1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 SmallVector<Instruction*, 256> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass(&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 bool SimplifyDivRemOfSelect(BinaryOperator &I);
176 Instruction *commonRemTransforms(BinaryOperator &I);
177 Instruction *commonIRemTransforms(BinaryOperator &I);
178 Instruction *commonDivTransforms(BinaryOperator &I);
179 Instruction *commonIDivTransforms(BinaryOperator &I);
180 Instruction *visitUDiv(BinaryOperator &I);
181 Instruction *visitSDiv(BinaryOperator &I);
182 Instruction *visitFDiv(BinaryOperator &I);
183 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
184 Instruction *visitAnd(BinaryOperator &I);
185 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
186 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
187 Value *A, Value *B, Value *C);
188 Instruction *visitOr (BinaryOperator &I);
189 Instruction *visitXor(BinaryOperator &I);
190 Instruction *visitShl(BinaryOperator &I);
191 Instruction *visitAShr(BinaryOperator &I);
192 Instruction *visitLShr(BinaryOperator &I);
193 Instruction *commonShiftTransforms(BinaryOperator &I);
194 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
196 Instruction *visitFCmpInst(FCmpInst &I);
197 Instruction *visitICmpInst(ICmpInst &I);
198 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
199 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
202 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
203 ConstantInt *DivRHS);
205 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
206 ICmpInst::Predicate Cond, Instruction &I);
207 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
209 Instruction *commonCastTransforms(CastInst &CI);
210 Instruction *commonIntCastTransforms(CastInst &CI);
211 Instruction *commonPointerCastTransforms(CastInst &CI);
212 Instruction *visitTrunc(TruncInst &CI);
213 Instruction *visitZExt(ZExtInst &CI);
214 Instruction *visitSExt(SExtInst &CI);
215 Instruction *visitFPTrunc(FPTruncInst &CI);
216 Instruction *visitFPExt(CastInst &CI);
217 Instruction *visitFPToUI(FPToUIInst &FI);
218 Instruction *visitFPToSI(FPToSIInst &FI);
219 Instruction *visitUIToFP(CastInst &CI);
220 Instruction *visitSIToFP(CastInst &CI);
221 Instruction *visitPtrToInt(PtrToIntInst &CI);
222 Instruction *visitIntToPtr(IntToPtrInst &CI);
223 Instruction *visitBitCast(BitCastInst &CI);
224 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
226 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
227 Instruction *visitSelectInst(SelectInst &SI);
228 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
229 Instruction *visitCallInst(CallInst &CI);
230 Instruction *visitInvokeInst(InvokeInst &II);
231 Instruction *visitPHINode(PHINode &PN);
232 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
233 Instruction *visitAllocationInst(AllocationInst &AI);
234 Instruction *visitFreeInst(FreeInst &FI);
235 Instruction *visitLoadInst(LoadInst &LI);
236 Instruction *visitStoreInst(StoreInst &SI);
237 Instruction *visitBranchInst(BranchInst &BI);
238 Instruction *visitSwitchInst(SwitchInst &SI);
239 Instruction *visitInsertElementInst(InsertElementInst &IE);
240 Instruction *visitExtractElementInst(ExtractElementInst &EI);
241 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
242 Instruction *visitExtractValueInst(ExtractValueInst &EV);
244 // visitInstruction - Specify what to return for unhandled instructions...
245 Instruction *visitInstruction(Instruction &I) { return 0; }
248 Instruction *visitCallSite(CallSite CS);
249 bool transformConstExprCastCall(CallSite CS);
250 Instruction *transformCallThroughTrampoline(CallSite CS);
251 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
252 bool DoXform = true);
253 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
254 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
258 // InsertNewInstBefore - insert an instruction New before instruction Old
259 // in the program. Add the new instruction to the worklist.
261 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
262 assert(New && New->getParent() == 0 &&
263 "New instruction already inserted into a basic block!");
264 BasicBlock *BB = Old.getParent();
265 BB->getInstList().insert(&Old, New); // Insert inst
270 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
271 /// This also adds the cast to the worklist. Finally, this returns the
273 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
275 if (V->getType() == Ty) return V;
277 if (Constant *CV = dyn_cast<Constant>(V))
278 return ConstantExpr::getCast(opc, CV, Ty);
280 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
285 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
286 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
290 // ReplaceInstUsesWith - This method is to be used when an instruction is
291 // found to be dead, replacable with another preexisting expression. Here
292 // we add all uses of I to the worklist, replace all uses of I with the new
293 // value, then return I, so that the inst combiner will know that I was
296 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
297 AddUsersToWorkList(I); // Add all modified instrs to worklist
299 I.replaceAllUsesWith(V);
302 // If we are replacing the instruction with itself, this must be in a
303 // segment of unreachable code, so just clobber the instruction.
304 I.replaceAllUsesWith(UndefValue::get(I.getType()));
309 // EraseInstFromFunction - When dealing with an instruction that has side
310 // effects or produces a void value, we can't rely on DCE to delete the
311 // instruction. Instead, visit methods should return the value returned by
313 Instruction *EraseInstFromFunction(Instruction &I) {
314 assert(I.use_empty() && "Cannot erase instruction that is used!");
315 AddUsesToWorkList(I);
316 RemoveFromWorkList(&I);
318 return 0; // Don't do anything with FI
321 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
322 APInt &KnownOne, unsigned Depth = 0) const {
323 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
326 bool MaskedValueIsZero(Value *V, const APInt &Mask,
327 unsigned Depth = 0) const {
328 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
330 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
331 return llvm::ComputeNumSignBits(Op, TD, Depth);
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 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
345 /// based on the demanded bits.
346 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
347 APInt& KnownZero, APInt& KnownOne,
349 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
350 APInt& KnownZero, APInt& KnownOne,
353 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
354 /// SimplifyDemandedBits knows about. See if the instruction has any
355 /// properties that allow us to simplify its operands.
356 bool SimplifyDemandedInstructionBits(Instruction &Inst);
358 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
359 APInt& UndefElts, unsigned Depth = 0);
361 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
362 // PHI node as operand #0, see if we can fold the instruction into the PHI
363 // (which is only possible if all operands to the PHI are constants).
364 Instruction *FoldOpIntoPhi(Instruction &I);
366 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
367 // operator and they all are only used by the PHI, PHI together their
368 // inputs, and do the operation once, to the result of the PHI.
369 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
370 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
371 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
374 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
375 ConstantInt *AndRHS, BinaryOperator &TheAnd);
377 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
378 bool isSub, Instruction &I);
379 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
380 bool isSigned, bool Inside, Instruction &IB);
381 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
382 Instruction *MatchBSwap(BinaryOperator &I);
383 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
384 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
385 Instruction *SimplifyMemSet(MemSetInst *MI);
388 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
390 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
391 unsigned CastOpc, int &NumCastsRemoved);
392 unsigned GetOrEnforceKnownAlignment(Value *V,
393 unsigned PrefAlign = 0);
398 char InstCombiner::ID = 0;
399 static RegisterPass<InstCombiner>
400 X("instcombine", "Combine redundant instructions");
402 // getComplexity: Assign a complexity or rank value to LLVM Values...
403 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
404 static unsigned getComplexity(Value *V) {
405 if (isa<Instruction>(V)) {
406 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
410 if (isa<Argument>(V)) return 3;
411 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
414 // isOnlyUse - Return true if this instruction will be deleted if we stop using
416 static bool isOnlyUse(Value *V) {
417 return V->hasOneUse() || isa<Constant>(V);
420 // getPromotedType - Return the specified type promoted as it would be to pass
421 // though a va_arg area...
422 static const Type *getPromotedType(const Type *Ty) {
423 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
424 if (ITy->getBitWidth() < 32)
425 return Type::Int32Ty;
430 /// getBitCastOperand - If the specified operand is a CastInst, a constant
431 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
432 /// operand value, otherwise return null.
433 static Value *getBitCastOperand(Value *V) {
434 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
436 return I->getOperand(0);
437 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
438 // GetElementPtrInst?
439 if (GEP->hasAllZeroIndices())
440 return GEP->getOperand(0);
441 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
442 if (CE->getOpcode() == Instruction::BitCast)
443 // BitCast ConstantExp?
444 return CE->getOperand(0);
445 else if (CE->getOpcode() == Instruction::GetElementPtr) {
446 // GetElementPtr ConstantExp?
447 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
449 ConstantInt *CI = dyn_cast<ConstantInt>(I);
450 if (!CI || !CI->isZero())
451 // Any non-zero indices? Not cast-like.
454 // All-zero indices? This is just like casting.
455 return CE->getOperand(0);
461 /// This function is a wrapper around CastInst::isEliminableCastPair. It
462 /// simply extracts arguments and returns what that function returns.
463 static Instruction::CastOps
464 isEliminableCastPair(
465 const CastInst *CI, ///< The first cast instruction
466 unsigned opcode, ///< The opcode of the second cast instruction
467 const Type *DstTy, ///< The target type for the second cast instruction
468 TargetData *TD ///< The target data for pointer size
471 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
472 const Type *MidTy = CI->getType(); // B from above
474 // Get the opcodes of the two Cast instructions
475 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
476 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
478 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
479 DstTy, TD->getIntPtrType());
481 // We don't want to form an inttoptr or ptrtoint that converts to an integer
482 // type that differs from the pointer size.
483 if ((Res == Instruction::IntToPtr && SrcTy != TD->getIntPtrType()) ||
484 (Res == Instruction::PtrToInt && DstTy != TD->getIntPtrType()))
487 return Instruction::CastOps(Res);
490 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
491 /// in any code being generated. It does not require codegen if V is simple
492 /// enough or if the cast can be folded into other casts.
493 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
494 const Type *Ty, TargetData *TD) {
495 if (V->getType() == Ty || isa<Constant>(V)) return false;
497 // If this is another cast that can be eliminated, it isn't codegen either.
498 if (const CastInst *CI = dyn_cast<CastInst>(V))
499 if (isEliminableCastPair(CI, opcode, Ty, TD))
504 // SimplifyCommutative - This performs a few simplifications for commutative
507 // 1. Order operands such that they are listed from right (least complex) to
508 // left (most complex). This puts constants before unary operators before
511 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
512 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
514 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
515 bool Changed = false;
516 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
517 Changed = !I.swapOperands();
519 if (!I.isAssociative()) return Changed;
520 Instruction::BinaryOps Opcode = I.getOpcode();
521 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
522 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
523 if (isa<Constant>(I.getOperand(1))) {
524 Constant *Folded = ConstantExpr::get(I.getOpcode(),
525 cast<Constant>(I.getOperand(1)),
526 cast<Constant>(Op->getOperand(1)));
527 I.setOperand(0, Op->getOperand(0));
528 I.setOperand(1, Folded);
530 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
531 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
532 isOnlyUse(Op) && isOnlyUse(Op1)) {
533 Constant *C1 = cast<Constant>(Op->getOperand(1));
534 Constant *C2 = cast<Constant>(Op1->getOperand(1));
536 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
537 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
538 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
542 I.setOperand(0, New);
543 I.setOperand(1, Folded);
550 /// SimplifyCompare - For a CmpInst this function just orders the operands
551 /// so that theyare listed from right (least complex) to left (most complex).
552 /// This puts constants before unary operators before binary operators.
553 bool InstCombiner::SimplifyCompare(CmpInst &I) {
554 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
557 // Compare instructions are not associative so there's nothing else we can do.
561 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
562 // if the LHS is a constant zero (which is the 'negate' form).
564 static inline Value *dyn_castNegVal(Value *V) {
565 if (BinaryOperator::isNeg(V))
566 return BinaryOperator::getNegArgument(V);
568 // Constants can be considered to be negated values if they can be folded.
569 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
570 return ConstantExpr::getNeg(C);
572 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
573 if (C->getType()->getElementType()->isInteger())
574 return ConstantExpr::getNeg(C);
579 static inline Value *dyn_castNotVal(Value *V) {
580 if (BinaryOperator::isNot(V))
581 return BinaryOperator::getNotArgument(V);
583 // Constants can be considered to be not'ed values...
584 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
585 return ConstantInt::get(~C->getValue());
589 // dyn_castFoldableMul - If this value is a multiply that can be folded into
590 // other computations (because it has a constant operand), return the
591 // non-constant operand of the multiply, and set CST to point to the multiplier.
592 // Otherwise, return null.
594 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
595 if (V->hasOneUse() && V->getType()->isInteger())
596 if (Instruction *I = dyn_cast<Instruction>(V)) {
597 if (I->getOpcode() == Instruction::Mul)
598 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
599 return I->getOperand(0);
600 if (I->getOpcode() == Instruction::Shl)
601 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
602 // The multiplier is really 1 << CST.
603 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
604 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
605 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
606 return I->getOperand(0);
612 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
613 /// expression, return it.
614 static User *dyn_castGetElementPtr(Value *V) {
615 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
616 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
617 if (CE->getOpcode() == Instruction::GetElementPtr)
618 return cast<User>(V);
622 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
623 /// opcode value. Otherwise return UserOp1.
624 static unsigned getOpcode(const Value *V) {
625 if (const Instruction *I = dyn_cast<Instruction>(V))
626 return I->getOpcode();
627 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
628 return CE->getOpcode();
629 // Use UserOp1 to mean there's no opcode.
630 return Instruction::UserOp1;
633 /// AddOne - Add one to a ConstantInt
634 static ConstantInt *AddOne(ConstantInt *C) {
635 APInt Val(C->getValue());
636 return ConstantInt::get(++Val);
638 /// SubOne - Subtract one from a ConstantInt
639 static ConstantInt *SubOne(ConstantInt *C) {
640 APInt Val(C->getValue());
641 return ConstantInt::get(--Val);
643 /// Add - Add two ConstantInts together
644 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
645 return ConstantInt::get(C1->getValue() + C2->getValue());
647 /// And - Bitwise AND two ConstantInts together
648 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
649 return ConstantInt::get(C1->getValue() & C2->getValue());
651 /// Subtract - Subtract one ConstantInt from another
652 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
653 return ConstantInt::get(C1->getValue() - C2->getValue());
655 /// Multiply - Multiply two ConstantInts together
656 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
657 return ConstantInt::get(C1->getValue() * C2->getValue());
659 /// MultiplyOverflows - True if the multiply can not be expressed in an int
661 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
662 uint32_t W = C1->getBitWidth();
663 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
672 APInt MulExt = LHSExt * RHSExt;
675 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
676 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
677 return MulExt.slt(Min) || MulExt.sgt(Max);
679 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
683 /// ShrinkDemandedConstant - Check to see if the specified operand of the
684 /// specified instruction is a constant integer. If so, check to see if there
685 /// are any bits set in the constant that are not demanded. If so, shrink the
686 /// constant and return true.
687 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
689 assert(I && "No instruction?");
690 assert(OpNo < I->getNumOperands() && "Operand index too large");
692 // If the operand is not a constant integer, nothing to do.
693 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
694 if (!OpC) return false;
696 // If there are no bits set that aren't demanded, nothing to do.
697 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
698 if ((~Demanded & OpC->getValue()) == 0)
701 // This instruction is producing bits that are not demanded. Shrink the RHS.
702 Demanded &= OpC->getValue();
703 I->setOperand(OpNo, ConstantInt::get(Demanded));
707 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
708 // set of known zero and one bits, compute the maximum and minimum values that
709 // could have the specified known zero and known one bits, returning them in
711 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
712 const APInt& KnownZero,
713 const APInt& KnownOne,
714 APInt& Min, APInt& Max) {
715 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
716 assert(KnownZero.getBitWidth() == BitWidth &&
717 KnownOne.getBitWidth() == BitWidth &&
718 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
719 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
720 APInt UnknownBits = ~(KnownZero|KnownOne);
722 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
723 // bit if it is unknown.
725 Max = KnownOne|UnknownBits;
727 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
729 Max.clear(BitWidth-1);
733 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
734 // a set of known zero and one bits, compute the maximum and minimum values that
735 // could have the specified known zero and known one bits, returning them in
737 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
738 const APInt &KnownZero,
739 const APInt &KnownOne,
740 APInt &Min, APInt &Max) {
741 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
742 assert(KnownZero.getBitWidth() == BitWidth &&
743 KnownOne.getBitWidth() == BitWidth &&
744 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
745 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
746 APInt UnknownBits = ~(KnownZero|KnownOne);
748 // The minimum value is when the unknown bits are all zeros.
750 // The maximum value is when the unknown bits are all ones.
751 Max = KnownOne|UnknownBits;
754 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
755 /// SimplifyDemandedBits knows about. See if the instruction has any
756 /// properties that allow us to simplify its operands.
757 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
758 unsigned BitWidth = cast<IntegerType>(Inst.getType())->getBitWidth();
759 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
760 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
762 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
763 KnownZero, KnownOne, 0);
764 if (V == 0) return false;
765 if (V == &Inst) return true;
766 ReplaceInstUsesWith(Inst, V);
770 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
771 /// specified instruction operand if possible, updating it in place. It returns
772 /// true if it made any change and false otherwise.
773 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
774 APInt &KnownZero, APInt &KnownOne,
776 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
777 KnownZero, KnownOne, Depth);
778 if (NewVal == 0) return false;
784 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
785 /// value based on the demanded bits. When this function is called, it is known
786 /// that only the bits set in DemandedMask of the result of V are ever used
787 /// downstream. Consequently, depending on the mask and V, it may be possible
788 /// to replace V with a constant or one of its operands. In such cases, this
789 /// function does the replacement and returns true. In all other cases, it
790 /// returns false after analyzing the expression and setting KnownOne and known
791 /// to be one in the expression. KnownZero contains all the bits that are known
792 /// to be zero in the expression. These are provided to potentially allow the
793 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
794 /// the expression. KnownOne and KnownZero always follow the invariant that
795 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
796 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
797 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
798 /// and KnownOne must all be the same.
800 /// This returns null if it did not change anything and it permits no
801 /// simplification. This returns V itself if it did some simplification of V's
802 /// operands based on the information about what bits are demanded. This returns
803 /// some other non-null value if it found out that V is equal to another value
804 /// in the context where the specified bits are demanded, but not for all users.
805 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
806 APInt &KnownZero, APInt &KnownOne,
808 assert(V != 0 && "Null pointer of Value???");
809 assert(Depth <= 6 && "Limit Search Depth");
810 uint32_t BitWidth = DemandedMask.getBitWidth();
811 const IntegerType *VTy = cast<IntegerType>(V->getType());
812 assert(VTy->getBitWidth() == BitWidth &&
813 KnownZero.getBitWidth() == BitWidth &&
814 KnownOne.getBitWidth() == BitWidth &&
815 "Value *V, DemandedMask, KnownZero and KnownOne \
816 must have same BitWidth");
817 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
818 // We know all of the bits for a constant!
819 KnownOne = CI->getValue() & DemandedMask;
820 KnownZero = ~KnownOne & DemandedMask;
826 if (DemandedMask == 0) { // Not demanding any bits from V.
827 if (isa<UndefValue>(V))
829 return UndefValue::get(VTy);
832 if (Depth == 6) // Limit search depth.
835 Instruction *I = dyn_cast<Instruction>(V);
836 if (!I) return 0; // Only analyze instructions.
838 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
839 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
841 // If there are multiple uses of this value and we aren't at the root, then
842 // we can't do any simplifications of the operands, because DemandedMask
843 // only reflects the bits demanded by *one* of the users.
844 if (Depth != 0 && !I->hasOneUse()) {
845 // Despite the fact that we can't simplify this instruction in all User's
846 // context, we can at least compute the knownzero/knownone bits, and we can
847 // do simplifications that apply to *just* the one user if we know that
848 // this instruction has a simpler value in that context.
849 if (I->getOpcode() == Instruction::And) {
850 // If either the LHS or the RHS are Zero, the result is zero.
851 ComputeMaskedBits(I->getOperand(1), DemandedMask,
852 RHSKnownZero, RHSKnownOne, Depth+1);
853 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
854 LHSKnownZero, LHSKnownOne, Depth+1);
856 // If all of the demanded bits are known 1 on one side, return the other.
857 // These bits cannot contribute to the result of the 'and' in this
859 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
860 (DemandedMask & ~LHSKnownZero))
861 return I->getOperand(0);
862 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
863 (DemandedMask & ~RHSKnownZero))
864 return I->getOperand(1);
866 // If all of the demanded bits in the inputs are known zeros, return zero.
867 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
868 return Constant::getNullValue(VTy);
870 } else if (I->getOpcode() == Instruction::Or) {
871 // We can simplify (X|Y) -> X or Y in the user's context if we know that
872 // only bits from X or Y are demanded.
874 // If either the LHS or the RHS are One, the result is One.
875 ComputeMaskedBits(I->getOperand(1), DemandedMask,
876 RHSKnownZero, RHSKnownOne, Depth+1);
877 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
878 LHSKnownZero, LHSKnownOne, Depth+1);
880 // If all of the demanded bits are known zero on one side, return the
881 // other. These bits cannot contribute to the result of the 'or' in this
883 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
884 (DemandedMask & ~LHSKnownOne))
885 return I->getOperand(0);
886 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
887 (DemandedMask & ~RHSKnownOne))
888 return I->getOperand(1);
890 // If all of the potentially set bits on one side are known to be set on
891 // the other side, just use the 'other' side.
892 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
893 (DemandedMask & (~RHSKnownZero)))
894 return I->getOperand(0);
895 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
896 (DemandedMask & (~LHSKnownZero)))
897 return I->getOperand(1);
900 // Compute the KnownZero/KnownOne bits to simplify things downstream.
901 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
905 // If this is the root being simplified, allow it to have multiple uses,
906 // just set the DemandedMask to all bits so that we can try to simplify the
907 // operands. This allows visitTruncInst (for example) to simplify the
908 // operand of a trunc without duplicating all the logic below.
909 if (Depth == 0 && !V->hasOneUse())
910 DemandedMask = APInt::getAllOnesValue(BitWidth);
912 switch (I->getOpcode()) {
914 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
916 case Instruction::And:
917 // If either the LHS or the RHS are Zero, the result is zero.
918 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
919 RHSKnownZero, RHSKnownOne, Depth+1) ||
920 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
921 LHSKnownZero, LHSKnownOne, Depth+1))
923 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
924 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
926 // If all of the demanded bits are known 1 on one side, return the other.
927 // These bits cannot contribute to the result of the 'and'.
928 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
929 (DemandedMask & ~LHSKnownZero))
930 return I->getOperand(0);
931 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
932 (DemandedMask & ~RHSKnownZero))
933 return I->getOperand(1);
935 // If all of the demanded bits in the inputs are known zeros, return zero.
936 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
937 return Constant::getNullValue(VTy);
939 // If the RHS is a constant, see if we can simplify it.
940 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
943 // Output known-1 bits are only known if set in both the LHS & RHS.
944 RHSKnownOne &= LHSKnownOne;
945 // Output known-0 are known to be clear if zero in either the LHS | RHS.
946 RHSKnownZero |= LHSKnownZero;
948 case Instruction::Or:
949 // If either the LHS or the RHS are One, the result is One.
950 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
951 RHSKnownZero, RHSKnownOne, Depth+1) ||
952 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
953 LHSKnownZero, LHSKnownOne, Depth+1))
955 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
956 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
958 // If all of the demanded bits are known zero on one side, return the other.
959 // These bits cannot contribute to the result of the 'or'.
960 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
961 (DemandedMask & ~LHSKnownOne))
962 return I->getOperand(0);
963 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
964 (DemandedMask & ~RHSKnownOne))
965 return I->getOperand(1);
967 // If all of the potentially set bits on one side are known to be set on
968 // the other side, just use the 'other' side.
969 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
970 (DemandedMask & (~RHSKnownZero)))
971 return I->getOperand(0);
972 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
973 (DemandedMask & (~LHSKnownZero)))
974 return I->getOperand(1);
976 // If the RHS is a constant, see if we can simplify it.
977 if (ShrinkDemandedConstant(I, 1, DemandedMask))
980 // Output known-0 bits are only known if clear in both the LHS & RHS.
981 RHSKnownZero &= LHSKnownZero;
982 // Output known-1 are known to be set if set in either the LHS | RHS.
983 RHSKnownOne |= LHSKnownOne;
985 case Instruction::Xor: {
986 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
987 RHSKnownZero, RHSKnownOne, Depth+1) ||
988 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
989 LHSKnownZero, LHSKnownOne, Depth+1))
991 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
992 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
994 // If all of the demanded bits are known zero on one side, return the other.
995 // These bits cannot contribute to the result of the 'xor'.
996 if ((DemandedMask & RHSKnownZero) == DemandedMask)
997 return I->getOperand(0);
998 if ((DemandedMask & LHSKnownZero) == DemandedMask)
999 return I->getOperand(1);
1001 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1002 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1003 (RHSKnownOne & LHSKnownOne);
1004 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1005 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1006 (RHSKnownOne & LHSKnownZero);
1008 // If all of the demanded bits are known to be zero on one side or the
1009 // other, turn this into an *inclusive* or.
1010 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1011 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1013 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1015 return InsertNewInstBefore(Or, *I);
1018 // If all of the demanded bits on one side are known, and all of the set
1019 // bits on that side are also known to be set on the other side, turn this
1020 // into an AND, as we know the bits will be cleared.
1021 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1022 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1024 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1025 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1027 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1028 return InsertNewInstBefore(And, *I);
1032 // If the RHS is a constant, see if we can simplify it.
1033 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1034 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1037 RHSKnownZero = KnownZeroOut;
1038 RHSKnownOne = KnownOneOut;
1041 case Instruction::Select:
1042 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1043 RHSKnownZero, RHSKnownOne, Depth+1) ||
1044 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1045 LHSKnownZero, LHSKnownOne, Depth+1))
1047 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1048 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1050 // If the operands are constants, see if we can simplify them.
1051 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1052 ShrinkDemandedConstant(I, 2, DemandedMask))
1055 // Only known if known in both the LHS and RHS.
1056 RHSKnownOne &= LHSKnownOne;
1057 RHSKnownZero &= LHSKnownZero;
1059 case Instruction::Trunc: {
1060 unsigned truncBf = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1061 DemandedMask.zext(truncBf);
1062 RHSKnownZero.zext(truncBf);
1063 RHSKnownOne.zext(truncBf);
1064 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1065 RHSKnownZero, RHSKnownOne, Depth+1))
1067 DemandedMask.trunc(BitWidth);
1068 RHSKnownZero.trunc(BitWidth);
1069 RHSKnownOne.trunc(BitWidth);
1070 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1073 case Instruction::BitCast:
1074 if (!I->getOperand(0)->getType()->isInteger())
1075 return false; // vector->int or fp->int?
1076 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1077 RHSKnownZero, RHSKnownOne, Depth+1))
1079 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1081 case Instruction::ZExt: {
1082 // Compute the bits in the result that are not present in the input.
1083 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1085 DemandedMask.trunc(SrcBitWidth);
1086 RHSKnownZero.trunc(SrcBitWidth);
1087 RHSKnownOne.trunc(SrcBitWidth);
1088 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1089 RHSKnownZero, RHSKnownOne, Depth+1))
1091 DemandedMask.zext(BitWidth);
1092 RHSKnownZero.zext(BitWidth);
1093 RHSKnownOne.zext(BitWidth);
1094 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1095 // The top bits are known to be zero.
1096 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1099 case Instruction::SExt: {
1100 // Compute the bits in the result that are not present in the input.
1101 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1103 APInt InputDemandedBits = DemandedMask &
1104 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1106 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1107 // If any of the sign extended bits are demanded, we know that the sign
1109 if ((NewBits & DemandedMask) != 0)
1110 InputDemandedBits.set(SrcBitWidth-1);
1112 InputDemandedBits.trunc(SrcBitWidth);
1113 RHSKnownZero.trunc(SrcBitWidth);
1114 RHSKnownOne.trunc(SrcBitWidth);
1115 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1116 RHSKnownZero, RHSKnownOne, Depth+1))
1118 InputDemandedBits.zext(BitWidth);
1119 RHSKnownZero.zext(BitWidth);
1120 RHSKnownOne.zext(BitWidth);
1121 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1123 // If the sign bit of the input is known set or clear, then we know the
1124 // top bits of the result.
1126 // If the input sign bit is known zero, or if the NewBits are not demanded
1127 // convert this into a zero extension.
1128 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1129 // Convert to ZExt cast
1130 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1131 return InsertNewInstBefore(NewCast, *I);
1132 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1133 RHSKnownOne |= NewBits;
1137 case Instruction::Add: {
1138 // Figure out what the input bits are. If the top bits of the and result
1139 // are not demanded, then the add doesn't demand them from its input
1141 unsigned NLZ = DemandedMask.countLeadingZeros();
1143 // If there is a constant on the RHS, there are a variety of xformations
1145 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1146 // If null, this should be simplified elsewhere. Some of the xforms here
1147 // won't work if the RHS is zero.
1151 // If the top bit of the output is demanded, demand everything from the
1152 // input. Otherwise, we demand all the input bits except NLZ top bits.
1153 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1155 // Find information about known zero/one bits in the input.
1156 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1157 LHSKnownZero, LHSKnownOne, Depth+1))
1160 // If the RHS of the add has bits set that can't affect the input, reduce
1162 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1165 // Avoid excess work.
1166 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1169 // Turn it into OR if input bits are zero.
1170 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1172 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1174 return InsertNewInstBefore(Or, *I);
1177 // We can say something about the output known-zero and known-one bits,
1178 // depending on potential carries from the input constant and the
1179 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1180 // bits set and the RHS constant is 0x01001, then we know we have a known
1181 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1183 // To compute this, we first compute the potential carry bits. These are
1184 // the bits which may be modified. I'm not aware of a better way to do
1186 const APInt &RHSVal = RHS->getValue();
1187 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1189 // Now that we know which bits have carries, compute the known-1/0 sets.
1191 // Bits are known one if they are known zero in one operand and one in the
1192 // other, and there is no input carry.
1193 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1194 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1196 // Bits are known zero if they are known zero in both operands and there
1197 // is no input carry.
1198 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1200 // If the high-bits of this ADD are not demanded, then it does not demand
1201 // the high bits of its LHS or RHS.
