cases like this: ((X & 27) ? 27 : 0)
8693 if (TrueValC->isZero() || FalseValC->isZero())
8694 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8695 cast<Constant>(IC->getOperand(1))->isNullValue())
8696 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8697 if (ICA->getOpcode() == Instruction::And &&
8698 isa<ConstantInt>(ICA->getOperand(1)) &&
8699 (ICA->getOperand(1) == TrueValC ||
8700 ICA->getOperand(1) == FalseValC) &&
8701 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8702 // Okay, now we know that everything is set up, we just don't
8703 // know whether we have a icmp_ne or icmp_eq and whether the
8704 // true or false val is the zero.
8705 bool ShouldNotVal = !TrueValC->isZero();
8706 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8709 V = InsertNewInstBefore(BinaryOperator::Create(
8710 Instruction::Xor, V, ICA->getOperand(1)), SI);
8711 return ReplaceInstUsesWith(SI, V);
8716 // See if we are selecting two values based on a comparison of the two values.
8717 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8718 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8719 // Transform (X == Y) ? X : Y -> Y
8720 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8721 // This is not safe in general for floating point:
8722 // consider X== -0, Y== +0.
8723 // It becomes safe if either operand is a nonzero constant.
8724 ConstantFP *CFPt, *CFPf;
8725 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8726 !CFPt->getValueAPF().isZero()) ||
8727 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8728 !CFPf->getValueAPF().isZero()))
8729 return ReplaceInstUsesWith(SI, FalseVal);
8731 // Transform (X != Y) ? X : Y -> X
8732 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8733 return ReplaceInstUsesWith(SI, TrueVal);
8734 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8736 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8737 // Transform (X == Y) ? Y : X -> X
8738 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8739 // This is not safe in general for floating point:
8740 // consider X== -0, Y== +0.
8741 // It becomes safe if either operand is a nonzero constant.
8742 ConstantFP *CFPt, *CFPf;
8743 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8744 !CFPt->getValueAPF().isZero()) ||
8745 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8746 !CFPf->getValueAPF().isZero()))
8747 return ReplaceInstUsesWith(SI, FalseVal);
8749 // Transform (X != Y) ? Y : X -> Y
8750 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8751 return ReplaceInstUsesWith(SI, TrueVal);
8752 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8756 // See if we are selecting two values based on a comparison of the two values.
8757 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8758 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8759 // Transform (X == Y) ? X : Y -> Y
8760 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8761 return ReplaceInstUsesWith(SI, FalseVal);
8762 // Transform (X != Y) ? X : Y -> X
8763 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8764 return ReplaceInstUsesWith(SI, TrueVal);
8765 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8767 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8768 // Transform (X == Y) ? Y : X -> X
8769 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8770 return ReplaceInstUsesWith(SI, FalseVal);
8771 // Transform (X != Y) ? Y : X -> Y
8772 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8773 return ReplaceInstUsesWith(SI, TrueVal);
8774 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8778 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8779 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8780 if (TI->hasOneUse() && FI->hasOneUse()) {
8781 Instruction *AddOp = 0, *SubOp = 0;
8783 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8784 if (TI->getOpcode() == FI->getOpcode())
8785 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8788 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8789 // even legal for FP.
8790 if (TI->getOpcode() == Instruction::Sub &&
8791 FI->getOpcode() == Instruction::Add) {
8792 AddOp = FI; SubOp = TI;
8793 } else if (FI->getOpcode() == Instruction::Sub &&
8794 TI->getOpcode() == Instruction::Add) {
8795 AddOp = TI; SubOp = FI;
8799 Value *OtherAddOp = 0;
8800 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8801 OtherAddOp = AddOp->getOperand(1);
8802 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8803 OtherAddOp = AddOp->getOperand(0);
8807 // So at this point we know we have (Y -> OtherAddOp):
8808 // select C, (add X, Y), (sub X, Z)
8809 Value *NegVal; // Compute -Z
8810 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8811 NegVal = ConstantExpr::getNeg(C);
8813 NegVal = InsertNewInstBefore(
8814 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8817 Value *NewTrueOp = OtherAddOp;
8818 Value *NewFalseOp = NegVal;
8820 std::swap(NewTrueOp, NewFalseOp);
8821 Instruction *NewSel =
8822 SelectInst::Create(CondVal, NewTrueOp,
8823 NewFalseOp, SI.getName() + ".p");
8825 NewSel = InsertNewInstBefore(NewSel, SI);
8826 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8831 // See if we can fold the select into one of our operands.
8832 if (SI.getType()->isInteger()) {
8833 // See the comment above GetSelectFoldableOperands for a description of the
8834 // transformation we are doing here.
8835 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8836 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8837 !isa<Constant>(FalseVal))
8838 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8839 unsigned OpToFold = 0;
8840 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8842 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8847 Constant *C = GetSelectFoldableConstant(TVI);
8848 Instruction *NewSel =
8849 SelectInst::Create(SI.getCondition(),
8850 TVI->getOperand(2-OpToFold), C);
8851 InsertNewInstBefore(NewSel, SI);
8852 NewSel->takeName(TVI);
8853 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8854 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8856 assert(0 && "Unknown instruction!!");
8861 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8862 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8863 !isa<Constant>(TrueVal))
8864 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8865 unsigned OpToFold = 0;
8866 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8868 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8873 Constant *C = GetSelectFoldableConstant(FVI);
8874 Instruction *NewSel =
8875 SelectInst::Create(SI.getCondition(), C,
8876 FVI->getOperand(2-OpToFold));
8877 InsertNewInstBefore(NewSel, SI);
8878 NewSel->takeName(FVI);
8879 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8880 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8882 assert(0 && "Unknown instruction!!");
8887 if (BinaryOperator::isNot(CondVal)) {
8888 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8889 SI.setOperand(1, FalseVal);
8890 SI.setOperand(2, TrueVal);
8897 /// EnforceKnownAlignment - If the specified pointer points to an object that
8898 /// we control, modify the object's alignment to PrefAlign. This isn't
8899 /// often possible though. If alignment is important, a more reliable approach
8900 /// is to simply align all global variables and allocation instructions to
8901 /// their preferred alignment from the beginning.
8903 static unsigned EnforceKnownAlignment(Value *V,
8904 unsigned Align, unsigned PrefAlign) {
8906 User *U = dyn_cast<User>(V);
8907 if (!U) return Align;
8909 switch (getOpcode(U)) {
8911 case Instruction::BitCast:
8912 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8913 case Instruction::GetElementPtr: {
8914 // If all indexes are zero, it is just the alignment of the base pointer.
8915 bool AllZeroOperands = true;
8916 for (unsigned i = 1, e = U->getNumOperands(); i != e; ++i)
8917 if (!isa<Constant>(U->getOperand(i)) ||
8918 !cast<Constant>(U->getOperand(i))->isNullValue()) {
8919 AllZeroOperands = false;
8923 if (AllZeroOperands) {
8924 // Treat this like a bitcast.
8925 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8931 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8932 // If there is a large requested alignment and we can, bump up the alignment
8934 if (!GV->isDeclaration()) {
8935 GV->setAlignment(PrefAlign);
8938 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8939 // If there is a requested alignment and if this is an alloca, round up. We
8940 // don't do this for malloc, because some systems can't respect the request.
8941 if (isa<AllocaInst>(AI)) {
8942 AI->setAlignment(PrefAlign);
8950 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8951 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8952 /// and it is more than the alignment of the ultimate object, see if we can
8953 /// increase the alignment of the ultimate object, making this check succeed.
8954 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8955 unsigned PrefAlign) {
8956 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8957 sizeof(PrefAlign) * CHAR_BIT;
8958 APInt Mask = APInt::getAllOnesValue(BitWidth);
8959 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8960 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8961 unsigned TrailZ = KnownZero.countTrailingOnes();
8962 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8964 if (PrefAlign > Align)
8965 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8967 // We don't need to make any adjustment.
8971 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8972 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8973 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8974 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8975 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8977 if (CopyAlign < MinAlign) {
8978 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8982 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8984 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8985 if (MemOpLength == 0) return 0;
8987 // Source and destination pointer types are always "i8*" for intrinsic. See
8988 // if the size is something we can handle with a single primitive load/store.
8989 // A single load+store correctly handles overlapping memory in the memmove
8991 unsigned Size = MemOpLength->getZExtValue();
8992 if (Size == 0) return MI; // Delete this mem transfer.
8994 if (Size > 8 || (Size&(Size-1)))
8995 return 0; // If not 1/2/4/8 bytes, exit.
8997 // Use an integer load+store unless we can find something better.
8998 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9000 // Memcpy forces the use of i8* for the source and destination. That means
9001 // that if you're using memcpy to move one double around, you'll get a cast
9002 // from double* to i8*. We'd much rather use a double load+store rather than
9003 // an i64 load+store, here because this improves the odds that the source or
9004 // dest address will be promotable. See if we can find a better type than the
9005 // integer datatype.
9006 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9007 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9008 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9009 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9010 // down through these levels if so.
9011 while (!SrcETy->isSingleValueType()) {
9012 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9013 if (STy->getNumElements() == 1)
9014 SrcETy = STy->getElementType(0);
9017 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9018 if (ATy->getNumElements() == 1)
9019 SrcETy = ATy->getElementType();
9026 if (SrcETy->isSingleValueType())
9027 NewPtrTy = PointerType::getUnqual(SrcETy);
9032 // If the memcpy/memmove provides better alignment info than we can
9034 SrcAlign = std::max(SrcAlign, CopyAlign);
9035 DstAlign = std::max(DstAlign, CopyAlign);
9037 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9038 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9039 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9040 InsertNewInstBefore(L, *MI);
9041 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9043 // Set the size of the copy to 0, it will be deleted on the next iteration.
9044 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9048 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9049 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9050 if (MI->getAlignment()->getZExtValue() < Alignment) {
9051 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9055 // Extract the length and alignment and fill if they are constant.
9056 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9057 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9058 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9060 uint64_t Len = LenC->getZExtValue();
9061 Alignment = MI->getAlignment()->getZExtValue();
9063 // If the length is zero, this is a no-op
9064 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9066 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9067 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9068 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9070 Value *Dest = MI->getDest();
9071 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9073 // Alignment 0 is identity for alignment 1 for memset, but not store.
9074 if (Alignment == 0) Alignment = 1;
9076 // Extract the fill value and store.
9077 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9078 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9081 // Set the size of the copy to 0, it will be deleted on the next iteration.
9082 MI->setLength(Constant::getNullValue(LenC->getType()));
9090 /// visitCallInst - CallInst simplification. This mostly only handles folding
9091 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9092 /// the heavy lifting.
9094 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9095 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9096 if (!II) return visitCallSite(&CI);
9098 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9100 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9101 bool Changed = false;
9103 // memmove/cpy/set of zero bytes is a noop.
9104 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9105 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9107 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9108 if (CI->getZExtValue() == 1) {
9109 // Replace the instruction with just byte operations. We would
9110 // transform other cases to loads/stores, but we don't know if
9111 // alignment is sufficient.