1202 if (DemandedMask[BitWidth-1] == 0) {
1203 // Right fill the mask of bits for this ADD to demand the most
1204 // significant bit and all those below it.
1205 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1206 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1207 LHSKnownZero, LHSKnownOne, Depth+1) ||
1208 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1209 LHSKnownZero, LHSKnownOne, Depth+1))
1215 case Instruction::Sub:
1216 // If the high-bits of this SUB are not demanded, then it does not demand
1217 // the high bits of its LHS or RHS.
1218 if (DemandedMask[BitWidth-1] == 0) {
1219 // Right fill the mask of bits for this SUB to demand the most
1220 // significant bit and all those below it.
1221 uint32_t NLZ = DemandedMask.countLeadingZeros();
1222 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1223 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1224 LHSKnownZero, LHSKnownOne, Depth+1) ||
1225 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1226 LHSKnownZero, LHSKnownOne, Depth+1))
1229 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1230 // the known zeros and ones.
1231 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1233 case Instruction::Shl:
1234 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1235 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1236 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1237 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1238 RHSKnownZero, RHSKnownOne, Depth+1))
1240 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1241 RHSKnownZero <<= ShiftAmt;
1242 RHSKnownOne <<= ShiftAmt;
1243 // low bits known zero.
1245 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1248 case Instruction::LShr:
1249 // For a logical shift right
1250 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1251 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1253 // Unsigned shift right.
1254 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1255 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1256 RHSKnownZero, RHSKnownOne, Depth+1))
1258 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1259 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1260 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1262 // Compute the new bits that are at the top now.
1263 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1264 RHSKnownZero |= HighBits; // high bits known zero.
1268 case Instruction::AShr:
1269 // If this is an arithmetic shift right and only the low-bit is set, we can
1270 // always convert this into a logical shr, even if the shift amount is
1271 // variable. The low bit of the shift cannot be an input sign bit unless
1272 // the shift amount is >= the size of the datatype, which is undefined.
1273 if (DemandedMask == 1) {
1274 // Perform the logical shift right.
1275 Instruction *NewVal = BinaryOperator::CreateLShr(
1276 I->getOperand(0), I->getOperand(1), I->getName());
1277 return InsertNewInstBefore(NewVal, *I);
1280 // If the sign bit is the only bit demanded by this ashr, then there is no
1281 // need to do it, the shift doesn't change the high bit.
1282 if (DemandedMask.isSignBit())
1283 return I->getOperand(0);
1285 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1286 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1288 // Signed shift right.
1289 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1290 // If any of the "high bits" are demanded, we should set the sign bit as
1292 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1293 DemandedMaskIn.set(BitWidth-1);
1294 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1295 RHSKnownZero, RHSKnownOne, Depth+1))
1297 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1298 // Compute the new bits that are at the top now.
1299 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1300 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1301 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1303 // Handle the sign bits.
1304 APInt SignBit(APInt::getSignBit(BitWidth));
1305 // Adjust to where it is now in the mask.
1306 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1308 // If the input sign bit is known to be zero, or if none of the top bits
1309 // are demanded, turn this into an unsigned shift right.
1310 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1311 (HighBits & ~DemandedMask) == HighBits) {
1312 // Perform the logical shift right.
1313 Instruction *NewVal = BinaryOperator::CreateLShr(
1314 I->getOperand(0), SA, I->getName());
1315 return InsertNewInstBefore(NewVal, *I);
1316 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1317 RHSKnownOne |= HighBits;
1321 case Instruction::SRem:
1322 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1323 APInt RA = Rem->getValue().abs();
1324 if (RA.isPowerOf2()) {
1325 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1326 return I->getOperand(0);
1328 APInt LowBits = RA - 1;
1329 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1330 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1331 LHSKnownZero, LHSKnownOne, Depth+1))
1334 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1335 LHSKnownZero |= ~LowBits;
1337 KnownZero |= LHSKnownZero & DemandedMask;
1339 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1343 case Instruction::URem: {
1344 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1345 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1346 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1347 KnownZero2, KnownOne2, Depth+1) ||
1348 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1349 KnownZero2, KnownOne2, Depth+1))
1352 unsigned Leaders = KnownZero2.countLeadingOnes();
1353 Leaders = std::max(Leaders,
1354 KnownZero2.countLeadingOnes());
1355 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1358 case Instruction::Call:
1359 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1360 switch (II->getIntrinsicID()) {
1362 case Intrinsic::bswap: {
1363 // If the only bits demanded come from one byte of the bswap result,
1364 // just shift the input byte into position to eliminate the bswap.
1365 unsigned NLZ = DemandedMask.countLeadingZeros();
1366 unsigned NTZ = DemandedMask.countTrailingZeros();
1368 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1369 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1370 // have 14 leading zeros, round to 8.
1373 // If we need exactly one byte, we can do this transformation.
1374 if (BitWidth-NLZ-NTZ == 8) {
1375 unsigned ResultBit = NTZ;
1376 unsigned InputBit = BitWidth-NTZ-8;
1378 // Replace this with either a left or right shift to get the byte into
1380 Instruction *NewVal;
1381 if (InputBit > ResultBit)
1382 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1383 ConstantInt::get(I->getType(), InputBit-ResultBit));
1385 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1386 ConstantInt::get(I->getType(), ResultBit-InputBit));
1387 NewVal->takeName(I);
1388 return InsertNewInstBefore(NewVal, *I);
1391 // TODO: Could compute known zero/one bits based on the input.
1396 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1400 // If the client is only demanding bits that we know, return the known
1402 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1403 return ConstantInt::get(RHSKnownOne);
1408 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1409 /// any number of elements. DemandedElts contains the set of elements that are
1410 /// actually used by the caller. This method analyzes which elements of the
1411 /// operand are undef and returns that information in UndefElts.
1413 /// If the information about demanded elements can be used to simplify the
1414 /// operation, the operation is simplified, then the resultant value is
1415 /// returned. This returns null if no change was made.
1416 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1419 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1420 APInt EltMask(APInt::getAllOnesValue(VWidth));
1421 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1423 if (isa<UndefValue>(V)) {
1424 // If the entire vector is undefined, just return this info.
1425 UndefElts = EltMask;
1427 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1428 UndefElts = EltMask;
1429 return UndefValue::get(V->getType());
1433 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1434 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1435 Constant *Undef = UndefValue::get(EltTy);
1437 std::vector<Constant*> Elts;
1438 for (unsigned i = 0; i != VWidth; ++i)
1439 if (!DemandedElts[i]) { // If not demanded, set to undef.
1440 Elts.push_back(Undef);
1442 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1443 Elts.push_back(Undef);
1445 } else { // Otherwise, defined.
1446 Elts.push_back(CP->getOperand(i));
1449 // If we changed the constant, return it.
1450 Constant *NewCP = ConstantVector::get(Elts);
1451 return NewCP != CP ? NewCP : 0;
1452 } else if (isa<ConstantAggregateZero>(V)) {
1453 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1456 // Check if this is identity. If so, return 0 since we are not simplifying
1458 if (DemandedElts == ((1ULL << VWidth) -1))
1461 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1462 Constant *Zero = Constant::getNullValue(EltTy);
1463 Constant *Undef = UndefValue::get(EltTy);
1464 std::vector<Constant*> Elts;
1465 for (unsigned i = 0; i != VWidth; ++i) {
1466 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1467 Elts.push_back(Elt);
1469 UndefElts = DemandedElts ^ EltMask;
1470 return ConstantVector::get(Elts);
1473 // Limit search depth.
1477 // If multiple users are using the root value, procede with
1478 // simplification conservatively assuming that all elements
1480 if (!V->hasOneUse()) {
1481 // Quit if we find multiple users of a non-root value though.
1482 // They'll be handled when it's their turn to be visited by
1483 // the main instcombine process.
1485 // TODO: Just compute the UndefElts information recursively.
1488 // Conservatively assume that all elements are needed.
1489 DemandedElts = EltMask;
1492 Instruction *I = dyn_cast<Instruction>(V);
1493 if (!I) return false; // Only analyze instructions.
1495 bool MadeChange = false;
1496 APInt UndefElts2(VWidth, 0);
1498 switch (I->getOpcode()) {
1501 case Instruction::InsertElement: {
1502 // If this is a variable index, we don't know which element it overwrites.
1503 // demand exactly the same input as we produce.
1504 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1506 // Note that we can't propagate undef elt info, because we don't know
1507 // which elt is getting updated.
1508 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1509 UndefElts2, Depth+1);
1510 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1514 // If this is inserting an element that isn't demanded, remove this
1516 unsigned IdxNo = Idx->getZExtValue();
1517 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1518 return AddSoonDeadInstToWorklist(*I, 0);
1520 // Otherwise, the element inserted overwrites whatever was there, so the
1521 // input demanded set is simpler than the output set.
1522 APInt DemandedElts2 = DemandedElts;
1523 DemandedElts2.clear(IdxNo);
1524 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1525 UndefElts, Depth+1);
1526 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1528 // The inserted element is defined.
1529 UndefElts.clear(IdxNo);
1532 case Instruction::ShuffleVector: {
1533 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1534 uint64_t LHSVWidth =
1535 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1536 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1537 for (unsigned i = 0; i < VWidth; i++) {
1538 if (DemandedElts[i]) {
1539 unsigned MaskVal = Shuffle->getMaskValue(i);
1540 if (MaskVal != -1u) {
1541 assert(MaskVal < LHSVWidth * 2 &&
1542 "shufflevector mask index out of range!");
1543 if (MaskVal < LHSVWidth)
1544 LeftDemanded.set(MaskVal);
1546 RightDemanded.set(MaskVal - LHSVWidth);
1551 APInt UndefElts4(LHSVWidth, 0);
1552 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1553 UndefElts4, Depth+1);
1554 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1556 APInt UndefElts3(LHSVWidth, 0);
1557 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1558 UndefElts3, Depth+1);
1559 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1561 bool NewUndefElts = false;
1562 for (unsigned i = 0; i < VWidth; i++) {
1563 unsigned MaskVal = Shuffle->getMaskValue(i);
1564 if (MaskVal == -1u) {
1566 } else if (MaskVal < LHSVWidth) {
1567 if (UndefElts4[MaskVal]) {
1568 NewUndefElts = true;
1572 if (UndefElts3[MaskVal - LHSVWidth]) {
1573 NewUndefElts = true;
1580 // Add additional discovered undefs.
1581 std::vector<Constant*> Elts;
1582 for (unsigned i = 0; i < VWidth; ++i) {
1584 Elts.push_back(UndefValue::get(Type::Int32Ty));
1586 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1587 Shuffle->getMaskValue(i)));
1589 I->setOperand(2, ConstantVector::get(Elts));
1594 case Instruction::BitCast: {
1595 // Vector->vector casts only.
1596 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1598 unsigned InVWidth = VTy->getNumElements();
1599 APInt InputDemandedElts(InVWidth, 0);
1602 if (VWidth == InVWidth) {
1603 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1604 // elements as are demanded of us.
1606 InputDemandedElts = DemandedElts;
1607 } else if (VWidth > InVWidth) {
1611 // If there are more elements in the result than there are in the source,
1612 // then an input element is live if any of the corresponding output
1613 // elements are live.
1614 Ratio = VWidth/InVWidth;
1615 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1616 if (DemandedElts[OutIdx])
1617 InputDemandedElts.set(OutIdx/Ratio);
1623 // If there are more elements in the source than there are in the result,
1624 // then an input element is live if the corresponding output element is
1626 Ratio = InVWidth/VWidth;
1627 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1628 if (DemandedElts[InIdx/Ratio])
1629 InputDemandedElts.set(InIdx);
1632 // div/rem demand all inputs, because they don't want divide by zero.
1633 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1634 UndefElts2, Depth+1);
1636 I->setOperand(0, TmpV);
1640 UndefElts = UndefElts2;
1641 if (VWidth > InVWidth) {
1642 assert(0 && "Unimp");
1643 // If there are more elements in the result than there are in the source,
1644 // then an output element is undef if the corresponding input element is
1646 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1647 if (UndefElts2[OutIdx/Ratio])
1648 UndefElts.set(OutIdx);
1649 } else if (VWidth < InVWidth) {
1650 assert(0 && "Unimp");
1651 // If there are more elements in the source than there are in the result,
1652 // then a result element is undef if all of the corresponding input
1653 // elements are undef.
1654 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1655 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1656 if (!UndefElts2[InIdx]) // Not undef?
1657 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1661 case Instruction::And:
1662 case Instruction::Or:
1663 case Instruction::Xor:
1664 case Instruction::Add:
1665 case Instruction::Sub:
1666 case Instruction::Mul:
1667 // div/rem demand all inputs, because they don't want divide by zero.
1668 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1669 UndefElts, Depth+1);
1670 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1671 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1672 UndefElts2, Depth+1);
1673 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1675 // Output elements are undefined if both are undefined. Consider things
1676 // like undef&0. The result is known zero, not undef.
1677 UndefElts &= UndefElts2;
1680 case Instruction::Call: {
1681 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1683 switch (II->getIntrinsicID()) {
1686 // Binary vector operations that work column-wise. A dest element is a
1687 // function of the corresponding input elements from the two inputs.
1688 case Intrinsic::x86_sse_sub_ss:
1689 case Intrinsic::x86_sse_mul_ss:
1690 case Intrinsic::x86_sse_min_ss:
1691 case Intrinsic::x86_sse_max_ss:
1692 case Intrinsic::x86_sse2_sub_sd:
1693 case Intrinsic::x86_sse2_mul_sd:
1694 case Intrinsic::x86_sse2_min_sd:
1695 case Intrinsic::x86_sse2_max_sd:
1696 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1697 UndefElts, Depth+1);
1698 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1699 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1700 UndefElts2, Depth+1);
1701 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1703 // If only the low elt is demanded and this is a scalarizable intrinsic,
1704 // scalarize it now.
1705 if (DemandedElts == 1) {
1706 switch (II->getIntrinsicID()) {
1708 case Intrinsic::x86_sse_sub_ss:
1709 case Intrinsic::x86_sse_mul_ss:
1710 case Intrinsic::x86_sse2_sub_sd:
1711 case Intrinsic::x86_sse2_mul_sd:
1712 // TODO: Lower MIN/MAX/ABS/etc
1713 Value *LHS = II->getOperand(1);
1714 Value *RHS = II->getOperand(2);
1715 // Extract the element as scalars.
1716 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1717 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1719 switch (II->getIntrinsicID()) {
1720 default: assert(0 && "Case stmts out of sync!");
1721 case Intrinsic::x86_sse_sub_ss:
1722 case Intrinsic::x86_sse2_sub_sd:
1723 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1724 II->getName()), *II);
1726 case Intrinsic::x86_sse_mul_ss:
1727 case Intrinsic::x86_sse2_mul_sd:
1728 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1729 II->getName()), *II);
1734 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1736 InsertNewInstBefore(New, *II);
1737 AddSoonDeadInstToWorklist(*II, 0);
1742 // Output elements are undefined if both are undefined. Consider things
1743 // like undef&0. The result is known zero, not undef.
1744 UndefElts &= UndefElts2;
1750 return MadeChange ? I : 0;
1754 /// AssociativeOpt - Perform an optimization on an associative operator. This
1755 /// function is designed to check a chain of associative operators for a
1756 /// potential to apply a certain optimization. Since the optimization may be
1757 /// applicable if the expression was reassociated, this checks the chain, then
1758 /// reassociates the expression as necessary to expose the optimization
1759 /// opportunity. This makes use of a special Functor, which must define
1760 /// 'shouldApply' and 'apply' methods.
1762 template<typename Functor>
1763 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1764 unsigned Opcode = Root.getOpcode();
1765 Value *LHS = Root.getOperand(0);
1767 // Quick check, see if the immediate LHS matches...
1768 if (F.shouldApply(LHS))
1769 return F.apply(Root);
1771 // Otherwise, if the LHS is not of the same opcode as the root, return.
1772 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1773 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1774 // Should we apply this transform to the RHS?
1775 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1777 // If not to the RHS, check to see if we should apply to the LHS...
1778 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1779 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1783 // If the functor wants to apply the optimization to the RHS of LHSI,
1784 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1786 // Now all of the instructions are in the current basic block, go ahead
1787 // and perform the reassociation.
1788 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1790 // First move the selected RHS to the LHS of the root...
1791 Root.setOperand(0, LHSI->getOperand(1));
1793 // Make what used to be the LHS of the root be the user of the root...
1794 Value *ExtraOperand = TmpLHSI->getOperand(1);
1795 if (&Root == TmpLHSI) {
1796 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1799 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1800 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1801 BasicBlock::iterator ARI = &Root; ++ARI;
1802 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1805 // Now propagate the ExtraOperand down the chain of instructions until we
1807 while (TmpLHSI != LHSI) {
1808 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1809 // Move the instruction to immediately before the chain we are
1810 // constructing to avoid breaking dominance properties.
1811 NextLHSI->moveBefore(ARI);
1814 Value *NextOp = NextLHSI->getOperand(1);
1815 NextLHSI->setOperand(1, ExtraOperand);
1817 ExtraOperand = NextOp;
1820 // Now that the instructions are reassociated, have the functor perform
1821 // the transformation...
1822 return F.apply(Root);
1825 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1832 // AddRHS - Implements: X + X --> X << 1
1835 AddRHS(Value *rhs) : RHS(rhs) {}
1836 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1837 Instruction *apply(BinaryOperator &Add) const {
1838 return BinaryOperator::CreateShl(Add.getOperand(0),
1839 ConstantInt::get(Add.getType(), 1));
1843 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1845 struct AddMaskingAnd {
1847 AddMaskingAnd(Constant *c) : C2(c) {}
1848 bool shouldApply(Value *LHS) const {
1850 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1851 ConstantExpr::getAnd(C1, C2)->isNullValue();
1853 Instruction *apply(BinaryOperator &Add) const {
1854 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1860 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1862 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1863 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1866 // Figure out if the constant is the left or the right argument.
1867 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1868 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1870 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1872 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1873 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1876 Value *Op0 = SO, *Op1 = ConstOperand;
1878 std::swap(Op0, Op1);
1880 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1881 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1882 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1883 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1884 SO->getName()+".cmp");
1886 assert(0 && "Unknown binary instruction type!");
1889 return IC->InsertNewInstBefore(New, I);
1892 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1893 // constant as the other operand, try to fold the binary operator into the
1894 // select arguments. This also works for Cast instructions, which obviously do
1895 // not have a second operand.
1896 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1898 // Don't modify shared select instructions
1899 if (!SI->hasOneUse()) return 0;
1900 Value *TV = SI->getOperand(1);
1901 Value *FV = SI->getOperand(2);
1903 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1904 // Bool selects with constant operands can be folded to logical ops.
1905 if (SI->getType() == Type::Int1Ty) return 0;
1907 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1908 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1910 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1917 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1918 /// node as operand #0, see if we can fold the instruction into the PHI (which
1919 /// is only possible if all operands to the PHI are constants).
1920 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1921 PHINode *PN = cast<PHINode>(I.getOperand(0));
1922 unsigned NumPHIValues = PN->getNumIncomingValues();
1923 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1925 // Check to see if all of the operands of the PHI are constants. If there is
1926 // one non-constant value, remember the BB it is. If there is more than one
1927 // or if *it* is a PHI, bail out.
1928 BasicBlock *NonConstBB = 0;
1929 for (unsigned i = 0; i != NumPHIValues; ++i)
1930 if (!isa<Constant>(PN->getIncomingValue(i))) {
1931 if (NonConstBB) return 0; // More than one non-const value.
1932 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1933 NonConstBB = PN->getIncomingBlock(i);
1935 // If the incoming non-constant value is in I's block, we have an infinite
1937 if (NonConstBB == I.getParent())
1941 // If there is exactly one non-constant value, we can insert a copy of the
1942 // operation in that block. However, if this is a critical edge, we would be
1943 // inserting the computation one some other paths (e.g. inside a loop). Only
1944 // do this if the pred block is unconditionally branching into the phi block.
1946 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1947 if (!BI || !BI->isUnconditional()) return 0;
1950 // Okay, we can do the transformation: create the new PHI node.
1951 PHINode *NewPN = PHINode::Create(I.getType(), "");
1952 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1953 InsertNewInstBefore(NewPN, *PN);
1954 NewPN->takeName(PN);
1956 // Next, add all of the operands to the PHI.
1957 if (I.getNumOperands() == 2) {
1958 Constant *C = cast<Constant>(I.getOperand(1));
1959 for (unsigned i = 0; i != NumPHIValues; ++i) {
1961 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1962 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1963 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1965 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1967 assert(PN->getIncomingBlock(i) == NonConstBB);
1968 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1969 InV = BinaryOperator::Create(BO->getOpcode(),
1970 PN->getIncomingValue(i), C, "phitmp",
1971 NonConstBB->getTerminator());
1972 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1973 InV = CmpInst::Create(CI->getOpcode(),
1975 PN->getIncomingValue(i), C, "phitmp",
1976 NonConstBB->getTerminator());
1978 assert(0 && "Unknown binop!");
1980 AddToWorkList(cast<Instruction>(InV));
1982 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1985 CastInst *CI = cast<CastInst>(&I);
1986 const Type *RetTy = CI->getType();
1987 for (unsigned i = 0; i != NumPHIValues; ++i) {
1989 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1990 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1992 assert(PN->getIncomingBlock(i) == NonConstBB);
1993 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1994 I.getType(), "phitmp",
1995 NonConstBB->getTerminator());
1996 AddToWorkList(cast<Instruction>(InV));
1998 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2001 return ReplaceInstUsesWith(I, NewPN);
2005 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2006 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2007 /// This basically requires proving that the add in the original type would not
2008 /// overflow to change the sign bit or have a carry out.
2009 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2010 // There are different heuristics we can use for this. Here are some simple
2013 // Add has the property that adding any two 2's complement numbers can only
2014 // have one carry bit which can change a sign. As such, if LHS and RHS each
2015 // have at least two sign bits, we know that the addition of the two values will
2016 // sign extend fine.
2017 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2021 // If one of the operands only has one non-zero bit, and if the other operand
2022 // has a known-zero bit in a more significant place than it (not including the
2023 // sign bit) the ripple may go up to and fill the zero, but won't change the
2024 // sign. For example, (X & ~4) + 1.
2032 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2033 bool Changed = SimplifyCommutative(I);
2034 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2036 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2037 // X + undef -> undef
2038 if (isa<UndefValue>(RHS))
2039 return ReplaceInstUsesWith(I, RHS);
2042 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2043 if (RHSC->isNullValue())
2044 return ReplaceInstUsesWith(I, LHS);
2045 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2046 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2047 (I.getType())->getValueAPF()))
2048 return ReplaceInstUsesWith(I, LHS);
2051 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2052 // X + (signbit) --> X ^ signbit
2053 const APInt& Val = CI->getValue();
2054 uint32_t BitWidth = Val.getBitWidth();
2055 if (Val == APInt::getSignBit(BitWidth))
2056 return BinaryOperator::CreateXor(LHS, RHS);
2058 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2059 // (X & 254)+1 -> (X&254)|1
2060 if (!isa<VectorType>(I.getType()) && SimplifyDemandedInstructionBits(I))
2063 // zext(i1) - 1 -> select i1, 0, -1
2064 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2065 if (CI->isAllOnesValue() &&
2066 ZI->getOperand(0)->getType() == Type::Int1Ty)
2067 return SelectInst::Create(ZI->getOperand(0),
2068 Constant::getNullValue(I.getType()),
2069 ConstantInt::getAllOnesValue(I.getType()));
2072 if (isa<PHINode>(LHS))
2073 if (Instruction *NV = FoldOpIntoPhi(I))
2076 ConstantInt *XorRHS = 0;
2078 if (isa<ConstantInt>(RHSC) &&
2079 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2080 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2081 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2083 uint32_t Size = TySizeBits / 2;
2084 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2085 APInt CFF80Val(-C0080Val);
2087 if (TySizeBits > Size) {
2088 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2089 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2090 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2091 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2092 // This is a sign extend if the top bits are known zero.
2093 if (!MaskedValueIsZero(XorLHS,
2094 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2095 Size = 0; // Not a sign ext, but can't be any others either.
2100 C0080Val = APIntOps::lshr(C0080Val, Size);
2101 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2102 } while (Size >= 1);
2104 // FIXME: This shouldn't be necessary. When the backends can handle types
2105 // with funny bit widths then this switch statement should be removed. It
2106 // is just here to get the size of the "middle" type back up to something
2107 // that the back ends can handle.
2108 const Type *MiddleType = 0;
2111 case 32: MiddleType = Type::Int32Ty; break;
2112 case 16: MiddleType = Type::Int16Ty; break;
2113 case 8: MiddleType = Type::Int8Ty; break;
2116 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2117 InsertNewInstBefore(NewTrunc, I);
2118 return new SExtInst(NewTrunc, I.getType(), I.getName());
2123 if (I.getType() == Type::Int1Ty)
2124 return BinaryOperator::CreateXor(LHS, RHS);
2127 if (I.getType()->isInteger()) {
2128 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2130 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2131 if (RHSI->getOpcode() == Instruction::Sub)
2132 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2133 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2135 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2136 if (LHSI->getOpcode() == Instruction::Sub)
2137 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2138 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2143 // -A + -B --> -(A + B)
2144 if (Value *LHSV = dyn_castNegVal(LHS)) {
2145 if (LHS->getType()->isIntOrIntVector()) {
2146 if (Value *RHSV = dyn_castNegVal(RHS)) {
2147 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2148 InsertNewInstBefore(NewAdd, I);
2149 return BinaryOperator::CreateNeg(NewAdd);
2153 return BinaryOperator::CreateSub(RHS, LHSV);
2157 if (!isa<Constant>(RHS))
2158 if (Value *V = dyn_castNegVal(RHS))
2159 return BinaryOperator::CreateSub(LHS, V);
2163 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2164 if (X == RHS) // X*C + X --> X * (C+1)
2165 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2167 // X*C1 + X*C2 --> X * (C1+C2)
2169 if (X == dyn_castFoldableMul(RHS, C1))
2170 return BinaryOperator::CreateMul(X, Add(C1, C2));
2173 // X + X*C --> X * (C+1)
2174 if (dyn_castFoldableMul(RHS, C2) == LHS)
2175 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2177 // X + ~X --> -1 since ~X = -X-1
2178 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2179 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2182 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2183 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2184 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2187 // A+B --> A|B iff A and B have no bits set in common.
2188 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2189 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2190 APInt LHSKnownOne(IT->getBitWidth(), 0);
2191 APInt LHSKnownZero(IT->getBitWidth(), 0);
2192 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2193 if (LHSKnownZero != 0) {
2194 APInt RHSKnownOne(IT->getBitWidth(), 0);
2195 APInt RHSKnownZero(IT->getBitWidth(), 0);
2196 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2198 // No bits in common -> bitwise or.
2199 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2200 return BinaryOperator::CreateOr(LHS, RHS);
2204 // W*X + Y*Z --> W * (X+Z) iff W == Y
2205 if (I.getType()->isIntOrIntVector()) {
2206 Value *W, *X, *Y, *Z;
2207 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2208 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2212 } else if (Y == X) {
2214 } else if (X == Z) {
2221 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2222 LHS->getName()), I);
2223 return BinaryOperator::CreateMul(W, NewAdd);
2228 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2230 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2231 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2233 // (X & FF00) + xx00 -> (X+xx00) & FF00
2234 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2235 Constant *Anded = And(CRHS, C2);
2236 if (Anded == CRHS) {
2237 // See if all bits from the first bit set in the Add RHS up are included
2238 // in the mask. First, get the rightmost bit.
2239 const APInt& AddRHSV = CRHS->getValue();
2241 // Form a mask of all bits from the lowest bit added through the top.
2242 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2244 // See if the and mask includes all of these bits.
2245 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2247 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2248 // Okay, the xform is safe. Insert the new add pronto.
2249 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2250 LHS->getName()), I);
2251 return BinaryOperator::CreateAnd(NewAdd, C2);
2256 // Try to fold constant add into select arguments.
2257 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2258 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2262 // add (cast *A to intptrtype) B ->
2263 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2265 CastInst *CI = dyn_cast<CastInst>(LHS);
2268 CI = dyn_cast<CastInst>(RHS);
2271 if (CI && CI->getType()->isSized() &&
2272 (CI->getType()->getPrimitiveSizeInBits() ==
2273 TD->getIntPtrType()->getPrimitiveSizeInBits())
2274 && isa<PointerType>(CI->getOperand(0)->getType())) {
2276 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2277 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2278 PointerType::get(Type::Int8Ty, AS), I);
2279 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2280 return new PtrToIntInst(I2, CI->getType());
2284 // add (select X 0 (sub n A)) A --> select X A n
2286 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2289 SI = dyn_cast<SelectInst>(RHS);
2292 if (SI && SI->hasOneUse()) {
2293 Value *TV = SI->getTrueValue();
2294 Value *FV = SI->getFalseValue();
2297 // Can we fold the add into the argument of the select?
2298 // We check both true and false select arguments for a matching subtract.
2299 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2300 // Fold the add into the true select value.
2301 return SelectInst::Create(SI->getCondition(), N, A);
2302 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2303 // Fold the add into the false select value.
2304 return SelectInst::Create(SI->getCondition(), A, N);
2308 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2309 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2310 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2311 return ReplaceInstUsesWith(I, LHS);
2313 // Check for (add (sext x), y), see if we can merge this into an
2314 // integer add followed by a sext.
2315 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2316 // (add (sext x), cst) --> (sext (add x, cst'))
2317 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2319 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2320 if (LHSConv->hasOneUse() &&
2321 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2322 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2323 // Insert the new, smaller add.
2324 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2326 InsertNewInstBefore(NewAdd, I);
2327 return new SExtInst(NewAdd, I.getType());
2331 // (add (sext x), (sext y)) --> (sext (add int x, y))
2332 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2333 // Only do this if x/y have the same type, if at last one of them has a
2334 // single use (so we don't increase the number of sexts), and if the
2335 // integer add will not overflow.
2336 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2337 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2338 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2339 RHSConv->getOperand(0))) {
2340 // Insert the new integer add.
2341 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2342 RHSConv->getOperand(0),
2344 InsertNewInstBefore(NewAdd, I);
2345 return new SExtInst(NewAdd, I.getType());
2350 // Check for (add double (sitofp x), y), see if we can merge this into an
2351 // integer add followed by a promotion.
2352 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2353 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2354 // ... if the constant fits in the integer value. This is useful for things
2355 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2356 // requires a constant pool load, and generally allows the add to be better
2358 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2360 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2361 if (LHSConv->hasOneUse() &&
2362 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2363 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2364 // Insert the new integer add.
2365 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2367 InsertNewInstBefore(NewAdd, I);
2368 return new SIToFPInst(NewAdd, I.getType());
2372 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2373 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2374 // Only do this if x/y have the same type, if at last one of them has a
2375 // single use (so we don't increase the number of int->fp conversions),
2376 // and if the integer add will not overflow.
2377 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2378 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2379 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2380 RHSConv->getOperand(0))) {
2381 // Insert the new integer add.
2382 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2383 RHSConv->getOperand(0),
2385 InsertNewInstBefore(NewAdd, I);
2386 return new SIToFPInst(NewAdd, I.getType());
2391 return Changed ? &I : 0;
2394 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2395 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2397 if (Op0 == Op1 && // sub X, X -> 0
2398 !I.getType()->isFPOrFPVector())
2399 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2401 // If this is a 'B = x-(-A)', change to B = x+A...
2402 if (Value *V = dyn_castNegVal(Op1))
2403 return BinaryOperator::CreateAdd(Op0, V);
2405 if (isa<UndefValue>(Op0))
2406 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2407 if (isa<UndefValue>(Op1))
2408 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2410 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2411 // Replace (-1 - A) with (~A)...
2412 if (C->isAllOnesValue())
2413 return BinaryOperator::CreateNot(Op1);
2415 // C - ~X == X + (1+C)
2417 if (match(Op1, m_Not(m_Value(X))))
2418 return BinaryOperator::CreateAdd(X, AddOne(C));
2420 // -(X >>u 31) -> (X >>s 31)
2421 // -(X >>s 31) -> (X >>u 31)
2423 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2424 if (SI->getOpcode() == Instruction::LShr) {
2425 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2426 // Check to see if we are shifting out everything but the sign bit.
2427 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2428 SI->getType()->getPrimitiveSizeInBits()-1) {
2429 // Ok, the transformation is safe. Insert AShr.
2430 return BinaryOperator::Create(Instruction::AShr,
2431 SI->getOperand(0), CU, SI->getName());
2435 else if (SI->getOpcode() == Instruction::AShr) {
2436 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2437 // Check to see if we are shifting out everything but the sign bit.
2438 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2439 SI->getType()->getPrimitiveSizeInBits()-1) {
2440 // Ok, the transformation is safe. Insert LShr.
2441 return BinaryOperator::CreateLShr(
2442 SI->getOperand(0), CU, SI->getName());
2449 // Try to fold constant sub into select arguments.
2450 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2451 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2455 if (I.getType() == Type::Int1Ty)
2456 return BinaryOperator::CreateXor(Op0, Op1);
2458 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2459 if (Op1I->getOpcode() == Instruction::Add &&
2460 !Op0->getType()->isFPOrFPVector()) {
2461 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2462 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2463 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2464 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2465 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2466 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2467 // C1-(X+C2) --> (C1-C2)-X
2468 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2469 Op1I->getOperand(0));
2473 if (Op1I->hasOneUse()) {
2474 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2475 // is not used by anyone else...
2477 if (Op1I->getOpcode() == Instruction::Sub &&
2478 !Op1I->getType()->isFPOrFPVector()) {
2479 // Swap the two operands of the subexpr...
2480 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2481 Op1I->setOperand(0, IIOp1);
2482 Op1I->setOperand(1, IIOp0);
2484 // Create the new top level add instruction...
2485 return BinaryOperator::CreateAdd(Op0, Op1);
2488 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2490 if (Op1I->getOpcode() == Instruction::And &&
2491 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2492 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2495 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2496 return BinaryOperator::CreateAnd(Op0, NewNot);
2499 // 0 - (X sdiv C) -> (X sdiv -C)
2500 if (Op1I->getOpcode() == Instruction::SDiv)
2501 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2503 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2504 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2505 ConstantExpr::getNeg(DivRHS));
2507 // X - X*C --> X * (1-C)
2508 ConstantInt *C2 = 0;
2509 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2510 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2511 return BinaryOperator::CreateMul(Op0, CP1);
2516 if (!Op0->getType()->isFPOrFPVector())
2517 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2518 if (Op0I->getOpcode() == Instruction::Add) {
2519 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2520 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2521 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2522 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2523 } else if (Op0I->getOpcode() == Instruction::Sub) {
2524 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2525 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2530 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2531 if (X == Op1) // X*C - X --> X * (C-1)
2532 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2534 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2535 if (X == dyn_castFoldableMul(Op1, C2))
2536 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2541 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2542 /// comparison only checks the sign bit. If it only checks the sign bit, set
2543 /// TrueIfSigned if the result of the comparison is true when the input value is
2545 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2546 bool &TrueIfSigned) {
2548 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2549 TrueIfSigned = true;
2550 return RHS->isZero();
2551 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2552 TrueIfSigned = true;
2553 return RHS->isAllOnesValue();
2554 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2555 TrueIfSigned = false;
2556 return RHS->isAllOnesValue();
2557 case ICmpInst::ICMP_UGT:
2558 // True if LHS u> RHS and RHS == high-bit-mask - 1
2559 TrueIfSigned = true;
2560 return RHS->getValue() ==
2561 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2562 case ICmpInst::ICMP_UGE:
2563 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2564 TrueIfSigned = true;
2565 return RHS->getValue().isSignBit();
2571 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2572 bool Changed = SimplifyCommutative(I);
2573 Value *Op0 = I.getOperand(0);
2575 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2576 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2578 // Simplify mul instructions with a constant RHS...
2579 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2580 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2582 // ((X << C1)*C2) == (X * (C2 << C1))
2583 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2584 if (SI->getOpcode() == Instruction::Shl)
2585 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2586 return BinaryOperator::CreateMul(SI->getOperand(0),
2587 ConstantExpr::getShl(CI, ShOp));
2590 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2591 if (CI->equalsInt(1)) // X * 1 == X
2592 return ReplaceInstUsesWith(I, Op0);
2593 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2594 return BinaryOperator::CreateNeg(Op0, I.getName());
2596 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2597 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2598 return BinaryOperator::CreateShl(Op0,
2599 ConstantInt::get(Op0->getType(), Val.logBase2()));
2601 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2602 if (Op1F->isNullValue())
2603 return ReplaceInstUsesWith(I, Op1);
2605 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2606 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2607 if (Op1F->isExactlyValue(1.0))
2608 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2609 } else if (isa<VectorType>(Op1->getType())) {
2610 if (isa<ConstantAggregateZero>(Op1))
2611 return ReplaceInstUsesWith(I, Op1);
2613 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2614 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2615 return BinaryOperator::CreateNeg(Op0, I.getName());
2617 // As above, vector X*splat(1.0) -> X in all defined cases.
2618 if (Constant *Splat = Op1V->getSplatValue()) {
2619 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2620 if (F->isExactlyValue(1.0))
2621 return ReplaceInstUsesWith(I, Op0);
2622 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2623 if (CI->equalsInt(1))
2624 return ReplaceInstUsesWith(I, Op0);
2629 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2630 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2631 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2632 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2633 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2635 InsertNewInstBefore(Add, I);
2636 Value *C1C2 = ConstantExpr::getMul(Op1,
2637 cast<Constant>(Op0I->getOperand(1)));
2638 return BinaryOperator::CreateAdd(Add, C1C2);
2642 // Try to fold constant mul into select arguments.
2643 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2644 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2647 if (isa<PHINode>(Op0))
2648 if (Instruction *NV = FoldOpIntoPhi(I))
2652 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2653 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2654 return BinaryOperator::CreateMul(Op0v, Op1v);
2656 // (X / Y) * Y = X - (X % Y)
2657 // (X / Y) * -Y = (X % Y) - X
2659 Value *Op1 = I.getOperand(1);
2660 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2662 (BO->getOpcode() != Instruction::UDiv &&
2663 BO->getOpcode() != Instruction::SDiv)) {
2665 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2667 Value *Neg = dyn_castNegVal(Op1);
2668 if (BO && BO->hasOneUse() &&
2669 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2670 (BO->getOpcode() == Instruction::UDiv ||
2671 BO->getOpcode() == Instruction::SDiv)) {
2672 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2675 if (BO->getOpcode() == Instruction::UDiv)
2676 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2678 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2680 InsertNewInstBefore(Rem, I);
2684 return BinaryOperator::CreateSub(Op0BO, Rem);
2686 return BinaryOperator::CreateSub(Rem, Op0BO);
2690 if (I.getType() == Type::Int1Ty)
2691 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2693 // If one of the operands of the multiply is a cast from a boolean value, then
2694 // we know the bool is either zero or one, so this is a 'masking' multiply.
2695 // See if we can simplify things based on how the boolean was originally
2697 CastInst *BoolCast = 0;
2698 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2699 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2702 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2703 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2706 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2707 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2708 const Type *SCOpTy = SCIOp0->getType();
2711 // If the icmp is true iff the sign bit of X is set, then convert this
2712 // multiply into a shift/and combination.
2713 if (isa<ConstantInt>(SCIOp1) &&
2714 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2716 // Shift the X value right to turn it into "all signbits".
2717 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2718 SCOpTy->getPrimitiveSizeInBits()-1);
2720 InsertNewInstBefore(
2721 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2722 BoolCast->getOperand(0)->getName()+
2725 // If the multiply type is not the same as the source type, sign extend
2726 // or truncate to the multiply type.
2727 if (I.getType() != V->getType()) {
2728 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2729 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2730 Instruction::CastOps opcode =
2731 (SrcBits == DstBits ? Instruction::BitCast :
2732 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2733 V = InsertCastBefore(opcode, V, I.getType(), I);
2736 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2737 return BinaryOperator::CreateAnd(V, OtherOp);
2742 return Changed ? &I : 0;
2745 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2747 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2748 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2750 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2751 int NonNullOperand = -1;
2752 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2753 if (ST->isNullValue())
2755 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2756 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2757 if (ST->isNullValue())
2760 if (NonNullOperand == -1)
2763 Value *SelectCond = SI->getOperand(0);
2765 // Change the div/rem to use 'Y' instead of the select.
2766 I.setOperand(1, SI->getOperand(NonNullOperand));
2768 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2769 // problem. However, the select, or the condition of the select may have
2770 // multiple uses. Based on our knowledge that the operand must be non-zero,
2771 // propagate the known value for the select into other uses of it, and
2772 // propagate a known value of the condition into its other users.
2774 // If the select and condition only have a single use, don't bother with this,
2776 if (SI->use_empty() && SelectCond->hasOneUse())
2779 // Scan the current block backward, looking for other uses of SI.
2780 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2782 while (BBI != BBFront) {
2784 // If we found a call to a function, we can't assume it will return, so
2785 // information from below it cannot be propagated above it.
2786 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2789 // Replace uses of the select or its condition with the known values.
2790 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2793 *I = SI->getOperand(NonNullOperand);
2795 } else if (*I == SelectCond) {
2796 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2797 ConstantInt::getFalse();
2802 // If we past the instruction, quit looking for it.
2805 if (&*BBI == SelectCond)
2808 // If we ran out of things to eliminate, break out of the loop.
2809 if (SelectCond == 0 && SI == 0)
2817 /// This function implements the transforms on div instructions that work
2818 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2819 /// used by the visitors to those instructions.
2820 /// @brief Transforms common to all three div instructions
2821 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2822 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2824 // undef / X -> 0 for integer.
2825 // undef / X -> undef for FP (the undef could be a snan).
2826 if (isa<UndefValue>(Op0)) {
2827 if (Op0->getType()->isFPOrFPVector())
2828 return ReplaceInstUsesWith(I, Op0);
2829 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2832 // X / undef -> undef
2833 if (isa<UndefValue>(Op1))
2834 return ReplaceInstUsesWith(I, Op1);
2839 /// This function implements the transforms common to both integer division
2840 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2841 /// division instructions.
2842 /// @brief Common integer divide transforms
2843 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2844 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2846 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2848 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2849 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2850 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2851 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2854 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2855 return ReplaceInstUsesWith(I, CI);
2858 if (Instruction *Common = commonDivTransforms(I))
2861 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2862 // This does not apply for fdiv.
2863 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2866 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2868 if (RHS->equalsInt(1))
2869 return ReplaceInstUsesWith(I, Op0);
2871 // (X / C1) / C2 -> X / (C1*C2)
2872 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2873 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2874 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2875 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2876 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2878 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2879 Multiply(RHS, LHSRHS));
2882 if (!RHS->isZero()) { // avoid X udiv 0
2883 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2884 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2886 if (isa<PHINode>(Op0))
2887 if (Instruction *NV = FoldOpIntoPhi(I))
2892 // 0 / X == 0, we don't need to preserve faults!
2893 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2894 if (LHS->equalsInt(0))
2895 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2897 // It can't be division by zero, hence it must be division by one.
2898 if (I.getType() == Type::Int1Ty)
2899 return ReplaceInstUsesWith(I, Op0);
2901 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2902 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2905 return ReplaceInstUsesWith(I, Op0);
2911 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2912 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2914 // Handle the integer div common cases
2915 if (Instruction *Common = commonIDivTransforms(I))
2918 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2919 // X udiv C^2 -> X >> C
2920 // Check to see if this is an unsigned division with an exact power of 2,
2921 // if so, convert to a right shift.
2922 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2923 return BinaryOperator::CreateLShr(Op0,
2924 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2926 // X udiv C, where C >= signbit
2927 if (C->getValue().isNegative()) {
2928 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
2930 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
2931 ConstantInt::get(I.getType(), 1));
2935 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2936 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2937 if (RHSI->getOpcode() == Instruction::Shl &&
2938 isa<ConstantInt>(RHSI->getOperand(0))) {
2939 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2940 if (C1.isPowerOf2()) {
2941 Value *N = RHSI->getOperand(1);
2942 const Type *NTy = N->getType();
2943 if (uint32_t C2 = C1.logBase2()) {
2944 Constant *C2V = ConstantInt::get(NTy, C2);
2945 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2947 return BinaryOperator::CreateLShr(Op0, N);
2952 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2953 // where C1&C2 are powers of two.
2954 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2955 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2956 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2957 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2958 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2959 // Compute the shift amounts
2960 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2961 // Construct the "on true" case of the select
2962 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2963 Instruction *TSI = BinaryOperator::CreateLShr(
2964 Op0, TC, SI->getName()+".t");
2965 TSI = InsertNewInstBefore(TSI, I);
2967 // Construct the "on false" case of the select
2968 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2969 Instruction *FSI = BinaryOperator::CreateLShr(
2970 Op0, FC, SI->getName()+".f");
2971 FSI = InsertNewInstBefore(FSI, I);
2973 // construct the select instruction and return it.
2974 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2980 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2981 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2983 // Handle the integer div common cases
2984 if (Instruction *Common = commonIDivTransforms(I))
2987 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2989 if (RHS->isAllOnesValue())
2990 return BinaryOperator::CreateNeg(Op0);
2993 // If the sign bits of both operands are zero (i.e. we can prove they are
2994 // unsigned inputs), turn this into a udiv.
2995 if (I.getType()->isInteger()) {
2996 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2997 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2998 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2999 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3006 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3007 return commonDivTransforms(I);
3010 /// This function implements the transforms on rem instructions that work
3011 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3012 /// is used by the visitors to those instructions.
3013 /// @brief Transforms common to all three rem instructions
3014 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3015 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3017 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3018 if (I.getType()->isFPOrFPVector())
3019 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3020 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3022 if (isa<UndefValue>(Op1))
3023 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3025 // Handle cases involving: rem X, (select Cond, Y, Z)
3026 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3032 /// This function implements the transforms common to both integer remainder
3033 /// instructions (urem and srem). It is called by the visitors to those integer
3034 /// remainder instructions.
3035 /// @brief Common integer remainder transforms
3036 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3037 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3039 if (Instruction *common = commonRemTransforms(I))
3042 // 0 % X == 0 for integer, we don't need to preserve faults!
3043 if (Constant *LHS = dyn_cast<Constant>(Op0))
3044 if (LHS->isNullValue())
3045 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3047 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3048 // X % 0 == undef, we don't need to preserve faults!
3049 if (RHS->equalsInt(0))
3050 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3052 if (RHS->equalsInt(1)) // X % 1 == 0
3053 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3055 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3056 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3057 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3059 } else if (isa<PHINode>(Op0I)) {
3060 if (Instruction *NV = FoldOpIntoPhi(I))
3064 // See if we can fold away this rem instruction.
3065 if (SimplifyDemandedInstructionBits(I))
3073 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3074 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3076 if (Instruction *common = commonIRemTransforms(I))
3079 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3080 // X urem C^2 -> X and C
3081 // Check to see if this is an unsigned remainder with an exact power of 2,
3082 // if so, convert to a bitwise and.
3083 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3084 if (C->getValue().isPowerOf2())
3085 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3088 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3089 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3090 if (RHSI->getOpcode() == Instruction::Shl &&
3091 isa<ConstantInt>(RHSI->getOperand(0))) {
3092 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3093 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3094 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3096 return BinaryOperator::CreateAnd(Op0, Add);
3101 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3102 // where C1&C2 are powers of two.
3103 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3104 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3105 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3106 // STO == 0 and SFO == 0 handled above.
3107 if ((STO->getValue().isPowerOf2()) &&
3108 (SFO->getValue().isPowerOf2())) {
3109 Value *TrueAnd = InsertNewInstBefore(
3110 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3111 Value *FalseAnd = InsertNewInstBefore(
3112 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3113 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3121 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3122 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3124 // Handle the integer rem common cases
3125 if (Instruction *common = commonIRemTransforms(I))
3128 if (Value *RHSNeg = dyn_castNegVal(Op1))
3129 if (!isa<Constant>(RHSNeg) ||
3130 (isa<ConstantInt>(RHSNeg) &&
3131 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3133 AddUsesToWorkList(I);
3134 I.setOperand(1, RHSNeg);
3138 // If the sign bits of both operands are zero (i.e. we can prove they are
3139 // unsigned inputs), turn this into a urem.
3140 if (I.getType()->isInteger()) {
3141 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3142 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3143 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3144 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3148 // If it's a constant vector, flip any negative values positive.
3149 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3150 unsigned VWidth = RHSV->getNumOperands();
3152 bool hasNegative = false;
3153 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3154 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3155 if (RHS->getValue().isNegative())
3159 std::vector<Constant *> Elts(VWidth);
3160 for (unsigned i = 0; i != VWidth; ++i) {
3161 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3162 if (RHS->getValue().isNegative())
3163 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3169 Constant *NewRHSV = ConstantVector::get(Elts);
3170 if (NewRHSV != RHSV) {
3171 AddUsesToWorkList(I);
3172 I.setOperand(1, NewRHSV);
3181 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3182 return commonRemTransforms(I);
3185 // isOneBitSet - Return true if there is exactly one bit set in the specified
3187 static bool isOneBitSet(const ConstantInt *CI) {
3188 return CI->getValue().isPowerOf2();
3191 // isHighOnes - Return true if the constant is of the form 1+0+.
3192 // This is the same as lowones(~X).
3193 static bool isHighOnes(const ConstantInt *CI) {
3194 return (~CI->getValue() + 1).isPowerOf2();
3197 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3198 /// are carefully arranged to allow folding of expressions such as:
3200 /// (A < B) | (A > B) --> (A != B)
3202 /// Note that this is only valid if the first and second predicates have the
3203 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3205 /// Three bits are used to represent the condition, as follows:
3210 /// <=> Value Definition
3211 /// 000 0 Always false
3218 /// 111 7 Always true
3220 static unsigned getICmpCode(const ICmpInst *ICI) {
3221 switch (ICI->getPredicate()) {
3223 case ICmpInst::ICMP_UGT: return 1; // 001
3224 case ICmpInst::ICMP_SGT: return 1; // 001
3225 case ICmpInst::ICMP_EQ: return 2; // 010
3226 case ICmpInst::ICMP_UGE: return 3; // 011
3227 case ICmpInst::ICMP_SGE: return 3; // 011
3228 case ICmpInst::ICMP_ULT: return 4; // 100
3229 case ICmpInst::ICMP_SLT: return 4; // 100
3230 case ICmpInst::ICMP_NE: return 5; // 101
3231 case ICmpInst::ICMP_ULE: return 6; // 110
3232 case ICmpInst::ICMP_SLE: return 6; // 110
3235 assert(0 && "Invalid ICmp predicate!");
3240 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3241 /// predicate into a three bit mask. It also returns whether it is an ordered
3242 /// predicate by reference.
3243 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3246 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3247 case FCmpInst::FCMP_UNO: return 0; // 000
3248 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3249 case FCmpInst::FCMP_UGT: return 1; // 001
3250 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3251 case FCmpInst::FCMP_UEQ: return 2; // 010
3252 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3253 case FCmpInst::FCMP_UGE: return 3; // 011
3254 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3255 case FCmpInst::FCMP_ULT: return 4; // 100
3256 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3257 case FCmpInst::FCMP_UNE: return 5; // 101
3258 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3259 case FCmpInst::FCMP_ULE: return 6; // 110
3262 // Not expecting FCMP_FALSE and FCMP_TRUE;
3263 assert(0 && "Unexpected FCmp predicate!");
3268 /// getICmpValue - This is the complement of getICmpCode, which turns an
3269 /// opcode and two operands into either a constant true or false, or a brand
3270 /// new ICmp instruction. The sign is passed in to determine which kind
3271 /// of predicate to use in the new icmp instruction.
3272 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3274 default: assert(0 && "Illegal ICmp code!");
3275 case 0: return ConstantInt::getFalse();
3278 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3280 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3281 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3284 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3286 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3289 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3291 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3292 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3295 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3297 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3298 case 7: return ConstantInt::getTrue();
3302 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3303 /// opcode and two operands into either a FCmp instruction. isordered is passed
3304 /// in to determine which kind of predicate to use in the new fcmp instruction.
3305 static Value *getFCmpValue(bool isordered, unsigned code,
3306 Value *LHS, Value *RHS) {
3308 default: assert(0 && "Illegal FCmp code!");
3311 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3313 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3316 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3318 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3321 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3323 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3326 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3328 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3331 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3333 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3336 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3338 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3341 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3343 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3344 case 7: return ConstantInt::getTrue();
3348 /// PredicatesFoldable - Return true if both predicates match sign or if at
3349 /// least one of them is an equality comparison (which is signless).
3350 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3351 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3352 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3353 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3357 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3358 struct FoldICmpLogical {
3361 ICmpInst::Predicate pred;
3362 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3363 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3364 pred(ICI->getPredicate()) {}
3365 bool shouldApply(Value *V) const {
3366 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3367 if (PredicatesFoldable(pred, ICI->getPredicate()))
3368 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3369 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3372 Instruction *apply(Instruction &Log) const {
3373 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3374 if (ICI->getOperand(0) != LHS) {
3375 assert(ICI->getOperand(1) == LHS);
3376 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3379 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3380 unsigned LHSCode = getICmpCode(ICI);
3381 unsigned RHSCode = getICmpCode(RHSICI);
3383 switch (Log.getOpcode()) {
3384 case Instruction::And: Code = LHSCode & RHSCode; break;
3385 case Instruction::Or: Code = LHSCode | RHSCode; break;
3386 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3387 default: assert(0 && "Illegal logical opcode!"); return 0;
3390 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3391 ICmpInst::isSignedPredicate(ICI->getPredicate());
3393 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3394 if (Instruction *I = dyn_cast<Instruction>(RV))
3396 // Otherwise, it's a constant boolean value...
3397 return IC.ReplaceInstUsesWith(Log, RV);
3400 } // end anonymous namespace
3402 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3403 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3404 // guaranteed to be a binary operator.
3405 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3407 ConstantInt *AndRHS,
3408 BinaryOperator &TheAnd) {
3409 Value *X = Op->getOperand(0);
3410 Constant *Together = 0;
3412 Together = And(AndRHS, OpRHS);
3414 switch (Op->getOpcode()) {
3415 case Instruction::Xor:
3416 if (Op->hasOneUse()) {
3417 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3418 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3419 InsertNewInstBefore(And, TheAnd);
3421 return BinaryOperator::CreateXor(And, Together);
3424 case Instruction::Or:
3425 if (Together == AndRHS) // (X | C) & C --> C
3426 return ReplaceInstUsesWith(TheAnd, AndRHS);
3428 if (Op->hasOneUse() && Together != OpRHS) {
3429 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3430 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3431 InsertNewInstBefore(Or, TheAnd);
3433 return BinaryOperator::CreateAnd(Or, AndRHS);
3436 case Instruction::Add:
3437 if (Op->hasOneUse()) {
3438 // Adding a one to a single bit bit-field should be turned into an XOR
3439 // of the bit. First thing to check is to see if this AND is with a
3440 // single bit constant.
3441 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3443 // If there is only one bit set...
3444 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3445 // Ok, at this point, we know that we are masking the result of the
3446 // ADD down to exactly one bit. If the constant we are adding has
3447 // no bits set below this bit, then we can eliminate the ADD.
3448 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3450 // Check to see if any bits below the one bit set in AndRHSV are set.
3451 if ((AddRHS & (AndRHSV-1)) == 0) {
3452 // If not, the only thing that can effect the output of the AND is
3453 // the bit specified by AndRHSV. If that bit is set, the effect of
3454 // the XOR is to toggle the bit. If it is clear, then the ADD has
3456 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3457 TheAnd.setOperand(0, X);
3460 // Pull the XOR out of the AND.
3461 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3462 InsertNewInstBefore(NewAnd, TheAnd);
3463 NewAnd->takeName(Op);
3464 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3471 case Instruction::Shl: {
3472 // We know that the AND will not produce any of the bits shifted in, so if
3473 // the anded constant includes them, clear them now!
3475 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3476 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3477 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3478 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3480 if (CI->getValue() == ShlMask) {
3481 // Masking out bits that the shift already masks
3482 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3483 } else if (CI != AndRHS) { // Reducing bits set in and.
3484 TheAnd.setOperand(1, CI);
3489 case Instruction::LShr:
3491 // We know that the AND will not produce any of the bits shifted in, so if
3492 // the anded constant includes them, clear them now! This only applies to
3493 // unsigned shifts, because a signed shr may bring in set bits!
3495 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3496 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3497 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3498 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3500 if (CI->getValue() == ShrMask) {
3501 // Masking out bits that the shift already masks.
3502 return ReplaceInstUsesWith(TheAnd, Op);
3503 } else if (CI != AndRHS) {
3504 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3509 case Instruction::AShr:
3511 // See if this is shifting in some sign extension, then masking it out
3513 if (Op->hasOneUse()) {
3514 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3515 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3516 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3517 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3518 if (C == AndRHS) { // Masking out bits shifted in.
3519 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3520 // Make the argument unsigned.
3521 Value *ShVal = Op->getOperand(0);
3522 ShVal = InsertNewInstBefore(
3523 BinaryOperator::CreateLShr(ShVal, OpRHS,
3524 Op->getName()), TheAnd);
3525 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3534 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3535 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3536 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3537 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3538 /// insert new instructions.
3539 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3540 bool isSigned, bool Inside,
3542 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3543 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3544 "Lo is not <= Hi in range emission code!");
3547 if (Lo == Hi) // Trivially false.
3548 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3550 // V >= Min && V < Hi --> V < Hi
3551 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3552 ICmpInst::Predicate pred = (isSigned ?
3553 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3554 return new ICmpInst(pred, V, Hi);
3557 // Emit V-Lo <u Hi-Lo
3558 Constant *NegLo = ConstantExpr::getNeg(Lo);
3559 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3560 InsertNewInstBefore(Add, IB);
3561 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3562 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3565 if (Lo == Hi) // Trivially true.
3566 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3568 // V < Min || V >= Hi -> V > Hi-1
3569 Hi = SubOne(cast<ConstantInt>(Hi));
3570 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3571 ICmpInst::Predicate pred = (isSigned ?
3572 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3573 return new ICmpInst(pred, V, Hi);
3576 // Emit V-Lo >u Hi-1-Lo
3577 // Note that Hi has already had one subtracted from it, above.
3578 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3579 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3580 InsertNewInstBefore(Add, IB);
3581 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3582 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3585 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3586 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3587 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3588 // not, since all 1s are not contiguous.
3589 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3590 const APInt& V = Val->getValue();
3591 uint32_t BitWidth = Val->getType()->getBitWidth();
3592 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3594 // look for the first zero bit after the run of ones
3595 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3596 // look for the first non-zero bit
3597 ME = V.getActiveBits();
3601 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3602 /// where isSub determines whether the operator is a sub. If we can fold one of
3603 /// the following xforms:
3605 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3606 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3607 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3609 /// return (A +/- B).
3611 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3612 ConstantInt *Mask, bool isSub,
3614 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3615 if (!LHSI || LHSI->getNumOperands() != 2 ||
3616 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3618 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3620 switch (LHSI->getOpcode()) {
3622 case Instruction::And:
3623 if (And(N, Mask) == Mask) {
3624 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3625 if ((Mask->getValue().countLeadingZeros() +
3626 Mask->getValue().countPopulation()) ==
3627 Mask->getValue().getBitWidth())
3630 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3631 // part, we don't need any explicit masks to take them out of A. If that
3632 // is all N is, ignore it.