9115 // If we have a memmove and the source operation is a constant global,
9116 // then the source and dest pointers can't alias, so we can change this
9117 // into a call to memcpy.
9118 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9119 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9120 if (GVSrc->isConstant()) {
9121 Module *M = CI.getParent()->getParent()->getParent();
9122 Intrinsic::ID MemCpyID;
9123 if (CI.getOperand(3)->getType() == Type::Int32Ty)
9124 MemCpyID = Intrinsic::memcpy_i32;
9126 MemCpyID = Intrinsic::memcpy_i64;
9127 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
9132 // If we can determine a pointer alignment that is bigger than currently
9133 // set, update the alignment.
9134 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9135 if (Instruction *I = SimplifyMemTransfer(MI))
9137 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9138 if (Instruction *I = SimplifyMemSet(MSI))
9142 if (Changed) return II;
9144 switch (II->getIntrinsicID()) {
9146 case Intrinsic::ppc_altivec_lvx:
9147 case Intrinsic::ppc_altivec_lvxl:
9148 case Intrinsic::x86_sse_loadu_ps:
9149 case Intrinsic::x86_sse2_loadu_pd:
9150 case Intrinsic::x86_sse2_loadu_dq:
9151 // Turn PPC lvx -> load if the pointer is known aligned.
9152 // Turn X86 loadups -> load if the pointer is known aligned.
9153 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9154 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9155 PointerType::getUnqual(II->getType()),
9157 return new LoadInst(Ptr);
9160 case Intrinsic::ppc_altivec_stvx:
9161 case Intrinsic::ppc_altivec_stvxl:
9162 // Turn stvx -> store if the pointer is known aligned.
9163 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9164 const Type *OpPtrTy =
9165 PointerType::getUnqual(II->getOperand(1)->getType());
9166 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9167 return new StoreInst(II->getOperand(1), Ptr);
9170 case Intrinsic::x86_sse_storeu_ps:
9171 case Intrinsic::x86_sse2_storeu_pd:
9172 case Intrinsic::x86_sse2_storeu_dq:
9173 case Intrinsic::x86_sse2_storel_dq:
9174 // Turn X86 storeu -> store if the pointer is known aligned.
9175 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9176 const Type *OpPtrTy =
9177 PointerType::getUnqual(II->getOperand(2)->getType());
9178 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9179 return new StoreInst(II->getOperand(2), Ptr);
9183 case Intrinsic::x86_sse_cvttss2si: {
9184 // These intrinsics only demands the 0th element of its input vector. If
9185 // we can simplify the input based on that, do so now.
9187 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9189 II->setOperand(1, V);
9195 case Intrinsic::ppc_altivec_vperm:
9196 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9197 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9198 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9200 // Check that all of the elements are integer constants or undefs.
9201 bool AllEltsOk = true;
9202 for (unsigned i = 0; i != 16; ++i) {
9203 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9204 !isa<UndefValue>(Mask->getOperand(i))) {
9211 // Cast the input vectors to byte vectors.
9212 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9213 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9214 Value *Result = UndefValue::get(Op0->getType());
9216 // Only extract each element once.
9217 Value *ExtractedElts[32];
9218 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9220 for (unsigned i = 0; i != 16; ++i) {
9221 if (isa<UndefValue>(Mask->getOperand(i)))
9223 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9224 Idx &= 31; // Match the hardware behavior.
9226 if (ExtractedElts[Idx] == 0) {
9228 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9229 InsertNewInstBefore(Elt, CI);
9230 ExtractedElts[Idx] = Elt;
9233 // Insert this value into the result vector.
9234 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9236 InsertNewInstBefore(cast<Instruction>(Result), CI);
9238 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9243 case Intrinsic::stackrestore: {
9244 // If the save is right next to the restore, remove the restore. This can
9245 // happen when variable allocas are DCE'd.
9246 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9247 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9248 BasicBlock::iterator BI = SS;
9250 return EraseInstFromFunction(CI);
9254 // Scan down this block to see if there is another stack restore in the
9255 // same block without an intervening call/alloca.
9256 BasicBlock::iterator BI = II;
9257 TerminatorInst *TI = II->getParent()->getTerminator();
9258 bool CannotRemove = false;
9259 for (++BI; &*BI != TI; ++BI) {
9260 if (isa<AllocaInst>(BI)) {
9261 CannotRemove = true;
9264 if (isa<CallInst>(BI)) {
9265 if (!isa<IntrinsicInst>(BI)) {
9266 CannotRemove = true;
9269 // If there is a stackrestore below this one, remove this one.
9270 return EraseInstFromFunction(CI);
9274 // If the stack restore is in a return/unwind block and if there are no
9275 // allocas or calls between the restore and the return, nuke the restore.
9276 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9277 return EraseInstFromFunction(CI);
9283 return visitCallSite(II);
9286 // InvokeInst simplification
9288 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9289 return visitCallSite(&II);
9292 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9293 /// passed through the varargs area, we can eliminate the use of the cast.
9294 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9295 const CastInst * const CI,
9296 const TargetData * const TD,
9298 if (!CI->isLosslessCast())
9301 // The size of ByVal arguments is derived from the type, so we
9302 // can't change to a type with a different size. If the size were
9303 // passed explicitly we could avoid this check.
9304 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
9308 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9309 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9310 if (!SrcTy->isSized() || !DstTy->isSized())
9312 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9317 // visitCallSite - Improvements for call and invoke instructions.
9319 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9320 bool Changed = false;
9322 // If the callee is a constexpr cast of a function, attempt to move the cast
9323 // to the arguments of the call/invoke.
9324 if (transformConstExprCastCall(CS)) return 0;
9326 Value *Callee = CS.getCalledValue();
9328 if (Function *CalleeF = dyn_cast<Function>(Callee))
9329 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9330 Instruction *OldCall = CS.getInstruction();
9331 // If the call and callee calling conventions don't match, this call must
9332 // be unreachable, as the call is undefined.
9333 new StoreInst(ConstantInt::getTrue(),
9334 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9336 if (!OldCall->use_empty())
9337 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9338 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9339 return EraseInstFromFunction(*OldCall);
9343 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9344 // This instruction is not reachable, just remove it. We insert a store to
9345 // undef so that we know that this code is not reachable, despite the fact
9346 // that we can't modify the CFG here.
9347 new StoreInst(ConstantInt::getTrue(),
9348 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9349 CS.getInstruction());
9351 if (!CS.getInstruction()->use_empty())
9352 CS.getInstruction()->
9353 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9355 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9356 // Don't break the CFG, insert a dummy cond branch.
9357 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9358 ConstantInt::getTrue(), II);
9360 return EraseInstFromFunction(*CS.getInstruction());
9363 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9364 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9365 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9366 return transformCallThroughTrampoline(CS);
9368 const PointerType *PTy = cast<PointerType>(Callee->getType());
9369 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9370 if (FTy->isVarArg()) {
9371 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9372 // See if we can optimize any arguments passed through the varargs area of
9374 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9375 E = CS.arg_end(); I != E; ++I, ++ix) {
9376 CastInst *CI = dyn_cast<CastInst>(*I);
9377 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9378 *I = CI->getOperand(0);
9384 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9385 // Inline asm calls cannot throw - mark them 'nounwind'.
9386 CS.setDoesNotThrow();
9390 return Changed ? CS.getInstruction() : 0;
9393 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9394 // attempt to move the cast to the arguments of the call/invoke.
9396 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9397 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9398 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9399 if (CE->getOpcode() != Instruction::BitCast ||
9400 !isa<Function>(CE->getOperand(0)))
9402 Function *Callee = cast<Function>(CE->getOperand(0));
9403 Instruction *Caller = CS.getInstruction();
9404 const PAListPtr &CallerPAL = CS.getParamAttrs();
9406 // Okay, this is a cast from a function to a different type. Unless doing so
9407 // would cause a type conversion of one of our arguments, change this call to
9408 // be a direct call with arguments casted to the appropriate types.
9410 const FunctionType *FT = Callee->getFunctionType();
9411 const Type *OldRetTy = Caller->getType();
9413 if (isa<StructType>(FT->getReturnType()))
9414 return false; // TODO: Handle multiple return values.
9416 // Check to see if we are changing the return type...
9417 if (OldRetTy != FT->getReturnType()) {
9418 if (Callee->isDeclaration() &&
9419 // Conversion is ok if changing from pointer to int of same size.
9420 !(isa<PointerType>(FT->getReturnType()) &&
9421 TD->getIntPtrType() == OldRetTy))
9422 return false; // Cannot transform this return value.
9424 if (!Caller->use_empty() &&
9425 // void -> non-void is handled specially
9426 FT->getReturnType() != Type::VoidTy &&
9427 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
9428 return false; // Cannot transform this return value.
9430 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9431 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
9432 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
9433 return false; // Attribute not compatible with transformed value.
9436 // If the callsite is an invoke instruction, and the return value is used by
9437 // a PHI node in a successor, we cannot change the return type of the call
9438 // because there is no place to put the cast instruction (without breaking
9439 // the critical edge). Bail out in this case.
9440 if (!Caller->use_empty())
9441 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9442 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9444 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9445 if (PN->getParent() == II->getNormalDest() ||
9446 PN->getParent() == II->getUnwindDest())
9450 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9451 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9453 CallSite::arg_iterator AI = CS.arg_begin();
9454 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9455 const Type *ParamTy = FT->getParamType(i);
9456 const Type *ActTy = (*AI)->getType();
9458 if (!CastInst::isCastable(ActTy, ParamTy))
9459 return false; // Cannot transform this parameter value.
9461 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
9462 return false; // Attribute not compatible with transformed value.
9464 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
9465 // Some conversions are safe even if we do not have a body.
9466 // Either we can cast directly, or we can upconvert the argument
9467 bool isConvertible = ActTy == ParamTy ||
9468 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
9469 (ParamTy->isInteger() && ActTy->isInteger() &&
9470 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
9471 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
9472 && c->getValue().isStrictlyPositive());
9473 if (Callee->isDeclaration() && !isConvertible) return false;
9476 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9477 Callee->isDeclaration())
9478 return false; // Do not delete arguments unless we have a function body.
9480 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9481 !CallerPAL.isEmpty())
9482 // In this case we have more arguments than the new function type, but we
9483 // won't be dropping them. Check that these extra arguments have attributes
9484 // that are compatible with being a vararg call argument.
9485 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9486 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9488 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9489 if (PAttrs & ParamAttr::VarArgsIncompatible)
9493 // Okay, we decided that this is a safe thing to do: go ahead and start
9494 // inserting cast instructions as necessary...
9495 std::vector<Value*> Args;
9496 Args.reserve(NumActualArgs);
9497 SmallVector<ParamAttrsWithIndex, 8> attrVec;
9498 attrVec.reserve(NumCommonArgs);
9500 // Get any return attributes.
9501 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
9503 // If the return value is not being used, the type may not be compatible
9504 // with the existing attributes. Wipe out any problematic attributes.