3633 uint32_t MB = 0, ME = 0;
3634 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3635 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3636 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3637 if (MaskedValueIsZero(RHS, Mask))
3642 case Instruction::Or:
3643 case Instruction::Xor:
3644 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3645 if ((Mask->getValue().countLeadingZeros() +
3646 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3647 && And(N, Mask)->isZero())
3654 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3656 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3657 return InsertNewInstBefore(New, I);
3660 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3661 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3662 ICmpInst *LHS, ICmpInst *RHS) {
3664 ConstantInt *LHSCst, *RHSCst;
3665 ICmpInst::Predicate LHSCC, RHSCC;
3667 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3668 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3669 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3672 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3673 // where C is a power of 2
3674 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3675 LHSCst->getValue().isPowerOf2()) {
3676 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3677 InsertNewInstBefore(NewOr, I);
3678 return new ICmpInst(LHSCC, NewOr, LHSCst);
3681 // From here on, we only handle:
3682 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3683 if (Val != Val2) return 0;
3685 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3686 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3687 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3688 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3689 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3692 // We can't fold (ugt x, C) & (sgt x, C2).
3693 if (!PredicatesFoldable(LHSCC, RHSCC))
3696 // Ensure that the larger constant is on the RHS.
3698 if (ICmpInst::isSignedPredicate(LHSCC) ||
3699 (ICmpInst::isEquality(LHSCC) &&
3700 ICmpInst::isSignedPredicate(RHSCC)))
3701 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3703 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3706 std::swap(LHS, RHS);
3707 std::swap(LHSCst, RHSCst);
3708 std::swap(LHSCC, RHSCC);
3711 // At this point, we know we have have two icmp instructions
3712 // comparing a value against two constants and and'ing the result
3713 // together. Because of the above check, we know that we only have
3714 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3715 // (from the FoldICmpLogical check above), that the two constants
3716 // are not equal and that the larger constant is on the RHS
3717 assert(LHSCst != RHSCst && "Compares not folded above?");
3720 default: assert(0 && "Unknown integer condition code!");
3721 case ICmpInst::ICMP_EQ:
3723 default: assert(0 && "Unknown integer condition code!");
3724 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3725 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3726 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3727 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3728 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3729 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3730 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3731 return ReplaceInstUsesWith(I, LHS);
3733 case ICmpInst::ICMP_NE:
3735 default: assert(0 && "Unknown integer condition code!");
3736 case ICmpInst::ICMP_ULT:
3737 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3738 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3739 break; // (X != 13 & X u< 15) -> no change
3740 case ICmpInst::ICMP_SLT:
3741 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3742 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3743 break; // (X != 13 & X s< 15) -> no change
3744 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3745 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3746 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3747 return ReplaceInstUsesWith(I, RHS);
3748 case ICmpInst::ICMP_NE:
3749 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3750 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3751 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3752 Val->getName()+".off");
3753 InsertNewInstBefore(Add, I);
3754 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3755 ConstantInt::get(Add->getType(), 1));
3757 break; // (X != 13 & X != 15) -> no change
3760 case ICmpInst::ICMP_ULT:
3762 default: assert(0 && "Unknown integer condition code!");
3763 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3764 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3765 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3766 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3768 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3769 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3770 return ReplaceInstUsesWith(I, LHS);
3771 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3775 case ICmpInst::ICMP_SLT:
3777 default: assert(0 && "Unknown integer condition code!");
3778 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3779 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3780 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3781 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3783 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3784 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3785 return ReplaceInstUsesWith(I, LHS);
3786 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3790 case ICmpInst::ICMP_UGT:
3792 default: assert(0 && "Unknown integer condition code!");
3793 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3794 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3795 return ReplaceInstUsesWith(I, RHS);
3796 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3798 case ICmpInst::ICMP_NE:
3799 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3800 return new ICmpInst(LHSCC, Val, RHSCst);
3801 break; // (X u> 13 & X != 15) -> no change
3802 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3803 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3804 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3808 case ICmpInst::ICMP_SGT:
3810 default: assert(0 && "Unknown integer condition code!");
3811 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3812 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3813 return ReplaceInstUsesWith(I, RHS);
3814 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3816 case ICmpInst::ICMP_NE:
3817 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3818 return new ICmpInst(LHSCC, Val, RHSCst);
3819 break; // (X s> 13 & X != 15) -> no change
3820 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3821 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3822 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3832 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3833 bool Changed = SimplifyCommutative(I);
3834 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3836 if (isa<UndefValue>(Op1)) // X & undef -> 0
3837 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3841 return ReplaceInstUsesWith(I, Op1);
3843 // See if we can simplify any instructions used by the instruction whose sole
3844 // purpose is to compute bits we don't care about.
3845 if (!isa<VectorType>(I.getType())) {
3846 if (SimplifyDemandedInstructionBits(I))
3849 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3850 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3851 return ReplaceInstUsesWith(I, I.getOperand(0));
3852 } else if (isa<ConstantAggregateZero>(Op1)) {
3853 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3857 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3858 const APInt& AndRHSMask = AndRHS->getValue();
3859 APInt NotAndRHS(~AndRHSMask);
3861 // Optimize a variety of ((val OP C1) & C2) combinations...
3862 if (isa<BinaryOperator>(Op0)) {
3863 Instruction *Op0I = cast<Instruction>(Op0);
3864 Value *Op0LHS = Op0I->getOperand(0);
3865 Value *Op0RHS = Op0I->getOperand(1);
3866 switch (Op0I->getOpcode()) {
3867 case Instruction::Xor:
3868 case Instruction::Or:
3869 // If the mask is only needed on one incoming arm, push it up.
3870 if (Op0I->hasOneUse()) {
3871 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3872 // Not masking anything out for the LHS, move to RHS.
3873 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3874 Op0RHS->getName()+".masked");
3875 InsertNewInstBefore(NewRHS, I);
3876 return BinaryOperator::Create(
3877 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3879 if (!isa<Constant>(Op0RHS) &&
3880 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3881 // Not masking anything out for the RHS, move to LHS.
3882 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3883 Op0LHS->getName()+".masked");
3884 InsertNewInstBefore(NewLHS, I);
3885 return BinaryOperator::Create(
3886 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3891 case Instruction::Add:
3892 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3893 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3894 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3895 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3896 return BinaryOperator::CreateAnd(V, AndRHS);
3897 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3898 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3901 case Instruction::Sub:
3902 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3903 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3904 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3905 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3906 return BinaryOperator::CreateAnd(V, AndRHS);
3908 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3909 // has 1's for all bits that the subtraction with A might affect.
3910 if (Op0I->hasOneUse()) {
3911 uint32_t BitWidth = AndRHSMask.getBitWidth();
3912 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3913 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3915 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3916 if (!(A && A->isZero()) && // avoid infinite recursion.
3917 MaskedValueIsZero(Op0LHS, Mask)) {
3918 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3919 InsertNewInstBefore(NewNeg, I);
3920 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3925 case Instruction::Shl:
3926 case Instruction::LShr:
3927 // (1 << x) & 1 --> zext(x == 0)
3928 // (1 >> x) & 1 --> zext(x == 0)
3929 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3930 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3931 Constant::getNullValue(I.getType()));
3932 InsertNewInstBefore(NewICmp, I);
3933 return new ZExtInst(NewICmp, I.getType());
3938 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3939 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3941 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3942 // If this is an integer truncation or change from signed-to-unsigned, and
3943 // if the source is an and/or with immediate, transform it. This
3944 // frequently occurs for bitfield accesses.
3945 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3946 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3947 CastOp->getNumOperands() == 2)
3948 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3949 if (CastOp->getOpcode() == Instruction::And) {
3950 // Change: and (cast (and X, C1) to T), C2
3951 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3952 // This will fold the two constants together, which may allow
3953 // other simplifications.
3954 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3955 CastOp->getOperand(0), I.getType(),
3956 CastOp->getName()+".shrunk");
3957 NewCast = InsertNewInstBefore(NewCast, I);
3958 // trunc_or_bitcast(C1)&C2
3959 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3960 C3 = ConstantExpr::getAnd(C3, AndRHS);
3961 return BinaryOperator::CreateAnd(NewCast, C3);
3962 } else if (CastOp->getOpcode() == Instruction::Or) {
3963 // Change: and (cast (or X, C1) to T), C2
3964 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3965 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3966 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3967 return ReplaceInstUsesWith(I, AndRHS);
3973 // Try to fold constant and into select arguments.
3974 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3975 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3977 if (isa<PHINode>(Op0))
3978 if (Instruction *NV = FoldOpIntoPhi(I))
3982 Value *Op0NotVal = dyn_castNotVal(Op0);
3983 Value *Op1NotVal = dyn_castNotVal(Op1);
3985 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3986 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3988 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3989 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3990 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3991 I.getName()+".demorgan");
3992 InsertNewInstBefore(Or, I);
3993 return BinaryOperator::CreateNot(Or);
3997 Value *A = 0, *B = 0, *C = 0, *D = 0;
3998 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3999 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4000 return ReplaceInstUsesWith(I, Op1);
4002 // (A|B) & ~(A&B) -> A^B
4003 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4004 if ((A == C && B == D) || (A == D && B == C))
4005 return BinaryOperator::CreateXor(A, B);
4009 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4010 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4011 return ReplaceInstUsesWith(I, Op0);
4013 // ~(A&B) & (A|B) -> A^B
4014 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4015 if ((A == C && B == D) || (A == D && B == C))
4016 return BinaryOperator::CreateXor(A, B);
4020 if (Op0->hasOneUse() &&
4021 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4022 if (A == Op1) { // (A^B)&A -> A&(A^B)
4023 I.swapOperands(); // Simplify below
4024 std::swap(Op0, Op1);
4025 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4026 cast<BinaryOperator>(Op0)->swapOperands();
4027 I.swapOperands(); // Simplify below
4028 std::swap(Op0, Op1);
4032 if (Op1->hasOneUse() &&
4033 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4034 if (B == Op0) { // B&(A^B) -> B&(B^A)
4035 cast<BinaryOperator>(Op1)->swapOperands();
4038 if (A == Op0) { // A&(A^B) -> A & ~B
4039 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4040 InsertNewInstBefore(NotB, I);
4041 return BinaryOperator::CreateAnd(A, NotB);
4045 // (A&((~A)|B)) -> A&B
4046 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4047 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4048 return BinaryOperator::CreateAnd(A, Op1);
4049 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4050 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4051 return BinaryOperator::CreateAnd(A, Op0);
4054 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4055 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4056 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4059 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4060 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4064 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4065 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4066 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4067 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4068 const Type *SrcTy = Op0C->getOperand(0)->getType();
4069 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4070 // Only do this if the casts both really cause code to be generated.
4071 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4073 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4075 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4076 Op1C->getOperand(0),
4078 InsertNewInstBefore(NewOp, I);
4079 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4083 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4084 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4085 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4086 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4087 SI0->getOperand(1) == SI1->getOperand(1) &&
4088 (SI0->hasOneUse() || SI1->hasOneUse())) {
4089 Instruction *NewOp =
4090 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4092 SI0->getName()), I);
4093 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4094 SI1->getOperand(1));
4098 // If and'ing two fcmp, try combine them into one.
4099 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4100 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4101 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4102 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4103 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4104 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4105 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4106 // If either of the constants are nans, then the whole thing returns
4108 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4109 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4110 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4111 RHS->getOperand(0));
4114 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4115 FCmpInst::Predicate Op0CC, Op1CC;
4116 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4117 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4118 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4119 // Swap RHS operands to match LHS.
4120 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4121 std::swap(Op1LHS, Op1RHS);
4123 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4124 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4126 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4127 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4128 Op1CC == FCmpInst::FCMP_FALSE)
4129 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4130 else if (Op0CC == FCmpInst::FCMP_TRUE)
4131 return ReplaceInstUsesWith(I, Op1);
4132 else if (Op1CC == FCmpInst::FCMP_TRUE)
4133 return ReplaceInstUsesWith(I, Op0);
4136 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4137 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4139 std::swap(Op0, Op1);
4140 std::swap(Op0Pred, Op1Pred);
4141 std::swap(Op0Ordered, Op1Ordered);
4144 // uno && ueq -> uno && (uno || eq) -> ueq
4145 // ord && olt -> ord && (ord && lt) -> olt
4146 if (Op0Ordered == Op1Ordered)
4147 return ReplaceInstUsesWith(I, Op1);
4148 // uno && oeq -> uno && (ord && eq) -> false
4149 // uno && ord -> false
4151 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4152 // ord && ueq -> ord && (uno || eq) -> oeq
4153 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4162 return Changed ? &I : 0;
4165 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4166 /// capable of providing pieces of a bswap. The subexpression provides pieces
4167 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4168 /// the expression came from the corresponding "byte swapped" byte in some other
4169 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4170 /// we know that the expression deposits the low byte of %X into the high byte
4171 /// of the bswap result and that all other bytes are zero. This expression is
4172 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4175 /// This function returns true if the match was unsuccessful and false if so.
4176 /// On entry to the function the "OverallLeftShift" is a signed integer value
4177 /// indicating the number of bytes that the subexpression is later shifted. For
4178 /// example, if the expression is later right shifted by 16 bits, the
4179 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4180 /// byte of ByteValues is actually being set.
4182 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4183 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4184 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4185 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4186 /// always in the local (OverallLeftShift) coordinate space.
4188 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4189 SmallVector<Value*, 8> &ByteValues) {
4190 if (Instruction *I = dyn_cast<Instruction>(V)) {
4191 // If this is an or instruction, it may be an inner node of the bswap.
4192 if (I->getOpcode() == Instruction::Or) {
4193 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4195 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4199 // If this is a logical shift by a constant multiple of 8, recurse with
4200 // OverallLeftShift and ByteMask adjusted.
4201 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4203 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4204 // Ensure the shift amount is defined and of a byte value.
4205 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4208 unsigned ByteShift = ShAmt >> 3;
4209 if (I->getOpcode() == Instruction::Shl) {
4210 // X << 2 -> collect(X, +2)
4211 OverallLeftShift += ByteShift;
4212 ByteMask >>= ByteShift;
4214 // X >>u 2 -> collect(X, -2)
4215 OverallLeftShift -= ByteShift;
4216 ByteMask <<= ByteShift;
4217 ByteMask &= (~0U >> (32-ByteValues.size()));
4220 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4221 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4223 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4227 // If this is a logical 'and' with a mask that clears bytes, clear the
4228 // corresponding bytes in ByteMask.
4229 if (I->getOpcode() == Instruction::And &&
4230 isa<ConstantInt>(I->getOperand(1))) {
4231 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4232 unsigned NumBytes = ByteValues.size();
4233 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4234 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4236 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4237 // If this byte is masked out by a later operation, we don't care what
4239 if ((ByteMask & (1 << i)) == 0)
4242 // If the AndMask is all zeros for this byte, clear the bit.
4243 APInt MaskB = AndMask & Byte;
4245 ByteMask &= ~(1U << i);
4249 // If the AndMask is not all ones for this byte, it's not a bytezap.
4253 // Otherwise, this byte is kept.
4256 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4261 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4262 // the input value to the bswap. Some observations: 1) if more than one byte
4263 // is demanded from this input, then it could not be successfully assembled
4264 // into a byteswap. At least one of the two bytes would not be aligned with
4265 // their ultimate destination.
4266 if (!isPowerOf2_32(ByteMask)) return true;
4267 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4269 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4270 // is demanded, it needs to go into byte 0 of the result. This means that the
4271 // byte needs to be shifted until it lands in the right byte bucket. The
4272 // shift amount depends on the position: if the byte is coming from the high
4273 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4274 // low part, it must be shifted left.
4275 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4276 if (InputByteNo < ByteValues.size()/2) {
4277 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4280 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4284 // If the destination byte value is already defined, the values are or'd
4285 // together, which isn't a bswap (unless it's an or of the same bits).
4286 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4288 ByteValues[DestByteNo] = V;
4292 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4293 /// If so, insert the new bswap intrinsic and return it.
4294 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4295 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4296 if (!ITy || ITy->getBitWidth() % 16 ||
4297 // ByteMask only allows up to 32-byte values.
4298 ITy->getBitWidth() > 32*8)
4299 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4301 /// ByteValues - For each byte of the result, we keep track of which value
4302 /// defines each byte.
4303 SmallVector<Value*, 8> ByteValues;
4304 ByteValues.resize(ITy->getBitWidth()/8);
4306 // Try to find all the pieces corresponding to the bswap.
4307 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4308 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4311 // Check to see if all of the bytes come from the same value.
4312 Value *V = ByteValues[0];
4313 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4315 // Check to make sure that all of the bytes come from the same value.
4316 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4317 if (ByteValues[i] != V)
4319 const Type *Tys[] = { ITy };
4320 Module *M = I.getParent()->getParent()->getParent();
4321 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4322 return CallInst::Create(F, V);
4325 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4326 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4327 /// we can simplify this expression to "cond ? C : D or B".
4328 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4329 Value *C, Value *D) {
4330 // If A is not a select of -1/0, this cannot match.
4332 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4335 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4336 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4337 return SelectInst::Create(Cond, C, B);
4338 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4339 return SelectInst::Create(Cond, C, B);
4340 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4341 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4342 return SelectInst::Create(Cond, C, D);
4343 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4344 return SelectInst::Create(Cond, C, D);
4348 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4349 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4350 ICmpInst *LHS, ICmpInst *RHS) {
4352 ConstantInt *LHSCst, *RHSCst;
4353 ICmpInst::Predicate LHSCC, RHSCC;
4355 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4356 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4357 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4360 // From here on, we only handle:
4361 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4362 if (Val != Val2) return 0;
4364 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4365 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4366 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4367 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4368 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4371 // We can't fold (ugt x, C) | (sgt x, C2).
4372 if (!PredicatesFoldable(LHSCC, RHSCC))
4375 // Ensure that the larger constant is on the RHS.
4377 if (ICmpInst::isSignedPredicate(LHSCC) ||
4378 (ICmpInst::isEquality(LHSCC) &&
4379 ICmpInst::isSignedPredicate(RHSCC)))
4380 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4382 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4385 std::swap(LHS, RHS);
4386 std::swap(LHSCst, RHSCst);
4387 std::swap(LHSCC, RHSCC);
4390 // At this point, we know we have have two icmp instructions
4391 // comparing a value against two constants and or'ing the result
4392 // together. Because of the above check, we know that we only have
4393 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4394 // FoldICmpLogical check above), that the two constants are not
4396 assert(LHSCst != RHSCst && "Compares not folded above?");
4399 default: assert(0 && "Unknown integer condition code!");
4400 case ICmpInst::ICMP_EQ:
4402 default: assert(0 && "Unknown integer condition code!");
4403 case ICmpInst::ICMP_EQ:
4404 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4405 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4406 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4407 Val->getName()+".off");
4408 InsertNewInstBefore(Add, I);
4409 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4410 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4412 break; // (X == 13 | X == 15) -> no change
4413 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4414 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4416 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4417 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4418 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4419 return ReplaceInstUsesWith(I, RHS);
4422 case ICmpInst::ICMP_NE:
4424 default: assert(0 && "Unknown integer condition code!");
4425 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4426 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4427 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4428 return ReplaceInstUsesWith(I, LHS);
4429 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4430 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4431 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4432 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4435 case ICmpInst::ICMP_ULT:
4437 default: assert(0 && "Unknown integer condition code!");
4438 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4440 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4441 // If RHSCst is [us]MAXINT, it is always false. Not handling
4442 // this can cause overflow.
4443 if (RHSCst->isMaxValue(false))
4444 return ReplaceInstUsesWith(I, LHS);
4445 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4446 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4448 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4449 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4450 return ReplaceInstUsesWith(I, RHS);
4451 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4455 case ICmpInst::ICMP_SLT:
4457 default: assert(0 && "Unknown integer condition code!");
4458 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4460 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4461 // If RHSCst is [us]MAXINT, it is always false. Not handling
4462 // this can cause overflow.
4463 if (RHSCst->isMaxValue(true))
4464 return ReplaceInstUsesWith(I, LHS);
4465 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4466 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4468 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4469 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4470 return ReplaceInstUsesWith(I, RHS);
4471 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4475 case ICmpInst::ICMP_UGT:
4477 default: assert(0 && "Unknown integer condition code!");
4478 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4479 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4480 return ReplaceInstUsesWith(I, LHS);
4481 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4483 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4484 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4485 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4486 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4490 case ICmpInst::ICMP_SGT:
4492 default: assert(0 && "Unknown integer condition code!");
4493 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4494 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4495 return ReplaceInstUsesWith(I, LHS);
4496 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4498 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4499 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4500 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4501 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4509 /// FoldOrWithConstants - This helper function folds:
4511 /// ((A | B) & C1) | (B & C2)
4517 /// when the XOR of the two constants is "all ones" (-1).
4518 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4519 Value *A, Value *B, Value *C) {
4520 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4524 ConstantInt *CI2 = 0;
4525 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4527 APInt Xor = CI1->getValue() ^ CI2->getValue();
4528 if (!Xor.isAllOnesValue()) return 0;
4530 if (V1 == A || V1 == B) {
4531 Instruction *NewOp =
4532 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4533 return BinaryOperator::CreateOr(NewOp, V1);
4539 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4540 bool Changed = SimplifyCommutative(I);
4541 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4543 if (isa<UndefValue>(Op1)) // X | undef -> -1
4544 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4548 return ReplaceInstUsesWith(I, Op0);
4550 // See if we can simplify any instructions used by the instruction whose sole
4551 // purpose is to compute bits we don't care about.
4552 if (!isa<VectorType>(I.getType())) {
4553 if (SimplifyDemandedInstructionBits(I))
4555 } else if (isa<ConstantAggregateZero>(Op1)) {
4556 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4557 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4558 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4559 return ReplaceInstUsesWith(I, I.getOperand(1));
4565 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4566 ConstantInt *C1 = 0; Value *X = 0;
4567 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4568 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4569 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4570 InsertNewInstBefore(Or, I);
4572 return BinaryOperator::CreateAnd(Or,
4573 ConstantInt::get(RHS->getValue() | C1->getValue()));
4576 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4577 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4578 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4579 InsertNewInstBefore(Or, I);
4581 return BinaryOperator::CreateXor(Or,
4582 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4585 // Try to fold constant and into select arguments.
4586 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4587 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4589 if (isa<PHINode>(Op0))
4590 if (Instruction *NV = FoldOpIntoPhi(I))
4594 Value *A = 0, *B = 0;
4595 ConstantInt *C1 = 0, *C2 = 0;
4597 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4598 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4599 return ReplaceInstUsesWith(I, Op1);
4600 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4601 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4602 return ReplaceInstUsesWith(I, Op0);
4604 // (A | B) | C and A | (B | C) -> bswap if possible.
4605 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4606 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4607 match(Op1, m_Or(m_Value(), m_Value())) ||
4608 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4609 match(Op1, m_Shift(m_Value(), m_Value())))) {
4610 if (Instruction *BSwap = MatchBSwap(I))
4614 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4615 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4616 MaskedValueIsZero(Op1, C1->getValue())) {
4617 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4618 InsertNewInstBefore(NOr, I);
4620 return BinaryOperator::CreateXor(NOr, C1);
4623 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4624 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4625 MaskedValueIsZero(Op0, C1->getValue())) {
4626 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4627 InsertNewInstBefore(NOr, I);
4629 return BinaryOperator::CreateXor(NOr, C1);
4633 Value *C = 0, *D = 0;
4634 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4635 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4636 Value *V1 = 0, *V2 = 0, *V3 = 0;
4637 C1 = dyn_cast<ConstantInt>(C);
4638 C2 = dyn_cast<ConstantInt>(D);
4639 if (C1 && C2) { // (A & C1)|(B & C2)
4640 // If we have: ((V + N) & C1) | (V & C2)
4641 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4642 // replace with V+N.
4643 if (C1->getValue() == ~C2->getValue()) {
4644 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4645 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4646 // Add commutes, try both ways.
4647 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4648 return ReplaceInstUsesWith(I, A);
4649 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4650 return ReplaceInstUsesWith(I, A);
4652 // Or commutes, try both ways.
4653 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4654 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4655 // Add commutes, try both ways.
4656 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4657 return ReplaceInstUsesWith(I, B);
4658 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4659 return ReplaceInstUsesWith(I, B);
4662 V1 = 0; V2 = 0; V3 = 0;
4665 // Check to see if we have any common things being and'ed. If so, find the
4666 // terms for V1 & (V2|V3).
4667 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4668 if (A == B) // (A & C)|(A & D) == A & (C|D)
4669 V1 = A, V2 = C, V3 = D;
4670 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4671 V1 = A, V2 = B, V3 = C;
4672 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4673 V1 = C, V2 = A, V3 = D;
4674 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4675 V1 = C, V2 = A, V3 = B;
4679 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4680 return BinaryOperator::CreateAnd(V1, Or);
4684 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4685 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4687 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4689 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4691 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4694 // ((A&~B)|(~A&B)) -> A^B
4695 if ((match(C, m_Not(m_Specific(D))) &&
4696 match(B, m_Not(m_Specific(A)))))
4697 return BinaryOperator::CreateXor(A, D);
4698 // ((~B&A)|(~A&B)) -> A^B
4699 if ((match(A, m_Not(m_Specific(D))) &&
4700 match(B, m_Not(m_Specific(C)))))
4701 return BinaryOperator::CreateXor(C, D);
4702 // ((A&~B)|(B&~A)) -> A^B
4703 if ((match(C, m_Not(m_Specific(B))) &&
4704 match(D, m_Not(m_Specific(A)))))
4705 return BinaryOperator::CreateXor(A, B);
4706 // ((~B&A)|(B&~A)) -> A^B
4707 if ((match(A, m_Not(m_Specific(B))) &&
4708 match(D, m_Not(m_Specific(C)))))
4709 return BinaryOperator::CreateXor(C, B);
4712 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4713 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4714 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4715 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4716 SI0->getOperand(1) == SI1->getOperand(1) &&
4717 (SI0->hasOneUse() || SI1->hasOneUse())) {
4718 Instruction *NewOp =
4719 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4721 SI0->getName()), I);
4722 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4723 SI1->getOperand(1));
4727 // ((A|B)&1)|(B&-2) -> (A&1) | B
4728 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4729 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4730 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4731 if (Ret) return Ret;
4733 // (B&-2)|((A|B)&1) -> (A&1) | B
4734 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4735 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4736 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4737 if (Ret) return Ret;
4740 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4741 if (A == Op1) // ~A | A == -1
4742 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4746 // Note, A is still live here!
4747 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4749 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4751 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4752 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4753 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4754 I.getName()+".demorgan"), I);
4755 return BinaryOperator::CreateNot(And);
4759 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4760 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4761 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4764 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4765 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4769 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4770 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4771 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4772 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4773 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4774 !isa<ICmpInst>(Op1C->getOperand(0))) {
4775 const Type *SrcTy = Op0C->getOperand(0)->getType();
4776 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4777 // Only do this if the casts both really cause code to be
4779 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4781 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4783 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4784 Op1C->getOperand(0),
4786 InsertNewInstBefore(NewOp, I);
4787 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4794 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4795 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4796 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4797 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4798 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4799 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4800 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4801 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4802 // If either of the constants are nans, then the whole thing returns
4804 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4805 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4807 // Otherwise, no need to compare the two constants, compare the
4809 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4810 RHS->getOperand(0));
4813 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4814 FCmpInst::Predicate Op0CC, Op1CC;
4815 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4816 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4817 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4818 // Swap RHS operands to match LHS.
4819 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4820 std::swap(Op1LHS, Op1RHS);
4822 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4823 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4825 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4826 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4827 Op1CC == FCmpInst::FCMP_TRUE)
4828 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4829 else if (Op0CC == FCmpInst::FCMP_FALSE)
4830 return ReplaceInstUsesWith(I, Op1);
4831 else if (Op1CC == FCmpInst::FCMP_FALSE)
4832 return ReplaceInstUsesWith(I, Op0);
4835 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4836 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4837 if (Op0Ordered == Op1Ordered) {
4838 // If both are ordered or unordered, return a new fcmp with
4839 // or'ed predicates.
4840 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4842 if (Instruction *I = dyn_cast<Instruction>(RV))
4844 // Otherwise, it's a constant boolean value...
4845 return ReplaceInstUsesWith(I, RV);
4853 return Changed ? &I : 0;
4858 // XorSelf - Implements: X ^ X --> 0
4861 XorSelf(Value *rhs) : RHS(rhs) {}
4862 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4863 Instruction *apply(BinaryOperator &Xor) const {
4870 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4871 bool Changed = SimplifyCommutative(I);
4872 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4874 if (isa<UndefValue>(Op1)) {
4875 if (isa<UndefValue>(Op0))
4876 // Handle undef ^ undef -> 0 special case. This is a common
4878 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4879 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4882 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4883 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4884 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4885 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4888 // See if we can simplify any instructions used by the instruction whose sole
4889 // purpose is to compute bits we don't care about.
4890 if (!isa<VectorType>(I.getType())) {
4891 if (SimplifyDemandedInstructionBits(I))
4893 } else if (isa<ConstantAggregateZero>(Op1)) {
4894 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4897 // Is this a ~ operation?
4898 if (Value *NotOp = dyn_castNotVal(&I)) {
4899 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4900 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4901 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4902 if (Op0I->getOpcode() == Instruction::And ||
4903 Op0I->getOpcode() == Instruction::Or) {
4904 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4905 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4907 BinaryOperator::CreateNot(Op0I->getOperand(1),
4908 Op0I->getOperand(1)->getName()+".not");
4909 InsertNewInstBefore(NotY, I);
4910 if (Op0I->getOpcode() == Instruction::And)
4911 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4913 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4920 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4921 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4922 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4923 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4924 return new ICmpInst(ICI->getInversePredicate(),
4925 ICI->getOperand(0), ICI->getOperand(1));
4927 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4928 return new FCmpInst(FCI->getInversePredicate(),
4929 FCI->getOperand(0), FCI->getOperand(1));
4932 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4933 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4934 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4935 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4936 Instruction::CastOps Opcode = Op0C->getOpcode();
4937 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4938 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4939 Op0C->getDestTy())) {
4940 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4941 CI->getOpcode(), CI->getInversePredicate(),
4942 CI->getOperand(0), CI->getOperand(1)), I);
4943 NewCI->takeName(CI);
4944 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4951 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4952 // ~(c-X) == X-c-1 == X+(-c-1)
4953 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4954 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4955 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4956 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4957 ConstantInt::get(I.getType(), 1));
4958 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4961 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4962 if (Op0I->getOpcode() == Instruction::Add) {
4963 // ~(X-c) --> (-c-1)-X
4964 if (RHS->isAllOnesValue()) {
4965 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4966 return BinaryOperator::CreateSub(
4967 ConstantExpr::getSub(NegOp0CI,
4968 ConstantInt::get(I.getType(), 1)),
4969 Op0I->getOperand(0));
4970 } else if (RHS->getValue().isSignBit()) {
4971 // (X + C) ^ signbit -> (X + C + signbit)
4972 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4973 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4976 } else if (Op0I->getOpcode() == Instruction::Or) {
4977 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4978 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4979 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4980 // Anything in both C1 and C2 is known to be zero, remove it from
4982 Constant *CommonBits = And(Op0CI, RHS);
4983 NewRHS = ConstantExpr::getAnd(NewRHS,
4984 ConstantExpr::getNot(CommonBits));
4985 AddToWorkList(Op0I);
4986 I.setOperand(0, Op0I->getOperand(0));
4987 I.setOperand(1, NewRHS);
4994 // Try to fold constant and into select arguments.
4995 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4996 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4998 if (isa<PHINode>(Op0))
4999 if (Instruction *NV = FoldOpIntoPhi(I))
5003 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5005 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5007 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5009 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5012 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5015 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5016 if (A == Op0) { // B^(B|A) == (A|B)^B
5017 Op1I->swapOperands();
5019 std::swap(Op0, Op1);
5020 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5021 I.swapOperands(); // Simplified below.
5022 std::swap(Op0, Op1);
5024 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5025 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5026 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5027 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5028 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
5029 if (A == Op0) { // A^(A&B) -> A^(B&A)
5030 Op1I->swapOperands();
5033 if (B == Op0) { // A^(B&A) -> (B&A)^A
5034 I.swapOperands(); // Simplified below.
5035 std::swap(Op0, Op1);
5040 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5043 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
5044 if (A == Op1) // (B|A)^B == (A|B)^B
5046 if (B == Op1) { // (A|B)^B == A & ~B
5048 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5049 return BinaryOperator::CreateAnd(A, NotB);
5051 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5052 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5053 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5054 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5055 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5056 if (A == Op1) // (A&B)^A -> (B&A)^A
5058 if (B == Op1 && // (B&A)^A == ~B & A
5059 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5061 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5062 return BinaryOperator::CreateAnd(N, Op1);
5067 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5068 if (Op0I && Op1I && Op0I->isShift() &&
5069 Op0I->getOpcode() == Op1I->getOpcode() &&
5070 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5071 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5072 Instruction *NewOp =
5073 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5074 Op1I->getOperand(0),
5075 Op0I->getName()), I);
5076 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5077 Op1I->getOperand(1));
5081 Value *A, *B, *C, *D;
5082 // (A & B)^(A | B) -> A ^ B
5083 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5084 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5085 if ((A == C && B == D) || (A == D && B == C))
5086 return BinaryOperator::CreateXor(A, B);
5088 // (A | B)^(A & B) -> A ^ B
5089 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5090 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5091 if ((A == C && B == D) || (A == D && B == C))
5092 return BinaryOperator::CreateXor(A, B);
5096 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5097 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5098 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5099 // (X & Y)^(X & Y) -> (Y^Z) & X
5100 Value *X = 0, *Y = 0, *Z = 0;
5102 X = A, Y = B, Z = D;
5104 X = A, Y = B, Z = C;
5106 X = B, Y = A, Z = D;
5108 X = B, Y = A, Z = C;
5111 Instruction *NewOp =
5112 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5113 return BinaryOperator::CreateAnd(NewOp, X);
5118 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5119 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5120 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5123 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5124 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5125 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5126 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5127 const Type *SrcTy = Op0C->getOperand(0)->getType();
5128 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5129 // Only do this if the casts both really cause code to be generated.
5130 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5132 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5134 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5135 Op1C->getOperand(0),
5137 InsertNewInstBefore(NewOp, I);
5138 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5143 return Changed ? &I : 0;
5146 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5147 /// overflowed for this type.
5148 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5149 ConstantInt *In2, bool IsSigned = false) {
5150 Result = cast<ConstantInt>(Add(In1, In2));
5153 if (In2->getValue().isNegative())
5154 return Result->getValue().sgt(In1->getValue());
5156 return Result->getValue().slt(In1->getValue());
5158 return Result->getValue().ult(In1->getValue());
5161 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5162 /// overflowed for this type.
5163 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5164 ConstantInt *In2, bool IsSigned = false) {
5165 Result = cast<ConstantInt>(Subtract(In1, In2));
5168 if (In2->getValue().isNegative())
5169 return Result->getValue().slt(In1->getValue());
5171 return Result->getValue().sgt(In1->getValue());
5173 return Result->getValue().ugt(In1->getValue());
5176 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5177 /// code necessary to compute the offset from the base pointer (without adding
5178 /// in the base pointer). Return the result as a signed integer of intptr size.
5179 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5180 TargetData &TD = IC.getTargetData();
5181 gep_type_iterator GTI = gep_type_begin(GEP);
5182 const Type *IntPtrTy = TD.getIntPtrType();
5183 Value *Result = Constant::getNullValue(IntPtrTy);
5185 // Build a mask for high order bits.
5186 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5187 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5189 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5192 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType()) & PtrSizeMask;
5193 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5194 if (OpC->isZero()) continue;
5196 // Handle a struct index, which adds its field offset to the pointer.
5197 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5198 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5200 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5201 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5203 Result = IC.InsertNewInstBefore(
5204 BinaryOperator::CreateAdd(Result,
5205 ConstantInt::get(IntPtrTy, Size),
5206 GEP->getName()+".offs"), I);
5210 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5211 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5212 Scale = ConstantExpr::getMul(OC, Scale);
5213 if (Constant *RC = dyn_cast<Constant>(Result))
5214 Result = ConstantExpr::getAdd(RC, Scale);
5216 // Emit an add instruction.
5217 Result = IC.InsertNewInstBefore(
5218 BinaryOperator::CreateAdd(Result, Scale,
5219 GEP->getName()+".offs"), I);
5223 // Convert to correct type.
5224 if (Op->getType() != IntPtrTy) {
5225 if (Constant *OpC = dyn_cast<Constant>(Op))
5226 Op = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true);
5228 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5230 Op->getName()+".c"), I);
5233 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5234 if (Constant *OpC = dyn_cast<Constant>(Op))
5235 Op = ConstantExpr::getMul(OpC, Scale);
5236 else // We'll let instcombine(mul) convert this to a shl if possible.
5237 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5238 GEP->getName()+".idx"), I);
5241 // Emit an add instruction.
5242 if (isa<Constant>(Op) && isa<Constant>(Result))
5243 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5244 cast<Constant>(Result));
5246 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5247 GEP->getName()+".offs"), I);
5253 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5254 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5255 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5256 /// complex, and scales are involved. The above expression would also be legal
5257 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5258 /// later form is less amenable to optimization though, and we are allowed to
5259 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5261 /// If we can't emit an optimized form for this expression, this returns null.
5263 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5265 TargetData &TD = IC.getTargetData();
5266 gep_type_iterator GTI = gep_type_begin(GEP);
5268 // Check to see if this gep only has a single variable index. If so, and if
5269 // any constant indices are a multiple of its scale, then we can compute this
5270 // in terms of the scale of the variable index. For example, if the GEP
5271 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5272 // because the expression will cross zero at the same point.
5273 unsigned i, e = GEP->getNumOperands();
5275 for (i = 1; i != e; ++i, ++GTI) {
5276 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5277 // Compute the aggregate offset of constant indices.
5278 if (CI->isZero()) continue;
5280 // Handle a struct index, which adds its field offset to the pointer.
5281 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5282 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5284 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5285 Offset += Size*CI->getSExtValue();
5288 // Found our variable index.
5293 // If there are no variable indices, we must have a constant offset, just
5294 // evaluate it the general way.
5295 if (i == e) return 0;
5297 Value *VariableIdx = GEP->getOperand(i);
5298 // Determine the scale factor of the variable element. For example, this is
5299 // 4 if the variable index is into an array of i32.
5300 uint64_t VariableScale = TD.getTypePaddedSize(GTI.getIndexedType());
5302 // Verify that there are no other variable indices. If so, emit the hard way.
5303 for (++i, ++GTI; i != e; ++i, ++GTI) {
5304 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5307 // Compute the aggregate offset of constant indices.
5308 if (CI->isZero()) continue;
5310 // Handle a struct index, which adds its field offset to the pointer.
5311 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5312 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5314 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5315 Offset += Size*CI->getSExtValue();
5319 // Okay, we know we have a single variable index, which must be a
5320 // pointer/array/vector index. If there is no offset, life is simple, return
5322 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5324 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5325 // we don't need to bother extending: the extension won't affect where the
5326 // computation crosses zero.
5327 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5328 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5329 VariableIdx->getNameStart(), &I);
5333 // Otherwise, there is an index. The computation we will do will be modulo
5334 // the pointer size, so get it.
5335 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5337 Offset &= PtrSizeMask;
5338 VariableScale &= PtrSizeMask;
5340 // To do this transformation, any constant index must be a multiple of the
5341 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5342 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5343 // multiple of the variable scale.
5344 int64_t NewOffs = Offset / (int64_t)VariableScale;
5345 if (Offset != NewOffs*(int64_t)VariableScale)
5348 // Okay, we can do this evaluation. Start by converting the index to intptr.
5349 const Type *IntPtrTy = TD.getIntPtrType();
5350 if (VariableIdx->getType() != IntPtrTy)
5351 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5353 VariableIdx->getNameStart(), &I);
5354 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5355 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5359 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5360 /// else. At this point we know that the GEP is on the LHS of the comparison.
5361 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5362 ICmpInst::Predicate Cond,
5364 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5366 // Look through bitcasts.
5367 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5368 RHS = BCI->getOperand(0);
5370 Value *PtrBase = GEPLHS->getOperand(0);
5371 if (PtrBase == RHS) {
5372 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5373 // This transformation (ignoring the base and scales) is valid because we
5374 // know pointers can't overflow. See if we can output an optimized form.
5375 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5377 // If not, synthesize the offset the hard way.
5379 Offset = EmitGEPOffset(GEPLHS, I, *this);
5380 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5381 Constant::getNullValue(Offset->getType()));
5382 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5383 // If the base pointers are different, but the indices are the same, just
5384 // compare the base pointer.
5385 if (PtrBase != GEPRHS->getOperand(0)) {
5386 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5387 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5388 GEPRHS->getOperand(0)->getType();
5390 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5391 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5392 IndicesTheSame = false;
5396 // If all indices are the same, just compare the base pointers.
5398 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5399 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5401 // Otherwise, the base pointers are different and the indices are
5402 // different, bail out.
5406 // If one of the GEPs has all zero indices, recurse.
5407 bool AllZeros = true;
5408 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5409 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5410 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5415 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5416 ICmpInst::getSwappedPredicate(Cond), I);
5418 // If the other GEP has all zero indices, recurse.
5420 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5421 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5422 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5427 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5429 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5430 // If the GEPs only differ by one index, compare it.
5431 unsigned NumDifferences = 0; // Keep track of # differences.
5432 unsigned DiffOperand = 0; // The operand that differs.
5433 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5434 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5435 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5436 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5437 // Irreconcilable differences.
5441 if (NumDifferences++) break;
5446 if (NumDifferences == 0) // SAME GEP?
5447 return ReplaceInstUsesWith(I, // No comparison is needed here.
5448 ConstantInt::get(Type::Int1Ty,
5449 ICmpInst::isTrueWhenEqual(Cond)));
5451 else if (NumDifferences == 1) {
5452 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5453 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5454 // Make sure we do a signed comparison here.
5455 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5459 // Only lower this if the icmp is the only user of the GEP or if we expect
5460 // the result to fold to a constant!
5461 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5462 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5463 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5464 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5465 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5466 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5472 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5474 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5477 if (!isa<ConstantFP>(RHSC)) return 0;
5478 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5480 // Get the width of the mantissa. We don't want to hack on conversions that
5481 // might lose information from the integer, e.g. "i64 -> float"
5482 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5483 if (MantissaWidth == -1) return 0; // Unknown.
5485 // Check to see that the input is converted from an integer type that is small
5486 // enough that preserves all bits. TODO: check here for "known" sign bits.
5487 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5488 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5490 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5491 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5495 // If the conversion would lose info, don't hack on this.
5496 if ((int)InputSize > MantissaWidth)
5499 // Otherwise, we can potentially simplify the comparison. We know that it
5500 // will always come through as an integer value and we know the constant is
5501 // not a NAN (it would have been previously simplified).
5502 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5504 ICmpInst::Predicate Pred;
5505 switch (I.getPredicate()) {
5506 default: assert(0 && "Unexpected predicate!");
5507 case FCmpInst::FCMP_UEQ:
5508 case FCmpInst::FCMP_OEQ:
5509 Pred = ICmpInst::ICMP_EQ;
5511 case FCmpInst::FCMP_UGT:
5512 case FCmpInst::FCMP_OGT:
5513 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5515 case FCmpInst::FCMP_UGE:
5516 case FCmpInst::FCMP_OGE:
5517 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5519 case FCmpInst::FCMP_ULT:
5520 case FCmpInst::FCMP_OLT:
5521 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5523 case FCmpInst::FCMP_ULE:
5524 case FCmpInst::FCMP_OLE:
5525 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5527 case FCmpInst::FCMP_UNE:
5528 case FCmpInst::FCMP_ONE:
5529 Pred = ICmpInst::ICMP_NE;
5531 case FCmpInst::FCMP_ORD:
5532 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5533 case FCmpInst::FCMP_UNO:
5534 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5537 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5539 // Now we know that the APFloat is a normal number, zero or inf.
5541 // See if the FP constant is too large for the integer. For example,
5542 // comparing an i8 to 300.0.
5543 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5546 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5547 // and large values.
5548 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5549 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5550 APFloat::rmNearestTiesToEven);
5551 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5552 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5553 Pred == ICmpInst::ICMP_SLE)
5554 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5555 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5558 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5559 // +INF and large values.
5560 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5561 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5562 APFloat::rmNearestTiesToEven);
5563 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5564 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5565 Pred == ICmpInst::ICMP_ULE)
5566 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5567 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5572 // See if the RHS value is < SignedMin.
5573 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5574 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5575 APFloat::rmNearestTiesToEven);
5576 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5577 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5578 Pred == ICmpInst::ICMP_SGE)
5579 return ReplaceInstUsesWith(I,ConstantInt::getTrue());
5580 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5584 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5585 // [0, UMAX], but it may still be fractional. See if it is fractional by
5586 // casting the FP value to the integer value and back, checking for equality.
5587 // Don't do this for zero, because -0.0 is not fractional.
5588 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5589 if (!RHS.isZero() &&
5590 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5591 // If we had a comparison against a fractional value, we have to adjust the
5592 // compare predicate and sometimes the value. RHSC is rounded towards zero
5595 default: assert(0 && "Unexpected integer comparison!");
5596 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5597 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5598 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5599 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5600 case ICmpInst::ICMP_ULE:
5601 // (float)int <= 4.4 --> int <= 4
5602 // (float)int <= -4.4 --> false
5603 if (RHS.isNegative())
5604 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5606 case ICmpInst::ICMP_SLE:
5607 // (float)int <= 4.4 --> int <= 4
5608 // (float)int <= -4.4 --> int < -4
5609 if (RHS.isNegative())
5610 Pred = ICmpInst::ICMP_SLT;
5612 case ICmpInst::ICMP_ULT:
5613 // (float)int < -4.4 --> false
5614 // (float)int < 4.4 --> int <= 4
5615 if (RHS.isNegative())
5616 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5617 Pred = ICmpInst::ICMP_ULE;
5619 case ICmpInst::ICMP_SLT:
5620 // (float)int < -4.4 --> int < -4
5621 // (float)int < 4.4 --> int <= 4
5622 if (!RHS.isNegative())
5623 Pred = ICmpInst::ICMP_SLE;
5625 case ICmpInst::ICMP_UGT:
5626 // (float)int > 4.4 --> int > 4
5627 // (float)int > -4.4 --> true
5628 if (RHS.isNegative())
5629 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5631 case ICmpInst::ICMP_SGT:
5632 // (float)int > 4.4 --> int > 4
5633 // (float)int > -4.4 --> int >= -4
5634 if (RHS.isNegative())
5635 Pred = ICmpInst::ICMP_SGE;
5637 case ICmpInst::ICMP_UGE:
5638 // (float)int >= -4.4 --> true
5639 // (float)int >= 4.4 --> int > 4
5640 if (!RHS.isNegative())
5641 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5642 Pred = ICmpInst::ICMP_UGT;
5644 case ICmpInst::ICMP_SGE:
5645 // (float)int >= -4.4 --> int >= -4
5646 // (float)int >= 4.4 --> int > 4
5647 if (!RHS.isNegative())
5648 Pred = ICmpInst::ICMP_SGT;
5653 // Lower this FP comparison into an appropriate integer version of the
5655 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5658 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5659 bool Changed = SimplifyCompare(I);
5660 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5662 // Fold trivial predicates.
5663 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5664 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5665 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5666 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5668 // Simplify 'fcmp pred X, X'
5670 switch (I.getPredicate()) {
5671 default: assert(0 && "Unknown predicate!");
5672 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5673 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5674 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5675 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5676 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5677 case FCmpInst::FCMP_OLT: // True if ordered and less than
5678 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5679 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5681 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5682 case FCmpInst::FCMP_ULT: // True if unordered or less than
5683 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5684 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5685 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5686 I.setPredicate(FCmpInst::FCMP_UNO);
5687 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5690 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5691 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5692 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5693 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5694 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5695 I.setPredicate(FCmpInst::FCMP_ORD);
5696 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5701 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5702 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5704 // Handle fcmp with constant RHS
5705 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5706 // If the constant is a nan, see if we can fold the comparison based on it.
5707 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5708 if (CFP->getValueAPF().isNaN()) {
5709 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5710 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5711 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5712 "Comparison must be either ordered or unordered!");
5713 // True if unordered.
5714 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5718 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5719 switch (LHSI->getOpcode()) {
5720 case Instruction::PHI:
5721 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5722 // block. If in the same block, we're encouraging jump threading. If
5723 // not, we are just pessimizing the code by making an i1 phi.
5724 if (LHSI->getParent() == I.getParent())
5725 if (Instruction *NV = FoldOpIntoPhi(I))
5728 case Instruction::SIToFP:
5729 case Instruction::UIToFP:
5730 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5733 case Instruction::Select:
5734 // If either operand of the select is a constant, we can fold the
5735 // comparison into the select arms, which will cause one to be
5736 // constant folded and the select turned into a bitwise or.
5737 Value *Op1 = 0, *Op2 = 0;
5738 if (LHSI->hasOneUse()) {
5739 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5740 // Fold the known value into the constant operand.
5741 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5742 // Insert a new FCmp of the other select operand.
5743 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5744 LHSI->getOperand(2), RHSC,
5746 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5747 // Fold the known value into the constant operand.
5748 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5749 // Insert a new FCmp of the other select operand.
5750 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5751 LHSI->getOperand(1), RHSC,
5757 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5762 return Changed ? &I : 0;
5765 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5766 bool Changed = SimplifyCompare(I);
5767 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5768 const Type *Ty = Op0->getType();
5772 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5773 I.isTrueWhenEqual()));
5775 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5776 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5778 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5779 // addresses never equal each other! We already know that Op0 != Op1.
5780 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5781 isa<ConstantPointerNull>(Op0)) &&
5782 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5783 isa<ConstantPointerNull>(Op1)))
5784 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5785 !I.isTrueWhenEqual()));
5787 // icmp's with boolean values can always be turned into bitwise operations
5788 if (Ty == Type::Int1Ty) {
5789 switch (I.getPredicate()) {
5790 default: assert(0 && "Invalid icmp instruction!");
5791 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5792 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5793 InsertNewInstBefore(Xor, I);
5794 return BinaryOperator::CreateNot(Xor);
5796 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5797 return BinaryOperator::CreateXor(Op0, Op1);
5799 case ICmpInst::ICMP_UGT:
5800 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5802 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5803 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5804 InsertNewInstBefore(Not, I);
5805 return BinaryOperator::CreateAnd(Not, Op1);
5807 case ICmpInst::ICMP_SGT:
5808 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5810 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5811 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5812 InsertNewInstBefore(Not, I);
5813 return BinaryOperator::CreateAnd(Not, Op0);
5815 case ICmpInst::ICMP_UGE:
5816 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5818 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5819 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5820 InsertNewInstBefore(Not, I);
5821 return BinaryOperator::CreateOr(Not, Op1);
5823 case ICmpInst::ICMP_SGE:
5824 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5826 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5827 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5828 InsertNewInstBefore(Not, I);
5829 return BinaryOperator::CreateOr(Not, Op0);
5834 // See if we are doing a comparison with a constant.
5835 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5836 Value *A = 0, *B = 0;
5838 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5839 if (I.isEquality() && CI->isNullValue() &&
5840 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5841 // (icmp cond A B) if cond is equality
5842 return new ICmpInst(I.getPredicate(), A, B);
5845 // If we have an icmp le or icmp ge instruction, turn it into the
5846 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5847 // them being folded in the code below.
5848 switch (I.getPredicate()) {
5850 case ICmpInst::ICMP_ULE:
5851 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5852 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5853 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5854 case ICmpInst::ICMP_SLE:
5855 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5856 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5857 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5858 case ICmpInst::ICMP_UGE:
5859 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5860 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5861 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5862 case ICmpInst::ICMP_SGE:
5863 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5864 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5865 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5868 // See if we can fold the comparison based on range information we can get
5869 // by checking whether bits are known to be zero or one in the input.
5870 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5871 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5873 // If this comparison is a normal comparison, it demands all
5874 // bits, if it is a sign bit comparison, it only demands the sign bit.
5876 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5878 if (SimplifyDemandedBits(I.getOperandUse(0),
5879 isSignBit ? APInt::getSignBit(BitWidth)
5880 : APInt::getAllOnesValue(BitWidth),
5881 KnownZero, KnownOne, 0))
5884 // Given the known and unknown bits, compute a range that the LHS could be
5885 // in. Compute the Min, Max and RHS values based on the known bits. For the
5886 // EQ and NE we use unsigned values.
5887 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5888 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5889 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5891 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5893 // If Min and Max are known to be the same, then SimplifyDemandedBits
5894 // figured out that the LHS is a constant. Just constant fold this now so
5895 // that code below can assume that Min != Max.
5897 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5898 ConstantInt::get(Min),
5901 // Based on the range information we know about the LHS, see if we can
5902 // simplify this comparison. For example, (x&4) < 8 is always true.
5903 const APInt &RHSVal = CI->getValue();
5904 switch (I.getPredicate()) { // LE/GE have been folded already.
5905 default: assert(0 && "Unknown icmp opcode!");
5906 case ICmpInst::ICMP_EQ:
5907 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5908 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5910 case ICmpInst::ICMP_NE:
5911 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5912 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5914 case ICmpInst::ICMP_ULT:
5915 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5916 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5917 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5918 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5919 if (RHSVal == Max) // A <u MAX -> A != MAX
5920 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5921 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5922 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5924 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5925 if (CI->isMinValue(true))
5926 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5927 ConstantInt::getAllOnesValue(Op0->getType()));
5929 case ICmpInst::ICMP_UGT:
5930 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5931 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5932 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5933 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5935 if (RHSVal == Min) // A >u MIN -> A != MIN
5936 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5937 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5938 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5940 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5941 if (CI->isMaxValue(true))
5942 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5943 ConstantInt::getNullValue(Op0->getType()));
5945 case ICmpInst::ICMP_SLT:
5946 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5947 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5948 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5949 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5950 if (RHSVal == Max) // A <s MAX -> A != MAX
5951 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5952 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5953 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5955 case ICmpInst::ICMP_SGT:
5956 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5957 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5958 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5959 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5961 if (RHSVal == Min) // A >s MIN -> A != MIN
5962 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5963 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5964 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5969 // Test if the ICmpInst instruction is used exclusively by a select as
5970 // part of a minimum or maximum operation. If so, refrain from doing
5971 // any other folding. This helps out other analyses which understand
5972 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5973 // and CodeGen. And in this case, at least one of the comparison
5974 // operands has at least one user besides the compare (the select),
5975 // which would often largely negate the benefit of folding anyway.
5977 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5978 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5979 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5982 // See if we are doing a comparison between a constant and an instruction that
5983 // can be folded into the comparison.
5984 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5985 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5986 // instruction, see if that instruction also has constants so that the
5987 // instruction can be folded into the icmp
5988 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5989 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5993 // Handle icmp with constant (but not simple integer constant) RHS
5994 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5995 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5996 switch (LHSI->getOpcode()) {
5997 case Instruction::GetElementPtr:
5998 if (RHSC->isNullValue()) {
5999 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6000 bool isAllZeros = true;
6001 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6002 if (!isa<Constant>(LHSI->getOperand(i)) ||
6003 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6008 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6009 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6013 case Instruction::PHI:
6014 // Only fold icmp into the PHI if the phi and fcmp are in the same
6015 // block. If in the same block, we're encouraging jump threading. If
6016 // not, we are just pessimizing the code by making an i1 phi.
6017 if (LHSI->getParent() == I.getParent())
6018 if (Instruction *NV = FoldOpIntoPhi(I))
6021 case Instruction::Select: {
6022 // If either operand of the select is a constant, we can fold the
6023 // comparison into the select arms, which will cause one to be
6024 // constant folded and the select turned into a bitwise or.
6025 Value *Op1 = 0, *Op2 = 0;
6026 if (LHSI->hasOneUse()) {
6027 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6028 // Fold the known value into the constant operand.
6029 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6030 // Insert a new ICmp of the other select operand.
6031 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6032 LHSI->getOperand(2), RHSC,
6034 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6035 // Fold the known value into the constant operand.
6036 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6037 // Insert a new ICmp of the other select operand.
6038 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6039 LHSI->getOperand(1), RHSC,
6045 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6048 case Instruction::Malloc:
6049 // If we have (malloc != null), and if the malloc has a single use, we
6050 // can assume it is successful and remove the malloc.
6051 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6052 AddToWorkList(LHSI);
6053 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6054 !I.isTrueWhenEqual()));
6060 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6061 if (User *GEP = dyn_castGetElementPtr(Op0))
6062 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6064 if (User *GEP = dyn_castGetElementPtr(Op1))
6065 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6066 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6069 // Test to see if the operands of the icmp are casted versions of other
6070 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6072 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6073 if (isa<PointerType>(Op0->getType()) &&
6074 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6075 // We keep moving the cast from the left operand over to the right
6076 // operand, where it can often be eliminated completely.
6077 Op0 = CI->getOperand(0);
6079 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6080 // so eliminate it as well.
6081 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6082 Op1 = CI2->getOperand(0);
6084 // If Op1 is a constant, we can fold the cast into the constant.
6085 if (Op0->getType() != Op1->getType()) {
6086 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6087 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6089 // Otherwise, cast the RHS right before the icmp
6090 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6093 return new ICmpInst(I.getPredicate(), Op0, Op1);
6097 if (isa<CastInst>(Op0)) {
6098 // Handle the special case of: icmp (cast bool to X), <cst>
6099 // This comes up when you have code like
6102 // For generality, we handle any zero-extension of any operand comparison
6103 // with a constant or another cast from the same type.
6104 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6105 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6109 // See if it's the same type of instruction on the left and right.
6110 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6111 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6112 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6113 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6114 switch (Op0I->getOpcode()) {
6116 case Instruction::Add:
6117 case Instruction::Sub:
6118 case Instruction::Xor:
6119 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6120 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6121 Op1I->getOperand(0));
6122 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6123 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6124 if (CI->getValue().isSignBit()) {
6125 ICmpInst::Predicate Pred = I.isSignedPredicate()
6126 ? I.getUnsignedPredicate()
6127 : I.getSignedPredicate();
6128 return new ICmpInst(Pred, Op0I->getOperand(0),
6129 Op1I->getOperand(0));
6132 if (CI->getValue().isMaxSignedValue()) {
6133 ICmpInst::Predicate Pred = I.isSignedPredicate()
6134 ? I.getUnsignedPredicate()
6135 : I.getSignedPredicate();
6136 Pred = I.getSwappedPredicate(Pred);
6137 return new ICmpInst(Pred, Op0I->getOperand(0),
6138 Op1I->getOperand(0));
6142 case Instruction::Mul:
6143 if (!I.isEquality())
6146 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6147 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6148 // Mask = -1 >> count-trailing-zeros(Cst).