9505 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
9507 // Add the new return attributes.
9509 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
9511 AI = CS.arg_begin();
9512 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9513 const Type *ParamTy = FT->getParamType(i);
9514 if ((*AI)->getType() == ParamTy) {
9515 Args.push_back(*AI);
9517 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9518 false, ParamTy, false);
9519 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9520 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9523 // Add any parameter attributes.
9524 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9525 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9528 // If the function takes more arguments than the call was taking, add them
9530 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9531 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9533 // If we are removing arguments to the function, emit an obnoxious warning...
9534 if (FT->getNumParams() < NumActualArgs) {
9535 if (!FT->isVarArg()) {
9536 cerr << "WARNING: While resolving call to function '"
9537 << Callee->getName() << "' arguments were dropped!\n";
9539 // Add all of the arguments in their promoted form to the arg list...
9540 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9541 const Type *PTy = getPromotedType((*AI)->getType());
9542 if (PTy != (*AI)->getType()) {
9543 // Must promote to pass through va_arg area!
9544 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9546 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9547 InsertNewInstBefore(Cast, *Caller);
9548 Args.push_back(Cast);
9550 Args.push_back(*AI);
9553 // Add any parameter attributes.
9554 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9555 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9560 if (FT->getReturnType() == Type::VoidTy)
9561 Caller->setName(""); // Void type should not have a name.
9563 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9566 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9567 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9568 Args.begin(), Args.end(),
9569 Caller->getName(), Caller);
9570 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9571 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9573 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9574 Caller->getName(), Caller);
9575 CallInst *CI = cast<CallInst>(Caller);
9576 if (CI->isTailCall())
9577 cast<CallInst>(NC)->setTailCall();
9578 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9579 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9582 // Insert a cast of the return type as necessary.
9584 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9585 if (NV->getType() != Type::VoidTy) {
9586 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9588 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9590 // If this is an invoke instruction, we should insert it after the first
9591 // non-phi, instruction in the normal successor block.
9592 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9593 BasicBlock::iterator I = II->getNormalDest()->begin();
9594 while (isa<PHINode>(I)) ++I;
9595 InsertNewInstBefore(NC, *I);
9597 // Otherwise, it's a call, just insert cast right after the call instr
9598 InsertNewInstBefore(NC, *Caller);
9600 AddUsersToWorkList(*Caller);
9602 NV = UndefValue::get(Caller->getType());
9606 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9607 Caller->replaceAllUsesWith(NV);
9608 Caller->eraseFromParent();
9609 RemoveFromWorkList(Caller);
9613 // transformCallThroughTrampoline - Turn a call to a function created by the
9614 // init_trampoline intrinsic into a direct call to the underlying function.
9616 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9617 Value *Callee = CS.getCalledValue();
9618 const PointerType *PTy = cast<PointerType>(Callee->getType());
9619 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9620 const PAListPtr &Attrs = CS.getParamAttrs();
9622 // If the call already has the 'nest' attribute somewhere then give up -
9623 // otherwise 'nest' would occur twice after splicing in the chain.
9624 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9627 IntrinsicInst *Tramp =
9628 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9630 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9631 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9632 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9634 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9635 if (!NestAttrs.isEmpty()) {
9636 unsigned NestIdx = 1;
9637 const Type *NestTy = 0;
9638 ParameterAttributes NestAttr = ParamAttr::None;
9640 // Look for a parameter marked with the 'nest' attribute.
9641 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9642 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9643 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9644 // Record the parameter type and any other attributes.
9646 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9651 Instruction *Caller = CS.getInstruction();
9652 std::vector<Value*> NewArgs;
9653 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9655 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9656 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9658 // Insert the nest argument into the call argument list, which may
9659 // mean appending it. Likewise for attributes.
9661 // Add any function result attributes.
9662 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9663 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9667 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9669 if (Idx == NestIdx) {
9670 // Add the chain argument and attributes.
9671 Value *NestVal = Tramp->getOperand(3);
9672 if (NestVal->getType() != NestTy)
9673 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9674 NewArgs.push_back(NestVal);
9675 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9681 // Add the original argument and attributes.
9682 NewArgs.push_back(*I);
9683 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9685 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9691 // The trampoline may have been bitcast to a bogus type (FTy).
9692 // Handle this by synthesizing a new function type, equal to FTy
9693 // with the chain parameter inserted.
9695 std::vector<const Type*> NewTypes;
9696 NewTypes.reserve(FTy->getNumParams()+1);
9698 // Insert the chain's type into the list of parameter types, which may
9699 // mean appending it.
9702 FunctionType::param_iterator I = FTy->param_begin(),
9703 E = FTy->param_end();
9707 // Add the chain's type.
9708 NewTypes.push_back(NestTy);
9713 // Add the original type.
9714 NewTypes.push_back(*I);
9720 // Replace the trampoline call with a direct call. Let the generic
9721 // code sort out any function type mismatches.
9722 FunctionType *NewFTy =
9723 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9724 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9725 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9726 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9728 Instruction *NewCaller;
9729 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9730 NewCaller = InvokeInst::Create(NewCallee,
9731 II->getNormalDest(), II->getUnwindDest(),
9732 NewArgs.begin(), NewArgs.end(),
9733 Caller->getName(), Caller);
9734 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9735 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9737 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9738 Caller->getName(), Caller);
9739 if (cast<CallInst>(Caller)->isTailCall())
9740 cast<CallInst>(NewCaller)->setTailCall();
9741 cast<CallInst>(NewCaller)->
9742 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9743 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9745 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9746 Caller->replaceAllUsesWith(NewCaller);
9747 Caller->eraseFromParent();
9748 RemoveFromWorkList(Caller);
9753 // Replace the trampoline call with a direct call. Since there is no 'nest'
9754 // parameter, there is no need to adjust the argument list. Let the generic
9755 // code sort out any function type mismatches.
9756 Constant *NewCallee =
9757 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9758 CS.setCalledFunction(NewCallee);
9759 return CS.getInstruction();
9762 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9763 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9764 /// and a single binop.
9765 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9766 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9767 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9768 isa<CmpInst>(FirstInst));
9769 unsigned Opc = FirstInst->getOpcode();
9770 Value *LHSVal = FirstInst->getOperand(0);
9771 Value *RHSVal = FirstInst->getOperand(1);
9773 const Type *LHSType = LHSVal->getType();
9774 const Type *RHSType = RHSVal->getType();
9776 // Scan to see if all operands are the same opcode, all have one use, and all
9777 // kill their operands (i.e. the operands have one use).
9778 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9779 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9780 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9781 // Verify type of the LHS matches so we don't fold cmp's of different
9782 // types or GEP's with different index types.
9783 I->getOperand(0)->getType() != LHSType ||
9784 I->getOperand(1)->getType() != RHSType)
9787 // If they are CmpInst instructions, check their predicates
9788 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9789 if (cast<CmpInst>(I)->getPredicate() !=
9790 cast<CmpInst>(FirstInst)->getPredicate())
9793 // Keep track of which operand needs a phi node.
9794 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9795 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9798 // Otherwise, this is safe to transform, determine if it is profitable.
9800 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9801 // Indexes are often folded into load/store instructions, so we don't want to
9802 // hide them behind a phi.
9803 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9806 Value *InLHS = FirstInst->getOperand(0);
9807 Value *InRHS = FirstInst->getOperand(1);
9808 PHINode *NewLHS = 0, *NewRHS = 0;
9810 NewLHS = PHINode::Create(LHSType,
9811 FirstInst->getOperand(0)->getName() + ".pn");
9812 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9813 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9814 InsertNewInstBefore(NewLHS, PN);
9819 NewRHS = PHINode::Create(RHSType,
9820 FirstInst->getOperand(1)->getName() + ".pn");
9821 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9822 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9823 InsertNewInstBefore(NewRHS, PN);
9827 // Add all operands to the new PHIs.
9828 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9830 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9831 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9834 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9835 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9839 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9840 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9841 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9842 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9845 assert(isa<GetElementPtrInst>(FirstInst));
9846 return GetElementPtrInst::Create(LHSVal, RHSVal);
9850 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9851 /// of the block that defines it. This means that it must be obvious the value
9852 /// of the load is not changed from the point of the load to the end of the
9855 /// Finally, it is safe, but not profitable, to sink a load targetting a
9856 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9858 static bool isSafeToSinkLoad(LoadInst *L) {
9859 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9861 for (++BBI; BBI != E; ++BBI)
9862 if (BBI->mayWriteToMemory())
9865 // Check for non-address taken alloca. If not address-taken already, it isn't
9866 // profitable to do this xform.
9867 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9868 bool isAddressTaken = false;
9869 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9871 if (isa<LoadInst>(UI)) continue;
9872 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9873 // If storing TO the alloca, then the address isn't taken.
9874 if (SI->getOperand(1) == AI) continue;
9876 isAddressTaken = true;
9880 if (!isAddressTaken)
9888 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9889 // operator and they all are only used by the PHI, PHI together their
9890 // inputs, and do the operation once, to the result of the PHI.
9891 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9892 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9894 // Scan the instruction, looking for input operations that can be folded away.
9895 // If all input operands to the phi are the same instruction (e.g. a cast from
9896 // the same type or "+42") we can pull the operation through the PHI, reducing
9897 // code size and simplifying code.
9898 Constant *ConstantOp = 0;
9899 const Type *CastSrcTy = 0;
9900 bool isVolatile = false;
9901 if (isa<CastInst>(FirstInst)) {
9902 CastSrcTy = FirstInst->getOperand(0)->getType();
9903 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9904 // Can fold binop, compare or shift here if the RHS is a constant,
9905 // otherwise call FoldPHIArgBinOpIntoPHI.
9906 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9907 if (ConstantOp == 0)
9908 return FoldPHIArgBinOpIntoPHI(PN);
9909 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9910 isVolatile = LI->isVolatile();
9911 // We can't sink the load if the loaded value could be modified between the
9912 // load and the PHI.
9913 if (LI->getParent() != PN.getIncomingBlock(0) ||
9914 !isSafeToSinkLoad(LI))
9916 } else if (isa<GetElementPtrInst>(FirstInst)) {
9917 if (FirstInst->getNumOperands() == 2)
9918 return FoldPHIArgBinOpIntoPHI(PN);
9919 // Can't handle general GEPs yet.
9922 return 0; // Cannot fold this operation.
9925 // Check to see if all arguments are the same operation.
9926 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9927 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9928 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9929 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9932 if (I->getOperand(0)->getType() != CastSrcTy)
9933 return 0; // Cast operation must match.
9934 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9935 // We can't sink the load if the loaded value could be modified between
9936 // the load and the PHI.
9937 if (LI->isVolatile() != isVolatile ||
9938 LI->getParent() != PN.getIncomingBlock(i) ||
9939 !isSafeToSinkLoad(LI))
9942 // If the PHI is volatile and its block has multiple successors, sinking
9943 // it would remove a load of the volatile value from the path through the
9946 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9950 } else if (I->getOperand(1) != ConstantOp) {
9955 // Okay, they are all the same operation. Create a new PHI node of the
9956 // correct type, and PHI together all of the LHS's of the instructions.
9957 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9958 PN.getName()+".in");
9959 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9961 Value *InVal = FirstInst->getOperand(0);
9962 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9964 // Add all operands to the new PHI.
9965 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9966 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9967 if (NewInVal != InVal)
9969 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9974 // The new PHI unions all of the same values together. This is really
9975 // common, so we handle it intelligently here for compile-time speed.
9979 InsertNewInstBefore(NewPN, PN);
9983 // Insert and return the new operation.
9984 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9985 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9986 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9987 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9988 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9989 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9990 PhiVal, ConstantOp);
9991 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9993 // If this was a volatile load that we are merging, make sure to loop through
9994 // and mark all the input loads as non-volatile. If we don't do this, we will
9995 // insert a new volatile load and the old ones will not be deletable.
9997 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9998 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10000 return new LoadInst(PhiVal, "", isVolatile);
10003 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10005 static bool DeadPHICycle(PHINode *PN,
10006 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10007 if (PN->use_empty()) return true;
10008 if (!PN->hasOneUse()) return false;
10010 // Remember this node, and if we find the cycle, return.
10011 if (!PotentiallyDeadPHIs.insert(PN))
10014 // Don't scan crazily complex things.
10015 if (PotentiallyDeadPHIs.size() == 16)
10018 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10019 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10024 /// PHIsEqualValue - Return true if this phi node is always equal to
10025 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10026 /// z = some value; x = phi (y, z); y = phi (x, z)
10027 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10028 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10029 // See if we already saw this PHI node.
10030 if (!ValueEqualPHIs.insert(PN))
10033 // Don't scan crazily complex things.
10034 if (ValueEqualPHIs.size() == 16)
10037 // Scan the operands to see if they are either phi nodes or are equal to
10039 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10040 Value *Op = PN->getIncomingValue(i);
10041 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10042 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10044 } else if (Op != NonPhiInVal)
10052 // PHINode simplification
10054 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10055 // If LCSSA is around, don't mess with Phi nodes
10056 if (MustPreserveLCSSA) return 0;
10058 if (Value *V = PN.hasConstantValue())
10059 return ReplaceInstUsesWith(PN, V);
10061 // If all PHI operands are the same operation, pull them through the PHI,
10062 // reducing code size.
10063 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10064 PN.getIncomingValue(0)->hasOneUse())
10065 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10068 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10069 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10070 // PHI)... break the cycle.
10071 if (PN.hasOneUse()) {
10072 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10073 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10074 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10075 PotentiallyDeadPHIs.insert(&PN);
10076 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10077 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10080 // If this phi has a single use, and if that use just computes a value for
10081 // the next iteration of a loop, delete the phi. This occurs with unused
10082 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10083 // common case here is good because the only other things that catch this
10084 // are induction variable analysis (sometimes) and ADCE, which is only run
10086 if (PHIUser->hasOneUse() &&
10087 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10088 PHIUser->use_back() == &PN) {
10089 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10093 // We sometimes end up with phi cycles that non-obviously end up being the
10094 // same value, for example:
10095 // z = some value; x = phi (y, z); y = phi (x, z)
10096 // where the phi nodes don't necessarily need to be in the same block. Do a
10097 // quick check to see if the PHI node only contains a single non-phi value, if
10098 // so, scan to see if the phi cycle is actually equal to that value.
10100 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10101 // Scan for the first non-phi operand.
10102 while (InValNo != NumOperandVals &&
10103 isa<PHINode>(PN.getIncomingValue(InValNo)))
10106 if (InValNo != NumOperandVals) {
10107 Value *NonPhiInVal = PN.getOperand(InValNo);
10109 // Scan the rest of the operands to see if there are any conflicts, if so
10110 // there is no need to recursively scan other phis.
10111 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10112 Value *OpVal = PN.getIncomingValue(InValNo);
10113 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10117 // If we scanned over all operands, then we have one unique value plus
10118 // phi values. Scan PHI nodes to see if they all merge in each other or
10120 if (InValNo == NumOperandVals) {
10121 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10122 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10123 return ReplaceInstUsesWith(PN, NonPhiInVal);
10130 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10131 Instruction *InsertPoint,
10132 InstCombiner *IC) {
10133 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10134 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10135 // We must cast correctly to the pointer type. Ensure that we
10136 // sign extend the integer value if it is smaller as this is
10137 // used for address computation.
10138 Instruction::CastOps opcode =
10139 (VTySize < PtrSize ? Instruction::SExt :
10140 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10141 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10145 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10146 Value *PtrOp = GEP.getOperand(0);
10147 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10148 // If so, eliminate the noop.
10149 if (GEP.getNumOperands() == 1)
10150 return ReplaceInstUsesWith(GEP, PtrOp);
10152 if (isa<UndefValue>(GEP.getOperand(0)))
10153 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10155 bool HasZeroPointerIndex = false;
10156 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10157 HasZeroPointerIndex = C->isNullValue();
10159 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10160 return ReplaceInstUsesWith(GEP, PtrOp);
10162 // Eliminate unneeded casts for indices.
10163 bool MadeChange = false;
10165 gep_type_iterator GTI = gep_type_begin(GEP);
10166 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
10167 if (isa<SequentialType>(*GTI)) {
10168 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
10169 if (CI->getOpcode() == Instruction::ZExt ||
10170 CI->getOpcode() == Instruction::SExt) {
10171 const Type *SrcTy = CI->getOperand(0)->getType();
10172 // We can eliminate a cast from i32 to i64 iff the target
10173 // is a 32-bit pointer target.
10174 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10176 GEP.setOperand(i, CI->getOperand(0));
10180 // If we are using a wider index than needed for this platform, shrink it
10181 // to what we need. If the incoming value needs a cast instruction,
10182 // insert it. This explicit cast can make subsequent optimizations more
10184 Value *Op = GEP.getOperand(i);
10185 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10186 if (Constant *C = dyn_cast<Constant>(Op)) {
10187 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
10190 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10192 GEP.setOperand(i, Op);
10198 if (MadeChange) return &GEP;
10200 // If this GEP instruction doesn't move the pointer, and if the input operand
10201 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10202 // real input to the dest type.
10203 if (GEP.hasAllZeroIndices()) {
10204 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10205 // If the bitcast is of an allocation, and the allocation will be
10206 // converted to match the type of the cast, don't touch this.
10207 if (isa<AllocationInst>(BCI->getOperand(0))) {
10208 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10209 if (Instruction *I = visitBitCast(*BCI)) {
10212 BCI->getParent()->getInstList().insert(BCI, I);
10213 ReplaceInstUsesWith(*BCI, I);
10218 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10222 // Combine Indices - If the source pointer to this getelementptr instruction
10223 // is a getelementptr instruction, combine the indices of the two
10224 // getelementptr instructions into a single instruction.
10226 SmallVector<Value*, 8> SrcGEPOperands;
10227 if (User *Src = dyn_castGetElementPtr(PtrOp))
10228 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10230 if (!SrcGEPOperands.empty()) {
10231 // Note that if our source is a gep chain itself that we wait for that
10232 // chain to be resolved before we perform this transformation. This
10233 // avoids us creating a TON of code in some cases.
10235 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10236 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10237 return 0; // Wait until our source is folded to completion.
10239 SmallVector<Value*, 8> Indices;
10241 // Find out whether the last index in the source GEP is a sequential idx.
10242 bool EndsWithSequential = false;
10243 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10244 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10245 EndsWithSequential = !isa<StructType>(*I);
10247 // Can we combine the two pointer arithmetics offsets?
10248 if (EndsWithSequential) {
10249 // Replace: gep (gep %P, long B), long A, ...
10250 // With: T = long A+B; gep %P, T, ...
10252 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10253 if (SO1 == Constant::getNullValue(SO1->getType())) {
10255 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10258 // If they aren't the same type, convert both to an integer of the
10259 // target's pointer size.
10260 if (SO1->getType() != GO1->getType()) {
10261 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10262 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10263 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10264 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10266 unsigned PS = TD->getPointerSizeInBits();
10267 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10268 // Convert GO1 to SO1's type.
10269 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10271 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10272 // Convert SO1 to GO1's type.
10273 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10275 const Type *PT = TD->getIntPtrType();
10276 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10277 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10281 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10282 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10284 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10285 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10289 // Recycle the GEP we already have if possible.
10290 if (SrcGEPOperands.size() == 2) {
10291 GEP.setOperand(0, SrcGEPOperands[0]);
10292 GEP.setOperand(1, Sum);
10295 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10296 SrcGEPOperands.end()-1);
10297 Indices.push_back(Sum);
10298 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10300 } else if (isa<Constant>(*GEP.idx_begin()) &&
10301 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10302 SrcGEPOperands.size() != 1) {
10303 // Otherwise we can do the fold if the first index of the GEP is a zero
10304 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10305 SrcGEPOperands.end());
10306 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10309 if (!Indices.empty())
10310 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10311 Indices.end(), GEP.getName());
10313 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10314 // GEP of global variable. If all of the indices for this GEP are
10315 // constants, we can promote this to a constexpr instead of an instruction.
10317 // Scan for nonconstants...
10318 SmallVector<Constant*, 8> Indices;
10319 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10320 for (; I != E && isa<Constant>(*I); ++I)
10321 Indices.push_back(cast<Constant>(*I));
10323 if (I == E) { // If they are all constants...
10324 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10325 &Indices[0],Indices.size());
10327 // Replace all uses of the GEP with the new constexpr...
10328 return ReplaceInstUsesWith(GEP, CE);
10330 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10331 if (!isa<PointerType>(X->getType())) {
10332 // Not interesting. Source pointer must be a cast from pointer.
10333 } else if (HasZeroPointerIndex) {
10334 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10335 // into : GEP [10 x i8]* X, i32 0, ...
10337 // This occurs when the program declares an array extern like "int X[];"
10339 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10340 const PointerType *XTy = cast<PointerType>(X->getType());
10341 if (const ArrayType *XATy =
10342 dyn_cast<ArrayType>(XTy->getElementType()))
10343 if (const ArrayType *CATy =
10344 dyn_cast<ArrayType>(CPTy->getElementType()))
10345 if (CATy->getElementType() == XATy->getElementType()) {
10346 // At this point, we know that the cast source type is a pointer
10347 // to an array of the same type as the destination pointer
10348 // array. Because the array type is never stepped over (there
10349 // is a leading zero) we can fold the cast into this GEP.