6149 if (!CI->isZero() && !CI->isOne()) {
6150 const APInt &AP = CI->getValue();
6151 ConstantInt *Mask = ConstantInt::get(
6152 APInt::getLowBitsSet(AP.getBitWidth(),
6154 AP.countTrailingZeros()));
6155 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6157 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6159 InsertNewInstBefore(And1, I);
6160 InsertNewInstBefore(And2, I);
6161 return new ICmpInst(I.getPredicate(), And1, And2);
6170 // ~x < ~y --> y < x
6172 if (match(Op0, m_Not(m_Value(A))) &&
6173 match(Op1, m_Not(m_Value(B))))
6174 return new ICmpInst(I.getPredicate(), B, A);
6177 if (I.isEquality()) {
6178 Value *A, *B, *C, *D;
6180 // -x == -y --> x == y
6181 if (match(Op0, m_Neg(m_Value(A))) &&
6182 match(Op1, m_Neg(m_Value(B))))
6183 return new ICmpInst(I.getPredicate(), A, B);
6185 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6186 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6187 Value *OtherVal = A == Op1 ? B : A;
6188 return new ICmpInst(I.getPredicate(), OtherVal,
6189 Constant::getNullValue(A->getType()));
6192 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6193 // A^c1 == C^c2 --> A == C^(c1^c2)
6194 ConstantInt *C1, *C2;
6195 if (match(B, m_ConstantInt(C1)) &&
6196 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6197 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6198 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6199 return new ICmpInst(I.getPredicate(), A,
6200 InsertNewInstBefore(Xor, I));
6203 // A^B == A^D -> B == D
6204 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6205 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6206 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6207 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6211 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6212 (A == Op0 || B == Op0)) {
6213 // A == (A^B) -> B == 0
6214 Value *OtherVal = A == Op0 ? B : A;
6215 return new ICmpInst(I.getPredicate(), OtherVal,
6216 Constant::getNullValue(A->getType()));
6219 // (A-B) == A -> B == 0
6220 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6221 return new ICmpInst(I.getPredicate(), B,
6222 Constant::getNullValue(B->getType()));
6224 // A == (A-B) -> B == 0
6225 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6226 return new ICmpInst(I.getPredicate(), B,
6227 Constant::getNullValue(B->getType()));
6229 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6230 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6231 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6232 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6233 Value *X = 0, *Y = 0, *Z = 0;
6236 X = B; Y = D; Z = A;
6237 } else if (A == D) {
6238 X = B; Y = C; Z = A;
6239 } else if (B == C) {
6240 X = A; Y = D; Z = B;
6241 } else if (B == D) {
6242 X = A; Y = C; Z = B;
6245 if (X) { // Build (X^Y) & Z
6246 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6247 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6248 I.setOperand(0, Op1);
6249 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6254 return Changed ? &I : 0;
6258 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6259 /// and CmpRHS are both known to be integer constants.
6260 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6261 ConstantInt *DivRHS) {
6262 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6263 const APInt &CmpRHSV = CmpRHS->getValue();
6265 // FIXME: If the operand types don't match the type of the divide
6266 // then don't attempt this transform. The code below doesn't have the
6267 // logic to deal with a signed divide and an unsigned compare (and
6268 // vice versa). This is because (x /s C1) <s C2 produces different
6269 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6270 // (x /u C1) <u C2. Simply casting the operands and result won't
6271 // work. :( The if statement below tests that condition and bails
6273 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6274 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6276 if (DivRHS->isZero())
6277 return 0; // The ProdOV computation fails on divide by zero.
6278 if (DivIsSigned && DivRHS->isAllOnesValue())
6279 return 0; // The overflow computation also screws up here
6280 if (DivRHS->isOne())
6281 return 0; // Not worth bothering, and eliminates some funny cases
6284 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6285 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6286 // C2 (CI). By solving for X we can turn this into a range check
6287 // instead of computing a divide.
6288 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6290 // Determine if the product overflows by seeing if the product is
6291 // not equal to the divide. Make sure we do the same kind of divide
6292 // as in the LHS instruction that we're folding.
6293 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6294 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6296 // Get the ICmp opcode
6297 ICmpInst::Predicate Pred = ICI.getPredicate();
6299 // Figure out the interval that is being checked. For example, a comparison
6300 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6301 // Compute this interval based on the constants involved and the signedness of
6302 // the compare/divide. This computes a half-open interval, keeping track of
6303 // whether either value in the interval overflows. After analysis each
6304 // overflow variable is set to 0 if it's corresponding bound variable is valid
6305 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6306 int LoOverflow = 0, HiOverflow = 0;
6307 ConstantInt *LoBound = 0, *HiBound = 0;
6309 if (!DivIsSigned) { // udiv
6310 // e.g. X/5 op 3 --> [15, 20)
6312 HiOverflow = LoOverflow = ProdOV;
6314 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6315 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6316 if (CmpRHSV == 0) { // (X / pos) op 0
6317 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6318 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6320 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6321 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6322 HiOverflow = LoOverflow = ProdOV;
6324 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6325 } else { // (X / pos) op neg
6326 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6327 HiBound = AddOne(Prod);
6328 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6330 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6331 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6335 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6336 if (CmpRHSV == 0) { // (X / neg) op 0
6337 // e.g. X/-5 op 0 --> [-4, 5)
6338 LoBound = AddOne(DivRHS);
6339 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6340 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6341 HiOverflow = 1; // [INTMIN+1, overflow)
6342 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6344 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6345 // e.g. X/-5 op 3 --> [-19, -14)
6346 HiBound = AddOne(Prod);
6347 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6349 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6350 } else { // (X / neg) op neg
6351 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6352 LoOverflow = HiOverflow = ProdOV;
6354 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6357 // Dividing by a negative swaps the condition. LT <-> GT
6358 Pred = ICmpInst::getSwappedPredicate(Pred);
6361 Value *X = DivI->getOperand(0);
6363 default: assert(0 && "Unhandled icmp opcode!");
6364 case ICmpInst::ICMP_EQ:
6365 if (LoOverflow && HiOverflow)
6366 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6367 else if (HiOverflow)
6368 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6369 ICmpInst::ICMP_UGE, X, LoBound);
6370 else if (LoOverflow)
6371 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6372 ICmpInst::ICMP_ULT, X, HiBound);
6374 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6375 case ICmpInst::ICMP_NE:
6376 if (LoOverflow && HiOverflow)
6377 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6378 else if (HiOverflow)
6379 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6380 ICmpInst::ICMP_ULT, X, LoBound);
6381 else if (LoOverflow)
6382 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6383 ICmpInst::ICMP_UGE, X, HiBound);
6385 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6386 case ICmpInst::ICMP_ULT:
6387 case ICmpInst::ICMP_SLT:
6388 if (LoOverflow == +1) // Low bound is greater than input range.
6389 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6390 if (LoOverflow == -1) // Low bound is less than input range.
6391 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6392 return new ICmpInst(Pred, X, LoBound);
6393 case ICmpInst::ICMP_UGT:
6394 case ICmpInst::ICMP_SGT:
6395 if (HiOverflow == +1) // High bound greater than input range.
6396 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6397 else if (HiOverflow == -1) // High bound less than input range.
6398 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6399 if (Pred == ICmpInst::ICMP_UGT)
6400 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6402 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6407 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6409 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6412 const APInt &RHSV = RHS->getValue();
6414 switch (LHSI->getOpcode()) {
6415 case Instruction::Trunc:
6416 if (ICI.isEquality() && LHSI->hasOneUse()) {
6417 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6418 // of the high bits truncated out of x are known.
6419 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6420 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6421 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6422 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6423 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6425 // If all the high bits are known, we can do this xform.
6426 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6427 // Pull in the high bits from known-ones set.
6428 APInt NewRHS(RHS->getValue());
6429 NewRHS.zext(SrcBits);
6431 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6432 ConstantInt::get(NewRHS));
6437 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6438 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6439 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6441 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6442 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6443 Value *CompareVal = LHSI->getOperand(0);
6445 // If the sign bit of the XorCST is not set, there is no change to
6446 // the operation, just stop using the Xor.
6447 if (!XorCST->getValue().isNegative()) {
6448 ICI.setOperand(0, CompareVal);
6449 AddToWorkList(LHSI);
6453 // Was the old condition true if the operand is positive?
6454 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6456 // If so, the new one isn't.
6457 isTrueIfPositive ^= true;
6459 if (isTrueIfPositive)
6460 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6462 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6465 if (LHSI->hasOneUse()) {
6466 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6467 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6468 const APInt &SignBit = XorCST->getValue();
6469 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6470 ? ICI.getUnsignedPredicate()
6471 : ICI.getSignedPredicate();
6472 return new ICmpInst(Pred, LHSI->getOperand(0),
6473 ConstantInt::get(RHSV ^ SignBit));
6476 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6477 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6478 const APInt &NotSignBit = XorCST->getValue();
6479 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6480 ? ICI.getUnsignedPredicate()
6481 : ICI.getSignedPredicate();
6482 Pred = ICI.getSwappedPredicate(Pred);
6483 return new ICmpInst(Pred, LHSI->getOperand(0),
6484 ConstantInt::get(RHSV ^ NotSignBit));
6489 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6490 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6491 LHSI->getOperand(0)->hasOneUse()) {
6492 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6494 // If the LHS is an AND of a truncating cast, we can widen the
6495 // and/compare to be the input width without changing the value
6496 // produced, eliminating a cast.
6497 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6498 // We can do this transformation if either the AND constant does not
6499 // have its sign bit set or if it is an equality comparison.
6500 // Extending a relational comparison when we're checking the sign
6501 // bit would not work.
6502 if (Cast->hasOneUse() &&
6503 (ICI.isEquality() ||
6504 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6506 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6507 APInt NewCST = AndCST->getValue();
6508 NewCST.zext(BitWidth);
6510 NewCI.zext(BitWidth);
6511 Instruction *NewAnd =
6512 BinaryOperator::CreateAnd(Cast->getOperand(0),
6513 ConstantInt::get(NewCST),LHSI->getName());
6514 InsertNewInstBefore(NewAnd, ICI);
6515 return new ICmpInst(ICI.getPredicate(), NewAnd,
6516 ConstantInt::get(NewCI));
6520 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6521 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6522 // happens a LOT in code produced by the C front-end, for bitfield
6524 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6525 if (Shift && !Shift->isShift())
6529 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6530 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6531 const Type *AndTy = AndCST->getType(); // Type of the and.
6533 // We can fold this as long as we can't shift unknown bits
6534 // into the mask. This can only happen with signed shift
6535 // rights, as they sign-extend.
6537 bool CanFold = Shift->isLogicalShift();
6539 // To test for the bad case of the signed shr, see if any
6540 // of the bits shifted in could be tested after the mask.
6541 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6542 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6544 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6545 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6546 AndCST->getValue()) == 0)
6552 if (Shift->getOpcode() == Instruction::Shl)
6553 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6555 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6557 // Check to see if we are shifting out any of the bits being
6559 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6560 // If we shifted bits out, the fold is not going to work out.
6561 // As a special case, check to see if this means that the
6562 // result is always true or false now.
6563 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6564 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6565 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6566 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6568 ICI.setOperand(1, NewCst);
6569 Constant *NewAndCST;
6570 if (Shift->getOpcode() == Instruction::Shl)
6571 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6573 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6574 LHSI->setOperand(1, NewAndCST);
6575 LHSI->setOperand(0, Shift->getOperand(0));
6576 AddToWorkList(Shift); // Shift is dead.
6577 AddUsesToWorkList(ICI);
6583 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6584 // preferable because it allows the C<<Y expression to be hoisted out
6585 // of a loop if Y is invariant and X is not.
6586 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6587 ICI.isEquality() && !Shift->isArithmeticShift() &&
6588 !isa<Constant>(Shift->getOperand(0))) {
6591 if (Shift->getOpcode() == Instruction::LShr) {
6592 NS = BinaryOperator::CreateShl(AndCST,
6593 Shift->getOperand(1), "tmp");
6595 // Insert a logical shift.
6596 NS = BinaryOperator::CreateLShr(AndCST,
6597 Shift->getOperand(1), "tmp");
6599 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6601 // Compute X & (C << Y).
6602 Instruction *NewAnd =
6603 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6604 InsertNewInstBefore(NewAnd, ICI);
6606 ICI.setOperand(0, NewAnd);
6612 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6613 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6616 uint32_t TypeBits = RHSV.getBitWidth();
6618 // Check that the shift amount is in range. If not, don't perform
6619 // undefined shifts. When the shift is visited it will be
6621 if (ShAmt->uge(TypeBits))
6624 if (ICI.isEquality()) {
6625 // If we are comparing against bits always shifted out, the
6626 // comparison cannot succeed.
6628 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6629 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6630 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6631 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6632 return ReplaceInstUsesWith(ICI, Cst);
6635 if (LHSI->hasOneUse()) {
6636 // Otherwise strength reduce the shift into an and.
6637 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6639 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6642 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6643 Mask, LHSI->getName()+".mask");
6644 Value *And = InsertNewInstBefore(AndI, ICI);
6645 return new ICmpInst(ICI.getPredicate(), And,
6646 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6650 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6651 bool TrueIfSigned = false;
6652 if (LHSI->hasOneUse() &&
6653 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6654 // (X << 31) <s 0 --> (X&1) != 0
6655 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6656 (TypeBits-ShAmt->getZExtValue()-1));
6658 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6659 Mask, LHSI->getName()+".mask");
6660 Value *And = InsertNewInstBefore(AndI, ICI);
6662 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6663 And, Constant::getNullValue(And->getType()));
6668 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6669 case Instruction::AShr: {
6670 // Only handle equality comparisons of shift-by-constant.
6671 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6672 if (!ShAmt || !ICI.isEquality()) break;
6674 // Check that the shift amount is in range. If not, don't perform
6675 // undefined shifts. When the shift is visited it will be
6677 uint32_t TypeBits = RHSV.getBitWidth();
6678 if (ShAmt->uge(TypeBits))
6681 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6683 // If we are comparing against bits always shifted out, the
6684 // comparison cannot succeed.
6685 APInt Comp = RHSV << ShAmtVal;
6686 if (LHSI->getOpcode() == Instruction::LShr)
6687 Comp = Comp.lshr(ShAmtVal);
6689 Comp = Comp.ashr(ShAmtVal);
6691 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6692 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6693 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6694 return ReplaceInstUsesWith(ICI, Cst);
6697 // Otherwise, check to see if the bits shifted out are known to be zero.
6698 // If so, we can compare against the unshifted value:
6699 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6700 if (LHSI->hasOneUse() &&
6701 MaskedValueIsZero(LHSI->getOperand(0),
6702 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6703 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6704 ConstantExpr::getShl(RHS, ShAmt));
6707 if (LHSI->hasOneUse()) {
6708 // Otherwise strength reduce the shift into an and.
6709 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6710 Constant *Mask = ConstantInt::get(Val);
6713 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6714 Mask, LHSI->getName()+".mask");
6715 Value *And = InsertNewInstBefore(AndI, ICI);
6716 return new ICmpInst(ICI.getPredicate(), And,
6717 ConstantExpr::getShl(RHS, ShAmt));
6722 case Instruction::SDiv:
6723 case Instruction::UDiv:
6724 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6725 // Fold this div into the comparison, producing a range check.
6726 // Determine, based on the divide type, what the range is being
6727 // checked. If there is an overflow on the low or high side, remember
6728 // it, otherwise compute the range [low, hi) bounding the new value.
6729 // See: InsertRangeTest above for the kinds of replacements possible.
6730 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6731 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6736 case Instruction::Add:
6737 // Fold: icmp pred (add, X, C1), C2
6739 if (!ICI.isEquality()) {
6740 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6742 const APInt &LHSV = LHSC->getValue();
6744 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6747 if (ICI.isSignedPredicate()) {
6748 if (CR.getLower().isSignBit()) {
6749 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6750 ConstantInt::get(CR.getUpper()));
6751 } else if (CR.getUpper().isSignBit()) {
6752 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6753 ConstantInt::get(CR.getLower()));
6756 if (CR.getLower().isMinValue()) {
6757 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6758 ConstantInt::get(CR.getUpper()));
6759 } else if (CR.getUpper().isMinValue()) {
6760 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6761 ConstantInt::get(CR.getLower()));
6768 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6769 if (ICI.isEquality()) {
6770 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6772 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6773 // the second operand is a constant, simplify a bit.
6774 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6775 switch (BO->getOpcode()) {
6776 case Instruction::SRem:
6777 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6778 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6779 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6780 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6781 Instruction *NewRem =
6782 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6784 InsertNewInstBefore(NewRem, ICI);
6785 return new ICmpInst(ICI.getPredicate(), NewRem,
6786 Constant::getNullValue(BO->getType()));
6790 case Instruction::Add:
6791 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6792 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6793 if (BO->hasOneUse())
6794 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6795 Subtract(RHS, BOp1C));
6796 } else if (RHSV == 0) {
6797 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6798 // efficiently invertible, or if the add has just this one use.
6799 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6801 if (Value *NegVal = dyn_castNegVal(BOp1))
6802 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6803 else if (Value *NegVal = dyn_castNegVal(BOp0))
6804 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6805 else if (BO->hasOneUse()) {
6806 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6807 InsertNewInstBefore(Neg, ICI);
6809 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6813 case Instruction::Xor:
6814 // For the xor case, we can xor two constants together, eliminating
6815 // the explicit xor.
6816 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6817 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6818 ConstantExpr::getXor(RHS, BOC));
6821 case Instruction::Sub:
6822 // Replace (([sub|xor] A, B) != 0) with (A != B)
6824 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6828 case Instruction::Or:
6829 // If bits are being or'd in that are not present in the constant we
6830 // are comparing against, then the comparison could never succeed!
6831 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6832 Constant *NotCI = ConstantExpr::getNot(RHS);
6833 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6834 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6839 case Instruction::And:
6840 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6841 // If bits are being compared against that are and'd out, then the
6842 // comparison can never succeed!
6843 if ((RHSV & ~BOC->getValue()) != 0)
6844 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6847 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6848 if (RHS == BOC && RHSV.isPowerOf2())
6849 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6850 ICmpInst::ICMP_NE, LHSI,
6851 Constant::getNullValue(RHS->getType()));
6853 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6854 if (BOC->getValue().isSignBit()) {
6855 Value *X = BO->getOperand(0);
6856 Constant *Zero = Constant::getNullValue(X->getType());
6857 ICmpInst::Predicate pred = isICMP_NE ?
6858 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6859 return new ICmpInst(pred, X, Zero);
6862 // ((X & ~7) == 0) --> X < 8
6863 if (RHSV == 0 && isHighOnes(BOC)) {
6864 Value *X = BO->getOperand(0);
6865 Constant *NegX = ConstantExpr::getNeg(BOC);
6866 ICmpInst::Predicate pred = isICMP_NE ?
6867 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6868 return new ICmpInst(pred, X, NegX);
6873 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6874 // Handle icmp {eq|ne} <intrinsic>, intcst.
6875 if (II->getIntrinsicID() == Intrinsic::bswap) {
6877 ICI.setOperand(0, II->getOperand(1));
6878 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6886 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6887 /// We only handle extending casts so far.
6889 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6890 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6891 Value *LHSCIOp = LHSCI->getOperand(0);
6892 const Type *SrcTy = LHSCIOp->getType();
6893 const Type *DestTy = LHSCI->getType();
6896 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6897 // integer type is the same size as the pointer type.
6898 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6899 getTargetData().getPointerSizeInBits() ==
6900 cast<IntegerType>(DestTy)->getBitWidth()) {
6902 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6903 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6904 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6905 RHSOp = RHSC->getOperand(0);
6906 // If the pointer types don't match, insert a bitcast.
6907 if (LHSCIOp->getType() != RHSOp->getType())
6908 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6912 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6915 // The code below only handles extension cast instructions, so far.
6917 if (LHSCI->getOpcode() != Instruction::ZExt &&
6918 LHSCI->getOpcode() != Instruction::SExt)
6921 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6922 bool isSignedCmp = ICI.isSignedPredicate();
6924 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6925 // Not an extension from the same type?
6926 RHSCIOp = CI->getOperand(0);
6927 if (RHSCIOp->getType() != LHSCIOp->getType())
6930 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6931 // and the other is a zext), then we can't handle this.
6932 if (CI->getOpcode() != LHSCI->getOpcode())
6935 // Deal with equality cases early.
6936 if (ICI.isEquality())
6937 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6939 // A signed comparison of sign extended values simplifies into a
6940 // signed comparison.
6941 if (isSignedCmp && isSignedExt)
6942 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6944 // The other three cases all fold into an unsigned comparison.
6945 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6948 // If we aren't dealing with a constant on the RHS, exit early
6949 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6953 // Compute the constant that would happen if we truncated to SrcTy then
6954 // reextended to DestTy.
6955 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6956 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6958 // If the re-extended constant didn't change...
6960 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6961 // For example, we might have:
6962 // %A = sext short %X to uint
6963 // %B = icmp ugt uint %A, 1330
6964 // It is incorrect to transform this into
6965 // %B = icmp ugt short %X, 1330
6966 // because %A may have negative value.
6968 // However, we allow this when the compare is EQ/NE, because they are
6970 if (isSignedExt == isSignedCmp || ICI.isEquality())
6971 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6975 // The re-extended constant changed so the constant cannot be represented
6976 // in the shorter type. Consequently, we cannot emit a simple comparison.
6978 // First, handle some easy cases. We know the result cannot be equal at this
6979 // point so handle the ICI.isEquality() cases
6980 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6981 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6982 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6983 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6985 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6986 // should have been folded away previously and not enter in here.
6989 // We're performing a signed comparison.
6990 if (cast<ConstantInt>(CI)->getValue().isNegative())
6991 Result = ConstantInt::getFalse(); // X < (small) --> false
6993 Result = ConstantInt::getTrue(); // X < (large) --> true
6995 // We're performing an unsigned comparison.
6997 // We're performing an unsigned comp with a sign extended value.
6998 // This is true if the input is >= 0. [aka >s -1]
6999 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
7000 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
7001 NegOne, ICI.getName()), ICI);
7003 // Unsigned extend & unsigned compare -> always true.
7004 Result = ConstantInt::getTrue();
7008 // Finally, return the value computed.
7009 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7010 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7011 return ReplaceInstUsesWith(ICI, Result);
7013 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7014 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7015 "ICmp should be folded!");
7016 if (Constant *CI = dyn_cast<Constant>(Result))
7017 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7018 return BinaryOperator::CreateNot(Result);
7021 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7022 return commonShiftTransforms(I);
7025 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7026 return commonShiftTransforms(I);
7029 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7030 if (Instruction *R = commonShiftTransforms(I))
7033 Value *Op0 = I.getOperand(0);
7035 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7036 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7037 if (CSI->isAllOnesValue())
7038 return ReplaceInstUsesWith(I, CSI);
7040 // See if we can turn a signed shr into an unsigned shr.
7041 if (!isa<VectorType>(I.getType())) {
7042 if (MaskedValueIsZero(Op0,
7043 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
7044 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7046 // Arithmetic shifting an all-sign-bit value is a no-op.
7047 unsigned NumSignBits = ComputeNumSignBits(Op0);
7048 if (NumSignBits == Op0->getType()->getPrimitiveSizeInBits())
7049 return ReplaceInstUsesWith(I, Op0);
7055 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7056 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7057 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7059 // shl X, 0 == X and shr X, 0 == X
7060 // shl 0, X == 0 and shr 0, X == 0
7061 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7062 Op0 == Constant::getNullValue(Op0->getType()))
7063 return ReplaceInstUsesWith(I, Op0);
7065 if (isa<UndefValue>(Op0)) {
7066 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7067 return ReplaceInstUsesWith(I, Op0);
7068 else // undef << X -> 0, undef >>u X -> 0
7069 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7071 if (isa<UndefValue>(Op1)) {
7072 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7073 return ReplaceInstUsesWith(I, Op0);
7074 else // X << undef, X >>u undef -> 0
7075 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7078 // Try to fold constant and into select arguments.
7079 if (isa<Constant>(Op0))
7080 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7081 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7084 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7085 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7090 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7091 BinaryOperator &I) {
7092 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7094 // See if we can simplify any instructions used by the instruction whose sole
7095 // purpose is to compute bits we don't care about.
7096 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
7097 if (SimplifyDemandedInstructionBits(I))
7100 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
7101 // of a signed value.
7103 if (Op1->uge(TypeBits)) {
7104 if (I.getOpcode() != Instruction::AShr)
7105 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7107 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7112 // ((X*C1) << C2) == (X * (C1 << C2))
7113 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7114 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7115 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7116 return BinaryOperator::CreateMul(BO->getOperand(0),
7117 ConstantExpr::getShl(BOOp, Op1));
7119 // Try to fold constant and into select arguments.
7120 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7121 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7123 if (isa<PHINode>(Op0))
7124 if (Instruction *NV = FoldOpIntoPhi(I))
7127 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7128 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7129 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7130 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7131 // place. Don't try to do this transformation in this case. Also, we
7132 // require that the input operand is a shift-by-constant so that we have
7133 // confidence that the shifts will get folded together. We could do this
7134 // xform in more cases, but it is unlikely to be profitable.
7135 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7136 isa<ConstantInt>(TrOp->getOperand(1))) {
7137 // Okay, we'll do this xform. Make the shift of shift.
7138 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7139 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7141 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7143 // For logical shifts, the truncation has the effect of making the high
7144 // part of the register be zeros. Emulate this by inserting an AND to
7145 // clear the top bits as needed. This 'and' will usually be zapped by
7146 // other xforms later if dead.
7147 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7148 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7149 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7151 // The mask we constructed says what the trunc would do if occurring
7152 // between the shifts. We want to know the effect *after* the second
7153 // shift. We know that it is a logical shift by a constant, so adjust the
7154 // mask as appropriate.
7155 if (I.getOpcode() == Instruction::Shl)
7156 MaskV <<= Op1->getZExtValue();
7158 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7159 MaskV = MaskV.lshr(Op1->getZExtValue());
7162 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7164 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7166 // Return the value truncated to the interesting size.
7167 return new TruncInst(And, I.getType());
7171 if (Op0->hasOneUse()) {
7172 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7173 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7176 switch (Op0BO->getOpcode()) {
7178 case Instruction::Add:
7179 case Instruction::And:
7180 case Instruction::Or:
7181 case Instruction::Xor: {
7182 // These operators commute.
7183 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7184 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7185 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7186 Instruction *YS = BinaryOperator::CreateShl(
7187 Op0BO->getOperand(0), Op1,
7189 InsertNewInstBefore(YS, I); // (Y << C)
7191 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7192 Op0BO->getOperand(1)->getName());
7193 InsertNewInstBefore(X, I); // (X + (Y << C))
7194 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7195 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7196 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7199 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7200 Value *Op0BOOp1 = Op0BO->getOperand(1);
7201 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7203 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7204 m_ConstantInt(CC))) &&
7205 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7206 Instruction *YS = BinaryOperator::CreateShl(
7207 Op0BO->getOperand(0), Op1,
7209 InsertNewInstBefore(YS, I); // (Y << C)
7211 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7212 V1->getName()+".mask");
7213 InsertNewInstBefore(XM, I); // X & (CC << C)
7215 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7220 case Instruction::Sub: {
7221 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7222 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7223 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7224 Instruction *YS = BinaryOperator::CreateShl(
7225 Op0BO->getOperand(1), Op1,
7227 InsertNewInstBefore(YS, I); // (Y << C)
7229 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7230 Op0BO->getOperand(0)->getName());
7231 InsertNewInstBefore(X, I); // (X + (Y << C))
7232 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7233 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7234 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7237 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7238 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7239 match(Op0BO->getOperand(0),
7240 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7241 m_ConstantInt(CC))) && V2 == Op1 &&
7242 cast<BinaryOperator>(Op0BO->getOperand(0))
7243 ->getOperand(0)->hasOneUse()) {
7244 Instruction *YS = BinaryOperator::CreateShl(
7245 Op0BO->getOperand(1), Op1,
7247 InsertNewInstBefore(YS, I); // (Y << C)
7249 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7250 V1->getName()+".mask");
7251 InsertNewInstBefore(XM, I); // X & (CC << C)
7253 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7261 // If the operand is an bitwise operator with a constant RHS, and the
7262 // shift is the only use, we can pull it out of the shift.
7263 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7264 bool isValid = true; // Valid only for And, Or, Xor
7265 bool highBitSet = false; // Transform if high bit of constant set?
7267 switch (Op0BO->getOpcode()) {
7268 default: isValid = false; break; // Do not perform transform!
7269 case Instruction::Add:
7270 isValid = isLeftShift;
7272 case Instruction::Or:
7273 case Instruction::Xor:
7276 case Instruction::And:
7281 // If this is a signed shift right, and the high bit is modified
7282 // by the logical operation, do not perform the transformation.
7283 // The highBitSet boolean indicates the value of the high bit of
7284 // the constant which would cause it to be modified for this
7287 if (isValid && I.getOpcode() == Instruction::AShr)
7288 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7291 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7293 Instruction *NewShift =
7294 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7295 InsertNewInstBefore(NewShift, I);
7296 NewShift->takeName(Op0BO);
7298 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7305 // Find out if this is a shift of a shift by a constant.
7306 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7307 if (ShiftOp && !ShiftOp->isShift())
7310 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7311 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7312 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7313 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7314 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7315 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7316 Value *X = ShiftOp->getOperand(0);
7318 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7320 const IntegerType *Ty = cast<IntegerType>(I.getType());
7322 // Check for (X << c1) << c2 and (X >> c1) >> c2
7323 if (I.getOpcode() == ShiftOp->getOpcode()) {
7324 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7326 if (AmtSum >= TypeBits) {
7327 if (I.getOpcode() != Instruction::AShr)
7328 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7329 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7332 return BinaryOperator::Create(I.getOpcode(), X,
7333 ConstantInt::get(Ty, AmtSum));
7334 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7335 I.getOpcode() == Instruction::AShr) {
7336 if (AmtSum >= TypeBits)
7337 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7339 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7340 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7341 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7342 I.getOpcode() == Instruction::LShr) {
7343 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7344 if (AmtSum >= TypeBits)
7345 AmtSum = TypeBits-1;
7347 Instruction *Shift =
7348 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7349 InsertNewInstBefore(Shift, I);
7351 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7352 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7355 // Okay, if we get here, one shift must be left, and the other shift must be
7356 // right. See if the amounts are equal.
7357 if (ShiftAmt1 == ShiftAmt2) {
7358 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7359 if (I.getOpcode() == Instruction::Shl) {
7360 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7361 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7363 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7364 if (I.getOpcode() == Instruction::LShr) {
7365 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7366 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7368 // We can simplify ((X << C) >>s C) into a trunc + sext.
7369 // NOTE: we could do this for any C, but that would make 'unusual' integer
7370 // types. For now, just stick to ones well-supported by the code
7372 const Type *SExtType = 0;
7373 switch (Ty->getBitWidth() - ShiftAmt1) {
7380 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7385 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7386 InsertNewInstBefore(NewTrunc, I);
7387 return new SExtInst(NewTrunc, Ty);
7389 // Otherwise, we can't handle it yet.