10350 GEP.setOperand(0, X);
10353 } else if (GEP.getNumOperands() == 2) {
10354 // Transform things like:
10355 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10356 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10357 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10358 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10359 if (isa<ArrayType>(SrcElTy) &&
10360 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10361 TD->getABITypeSize(ResElTy)) {
10363 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10364 Idx[1] = GEP.getOperand(1);
10365 Value *V = InsertNewInstBefore(
10366 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10367 // V and GEP are both pointer types --> BitCast
10368 return new BitCastInst(V, GEP.getType());
10371 // Transform things like:
10372 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10373 // (where tmp = 8*tmp2) into:
10374 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10376 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10377 uint64_t ArrayEltSize =
10378 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10380 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10381 // allow either a mul, shift, or constant here.
10383 ConstantInt *Scale = 0;
10384 if (ArrayEltSize == 1) {
10385 NewIdx = GEP.getOperand(1);
10386 Scale = ConstantInt::get(NewIdx->getType(), 1);
10387 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10388 NewIdx = ConstantInt::get(CI->getType(), 1);
10390 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10391 if (Inst->getOpcode() == Instruction::Shl &&
10392 isa<ConstantInt>(Inst->getOperand(1))) {
10393 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10394 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10395 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10396 NewIdx = Inst->getOperand(0);
10397 } else if (Inst->getOpcode() == Instruction::Mul &&
10398 isa<ConstantInt>(Inst->getOperand(1))) {
10399 Scale = cast<ConstantInt>(Inst->getOperand(1));
10400 NewIdx = Inst->getOperand(0);
10404 // If the index will be to exactly the right offset with the scale taken
10405 // out, perform the transformation. Note, we don't know whether Scale is
10406 // signed or not. We'll use unsigned version of division/modulo
10407 // operation after making sure Scale doesn't have the sign bit set.
10408 if (Scale && Scale->getSExtValue() >= 0LL &&
10409 Scale->getZExtValue() % ArrayEltSize == 0) {
10410 Scale = ConstantInt::get(Scale->getType(),
10411 Scale->getZExtValue() / ArrayEltSize);
10412 if (Scale->getZExtValue() != 1) {
10413 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10415 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10416 NewIdx = InsertNewInstBefore(Sc, GEP);
10419 // Insert the new GEP instruction.
10421 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10423 Instruction *NewGEP =
10424 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10425 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10426 // The NewGEP must be pointer typed, so must the old one -> BitCast
10427 return new BitCastInst(NewGEP, GEP.getType());
10436 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10437 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10438 if (AI.isArrayAllocation()) { // Check C != 1
10439 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10440 const Type *NewTy =
10441 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10442 AllocationInst *New = 0;
10444 // Create and insert the replacement instruction...
10445 if (isa<MallocInst>(AI))
10446 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10448 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10449 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10452 InsertNewInstBefore(New, AI);
10454 // Scan to the end of the allocation instructions, to skip over a block of
10455 // allocas if possible...
10457 BasicBlock::iterator It = New;
10458 while (isa<AllocationInst>(*It)) ++It;
10460 // Now that I is pointing to the first non-allocation-inst in the block,
10461 // insert our getelementptr instruction...
10463 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10467 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10468 New->getName()+".sub", It);
10470 // Now make everything use the getelementptr instead of the original
10472 return ReplaceInstUsesWith(AI, V);
10473 } else if (isa<UndefValue>(AI.getArraySize())) {
10474 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10478 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10479 // Note that we only do this for alloca's, because malloc should allocate and
10480 // return a unique pointer, even for a zero byte allocation.
10481 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10482 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10483 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10488 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10489 Value *Op = FI.getOperand(0);
10491 // free undef -> unreachable.
10492 if (isa<UndefValue>(Op)) {
10493 // Insert a new store to null because we cannot modify the CFG here.
10494 new StoreInst(ConstantInt::getTrue(),
10495 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10496 return EraseInstFromFunction(FI);
10499 // If we have 'free null' delete the instruction. This can happen in stl code
10500 // when lots of inlining happens.
10501 if (isa<ConstantPointerNull>(Op))
10502 return EraseInstFromFunction(FI);
10504 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10505 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10506 FI.setOperand(0, CI->getOperand(0));
10510 // Change free (gep X, 0,0,0,0) into free(X)
10511 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10512 if (GEPI->hasAllZeroIndices()) {
10513 AddToWorkList(GEPI);
10514 FI.setOperand(0, GEPI->getOperand(0));
10519 // Change free(malloc) into nothing, if the malloc has a single use.
10520 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10521 if (MI->hasOneUse()) {
10522 EraseInstFromFunction(FI);
10523 return EraseInstFromFunction(*MI);
10530 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10531 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10532 const TargetData *TD) {
10533 User *CI = cast<User>(LI.getOperand(0));
10534 Value *CastOp = CI->getOperand(0);
10536 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10537 // Instead of loading constant c string, use corresponding integer value
10538 // directly if string length is small enough.
10539 const std::string &Str = CE->getOperand(0)->getStringValue();
10540 if (!Str.empty()) {
10541 unsigned len = Str.length();
10542 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10543 unsigned numBits = Ty->getPrimitiveSizeInBits();
10544 // Replace LI with immediate integer store.
10545 if ((numBits >> 3) == len + 1) {
10546 APInt StrVal(numBits, 0);
10547 APInt SingleChar(numBits, 0);
10548 if (TD->isLittleEndian()) {
10549 for (signed i = len-1; i >= 0; i--) {
10550 SingleChar = (uint64_t) Str[i];
10551 StrVal = (StrVal << 8) | SingleChar;
10554 for (unsigned i = 0; i < len; i++) {
10555 SingleChar = (uint64_t) Str[i];
10556 StrVal = (StrVal << 8) | SingleChar;
10558 // Append NULL at the end.
10560 StrVal = (StrVal << 8) | SingleChar;
10562 Value *NL = ConstantInt::get(StrVal);
10563 return IC.ReplaceInstUsesWith(LI, NL);
10568 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10569 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10570 const Type *SrcPTy = SrcTy->getElementType();
10572 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10573 isa<VectorType>(DestPTy)) {
10574 // If the source is an array, the code below will not succeed. Check to
10575 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10577 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10578 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10579 if (ASrcTy->getNumElements() != 0) {
10581 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10582 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10583 SrcTy = cast<PointerType>(CastOp->getType());
10584 SrcPTy = SrcTy->getElementType();
10587 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10588 isa<VectorType>(SrcPTy)) &&
10589 // Do not allow turning this into a load of an integer, which is then
10590 // casted to a pointer, this pessimizes pointer analysis a lot.
10591 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10592 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10593 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10595 // Okay, we are casting from one integer or pointer type to another of
10596 // the same size. Instead of casting the pointer before the load, cast
10597 // the result of the loaded value.
10598 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10600 LI.isVolatile()),LI);
10601 // Now cast the result of the load.
10602 return new BitCastInst(NewLoad, LI.getType());
10609 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10610 /// from this value cannot trap. If it is not obviously safe to load from the
10611 /// specified pointer, we do a quick local scan of the basic block containing
10612 /// ScanFrom, to determine if the address is already accessed.
10613 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10614 // If it is an alloca it is always safe to load from.
10615 if (isa<AllocaInst>(V)) return true;
10617 // If it is a global variable it is mostly safe to load from.
10618 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10619 // Don't try to evaluate aliases. External weak GV can be null.
10620 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10622 // Otherwise, be a little bit agressive by scanning the local block where we
10623 // want to check to see if the pointer is already being loaded or stored
10624 // from/to. If so, the previous load or store would have already trapped,
10625 // so there is no harm doing an extra load (also, CSE will later eliminate
10626 // the load entirely).
10627 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10632 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10633 if (LI->getOperand(0) == V) return true;
10634 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10635 if (SI->getOperand(1) == V) return true;
10641 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10642 /// until we find the underlying object a pointer is referring to or something
10643 /// we don't understand. Note that the returned pointer may be offset from the
10644 /// input, because we ignore GEP indices.
10645 static Value *GetUnderlyingObject(Value *Ptr) {
10647 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10648 if (CE->getOpcode() == Instruction::BitCast ||
10649 CE->getOpcode() == Instruction::GetElementPtr)
10650 Ptr = CE->getOperand(0);
10653 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10654 Ptr = BCI->getOperand(0);
10655 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10656 Ptr = GEP->getOperand(0);
10663 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10664 Value *Op = LI.getOperand(0);
10666 // Attempt to improve the alignment.
10667 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10669 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10670 LI.getAlignment()))
10671 LI.setAlignment(KnownAlign);
10673 // load (cast X) --> cast (load X) iff safe
10674 if (isa<CastInst>(Op))
10675 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10678 // None of the following transforms are legal for volatile loads.
10679 if (LI.isVolatile()) return 0;
10681 if (&LI.getParent()->front() != &LI) {
10682 BasicBlock::iterator BBI = &LI; --BBI;
10683 // If the instruction immediately before this is a store to the same
10684 // address, do a simple form of store->load forwarding.
10685 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10686 if (SI->getOperand(1) == LI.getOperand(0))
10687 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10688 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10689 if (LIB->getOperand(0) == LI.getOperand(0))
10690 return ReplaceInstUsesWith(LI, LIB);
10693 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10694 const Value *GEPI0 = GEPI->getOperand(0);
10695 // TODO: Consider a target hook for valid address spaces for this xform.
10696 if (isa<ConstantPointerNull>(GEPI0) &&
10697 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10698 // Insert a new store to null instruction before the load to indicate
10699 // that this code is not reachable. We do this instead of inserting
10700 // an unreachable instruction directly because we cannot modify the
10702 new StoreInst(UndefValue::get(LI.getType()),
10703 Constant::getNullValue(Op->getType()), &LI);
10704 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10708 if (Constant *C = dyn_cast<Constant>(Op)) {
10709 // load null/undef -> undef
10710 // TODO: Consider a target hook for valid address spaces for this xform.
10711 if (isa<UndefValue>(C) || (C->isNullValue() &&
10712 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10713 // Insert a new store to null instruction before the load to indicate that
10714 // this code is not reachable. We do this instead of inserting an
10715 // unreachable instruction directly because we cannot modify the CFG.
10716 new StoreInst(UndefValue::get(LI.getType()),
10717 Constant::getNullValue(Op->getType()), &LI);
10718 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10721 // Instcombine load (constant global) into the value loaded.
10722 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10723 if (GV->isConstant() && !GV->isDeclaration())
10724 return ReplaceInstUsesWith(LI, GV->getInitializer());
10726 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10727 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10728 if (CE->getOpcode() == Instruction::GetElementPtr) {
10729 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10730 if (GV->isConstant() && !GV->isDeclaration())
10732 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10733 return ReplaceInstUsesWith(LI, V);
10734 if (CE->getOperand(0)->isNullValue()) {
10735 // Insert a new store to null instruction before the load to indicate
10736 // that this code is not reachable. We do this instead of inserting
10737 // an unreachable instruction directly because we cannot modify the
10739 new StoreInst(UndefValue::get(LI.getType()),
10740 Constant::getNullValue(Op->getType()), &LI);
10741 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10744 } else if (CE->isCast()) {
10745 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10751 // If this load comes from anywhere in a constant global, and if the global
10752 // is all undef or zero, we know what it loads.
10753 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10754 if (GV->isConstant() && GV->hasInitializer()) {
10755 if (GV->getInitializer()->isNullValue())
10756 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10757 else if (isa<UndefValue>(GV->getInitializer()))
10758 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10762 if (Op->hasOneUse()) {
10763 // Change select and PHI nodes to select values instead of addresses: this
10764 // helps alias analysis out a lot, allows many others simplifications, and
10765 // exposes redundancy in the code.
10767 // Note that we cannot do the transformation unless we know that the
10768 // introduced loads cannot trap! Something like this is valid as long as
10769 // the condition is always false: load (select bool %C, int* null, int* %G),
10770 // but it would not be valid if we transformed it to load from null
10771 // unconditionally.
10773 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10774 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10775 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10776 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10777 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10778 SI->getOperand(1)->getName()+".val"), LI);
10779 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10780 SI->getOperand(2)->getName()+".val"), LI);
10781 return SelectInst::Create(SI->getCondition(), V1, V2);
10784 // load (select (cond, null, P)) -> load P
10785 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10786 if (C->isNullValue()) {
10787 LI.setOperand(0, SI->getOperand(2));
10791 // load (select (cond, P, null)) -> load P
10792 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10793 if (C->isNullValue()) {
10794 LI.setOperand(0, SI->getOperand(1));
10802 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10804 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10805 User *CI = cast<User>(SI.getOperand(1));
10806 Value *CastOp = CI->getOperand(0);
10808 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10809 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10810 const Type *SrcPTy = SrcTy->getElementType();
10812 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10813 // If the source is an array, the code below will not succeed. Check to
10814 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10816 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10817 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10818 if (ASrcTy->getNumElements() != 0) {
10820 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10821 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10822 SrcTy = cast<PointerType>(CastOp->getType());
10823 SrcPTy = SrcTy->getElementType();
10826 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10827 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10828 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10830 // Okay, we are casting from one integer or pointer type to another of
10831 // the same size. Instead of casting the pointer before
10832 // the store, cast the value to be stored.
10834 Value *SIOp0 = SI.getOperand(0);
10835 Instruction::CastOps opcode = Instruction::BitCast;
10836 const Type* CastSrcTy = SIOp0->getType();
10837 const Type* CastDstTy = SrcPTy;
10838 if (isa<PointerType>(CastDstTy)) {
10839 if (CastSrcTy->isInteger())
10840 opcode = Instruction::IntToPtr;
10841 } else if (isa<IntegerType>(CastDstTy)) {
10842 if (isa<PointerType>(SIOp0->getType()))
10843 opcode = Instruction::PtrToInt;
10845 if (Constant *C = dyn_cast<Constant>(SIOp0))
10846 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10848 NewCast = IC.InsertNewInstBefore(
10849 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10851 return new StoreInst(NewCast, CastOp);
10858 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10859 Value *Val = SI.getOperand(0);
10860 Value *Ptr = SI.getOperand(1);
10862 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10863 EraseInstFromFunction(SI);
10868 // If the RHS is an alloca with a single use, zapify the store, making the
10870 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10871 if (isa<AllocaInst>(Ptr)) {
10872 EraseInstFromFunction(SI);
10877 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10878 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10879 GEP->getOperand(0)->hasOneUse()) {
10880 EraseInstFromFunction(SI);
10886 // Attempt to improve the alignment.
10887 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10889 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10890 SI.getAlignment()))
10891 SI.setAlignment(KnownAlign);
10893 // Do really simple DSE, to catch cases where there are several consequtive
10894 // stores to the same location, separated by a few arithmetic operations. This
10895 // situation often occurs with bitfield accesses.
10896 BasicBlock::iterator BBI = &SI;
10897 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10901 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10902 // Prev store isn't volatile, and stores to the same location?
10903 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10906 EraseInstFromFunction(*PrevSI);
10912 // If this is a load, we have to stop. However, if the loaded value is from
10913 // the pointer we're loading and is producing the pointer we're storing,
10914 // then *this* store is dead (X = load P; store X -> P).
10915 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10916 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10917 EraseInstFromFunction(SI);
10921 // Otherwise, this is a load from some other location. Stores before it
10922 // may not be dead.
10926 // Don't skip over loads or things that can modify memory.
10927 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10932 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10934 // store X, null -> turns into 'unreachable' in SimplifyCFG
10935 if (isa<ConstantPointerNull>(Ptr)) {
10936 if (!isa<UndefValue>(Val)) {
10937 SI.setOperand(0, UndefValue::get(Val->getType()));
10938 if (Instruction *U = dyn_cast<Instruction>(Val))
10939 AddToWorkList(U); // Dropped a use.
10942 return 0; // Do not modify these!
10945 // store undef, Ptr -> noop
10946 if (isa<UndefValue>(Val)) {
10947 EraseInstFromFunction(SI);
10952 // If the pointer destination is a cast, see if we can fold the cast into the
10954 if (isa<CastInst>(Ptr))
10955 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10957 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10959 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10963 // If this store is the last instruction in the basic block, and if the block
10964 // ends with an unconditional branch, try to move it to the successor block.
10966 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10967 if (BI->isUnconditional())
10968 if (SimplifyStoreAtEndOfBlock(SI))
10969 return 0; // xform done!
10974 /// SimplifyStoreAtEndOfBlock - Turn things like:
10975 /// if () { *P = v1; } else { *P = v2 }
10976 /// into a phi node with a store in the successor.
10978 /// Simplify things like:
10979 /// *P = v1; if () { *P = v2; }
10980 /// into a phi node with a store in the successor.
10982 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10983 BasicBlock *StoreBB = SI.getParent();
10985 // Check to see if the successor block has exactly two incoming edges. If
10986 // so, see if the other predecessor contains a store to the same location.
10987 // if so, insert a PHI node (if needed) and move the stores down.
10988 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10990 // Determine whether Dest has exactly two predecessors and, if so, compute
10991 // the other predecessor.
10992 pred_iterator PI = pred_begin(DestBB);
10993 BasicBlock *OtherBB = 0;
10994 if (*PI != StoreBB)
10997 if (PI == pred_end(DestBB))
11000 if (*PI != StoreBB) {
11005 if (++PI != pred_end(DestBB))
11009 // Verify that the other block ends in a branch and is not otherwise empty.
11010 BasicBlock::iterator BBI = OtherBB->getTerminator();
11011 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11012 if (!OtherBr || BBI == OtherBB->begin())
11015 // If the other block ends in an unconditional branch, check for the 'if then
11016 // else' case. there is an instruction before the branch.
11017 StoreInst *OtherStore = 0;
11018 if (OtherBr->isUnconditional()) {
11019 // If this isn't a store, or isn't a store to the same location, bail out.
11021 OtherStore = dyn_cast<StoreInst>(BBI);
11022 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11025 // Otherwise, the other block ended with a conditional branch. If one of the
11026 // destinations is StoreBB, then we have the if/then case.
11027 if (OtherBr->getSuccessor(0) != StoreBB &&
11028 OtherBr->getSuccessor(1) != StoreBB)
11031 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11032 // if/then triangle. See if there is a store to the same ptr as SI that
11033 // lives in OtherBB.
11035 // Check to see if we find the matching store.
11036 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11037 if (OtherStore->getOperand(1) != SI.getOperand(1))
11041 // If we find something that may be using the stored value, or if we run
11042 // out of instructions, we can't do the xform.
11043 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
11044 BBI == OtherBB->begin())
11048 // In order to eliminate the store in OtherBr, we have to
11049 // make sure nothing reads the stored value in StoreBB.
11050 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11051 // FIXME: This should really be AA driven.
11052 if (isa<LoadInst>(I) || I->mayWriteToMemory())
11057 // Insert a PHI node now if we need it.
11058 Value *MergedVal = OtherStore->getOperand(0);
11059 if (MergedVal != SI.getOperand(0)) {
11060 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11061 PN->reserveOperandSpace(2);
11062 PN->addIncoming(SI.getOperand(0), SI.getParent());
11063 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11064 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11067 // Advance to a place where it is safe to insert the new store and
11069 BBI = DestBB->begin();
11070 while (isa<PHINode>(BBI)) ++BBI;
11071 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11072 OtherStore->isVolatile()), *BBI);
11074 // Nuke the old stores.
11075 EraseInstFromFunction(SI);
11076 EraseInstFromFunction(*OtherStore);
11082 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11083 // Change br (not X), label True, label False to: br X, label False, True
11085 BasicBlock *TrueDest;
11086 BasicBlock *FalseDest;
11087 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11088 !isa<Constant>(X)) {
11089 // Swap Destinations and condition...
11090 BI.setCondition(X);
11091 BI.setSuccessor(0, FalseDest);
11092 BI.setSuccessor(1, TrueDest);
11096 // Cannonicalize fcmp_one -> fcmp_oeq
11097 FCmpInst::Predicate FPred; Value *Y;
11098 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11099 TrueDest, FalseDest)))
11100 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11101 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11102 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11103 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11104 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11105 NewSCC->takeName(I);
11106 // Swap Destinations and condition...
11107 BI.setCondition(NewSCC);
11108 BI.setSuccessor(0, FalseDest);
11109 BI.setSuccessor(1, TrueDest);
11110 RemoveFromWorkList(I);
11111 I->eraseFromParent();
11112 AddToWorkList(NewSCC);
11116 // Cannonicalize icmp_ne -> icmp_eq
11117 ICmpInst::Predicate IPred;
11118 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11119 TrueDest, FalseDest)))
11120 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11121 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11122 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11123 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11124 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11125 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11126 NewSCC->takeName(I);
11127 // Swap Destinations and condition...
11128 BI.setCondition(NewSCC);
11129 BI.setSuccessor(0, FalseDest);
11130 BI.setSuccessor(1, TrueDest);
11131 RemoveFromWorkList(I);
11132 I->eraseFromParent();;
11133 AddToWorkList(NewSCC);
11140 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11141 Value *Cond = SI.getCondition();
11142 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11143 if (I->getOpcode() == Instruction::Add)
11144 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11145 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11146 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11147 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11149 SI.setOperand(0, I->getOperand(0));
11157 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11158 /// is to leave as a vector operation.
11159 static bool CheapToScalarize(Value *V, bool isConstant) {
11160 if (isa<ConstantAggregateZero>(V))
11162 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11163 if (isConstant) return true;
11164 // If all elts are the same, we can extract.
11165 Constant *Op0 = C->getOperand(0);
11166 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11167 if (C->getOperand(i) != Op0)
11171 Instruction *I = dyn_cast<Instruction>(V);
11172 if (!I) return false;
11174 // Insert element gets simplified to the inserted element or is deleted if
11175 // this is constant idx extract element and its a constant idx insertelt.
11176 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11177 isa<ConstantInt>(I->getOperand(2)))
11179 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11181 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11182 if (BO->hasOneUse() &&
11183 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11184 CheapToScalarize(BO->getOperand(1), isConstant)))
11186 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11187 if (CI->hasOneUse() &&
11188 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11189 CheapToScalarize(CI->getOperand(1), isConstant)))
11195 /// Read and decode a shufflevector mask.