7390 } else if (ShiftAmt1 < ShiftAmt2) {
7391 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7393 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7394 if (I.getOpcode() == Instruction::Shl) {
7395 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7396 ShiftOp->getOpcode() == Instruction::AShr);
7397 Instruction *Shift =
7398 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7399 InsertNewInstBefore(Shift, I);
7401 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7402 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7405 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7406 if (I.getOpcode() == Instruction::LShr) {
7407 assert(ShiftOp->getOpcode() == Instruction::Shl);
7408 Instruction *Shift =
7409 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7410 InsertNewInstBefore(Shift, I);
7412 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7413 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7416 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7418 assert(ShiftAmt2 < ShiftAmt1);
7419 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7421 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7422 if (I.getOpcode() == Instruction::Shl) {
7423 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7424 ShiftOp->getOpcode() == Instruction::AShr);
7425 Instruction *Shift =
7426 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7427 ConstantInt::get(Ty, ShiftDiff));
7428 InsertNewInstBefore(Shift, I);
7430 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7431 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7434 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7435 if (I.getOpcode() == Instruction::LShr) {
7436 assert(ShiftOp->getOpcode() == Instruction::Shl);
7437 Instruction *Shift =
7438 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7439 InsertNewInstBefore(Shift, I);
7441 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7442 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7445 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7452 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7453 /// expression. If so, decompose it, returning some value X, such that Val is
7456 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7458 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7459 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7460 Offset = CI->getZExtValue();
7462 return ConstantInt::get(Type::Int32Ty, 0);
7463 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7464 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7465 if (I->getOpcode() == Instruction::Shl) {
7466 // This is a value scaled by '1 << the shift amt'.
7467 Scale = 1U << RHS->getZExtValue();
7469 return I->getOperand(0);
7470 } else if (I->getOpcode() == Instruction::Mul) {
7471 // This value is scaled by 'RHS'.
7472 Scale = RHS->getZExtValue();
7474 return I->getOperand(0);
7475 } else if (I->getOpcode() == Instruction::Add) {
7476 // We have X+C. Check to see if we really have (X*C2)+C1,
7477 // where C1 is divisible by C2.
7480 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7481 Offset += RHS->getZExtValue();
7488 // Otherwise, we can't look past this.
7495 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7496 /// try to eliminate the cast by moving the type information into the alloc.
7497 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7498 AllocationInst &AI) {
7499 const PointerType *PTy = cast<PointerType>(CI.getType());
7501 // Remove any uses of AI that are dead.
7502 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7504 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7505 Instruction *User = cast<Instruction>(*UI++);
7506 if (isInstructionTriviallyDead(User)) {
7507 while (UI != E && *UI == User)
7508 ++UI; // If this instruction uses AI more than once, don't break UI.
7511 DOUT << "IC: DCE: " << *User;
7512 EraseInstFromFunction(*User);
7516 // Get the type really allocated and the type casted to.
7517 const Type *AllocElTy = AI.getAllocatedType();
7518 const Type *CastElTy = PTy->getElementType();
7519 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7521 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7522 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7523 if (CastElTyAlign < AllocElTyAlign) return 0;
7525 // If the allocation has multiple uses, only promote it if we are strictly
7526 // increasing the alignment of the resultant allocation. If we keep it the
7527 // same, we open the door to infinite loops of various kinds. (A reference
7528 // from a dbg.declare doesn't count as a use for this purpose.)
7529 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7530 CastElTyAlign == AllocElTyAlign) return 0;
7532 uint64_t AllocElTySize = TD->getTypePaddedSize(AllocElTy);
7533 uint64_t CastElTySize = TD->getTypePaddedSize(CastElTy);
7534 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7536 // See if we can satisfy the modulus by pulling a scale out of the array
7538 unsigned ArraySizeScale;
7540 Value *NumElements = // See if the array size is a decomposable linear expr.
7541 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7543 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7545 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7546 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7548 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7553 // If the allocation size is constant, form a constant mul expression
7554 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7555 if (isa<ConstantInt>(NumElements))
7556 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7557 // otherwise multiply the amount and the number of elements
7559 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7560 Amt = InsertNewInstBefore(Tmp, AI);
7564 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7565 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7566 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7567 Amt = InsertNewInstBefore(Tmp, AI);
7570 AllocationInst *New;
7571 if (isa<MallocInst>(AI))
7572 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7574 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7575 InsertNewInstBefore(New, AI);
7578 // If the allocation has one real use plus a dbg.declare, just remove the
7580 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7581 EraseInstFromFunction(*DI);
7583 // If the allocation has multiple real uses, insert a cast and change all
7584 // things that used it to use the new cast. This will also hack on CI, but it
7586 else if (!AI.hasOneUse()) {
7587 AddUsesToWorkList(AI);
7588 // New is the allocation instruction, pointer typed. AI is the original
7589 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7590 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7591 InsertNewInstBefore(NewCast, AI);
7592 AI.replaceAllUsesWith(NewCast);
7594 return ReplaceInstUsesWith(CI, New);
7597 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7598 /// and return it as type Ty without inserting any new casts and without
7599 /// changing the computed value. This is used by code that tries to decide
7600 /// whether promoting or shrinking integer operations to wider or smaller types
7601 /// will allow us to eliminate a truncate or extend.
7603 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7604 /// extension operation if Ty is larger.
7606 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7607 /// should return true if trunc(V) can be computed by computing V in the smaller
7608 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7609 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7610 /// efficiently truncated.
7612 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7613 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7614 /// the final result.
7615 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7617 int &NumCastsRemoved){
7618 // We can always evaluate constants in another type.
7619 if (isa<ConstantInt>(V))
7622 Instruction *I = dyn_cast<Instruction>(V);
7623 if (!I) return false;
7625 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7627 // If this is an extension or truncate, we can often eliminate it.
7628 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7629 // If this is a cast from the destination type, we can trivially eliminate
7630 // it, and this will remove a cast overall.
7631 if (I->getOperand(0)->getType() == Ty) {
7632 // If the first operand is itself a cast, and is eliminable, do not count
7633 // this as an eliminable cast. We would prefer to eliminate those two
7635 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7641 // We can't extend or shrink something that has multiple uses: doing so would
7642 // require duplicating the instruction in general, which isn't profitable.
7643 if (!I->hasOneUse()) return false;
7645 unsigned Opc = I->getOpcode();
7647 case Instruction::Add:
7648 case Instruction::Sub:
7649 case Instruction::Mul:
7650 case Instruction::And:
7651 case Instruction::Or:
7652 case Instruction::Xor:
7653 // These operators can all arbitrarily be extended or truncated.
7654 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7656 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7659 case Instruction::Shl:
7660 // If we are truncating the result of this SHL, and if it's a shift of a
7661 // constant amount, we can always perform a SHL in a smaller type.
7662 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7663 uint32_t BitWidth = Ty->getBitWidth();
7664 if (BitWidth < OrigTy->getBitWidth() &&
7665 CI->getLimitedValue(BitWidth) < BitWidth)
7666 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7670 case Instruction::LShr:
7671 // If this is a truncate of a logical shr, we can truncate it to a smaller
7672 // lshr iff we know that the bits we would otherwise be shifting in are
7674 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7675 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7676 uint32_t BitWidth = Ty->getBitWidth();
7677 if (BitWidth < OrigBitWidth &&
7678 MaskedValueIsZero(I->getOperand(0),
7679 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7680 CI->getLimitedValue(BitWidth) < BitWidth) {
7681 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7686 case Instruction::ZExt:
7687 case Instruction::SExt:
7688 case Instruction::Trunc:
7689 // If this is the same kind of case as our original (e.g. zext+zext), we
7690 // can safely replace it. Note that replacing it does not reduce the number
7691 // of casts in the input.
7695 // sext (zext ty1), ty2 -> zext ty2
7696 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7699 case Instruction::Select: {
7700 SelectInst *SI = cast<SelectInst>(I);
7701 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7703 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7706 case Instruction::PHI: {
7707 // We can change a phi if we can change all operands.
7708 PHINode *PN = cast<PHINode>(I);
7709 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7710 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7716 // TODO: Can handle more cases here.
7723 /// EvaluateInDifferentType - Given an expression that
7724 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7725 /// evaluate the expression.
7726 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7728 if (Constant *C = dyn_cast<Constant>(V))
7729 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7731 // Otherwise, it must be an instruction.
7732 Instruction *I = cast<Instruction>(V);
7733 Instruction *Res = 0;
7734 unsigned Opc = I->getOpcode();
7736 case Instruction::Add:
7737 case Instruction::Sub:
7738 case Instruction::Mul:
7739 case Instruction::And:
7740 case Instruction::Or:
7741 case Instruction::Xor:
7742 case Instruction::AShr:
7743 case Instruction::LShr:
7744 case Instruction::Shl: {
7745 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7746 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7747 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
7750 case Instruction::Trunc:
7751 case Instruction::ZExt:
7752 case Instruction::SExt:
7753 // If the source type of the cast is the type we're trying for then we can
7754 // just return the source. There's no need to insert it because it is not
7756 if (I->getOperand(0)->getType() == Ty)
7757 return I->getOperand(0);
7759 // Otherwise, must be the same type of cast, so just reinsert a new one.
7760 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7763 case Instruction::Select: {
7764 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7765 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7766 Res = SelectInst::Create(I->getOperand(0), True, False);
7769 case Instruction::PHI: {
7770 PHINode *OPN = cast<PHINode>(I);
7771 PHINode *NPN = PHINode::Create(Ty);
7772 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7773 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7774 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7780 // TODO: Can handle more cases here.
7781 assert(0 && "Unreachable!");
7786 return InsertNewInstBefore(Res, *I);
7789 /// @brief Implement the transforms common to all CastInst visitors.
7790 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7791 Value *Src = CI.getOperand(0);
7793 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7794 // eliminate it now.
7795 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7796 if (Instruction::CastOps opc =
7797 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7798 // The first cast (CSrc) is eliminable so we need to fix up or replace
7799 // the second cast (CI). CSrc will then have a good chance of being dead.
7800 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7804 // If we are casting a select then fold the cast into the select
7805 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7806 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7809 // If we are casting a PHI then fold the cast into the PHI
7810 if (isa<PHINode>(Src))
7811 if (Instruction *NV = FoldOpIntoPhi(CI))
7817 /// FindElementAtOffset - Given a type and a constant offset, determine whether
7818 /// or not there is a sequence of GEP indices into the type that will land us at
7819 /// the specified offset. If so, fill them into NewIndices and return the
7820 /// resultant element type, otherwise return null.
7821 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
7822 SmallVectorImpl<Value*> &NewIndices,
7823 const TargetData *TD) {
7824 if (!Ty->isSized()) return 0;
7826 // Start with the index over the outer type. Note that the type size
7827 // might be zero (even if the offset isn't zero) if the indexed type
7828 // is something like [0 x {int, int}]
7829 const Type *IntPtrTy = TD->getIntPtrType();
7830 int64_t FirstIdx = 0;
7831 if (int64_t TySize = TD->getTypePaddedSize(Ty)) {
7832 FirstIdx = Offset/TySize;
7833 Offset -= FirstIdx*TySize;
7835 // Handle hosts where % returns negative instead of values [0..TySize).
7839 assert(Offset >= 0);
7841 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
7844 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7846 // Index into the types. If we fail, set OrigBase to null.
7848 // Indexing into tail padding between struct/array elements.
7849 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
7852 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
7853 const StructLayout *SL = TD->getStructLayout(STy);
7854 assert(Offset < (int64_t)SL->getSizeInBytes() &&
7855 "Offset must stay within the indexed type");
7857 unsigned Elt = SL->getElementContainingOffset(Offset);
7858 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7860 Offset -= SL->getElementOffset(Elt);
7861 Ty = STy->getElementType(Elt);
7862 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
7863 uint64_t EltSize = TD->getTypePaddedSize(AT->getElementType());
7864 assert(EltSize && "Cannot index into a zero-sized array");
7865 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7867 Ty = AT->getElementType();
7869 // Otherwise, we can't index into the middle of this atomic type, bail.
7877 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7878 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7879 Value *Src = CI.getOperand(0);
7881 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7882 // If casting the result of a getelementptr instruction with no offset, turn
7883 // this into a cast of the original pointer!
7884 if (GEP->hasAllZeroIndices()) {
7885 // Changing the cast operand is usually not a good idea but it is safe
7886 // here because the pointer operand is being replaced with another
7887 // pointer operand so the opcode doesn't need to change.
7889 CI.setOperand(0, GEP->getOperand(0));
7893 // If the GEP has a single use, and the base pointer is a bitcast, and the
7894 // GEP computes a constant offset, see if we can convert these three
7895 // instructions into fewer. This typically happens with unions and other
7896 // non-type-safe code.
7897 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7898 if (GEP->hasAllConstantIndices()) {
7899 // We are guaranteed to get a constant from EmitGEPOffset.
7900 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7901 int64_t Offset = OffsetV->getSExtValue();
7903 // Get the base pointer input of the bitcast, and the type it points to.
7904 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7905 const Type *GEPIdxTy =
7906 cast<PointerType>(OrigBase->getType())->getElementType();
7907 SmallVector<Value*, 8> NewIndices;
7908 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD)) {
7909 // If we were able to index down into an element, create the GEP
7910 // and bitcast the result. This eliminates one bitcast, potentially
7912 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7914 NewIndices.end(), "");
7915 InsertNewInstBefore(NGEP, CI);
7916 NGEP->takeName(GEP);
7918 if (isa<BitCastInst>(CI))
7919 return new BitCastInst(NGEP, CI.getType());
7920 assert(isa<PtrToIntInst>(CI));
7921 return new PtrToIntInst(NGEP, CI.getType());
7927 return commonCastTransforms(CI);
7930 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
7931 /// type like i42. We don't want to introduce operations on random non-legal
7932 /// integer types where they don't already exist in the code. In the future,
7933 /// we should consider making this based off target-data, so that 32-bit targets
7934 /// won't get i64 operations etc.
7935 static bool isSafeIntegerType(const Type *Ty) {
7936 switch (Ty->getPrimitiveSizeInBits()) {
7947 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7948 /// integer types. This function implements the common transforms for all those
7950 /// @brief Implement the transforms common to CastInst with integer operands
7951 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7952 if (Instruction *Result = commonCastTransforms(CI))
7955 Value *Src = CI.getOperand(0);
7956 const Type *SrcTy = Src->getType();
7957 const Type *DestTy = CI.getType();
7958 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7959 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7961 // See if we can simplify any instructions used by the LHS whose sole
7962 // purpose is to compute bits we don't care about.
7963 if (SimplifyDemandedInstructionBits(CI))
7966 // If the source isn't an instruction or has more than one use then we
7967 // can't do anything more.
7968 Instruction *SrcI = dyn_cast<Instruction>(Src);
7969 if (!SrcI || !Src->hasOneUse())
7972 // Attempt to propagate the cast into the instruction for int->int casts.
7973 int NumCastsRemoved = 0;
7974 if (!isa<BitCastInst>(CI) &&
7975 // Only do this if the dest type is a simple type, don't convert the
7976 // expression tree to something weird like i93 unless the source is also
7978 (isSafeIntegerType(DestTy) || !isSafeIntegerType(SrcI->getType())) &&
7979 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7980 CI.getOpcode(), NumCastsRemoved)) {
7981 // If this cast is a truncate, evaluting in a different type always
7982 // eliminates the cast, so it is always a win. If this is a zero-extension,
7983 // we need to do an AND to maintain the clear top-part of the computation,
7984 // so we require that the input have eliminated at least one cast. If this
7985 // is a sign extension, we insert two new casts (to do the extension) so we
7986 // require that two casts have been eliminated.
7987 bool DoXForm = false;
7988 bool JustReplace = false;
7989 switch (CI.getOpcode()) {
7991 // All the others use floating point so we shouldn't actually
7992 // get here because of the check above.
7993 assert(0 && "Unknown cast type");
7994 case Instruction::Trunc:
7997 case Instruction::ZExt: {
7998 DoXForm = NumCastsRemoved >= 1;
7999 if (!DoXForm && 0) {
8000 // If it's unnecessary to issue an AND to clear the high bits, it's
8001 // always profitable to do this xform.
8002 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8003 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8004 if (MaskedValueIsZero(TryRes, Mask))
8005 return ReplaceInstUsesWith(CI, TryRes);
8007 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8008 if (TryI->use_empty())
8009 EraseInstFromFunction(*TryI);
8013 case Instruction::SExt: {
8014 DoXForm = NumCastsRemoved >= 2;
8015 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8016 // If we do not have to emit the truncate + sext pair, then it's always
8017 // profitable to do this xform.
8019 // It's not safe to eliminate the trunc + sext pair if one of the
8020 // eliminated cast is a truncate. e.g.
8021 // t2 = trunc i32 t1 to i16
8022 // t3 = sext i16 t2 to i32
8025 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8026 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8027 if (NumSignBits > (DestBitSize - SrcBitSize))
8028 return ReplaceInstUsesWith(CI, TryRes);
8030 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8031 if (TryI->use_empty())
8032 EraseInstFromFunction(*TryI);
8039 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8041 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8042 CI.getOpcode() == Instruction::SExt);
8044 // Just replace this cast with the result.
8045 return ReplaceInstUsesWith(CI, Res);
8047 assert(Res->getType() == DestTy);
8048 switch (CI.getOpcode()) {
8049 default: assert(0 && "Unknown cast type!");
8050 case Instruction::Trunc:
8051 case Instruction::BitCast:
8052 // Just replace this cast with the result.
8053 return ReplaceInstUsesWith(CI, Res);
8054 case Instruction::ZExt: {
8055 assert(SrcBitSize < DestBitSize && "Not a zext?");
8057 // If the high bits are already zero, just replace this cast with the
8059 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8060 if (MaskedValueIsZero(Res, Mask))
8061 return ReplaceInstUsesWith(CI, Res);
8063 // We need to emit an AND to clear the high bits.
8064 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
8066 return BinaryOperator::CreateAnd(Res, C);
8068 case Instruction::SExt: {
8069 // If the high bits are already filled with sign bit, just replace this
8070 // cast with the result.
8071 unsigned NumSignBits = ComputeNumSignBits(Res);
8072 if (NumSignBits > (DestBitSize - SrcBitSize))
8073 return ReplaceInstUsesWith(CI, Res);
8075 // We need to emit a cast to truncate, then a cast to sext.
8076 return CastInst::Create(Instruction::SExt,
8077 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8084 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8085 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8087 switch (SrcI->getOpcode()) {
8088 case Instruction::Add:
8089 case Instruction::Mul:
8090 case Instruction::And:
8091 case Instruction::Or:
8092 case Instruction::Xor:
8093 // If we are discarding information, rewrite.
8094 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
8095 // Don't insert two casts if they cannot be eliminated. We allow
8096 // two casts to be inserted if the sizes are the same. This could
8097 // only be converting signedness, which is a noop.
8098 if (DestBitSize == SrcBitSize ||
8099 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
8100 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8101 Instruction::CastOps opcode = CI.getOpcode();
8102 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8103 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8104 return BinaryOperator::Create(
8105 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8109 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8110 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8111 SrcI->getOpcode() == Instruction::Xor &&
8112 Op1 == ConstantInt::getTrue() &&
8113 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8114 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8115 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
8118 case Instruction::SDiv:
8119 case Instruction::UDiv:
8120 case Instruction::SRem:
8121 case Instruction::URem:
8122 // If we are just changing the sign, rewrite.
8123 if (DestBitSize == SrcBitSize) {
8124 // Don't insert two casts if they cannot be eliminated. We allow
8125 // two casts to be inserted if the sizes are the same. This could
8126 // only be converting signedness, which is a noop.
8127 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8128 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8129 Value *Op0c = InsertCastBefore(Instruction::BitCast,
8130 Op0, DestTy, *SrcI);
8131 Value *Op1c = InsertCastBefore(Instruction::BitCast,
8132 Op1, DestTy, *SrcI);
8133 return BinaryOperator::Create(
8134 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8139 case Instruction::Shl:
8140 // Allow changing the sign of the source operand. Do not allow
8141 // changing the size of the shift, UNLESS the shift amount is a
8142 // constant. We must not change variable sized shifts to a smaller
8143 // size, because it is undefined to shift more bits out than exist
8145 if (DestBitSize == SrcBitSize ||
8146 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
8147 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
8148 Instruction::BitCast : Instruction::Trunc);
8149 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8150 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8151 return BinaryOperator::CreateShl(Op0c, Op1c);
8154 case Instruction::AShr:
8155 // If this is a signed shr, and if all bits shifted in are about to be
8156 // truncated off, turn it into an unsigned shr to allow greater
8158 if (DestBitSize < SrcBitSize &&
8159 isa<ConstantInt>(Op1)) {
8160 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
8161 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
8162 // Insert the new logical shift right.
8163 return BinaryOperator::CreateLShr(Op0, Op1);
8171 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8172 if (Instruction *Result = commonIntCastTransforms(CI))
8175 Value *Src = CI.getOperand(0);
8176 const Type *Ty = CI.getType();
8177 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
8178 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
8180 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8181 if (DestBitWidth == 1) {
8182 Constant *One = ConstantInt::get(Src->getType(), 1);
8183 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8184 Value *Zero = Constant::getNullValue(Src->getType());
8185 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8188 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8189 ConstantInt *ShAmtV = 0;
8191 if (Src->hasOneUse() &&
8192 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8193 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8195 // Get a mask for the bits shifting in.
8196 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8197 if (MaskedValueIsZero(ShiftOp, Mask)) {
8198 if (ShAmt >= DestBitWidth) // All zeros.
8199 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8201 // Okay, we can shrink this. Truncate the input, then return a new
8203 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8204 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8205 return BinaryOperator::CreateLShr(V1, V2);
8212 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8213 /// in order to eliminate the icmp.
8214 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8216 // If we are just checking for a icmp eq of a single bit and zext'ing it
8217 // to an integer, then shift the bit to the appropriate place and then
8218 // cast to integer to avoid the comparison.
8219 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8220 const APInt &Op1CV = Op1C->getValue();
8222 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8223 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8224 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8225 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8226 if (!DoXform) return ICI;
8228 Value *In = ICI->getOperand(0);
8229 Value *Sh = ConstantInt::get(In->getType(),
8230 In->getType()->getPrimitiveSizeInBits()-1);
8231 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8232 In->getName()+".lobit"),
8234 if (In->getType() != CI.getType())
8235 In = CastInst::CreateIntegerCast(In, CI.getType(),
8236 false/*ZExt*/, "tmp", &CI);
8238 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8239 Constant *One = ConstantInt::get(In->getType(), 1);
8240 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8241 In->getName()+".not"),
8245 return ReplaceInstUsesWith(CI, In);
8250 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8251 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8252 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8253 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8254 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8255 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8256 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8257 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8258 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8259 // This only works for EQ and NE
8260 ICI->isEquality()) {
8261 // If Op1C some other power of two, convert:
8262 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8263 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8264 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8265 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8267 APInt KnownZeroMask(~KnownZero);
8268 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8269 if (!DoXform) return ICI;
8271 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8272 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8273 // (X&4) == 2 --> false
8274 // (X&4) != 2 --> true
8275 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8276 Res = ConstantExpr::getZExt(Res, CI.getType());
8277 return ReplaceInstUsesWith(CI, Res);
8280 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8281 Value *In = ICI->getOperand(0);
8283 // Perform a logical shr by shiftamt.
8284 // Insert the shift to put the result in the low bit.
8285 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8286 ConstantInt::get(In->getType(), ShiftAmt),
8287 In->getName()+".lobit"), CI);
8290 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8291 Constant *One = ConstantInt::get(In->getType(), 1);
8292 In = BinaryOperator::CreateXor(In, One, "tmp");
8293 InsertNewInstBefore(cast<Instruction>(In), CI);
8296 if (CI.getType() == In->getType())
8297 return ReplaceInstUsesWith(CI, In);
8299 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8307 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8308 // If one of the common conversion will work ..
8309 if (Instruction *Result = commonIntCastTransforms(CI))
8312 Value *Src = CI.getOperand(0);
8314 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8315 // types and if the sizes are just right we can convert this into a logical
8316 // 'and' which will be much cheaper than the pair of casts.
8317 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8318 // Get the sizes of the types involved. We know that the intermediate type
8319 // will be smaller than A or C, but don't know the relation between A and C.
8320 Value *A = CSrc->getOperand(0);
8321 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
8322 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8323 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8324 // If we're actually extending zero bits, then if
8325 // SrcSize < DstSize: zext(a & mask)
8326 // SrcSize == DstSize: a & mask
8327 // SrcSize > DstSize: trunc(a) & mask
8328 if (SrcSize < DstSize) {
8329 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8330 Constant *AndConst = ConstantInt::get(AndValue);
8332 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8333 InsertNewInstBefore(And, CI);
8334 return new ZExtInst(And, CI.getType());
8335 } else if (SrcSize == DstSize) {
8336 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8337 return BinaryOperator::CreateAnd(A, ConstantInt::get(AndValue));
8338 } else if (SrcSize > DstSize) {
8339 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8340 InsertNewInstBefore(Trunc, CI);
8341 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8342 return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(AndValue));
8346 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8347 return transformZExtICmp(ICI, CI);
8349 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8350 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8351 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8352 // of the (zext icmp) will be transformed.
8353 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8354 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8355 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8356 (transformZExtICmp(LHS, CI, false) ||
8357 transformZExtICmp(RHS, CI, false))) {
8358 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8359 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8360 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8367 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8368 if (Instruction *I = commonIntCastTransforms(CI))
8371 Value *Src = CI.getOperand(0);
8373 // Canonicalize sign-extend from i1 to a select.
8374 if (Src->getType() == Type::Int1Ty)
8375 return SelectInst::Create(Src,
8376 ConstantInt::getAllOnesValue(CI.getType()),
8377 Constant::getNullValue(CI.getType()));
8379 // See if the value being truncated is already sign extended. If so, just
8380 // eliminate the trunc/sext pair.
8381 if (getOpcode(Src) == Instruction::Trunc) {
8382 Value *Op = cast<User>(Src)->getOperand(0);
8383 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8384 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8385 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8386 unsigned NumSignBits = ComputeNumSignBits(Op);
8388 if (OpBits == DestBits) {
8389 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8390 // bits, it is already ready.
8391 if (NumSignBits > DestBits-MidBits)
8392 return ReplaceInstUsesWith(CI, Op);
8393 } else if (OpBits < DestBits) {
8394 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8395 // bits, just sext from i32.
8396 if (NumSignBits > OpBits-MidBits)
8397 return new SExtInst(Op, CI.getType(), "tmp");
8399 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8400 // bits, just truncate to i32.
8401 if (NumSignBits > OpBits-MidBits)
8402 return new TruncInst(Op, CI.getType(), "tmp");
8406 // If the input is a shl/ashr pair of a same constant, then this is a sign
8407 // extension from a smaller value. If we could trust arbitrary bitwidth
8408 // integers, we could turn this into a truncate to the smaller bit and then
8409 // use a sext for the whole extension. Since we don't, look deeper and check
8410 // for a truncate. If the source and dest are the same type, eliminate the
8411 // trunc and extend and just do shifts. For example, turn:
8412 // %a = trunc i32 %i to i8
8413 // %b = shl i8 %a, 6
8414 // %c = ashr i8 %b, 6
8415 // %d = sext i8 %c to i32
8417 // %a = shl i32 %i, 30
8418 // %d = ashr i32 %a, 30
8420 ConstantInt *BA = 0, *CA = 0;
8421 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8422 m_ConstantInt(CA))) &&
8423 BA == CA && isa<TruncInst>(A)) {
8424 Value *I = cast<TruncInst>(A)->getOperand(0);
8425 if (I->getType() == CI.getType()) {
8426 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8427 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8428 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8429 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8430 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8432 return BinaryOperator::CreateAShr(I, ShAmtV);
8439 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8440 /// in the specified FP type without changing its value.
8441 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8443 APFloat F = CFP->getValueAPF();
8444 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8446 return ConstantFP::get(F);
8450 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8451 /// through it until we get the source value.
8452 static Value *LookThroughFPExtensions(Value *V) {
8453 if (Instruction *I = dyn_cast<Instruction>(V))
8454 if (I->getOpcode() == Instruction::FPExt)
8455 return LookThroughFPExtensions(I->getOperand(0));
8457 // If this value is a constant, return the constant in the smallest FP type
8458 // that can accurately represent it. This allows us to turn
8459 // (float)((double)X+2.0) into x+2.0f.
8460 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8461 if (CFP->getType() == Type::PPC_FP128Ty)
8462 return V; // No constant folding of this.
8463 // See if the value can be truncated to float and then reextended.
8464 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8466 if (CFP->getType() == Type::DoubleTy)
8467 return V; // Won't shrink.
8468 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8470 // Don't try to shrink to various long double types.
8476 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8477 if (Instruction *I = commonCastTransforms(CI))
8480 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8481 // smaller than the destination type, we can eliminate the truncate by doing
8482 // the add as the smaller type. This applies to add/sub/mul/div as well as
8483 // many builtins (sqrt, etc).
8484 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8485 if (OpI && OpI->hasOneUse()) {
8486 switch (OpI->getOpcode()) {
8488 case Instruction::Add:
8489 case Instruction::Sub:
8490 case Instruction::Mul:
8491 case Instruction::FDiv:
8492 case Instruction::FRem:
8493 const Type *SrcTy = OpI->getType();
8494 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8495 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8496 if (LHSTrunc->getType() != SrcTy &&
8497 RHSTrunc->getType() != SrcTy) {
8498 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8499 // If the source types were both smaller than the destination type of
8500 // the cast, do this xform.
8501 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8502 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8503 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8505 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8507 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8516 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8517 return commonCastTransforms(CI);
8520 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8521 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8523 return commonCastTransforms(FI);
8525 // fptoui(uitofp(X)) --> X
8526 // fptoui(sitofp(X)) --> X
8527 // This is safe if the intermediate type has enough bits in its mantissa to
8528 // accurately represent all values of X. For example, do not do this with
8529 // i64->float->i64. This is also safe for sitofp case, because any negative
8530 // 'X' value would cause an undefined result for the fptoui.