11197 /// It turns undef elements into values that are larger than the number of
11198 /// elements in the input.
11199 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11200 unsigned NElts = SVI->getType()->getNumElements();
11201 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11202 return std::vector<unsigned>(NElts, 0);
11203 if (isa<UndefValue>(SVI->getOperand(2)))
11204 return std::vector<unsigned>(NElts, 2*NElts);
11206 std::vector<unsigned> Result;
11207 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11208 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
11209 if (isa<UndefValue>(CP->getOperand(i)))
11210 Result.push_back(NElts*2); // undef -> 8
11212 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
11216 /// FindScalarElement - Given a vector and an element number, see if the scalar
11217 /// value is already around as a register, for example if it were inserted then
11218 /// extracted from the vector.
11219 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11220 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11221 const VectorType *PTy = cast<VectorType>(V->getType());
11222 unsigned Width = PTy->getNumElements();
11223 if (EltNo >= Width) // Out of range access.
11224 return UndefValue::get(PTy->getElementType());
11226 if (isa<UndefValue>(V))
11227 return UndefValue::get(PTy->getElementType());
11228 else if (isa<ConstantAggregateZero>(V))
11229 return Constant::getNullValue(PTy->getElementType());
11230 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11231 return CP->getOperand(EltNo);
11232 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11233 // If this is an insert to a variable element, we don't know what it is.
11234 if (!isa<ConstantInt>(III->getOperand(2)))
11236 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11238 // If this is an insert to the element we are looking for, return the
11240 if (EltNo == IIElt)
11241 return III->getOperand(1);
11243 // Otherwise, the insertelement doesn't modify the value, recurse on its
11245 return FindScalarElement(III->getOperand(0), EltNo);
11246 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11247 unsigned InEl = getShuffleMask(SVI)[EltNo];
11249 return FindScalarElement(SVI->getOperand(0), InEl);
11250 else if (InEl < Width*2)
11251 return FindScalarElement(SVI->getOperand(1), InEl - Width);
11253 return UndefValue::get(PTy->getElementType());
11256 // Otherwise, we don't know.
11260 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11262 // If vector val is undef, replace extract with scalar undef.
11263 if (isa<UndefValue>(EI.getOperand(0)))
11264 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11266 // If vector val is constant 0, replace extract with scalar 0.
11267 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11268 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11270 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11271 // If vector val is constant with uniform operands, replace EI
11272 // with that operand
11273 Constant *op0 = C->getOperand(0);
11274 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11275 if (C->getOperand(i) != op0) {
11280 return ReplaceInstUsesWith(EI, op0);
11283 // If extracting a specified index from the vector, see if we can recursively
11284 // find a previously computed scalar that was inserted into the vector.
11285 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11286 unsigned IndexVal = IdxC->getZExtValue();
11287 unsigned VectorWidth =
11288 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11290 // If this is extracting an invalid index, turn this into undef, to avoid
11291 // crashing the code below.
11292 if (IndexVal >= VectorWidth)
11293 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11295 // This instruction only demands the single element from the input vector.
11296 // If the input vector has a single use, simplify it based on this use
11298 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11299 uint64_t UndefElts;
11300 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11303 EI.setOperand(0, V);
11308 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11309 return ReplaceInstUsesWith(EI, Elt);
11311 // If the this extractelement is directly using a bitcast from a vector of
11312 // the same number of elements, see if we can find the source element from
11313 // it. In this case, we will end up needing to bitcast the scalars.
11314 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11315 if (const VectorType *VT =
11316 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11317 if (VT->getNumElements() == VectorWidth)
11318 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11319 return new BitCastInst(Elt, EI.getType());
11323 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11324 if (I->hasOneUse()) {
11325 // Push extractelement into predecessor operation if legal and
11326 // profitable to do so
11327 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11328 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11329 if (CheapToScalarize(BO, isConstantElt)) {
11330 ExtractElementInst *newEI0 =
11331 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11332 EI.getName()+".lhs");
11333 ExtractElementInst *newEI1 =
11334 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11335 EI.getName()+".rhs");
11336 InsertNewInstBefore(newEI0, EI);
11337 InsertNewInstBefore(newEI1, EI);
11338 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11340 } else if (isa<LoadInst>(I)) {
11342 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11343 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11344 PointerType::get(EI.getType(), AS),EI);
11345 GetElementPtrInst *GEP =
11346 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11347 InsertNewInstBefore(GEP, EI);
11348 return new LoadInst(GEP);
11351 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11352 // Extracting the inserted element?
11353 if (IE->getOperand(2) == EI.getOperand(1))
11354 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11355 // If the inserted and extracted elements are constants, they must not
11356 // be the same value, extract from the pre-inserted value instead.
11357 if (isa<Constant>(IE->getOperand(2)) &&
11358 isa<Constant>(EI.getOperand(1))) {
11359 AddUsesToWorkList(EI);
11360 EI.setOperand(0, IE->getOperand(0));
11363 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11364 // If this is extracting an element from a shufflevector, figure out where
11365 // it came from and extract from the appropriate input element instead.
11366 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11367 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11369 if (SrcIdx < SVI->getType()->getNumElements())
11370 Src = SVI->getOperand(0);
11371 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
11372 SrcIdx -= SVI->getType()->getNumElements();
11373 Src = SVI->getOperand(1);
11375 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11377 return new ExtractElementInst(Src, SrcIdx);
11384 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11385 /// elements from either LHS or RHS, return the shuffle mask and true.
11386 /// Otherwise, return false.
11387 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11388 std::vector<Constant*> &Mask) {
11389 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11390 "Invalid CollectSingleShuffleElements");
11391 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11393 if (isa<UndefValue>(V)) {
11394 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11396 } else if (V == LHS) {
11397 for (unsigned i = 0; i != NumElts; ++i)
11398 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11400 } else if (V == RHS) {
11401 for (unsigned i = 0; i != NumElts; ++i)
11402 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11404 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11405 // If this is an insert of an extract from some other vector, include it.
11406 Value *VecOp = IEI->getOperand(0);
11407 Value *ScalarOp = IEI->getOperand(1);
11408 Value *IdxOp = IEI->getOperand(2);
11410 if (!isa<ConstantInt>(IdxOp))
11412 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11414 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11415 // Okay, we can handle this if the vector we are insertinting into is
11416 // transitively ok.
11417 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11418 // If so, update the mask to reflect the inserted undef.
11419 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11422 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11423 if (isa<ConstantInt>(EI->getOperand(1)) &&
11424 EI->getOperand(0)->getType() == V->getType()) {
11425 unsigned ExtractedIdx =
11426 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11428 // This must be extracting from either LHS or RHS.
11429 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11430 // Okay, we can handle this if the vector we are insertinting into is
11431 // transitively ok.
11432 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11433 // If so, update the mask to reflect the inserted value.
11434 if (EI->getOperand(0) == LHS) {
11435 Mask[InsertedIdx & (NumElts-1)] =
11436 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11438 assert(EI->getOperand(0) == RHS);
11439 Mask[InsertedIdx & (NumElts-1)] =
11440 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11449 // TODO: Handle shufflevector here!
11454 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11455 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11456 /// that computes V and the LHS value of the shuffle.
11457 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11459 assert(isa<VectorType>(V->getType()) &&
11460 (RHS == 0 || V->getType() == RHS->getType()) &&
11461 "Invalid shuffle!");
11462 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11464 if (isa<UndefValue>(V)) {
11465 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11467 } else if (isa<ConstantAggregateZero>(V)) {
11468 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11470 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11471 // If this is an insert of an extract from some other vector, include it.
11472 Value *VecOp = IEI->getOperand(0);
11473 Value *ScalarOp = IEI->getOperand(1);
11474 Value *IdxOp = IEI->getOperand(2);
11476 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11477 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11478 EI->getOperand(0)->getType() == V->getType()) {
11479 unsigned ExtractedIdx =
11480 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11481 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11483 // Either the extracted from or inserted into vector must be RHSVec,
11484 // otherwise we'd end up with a shuffle of three inputs.
11485 if (EI->getOperand(0) == RHS || RHS == 0) {
11486 RHS = EI->getOperand(0);
11487 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11488 Mask[InsertedIdx & (NumElts-1)] =
11489 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11493 if (VecOp == RHS) {
11494 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11495 // Everything but the extracted element is replaced with the RHS.
11496 for (unsigned i = 0; i != NumElts; ++i) {
11497 if (i != InsertedIdx)
11498 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11503 // If this insertelement is a chain that comes from exactly these two
11504 // vectors, return the vector and the effective shuffle.
11505 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11506 return EI->getOperand(0);
11511 // TODO: Handle shufflevector here!
11513 // Otherwise, can't do anything fancy. Return an identity vector.
11514 for (unsigned i = 0; i != NumElts; ++i)
11515 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11519 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11520 Value *VecOp = IE.getOperand(0);
11521 Value *ScalarOp = IE.getOperand(1);
11522 Value *IdxOp = IE.getOperand(2);
11524 // Inserting an undef or into an undefined place, remove this.
11525 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11526 ReplaceInstUsesWith(IE, VecOp);
11528 // If the inserted element was extracted from some other vector, and if the
11529 // indexes are constant, try to turn this into a shufflevector operation.
11530 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11531 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11532 EI->getOperand(0)->getType() == IE.getType()) {
11533 unsigned NumVectorElts = IE.getType()->getNumElements();
11534 unsigned ExtractedIdx =
11535 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11536 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11538 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11539 return ReplaceInstUsesWith(IE, VecOp);
11541 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11542 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11544 // If we are extracting a value from a vector, then inserting it right
11545 // back into the same place, just use the input vector.
11546 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11547 return ReplaceInstUsesWith(IE, VecOp);
11549 // We could theoretically do this for ANY input. However, doing so could
11550 // turn chains of insertelement instructions into a chain of shufflevector
11551 // instructions, and right now we do not merge shufflevectors. As such,
11552 // only do this in a situation where it is clear that there is benefit.
11553 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11554 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11555 // the values of VecOp, except then one read from EIOp0.
11556 // Build a new shuffle mask.
11557 std::vector<Constant*> Mask;
11558 if (isa<UndefValue>(VecOp))
11559 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11561 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11562 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11565 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11566 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11567 ConstantVector::get(Mask));
11570 // If this insertelement isn't used by some other insertelement, turn it
11571 // (and any insertelements it points to), into one big shuffle.
11572 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11573 std::vector<Constant*> Mask;
11575 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11576 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11577 // We now have a shuffle of LHS, RHS, Mask.
11578 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11587 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11588 Value *LHS = SVI.getOperand(0);
11589 Value *RHS = SVI.getOperand(1);
11590 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11592 bool MadeChange = false;
11594 // Undefined shuffle mask -> undefined value.
11595 if (isa<UndefValue>(SVI.getOperand(2)))
11596 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11598 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11599 // the undef, change them to undefs.
11600 if (isa<UndefValue>(SVI.getOperand(1))) {
11601 // Scan to see if there are any references to the RHS. If so, replace them
11602 // with undef element refs and set MadeChange to true.
11603 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11604 if (Mask[i] >= e && Mask[i] != 2*e) {
11611 // Remap any references to RHS to use LHS.
11612 std::vector<Constant*> Elts;
11613 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11614 if (Mask[i] == 2*e)
11615 Elts.push_back(UndefValue::get(Type::Int32Ty));
11617 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11619 SVI.setOperand(2, ConstantVector::get(Elts));
11623 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11624 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11625 if (LHS == RHS || isa<UndefValue>(LHS)) {
11626 if (isa<UndefValue>(LHS) && LHS == RHS) {
11627 // shuffle(undef,undef,mask) -> undef.
11628 return ReplaceInstUsesWith(SVI, LHS);
11631 // Remap any references to RHS to use LHS.
11632 std::vector<Constant*> Elts;
11633 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11634 if (Mask[i] >= 2*e)
11635 Elts.push_back(UndefValue::get(Type::Int32Ty));
11637 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11638 (Mask[i] < e && isa<UndefValue>(LHS)))
11639 Mask[i] = 2*e; // Turn into undef.
11641 Mask[i] &= (e-1); // Force to LHS.
11642 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11645 SVI.setOperand(0, SVI.getOperand(1));
11646 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11647 SVI.setOperand(2, ConstantVector::get(Elts));
11648 LHS = SVI.getOperand(0);
11649 RHS = SVI.getOperand(1);
11653 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11654 bool isLHSID = true, isRHSID = true;
11656 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11657 if (Mask[i] >= e*2) continue; // Ignore undef values.
11658 // Is this an identity shuffle of the LHS value?
11659 isLHSID &= (Mask[i] == i);
11661 // Is this an identity shuffle of the RHS value?
11662 isRHSID &= (Mask[i]-e == i);
11665 // Eliminate identity shuffles.
11666 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11667 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11669 // If the LHS is a shufflevector itself, see if we can combine it with this
11670 // one without producing an unusual shuffle. Here we are really conservative:
11671 // we are absolutely afraid of producing a shuffle mask not in the input
11672 // program, because the code gen may not be smart enough to turn a merged
11673 // shuffle into two specific shuffles: it may produce worse code. As such,
11674 // we only merge two shuffles if the result is one of the two input shuffle
11675 // masks. In this case, merging the shuffles just removes one instruction,
11676 // which we know is safe. This is good for things like turning:
11677 // (splat(splat)) -> splat.
11678 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11679 if (isa<UndefValue>(RHS)) {
11680 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11682 std::vector<unsigned> NewMask;
11683 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11684 if (Mask[i] >= 2*e)
11685 NewMask.push_back(2*e);
11687 NewMask.push_back(LHSMask[Mask[i]]);
11689 // If the result mask is equal to the src shuffle or this shuffle mask, do
11690 // the replacement.
11691 if (NewMask == LHSMask || NewMask == Mask) {
11692 std::vector<Constant*> Elts;
11693 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11694 if (NewMask[i] >= e*2) {
11695 Elts.push_back(UndefValue::get(Type::Int32Ty));
11697 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11700 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11701 LHSSVI->getOperand(1),
11702 ConstantVector::get(Elts));
11707 return MadeChange ? &SVI : 0;
11713 /// TryToSinkInstruction - Try to move the specified instruction from its
11714 /// current block into the beginning of DestBlock, which can only happen if it's
11715 /// safe to move the instruction past all of the instructions between it and the
11716 /// end of its block.
11717 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11718 assert(I->hasOneUse() && "Invariants didn't hold!");
11720 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11721 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11724 // Do not sink alloca instructions out of the entry block.
11725 if (isa<AllocaInst>(I) && I->getParent() ==
11726 &DestBlock->getParent()->getEntryBlock())
11729 // We can only sink load instructions if there is nothing between the load and
11730 // the end of block that could change the value.
11731 if (I->mayReadFromMemory()) {
11732 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11734 if (Scan->mayWriteToMemory())
11738 BasicBlock::iterator InsertPos = DestBlock->begin();
11739 while (isa<PHINode>(InsertPos)) ++InsertPos;
11741 I->moveBefore(InsertPos);
11747 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11748 /// all reachable code to the worklist.
11750 /// This has a couple of tricks to make the code faster and more powerful. In
11751 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11752 /// them to the worklist (this significantly speeds up instcombine on code where
11753 /// many instructions are dead or constant). Additionally, if we find a branch
11754 /// whose condition is a known constant, we only visit the reachable successors.
11756 static void AddReachableCodeToWorklist(BasicBlock *BB,
11757 SmallPtrSet<BasicBlock*, 64> &Visited,
11759 const TargetData *TD) {
11760 std::vector<BasicBlock*> Worklist;
11761 Worklist.push_back(BB);
11763 while (!Worklist.empty()) {
11764 BB = Worklist.back();
11765 Worklist.pop_back();
11767 // We have now visited this block! If we've already been here, ignore it.
11768 if (!Visited.insert(BB)) continue;
11770 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11771 Instruction *Inst = BBI++;
11773 // DCE instruction if trivially dead.
11774 if (isInstructionTriviallyDead(Inst)) {
11776 DOUT << "IC: DCE: " << *Inst;
11777 Inst->eraseFromParent();
11781 // ConstantProp instruction if trivially constant.
11782 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11783 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11784 Inst->replaceAllUsesWith(C);
11786 Inst->eraseFromParent();
11790 IC.AddToWorkList(Inst);
11793 // Recursively visit successors. If this is a branch or switch on a
11794 // constant, only visit the reachable successor.
11795 TerminatorInst *TI = BB->getTerminator();
11796 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11797 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11798 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11799 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11800 Worklist.push_back(ReachableBB);
11803 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11804 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11805 // See if this is an explicit destination.
11806 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11807 if (SI->getCaseValue(i) == Cond) {
11808 BasicBlock *ReachableBB = SI->getSuccessor(i);
11809 Worklist.push_back(ReachableBB);
11813 // Otherwise it is the default destination.
11814 Worklist.push_back(SI->getSuccessor(0));
11819 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11820 Worklist.push_back(TI->getSuccessor(i));
11824 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11825 bool Changed = false;
11826 TD = &getAnalysis<TargetData>();
11828 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11829 << F.getNameStr() << "\n");
11832 // Do a depth-first traversal of the function, populate the worklist with
11833 // the reachable instructions. Ignore blocks that are not reachable. Keep
11834 // track of which blocks we visit.
11835 SmallPtrSet<BasicBlock*, 64> Visited;
11836 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11838 // Do a quick scan over the function. If we find any blocks that are
11839 // unreachable, remove any instructions inside of them. This prevents
11840 // the instcombine code from having to deal with some bad special cases.
11841 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11842 if (!Visited.count(BB)) {
11843 Instruction *Term = BB->getTerminator();
11844 while (Term != BB->begin()) { // Remove instrs bottom-up
11845 BasicBlock::iterator I = Term; --I;
11847 DOUT << "IC: DCE: " << *I;
11850 if (!I->use_empty())
11851 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11852 I->eraseFromParent();
11857 while (!Worklist.empty()) {
11858 Instruction *I = RemoveOneFromWorkList();
11859 if (I == 0) continue; // skip null values.
11861 // Check to see if we can DCE the instruction.
11862 if (isInstructionTriviallyDead(I)) {
11863 // Add operands to the worklist.
11864 if (I->getNumOperands() < 4)
11865 AddUsesToWorkList(*I);
11868 DOUT << "IC: DCE: " << *I;
11870 I->eraseFromParent();
11871 RemoveFromWorkList(I);
11875 // Instruction isn't dead, see if we can constant propagate it.
11876 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11877 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11879 // Add operands to the worklist.
11880 AddUsesToWorkList(*I);
11881 ReplaceInstUsesWith(*I, C);
11884 I->eraseFromParent();
11885 RemoveFromWorkList(I);
11889 // See if we can trivially sink this instruction to a successor basic block.
11890 // FIXME: Remove GetResultInst test when first class support for aggregates
11892 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11893 BasicBlock *BB = I->getParent();
11894 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11895 if (UserParent != BB) {
11896 bool UserIsSuccessor = false;
11897 // See if the user is one of our successors.
11898 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11899 if (*SI == UserParent) {
11900 UserIsSuccessor = true;
11904 // If the user is one of our immediate successors, and if that successor
11905 // only has us as a predecessors (we'd have to split the critical edge
11906 // otherwise), we can keep going.
11907 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11908 next(pred_begin(UserParent)) == pred_end(UserParent))
11909 // Okay, the CFG is simple enough, try to sink this instruction.
11910 Changed |= TryToSinkInstruction(I, UserParent);
11914 // Now that we have an instruction, try combining it to simplify it...
11918 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11919 if (Instruction *Result = visit(*I)) {
11921 // Should we replace the old instruction with a new one?
11923 DOUT << "IC: Old = " << *I
11924 << " New = " << *Result;
11926 // Everything uses the new instruction now.
11927 I->replaceAllUsesWith(Result);
11929 // Push the new instruction and any users onto the worklist.
11930 AddToWorkList(Result);
11931 AddUsersToWorkList(*Result);
11933 // Move the name to the new instruction first.
11934 Result->takeName(I);
11936 // Insert the new instruction into the basic block...
11937 BasicBlock *InstParent = I->getParent();
11938 BasicBlock::iterator InsertPos = I;
11940 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11941 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11944 InstParent->getInstList().insert(InsertPos, Result);
11946 // Make sure that we reprocess all operands now that we reduced their
11948 AddUsesToWorkList(*I);
11950 // Instructions can end up on the worklist more than once. Make sure
11951 // we do not process an instruction that has been deleted.
11952 RemoveFromWorkList(I);
11954 // Erase the old instruction.
11955 InstParent->getInstList().erase(I);
11958 DOUT << "IC: Mod = " << OrigI
11959 << " New = " << *I;
11962 // If the instruction was modified, it's possible that it is now dead.
11963 // if so, remove it.
11964 if (isInstructionTriviallyDead(I)) {
11965 // Make sure we process all operands now that we are reducing their
11967 AddUsesToWorkList(*I);
11969 // Instructions may end up in the worklist more than once. Erase all
11970 // occurrences of this instruction.
11971 RemoveFromWorkList(I);
11972 I->eraseFromParent();
11975 AddUsersToWorkList(*I);
11982 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11984 // Do an explicit clear, this shrinks the map if needed.
11985 WorklistMap.clear();
11990 bool InstCombiner::runOnFunction(Function &F) {
11991 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11993 bool EverMadeChange = false;
11995 // Iterate while there is work to do.
11996 unsigned Iteration = 0;
11997 while (DoOneIteration(F, Iteration++))
11998 EverMadeChange = true;
11999 return EverMadeChange;
12002 FunctionPass *llvm::createInstructionCombiningPass() {
12003 return new InstCombiner();