8531 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8532 OpI->getOperand(0)->getType() == FI.getType() &&
8533 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8534 OpI->getType()->getFPMantissaWidth())
8535 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8537 return commonCastTransforms(FI);
8540 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8541 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8543 return commonCastTransforms(FI);
8545 // fptosi(sitofp(X)) --> X
8546 // fptosi(uitofp(X)) --> X
8547 // This is safe if the intermediate type has enough bits in its mantissa to
8548 // accurately represent all values of X. For example, do not do this with
8549 // i64->float->i64. This is also safe for sitofp case, because any negative
8550 // 'X' value would cause an undefined result for the fptoui.
8551 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8552 OpI->getOperand(0)->getType() == FI.getType() &&
8553 (int)FI.getType()->getPrimitiveSizeInBits() <=
8554 OpI->getType()->getFPMantissaWidth())
8555 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8557 return commonCastTransforms(FI);
8560 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8561 return commonCastTransforms(CI);
8564 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8565 return commonCastTransforms(CI);
8568 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8569 // If the destination integer type is smaller than the intptr_t type for
8570 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8571 // trunc to be exposed to other transforms. Don't do this for extending
8572 // ptrtoint's, because we don't know if the target sign or zero extends its
8574 if (CI.getType()->getPrimitiveSizeInBits() < TD->getPointerSizeInBits()) {
8575 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8576 TD->getIntPtrType(),
8578 return new TruncInst(P, CI.getType());
8581 return commonPointerCastTransforms(CI);
8584 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8585 // If the source integer type is larger than the intptr_t type for
8586 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8587 // allows the trunc to be exposed to other transforms. Don't do this for
8588 // extending inttoptr's, because we don't know if the target sign or zero
8589 // extends to pointers.
8590 if (CI.getOperand(0)->getType()->getPrimitiveSizeInBits() >
8591 TD->getPointerSizeInBits()) {
8592 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8593 TD->getIntPtrType(),
8595 return new IntToPtrInst(P, CI.getType());
8598 if (Instruction *I = commonCastTransforms(CI))
8601 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8602 if (!DestPointee->isSized()) return 0;
8604 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8607 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8608 m_ConstantInt(Cst)))) {
8609 // If the source and destination operands have the same type, see if this
8610 // is a single-index GEP.
8611 if (X->getType() == CI.getType()) {
8612 // Get the size of the pointee type.
8613 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8615 // Convert the constant to intptr type.
8616 APInt Offset = Cst->getValue();
8617 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8619 // If Offset is evenly divisible by Size, we can do this xform.
8620 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8621 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8622 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8625 // TODO: Could handle other cases, e.g. where add is indexing into field of
8627 } else if (CI.getOperand(0)->hasOneUse() &&
8628 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8629 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8630 // "inttoptr+GEP" instead of "add+intptr".
8632 // Get the size of the pointee type.
8633 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8635 // Convert the constant to intptr type.
8636 APInt Offset = Cst->getValue();
8637 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8639 // If Offset is evenly divisible by Size, we can do this xform.
8640 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8641 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8643 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8645 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8651 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8652 // If the operands are integer typed then apply the integer transforms,
8653 // otherwise just apply the common ones.
8654 Value *Src = CI.getOperand(0);
8655 const Type *SrcTy = Src->getType();
8656 const Type *DestTy = CI.getType();
8658 if (SrcTy->isInteger() && DestTy->isInteger()) {
8659 if (Instruction *Result = commonIntCastTransforms(CI))
8661 } else if (isa<PointerType>(SrcTy)) {
8662 if (Instruction *I = commonPointerCastTransforms(CI))
8665 if (Instruction *Result = commonCastTransforms(CI))
8670 // Get rid of casts from one type to the same type. These are useless and can
8671 // be replaced by the operand.
8672 if (DestTy == Src->getType())
8673 return ReplaceInstUsesWith(CI, Src);
8675 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8676 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8677 const Type *DstElTy = DstPTy->getElementType();
8678 const Type *SrcElTy = SrcPTy->getElementType();
8680 // If the address spaces don't match, don't eliminate the bitcast, which is
8681 // required for changing types.
8682 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8685 // If we are casting a malloc or alloca to a pointer to a type of the same
8686 // size, rewrite the allocation instruction to allocate the "right" type.
8687 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8688 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8691 // If the source and destination are pointers, and this cast is equivalent
8692 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8693 // This can enhance SROA and other transforms that want type-safe pointers.
8694 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8695 unsigned NumZeros = 0;
8696 while (SrcElTy != DstElTy &&
8697 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8698 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8699 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8703 // If we found a path from the src to dest, create the getelementptr now.
8704 if (SrcElTy == DstElTy) {
8705 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8706 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8707 ((Instruction*) NULL));
8711 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8712 if (SVI->hasOneUse()) {
8713 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8714 // a bitconvert to a vector with the same # elts.
8715 if (isa<VectorType>(DestTy) &&
8716 cast<VectorType>(DestTy)->getNumElements() ==
8717 SVI->getType()->getNumElements() &&
8718 SVI->getType()->getNumElements() ==
8719 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8721 // If either of the operands is a cast from CI.getType(), then
8722 // evaluating the shuffle in the casted destination's type will allow
8723 // us to eliminate at least one cast.
8724 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8725 Tmp->getOperand(0)->getType() == DestTy) ||
8726 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8727 Tmp->getOperand(0)->getType() == DestTy)) {
8728 Value *LHS = InsertCastBefore(Instruction::BitCast,
8729 SVI->getOperand(0), DestTy, CI);
8730 Value *RHS = InsertCastBefore(Instruction::BitCast,
8731 SVI->getOperand(1), DestTy, CI);
8732 // Return a new shuffle vector. Use the same element ID's, as we
8733 // know the vector types match #elts.
8734 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8742 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8744 /// %D = select %cond, %C, %A
8746 /// %C = select %cond, %B, 0
8749 /// Assuming that the specified instruction is an operand to the select, return
8750 /// a bitmask indicating which operands of this instruction are foldable if they
8751 /// equal the other incoming value of the select.
8753 static unsigned GetSelectFoldableOperands(Instruction *I) {
8754 switch (I->getOpcode()) {
8755 case Instruction::Add:
8756 case Instruction::Mul:
8757 case Instruction::And:
8758 case Instruction::Or:
8759 case Instruction::Xor:
8760 return 3; // Can fold through either operand.
8761 case Instruction::Sub: // Can only fold on the amount subtracted.
8762 case Instruction::Shl: // Can only fold on the shift amount.
8763 case Instruction::LShr:
8764 case Instruction::AShr:
8767 return 0; // Cannot fold
8771 /// GetSelectFoldableConstant - For the same transformation as the previous
8772 /// function, return the identity constant that goes into the select.
8773 static Constant *GetSelectFoldableConstant(Instruction *I) {
8774 switch (I->getOpcode()) {
8775 default: assert(0 && "This cannot happen!"); abort();
8776 case Instruction::Add:
8777 case Instruction::Sub:
8778 case Instruction::Or:
8779 case Instruction::Xor:
8780 case Instruction::Shl:
8781 case Instruction::LShr:
8782 case Instruction::AShr:
8783 return Constant::getNullValue(I->getType());
8784 case Instruction::And:
8785 return Constant::getAllOnesValue(I->getType());
8786 case Instruction::Mul:
8787 return ConstantInt::get(I->getType(), 1);
8791 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8792 /// have the same opcode and only one use each. Try to simplify this.
8793 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8795 if (TI->getNumOperands() == 1) {
8796 // If this is a non-volatile load or a cast from the same type,
8799 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8802 return 0; // unknown unary op.
8805 // Fold this by inserting a select from the input values.
8806 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8807 FI->getOperand(0), SI.getName()+".v");
8808 InsertNewInstBefore(NewSI, SI);
8809 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8813 // Only handle binary operators here.
8814 if (!isa<BinaryOperator>(TI))
8817 // Figure out if the operations have any operands in common.
8818 Value *MatchOp, *OtherOpT, *OtherOpF;
8820 if (TI->getOperand(0) == FI->getOperand(0)) {
8821 MatchOp = TI->getOperand(0);
8822 OtherOpT = TI->getOperand(1);
8823 OtherOpF = FI->getOperand(1);
8824 MatchIsOpZero = true;
8825 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8826 MatchOp = TI->getOperand(1);
8827 OtherOpT = TI->getOperand(0);
8828 OtherOpF = FI->getOperand(0);
8829 MatchIsOpZero = false;
8830 } else if (!TI->isCommutative()) {
8832 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8833 MatchOp = TI->getOperand(0);
8834 OtherOpT = TI->getOperand(1);
8835 OtherOpF = FI->getOperand(0);
8836 MatchIsOpZero = true;
8837 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8838 MatchOp = TI->getOperand(1);
8839 OtherOpT = TI->getOperand(0);
8840 OtherOpF = FI->getOperand(1);
8841 MatchIsOpZero = true;
8846 // If we reach here, they do have operations in common.
8847 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8848 OtherOpF, SI.getName()+".v");
8849 InsertNewInstBefore(NewSI, SI);
8851 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8853 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8855 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8857 assert(0 && "Shouldn't get here");
8861 static bool isSelect01(Constant *C1, Constant *C2) {
8862 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
8865 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
8868 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
8871 /// FoldSelectIntoOp - Try fold the select into one of the operands to
8872 /// facilitate further optimization.
8873 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
8875 // See the comment above GetSelectFoldableOperands for a description of the
8876 // transformation we are doing here.
8877 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
8878 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8879 !isa<Constant>(FalseVal)) {
8880 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8881 unsigned OpToFold = 0;
8882 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8884 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8889 Constant *C = GetSelectFoldableConstant(TVI);
8890 Value *OOp = TVI->getOperand(2-OpToFold);
8891 // Avoid creating select between 2 constants unless it's selecting
8893 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
8894 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
8895 InsertNewInstBefore(NewSel, SI);
8896 NewSel->takeName(TVI);
8897 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8898 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8899 assert(0 && "Unknown instruction!!");
8906 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
8907 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8908 !isa<Constant>(TrueVal)) {
8909 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8910 unsigned OpToFold = 0;
8911 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8913 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8918 Constant *C = GetSelectFoldableConstant(FVI);
8919 Value *OOp = FVI->getOperand(2-OpToFold);
8920 // Avoid creating select between 2 constants unless it's selecting
8922 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
8923 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
8924 InsertNewInstBefore(NewSel, SI);
8925 NewSel->takeName(FVI);
8926 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8927 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8928 assert(0 && "Unknown instruction!!");
8938 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8939 /// ICmpInst as its first operand.
8941 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8943 bool Changed = false;
8944 ICmpInst::Predicate Pred = ICI->getPredicate();
8945 Value *CmpLHS = ICI->getOperand(0);
8946 Value *CmpRHS = ICI->getOperand(1);
8947 Value *TrueVal = SI.getTrueValue();
8948 Value *FalseVal = SI.getFalseValue();
8950 // Check cases where the comparison is with a constant that
8951 // can be adjusted to fit the min/max idiom. We may edit ICI in
8952 // place here, so make sure the select is the only user.
8953 if (ICI->hasOneUse())
8954 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8957 case ICmpInst::ICMP_ULT:
8958 case ICmpInst::ICMP_SLT: {
8959 // X < MIN ? T : F --> F
8960 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8961 return ReplaceInstUsesWith(SI, FalseVal);
8962 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8963 Constant *AdjustedRHS = SubOne(CI);
8964 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8965 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8966 Pred = ICmpInst::getSwappedPredicate(Pred);
8967 CmpRHS = AdjustedRHS;
8968 std::swap(FalseVal, TrueVal);
8969 ICI->setPredicate(Pred);
8970 ICI->setOperand(1, CmpRHS);
8971 SI.setOperand(1, TrueVal);
8972 SI.setOperand(2, FalseVal);
8977 case ICmpInst::ICMP_UGT:
8978 case ICmpInst::ICMP_SGT: {
8979 // X > MAX ? T : F --> F
8980 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8981 return ReplaceInstUsesWith(SI, FalseVal);
8982 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8983 Constant *AdjustedRHS = AddOne(CI);
8984 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8985 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8986 Pred = ICmpInst::getSwappedPredicate(Pred);
8987 CmpRHS = AdjustedRHS;
8988 std::swap(FalseVal, TrueVal);
8989 ICI->setPredicate(Pred);
8990 ICI->setOperand(1, CmpRHS);
8991 SI.setOperand(1, TrueVal);
8992 SI.setOperand(2, FalseVal);
8999 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9000 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9001 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9002 if (match(TrueVal, m_ConstantInt<-1>()) &&
9003 match(FalseVal, m_ConstantInt<0>()))
9004 Pred = ICI->getPredicate();
9005 else if (match(TrueVal, m_ConstantInt<0>()) &&
9006 match(FalseVal, m_ConstantInt<-1>()))
9007 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9009 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9010 // If we are just checking for a icmp eq of a single bit and zext'ing it
9011 // to an integer, then shift the bit to the appropriate place and then
9012 // cast to integer to avoid the comparison.
9013 const APInt &Op1CV = CI->getValue();
9015 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9016 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9017 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9018 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9019 Value *In = ICI->getOperand(0);
9020 Value *Sh = ConstantInt::get(In->getType(),
9021 In->getType()->getPrimitiveSizeInBits()-1);
9022 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9023 In->getName()+".lobit"),
9025 if (In->getType() != SI.getType())
9026 In = CastInst::CreateIntegerCast(In, SI.getType(),
9027 true/*SExt*/, "tmp", ICI);
9029 if (Pred == ICmpInst::ICMP_SGT)
9030 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9031 In->getName()+".not"), *ICI);
9033 return ReplaceInstUsesWith(SI, In);
9038 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9039 // Transform (X == Y) ? X : Y -> Y
9040 if (Pred == ICmpInst::ICMP_EQ)
9041 return ReplaceInstUsesWith(SI, FalseVal);
9042 // Transform (X != Y) ? X : Y -> X
9043 if (Pred == ICmpInst::ICMP_NE)
9044 return ReplaceInstUsesWith(SI, TrueVal);
9045 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9047 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9048 // Transform (X == Y) ? Y : X -> X
9049 if (Pred == ICmpInst::ICMP_EQ)
9050 return ReplaceInstUsesWith(SI, FalseVal);
9051 // Transform (X != Y) ? Y : X -> Y
9052 if (Pred == ICmpInst::ICMP_NE)
9053 return ReplaceInstUsesWith(SI, TrueVal);
9054 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9057 /// NOTE: if we wanted to, this is where to detect integer ABS
9059 return Changed ? &SI : 0;
9062 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9063 Value *CondVal = SI.getCondition();
9064 Value *TrueVal = SI.getTrueValue();
9065 Value *FalseVal = SI.getFalseValue();
9067 // select true, X, Y -> X
9068 // select false, X, Y -> Y
9069 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9070 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9072 // select C, X, X -> X
9073 if (TrueVal == FalseVal)
9074 return ReplaceInstUsesWith(SI, TrueVal);
9076 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9077 return ReplaceInstUsesWith(SI, FalseVal);
9078 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9079 return ReplaceInstUsesWith(SI, TrueVal);
9080 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9081 if (isa<Constant>(TrueVal))
9082 return ReplaceInstUsesWith(SI, TrueVal);
9084 return ReplaceInstUsesWith(SI, FalseVal);
9087 if (SI.getType() == Type::Int1Ty) {
9088 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9089 if (C->getZExtValue()) {
9090 // Change: A = select B, true, C --> A = or B, C
9091 return BinaryOperator::CreateOr(CondVal, FalseVal);
9093 // Change: A = select B, false, C --> A = and !B, C
9095 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9096 "not."+CondVal->getName()), SI);
9097 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9099 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9100 if (C->getZExtValue() == false) {
9101 // Change: A = select B, C, false --> A = and B, C
9102 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9104 // Change: A = select B, C, true --> A = or !B, C
9106 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9107 "not."+CondVal->getName()), SI);
9108 return BinaryOperator::CreateOr(NotCond, TrueVal);
9112 // select a, b, a -> a&b
9113 // select a, a, b -> a|b
9114 if (CondVal == TrueVal)
9115 return BinaryOperator::CreateOr(CondVal, FalseVal);
9116 else if (CondVal == FalseVal)
9117 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9120 // Selecting between two integer constants?
9121 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9122 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9123 // select C, 1, 0 -> zext C to int
9124 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9125 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9126 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9127 // select C, 0, 1 -> zext !C to int
9129 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9130 "not."+CondVal->getName()), SI);
9131 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9134 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9136 // (x <s 0) ? -1 : 0 -> ashr x, 31
9137 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
9138 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
9139 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
9140 // The comparison constant and the result are not neccessarily the
9141 // same width. Make an all-ones value by inserting a AShr.
9142 Value *X = IC->getOperand(0);
9143 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
9144 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
9145 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
9147 InsertNewInstBefore(SRA, SI);
9149 // Then cast to the appropriate width.
9150 return CastInst::CreateIntegerCast(SRA, SI.getType(), true);
9155 // If one of the constants is zero (we know they can't both be) and we
9156 // have an icmp instruction with zero, and we have an 'and' with the
9157 // non-constant value, eliminate this whole mess. This corresponds to
9158 // cases like this: ((X & 27) ? 27 : 0)
9159 if (TrueValC->isZero() || FalseValC->isZero())
9160 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9161 cast<Constant>(IC->getOperand(1))->isNullValue())
9162 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9163 if (ICA->getOpcode() == Instruction::And &&
9164 isa<ConstantInt>(ICA->getOperand(1)) &&
9165 (ICA->getOperand(1) == TrueValC ||
9166 ICA->getOperand(1) == FalseValC) &&
9167 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9168 // Okay, now we know that everything is set up, we just don't
9169 // know whether we have a icmp_ne or icmp_eq and whether the
9170 // true or false val is the zero.
9171 bool ShouldNotVal = !TrueValC->isZero();
9172 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9175 V = InsertNewInstBefore(BinaryOperator::Create(
9176 Instruction::Xor, V, ICA->getOperand(1)), SI);
9177 return ReplaceInstUsesWith(SI, V);
9182 // See if we are selecting two values based on a comparison of the two values.
9183 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9184 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9185 // Transform (X == Y) ? X : Y -> Y
9186 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9187 // This is not safe in general for floating point:
9188 // consider X== -0, Y== +0.
9189 // It becomes safe if either operand is a nonzero constant.
9190 ConstantFP *CFPt, *CFPf;
9191 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9192 !CFPt->getValueAPF().isZero()) ||
9193 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9194 !CFPf->getValueAPF().isZero()))
9195 return ReplaceInstUsesWith(SI, FalseVal);
9197 // Transform (X != Y) ? X : Y -> X
9198 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9199 return ReplaceInstUsesWith(SI, TrueVal);
9200 // NOTE: if we wanted to, this is where to detect MIN/MAX
9202 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9203 // Transform (X == Y) ? Y : X -> X
9204 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9205 // This is not safe in general for floating point:
9206 // consider X== -0, Y== +0.
9207 // It becomes safe if either operand is a nonzero constant.
9208 ConstantFP *CFPt, *CFPf;
9209 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9210 !CFPt->getValueAPF().isZero()) ||
9211 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9212 !CFPf->getValueAPF().isZero()))
9213 return ReplaceInstUsesWith(SI, FalseVal);
9215 // Transform (X != Y) ? Y : X -> Y
9216 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9217 return ReplaceInstUsesWith(SI, TrueVal);
9218 // NOTE: if we wanted to, this is where to detect MIN/MAX
9220 // NOTE: if we wanted to, this is where to detect ABS
9223 // See if we are selecting two values based on a comparison of the two values.
9224 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9225 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9228 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9229 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9230 if (TI->hasOneUse() && FI->hasOneUse()) {
9231 Instruction *AddOp = 0, *SubOp = 0;
9233 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9234 if (TI->getOpcode() == FI->getOpcode())
9235 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9238 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9239 // even legal for FP.
9240 if (TI->getOpcode() == Instruction::Sub &&
9241 FI->getOpcode() == Instruction::Add) {
9242 AddOp = FI; SubOp = TI;
9243 } else if (FI->getOpcode() == Instruction::Sub &&
9244 TI->getOpcode() == Instruction::Add) {
9245 AddOp = TI; SubOp = FI;
9249 Value *OtherAddOp = 0;
9250 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9251 OtherAddOp = AddOp->getOperand(1);
9252 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9253 OtherAddOp = AddOp->getOperand(0);
9257 // So at this point we know we have (Y -> OtherAddOp):
9258 // select C, (add X, Y), (sub X, Z)
9259 Value *NegVal; // Compute -Z
9260 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9261 NegVal = ConstantExpr::getNeg(C);
9263 NegVal = InsertNewInstBefore(
9264 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
9267 Value *NewTrueOp = OtherAddOp;
9268 Value *NewFalseOp = NegVal;
9270 std::swap(NewTrueOp, NewFalseOp);
9271 Instruction *NewSel =
9272 SelectInst::Create(CondVal, NewTrueOp,
9273 NewFalseOp, SI.getName() + ".p");
9275 NewSel = InsertNewInstBefore(NewSel, SI);
9276 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9281 // See if we can fold the select into one of our operands.
9282 if (SI.getType()->isInteger()) {
9283 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9288 if (BinaryOperator::isNot(CondVal)) {
9289 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9290 SI.setOperand(1, FalseVal);
9291 SI.setOperand(2, TrueVal);
9298 /// EnforceKnownAlignment - If the specified pointer points to an object that
9299 /// we control, modify the object's alignment to PrefAlign. This isn't
9300 /// often possible though. If alignment is important, a more reliable approach
9301 /// is to simply align all global variables and allocation instructions to
9302 /// their preferred alignment from the beginning.
9304 static unsigned EnforceKnownAlignment(Value *V,
9305 unsigned Align, unsigned PrefAlign) {
9307 User *U = dyn_cast<User>(V);
9308 if (!U) return Align;
9310 switch (getOpcode(U)) {
9312 case Instruction::BitCast:
9313 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9314 case Instruction::GetElementPtr: {
9315 // If all indexes are zero, it is just the alignment of the base pointer.
9316 bool AllZeroOperands = true;
9317 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9318 if (!isa<Constant>(*i) ||
9319 !cast<Constant>(*i)->isNullValue()) {
9320 AllZeroOperands = false;
9324 if (AllZeroOperands) {
9325 // Treat this like a bitcast.
9326 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9332 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9333 // If there is a large requested alignment and we can, bump up the alignment
9335 if (!GV->isDeclaration()) {
9336 if (GV->getAlignment() >= PrefAlign)
9337 Align = GV->getAlignment();
9339 GV->setAlignment(PrefAlign);
9343 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9344 // If there is a requested alignment and if this is an alloca, round up. We
9345 // don't do this for malloc, because some systems can't respect the request.
9346 if (isa<AllocaInst>(AI)) {
9347 if (AI->getAlignment() >= PrefAlign)
9348 Align = AI->getAlignment();
9350 AI->setAlignment(PrefAlign);
9359 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9360 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9361 /// and it is more than the alignment of the ultimate object, see if we can
9362 /// increase the alignment of the ultimate object, making this check succeed.
9363 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9364 unsigned PrefAlign) {
9365 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9366 sizeof(PrefAlign) * CHAR_BIT;
9367 APInt Mask = APInt::getAllOnesValue(BitWidth);
9368 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9369 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9370 unsigned TrailZ = KnownZero.countTrailingOnes();
9371 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9373 if (PrefAlign > Align)
9374 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9376 // We don't need to make any adjustment.
9380 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9381 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9382 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9383 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9384 unsigned CopyAlign = MI->getAlignment();
9386 if (CopyAlign < MinAlign) {
9387 MI->setAlignment(MinAlign);
9391 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9393 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9394 if (MemOpLength == 0) return 0;
9396 // Source and destination pointer types are always "i8*" for intrinsic. See
9397 // if the size is something we can handle with a single primitive load/store.
9398 // A single load+store correctly handles overlapping memory in the memmove
9400 unsigned Size = MemOpLength->getZExtValue();
9401 if (Size == 0) return MI; // Delete this mem transfer.
9403 if (Size > 8 || (Size&(Size-1)))
9404 return 0; // If not 1/2/4/8 bytes, exit.
9406 // Use an integer load+store unless we can find something better.
9407 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9409 // Memcpy forces the use of i8* for the source and destination. That means
9410 // that if you're using memcpy to move one double around, you'll get a cast
9411 // from double* to i8*. We'd much rather use a double load+store rather than
9412 // an i64 load+store, here because this improves the odds that the source or
9413 // dest address will be promotable. See if we can find a better type than the
9414 // integer datatype.
9415 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9416 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9417 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9418 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9419 // down through these levels if so.
9420 while (!SrcETy->isSingleValueType()) {
9421 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9422 if (STy->getNumElements() == 1)
9423 SrcETy = STy->getElementType(0);
9426 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9427 if (ATy->getNumElements() == 1)
9428 SrcETy = ATy->getElementType();
9435 if (SrcETy->isSingleValueType())
9436 NewPtrTy = PointerType::getUnqual(SrcETy);
9441 // If the memcpy/memmove provides better alignment info than we can
9443 SrcAlign = std::max(SrcAlign, CopyAlign);
9444 DstAlign = std::max(DstAlign, CopyAlign);
9446 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9447 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9448 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9449 InsertNewInstBefore(L, *MI);
9450 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9452 // Set the size of the copy to 0, it will be deleted on the next iteration.
9453 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9457 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9458 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9459 if (MI->getAlignment() < Alignment) {
9460 MI->setAlignment(Alignment);
9464 // Extract the length and alignment and fill if they are constant.
9465 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9466 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9467 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9469 uint64_t Len = LenC->getZExtValue();
9470 Alignment = MI->getAlignment();
9472 // If the length is zero, this is a no-op
9473 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9475 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9476 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9477 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9479 Value *Dest = MI->getDest();
9480 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9482 // Alignment 0 is identity for alignment 1 for memset, but not store.
9483 if (Alignment == 0) Alignment = 1;
9485 // Extract the fill value and store.
9486 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9487 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9490 // Set the size of the copy to 0, it will be deleted on the next iteration.
9491 MI->setLength(Constant::getNullValue(LenC->getType()));
9499 /// visitCallInst - CallInst simplification. This mostly only handles folding
9500 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9501 /// the heavy lifting.
9503 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9504 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9505 if (!II) return visitCallSite(&CI);
9507 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9509 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9510 bool Changed = false;
9512 // memmove/cpy/set of zero bytes is a noop.
9513 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9514 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9516 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9517 if (CI->getZExtValue() == 1) {
9518 // Replace the instruction with just byte operations. We would
9519 // transform other cases to loads/stores, but we don't know if
9520 // alignment is sufficient.
9524 // If we have a memmove and the source operation is a constant global,
9525 // then the source and dest pointers can't alias, so we can change this
9526 // into a call to memcpy.
9527 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9528 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9529 if (GVSrc->isConstant()) {
9530 Module *M = CI.getParent()->getParent()->getParent();
9531 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9533 Tys[0] = CI.getOperand(3)->getType();
9535 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9539 // memmove(x,x,size) -> noop.
9540 if (MMI->getSource() == MMI->getDest())
9541 return EraseInstFromFunction(CI);
9544 // If we can determine a pointer alignment that is bigger than currently
9545 // set, update the alignment.
9546 if (isa<MemTransferInst>(MI)) {
9547 if (Instruction *I = SimplifyMemTransfer(MI))
9549 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9550 if (Instruction *I = SimplifyMemSet(MSI))
9554 if (Changed) return II;
9557 switch (II->getIntrinsicID()) {
9559 case Intrinsic::bswap:
9560 // bswap(bswap(x)) -> x
9561 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9562 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9563 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9565 case Intrinsic::ppc_altivec_lvx:
9566 case Intrinsic::ppc_altivec_lvxl:
9567 case Intrinsic::x86_sse_loadu_ps:
9568 case Intrinsic::x86_sse2_loadu_pd:
9569 case Intrinsic::x86_sse2_loadu_dq:
9570 // Turn PPC lvx -> load if the pointer is known aligned.
9571 // Turn X86 loadups -> load if the pointer is known aligned.
9572 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9573 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9574 PointerType::getUnqual(II->getType()),
9576 return new LoadInst(Ptr);
9579 case Intrinsic::ppc_altivec_stvx:
9580 case Intrinsic::ppc_altivec_stvxl:
9581 // Turn stvx -> store if the pointer is known aligned.
9582 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9583 const Type *OpPtrTy =
9584 PointerType::getUnqual(II->getOperand(1)->getType());
9585 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9586 return new StoreInst(II->getOperand(1), Ptr);
9589 case Intrinsic::x86_sse_storeu_ps:
9590 case Intrinsic::x86_sse2_storeu_pd:
9591 case Intrinsic::x86_sse2_storeu_dq:
9592 // Turn X86 storeu -> store if the pointer is known aligned.
9593 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9594 const Type *OpPtrTy =
9595 PointerType::getUnqual(II->getOperand(2)->getType());
9596 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9597 return new StoreInst(II->getOperand(2), Ptr);
9601 case Intrinsic::x86_sse_cvttss2si: {
9602 // These intrinsics only demands the 0th element of its input vector. If
9603 // we can simplify the input based on that, do so now.
9605 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9606 APInt DemandedElts(VWidth, 1);
9607 APInt UndefElts(VWidth, 0);
9608 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9610 II->setOperand(1, V);
9616 case Intrinsic::ppc_altivec_vperm:
9617 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9618 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9619 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9621 // Check that all of the elements are integer constants or undefs.
9622 bool AllEltsOk = true;
9623 for (unsigned i = 0; i != 16; ++i) {
9624 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9625 !isa<UndefValue>(Mask->getOperand(i))) {
9632 // Cast the input vectors to byte vectors.
9633 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9634 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9635 Value *Result = UndefValue::get(Op0->getType());
9637 // Only extract each element once.
9638 Value *ExtractedElts[32];
9639 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9641 for (unsigned i = 0; i != 16; ++i) {
9642 if (isa<UndefValue>(Mask->getOperand(i)))
9644 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9645 Idx &= 31; // Match the hardware behavior.
9647 if (ExtractedElts[Idx] == 0) {
9649 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9650 InsertNewInstBefore(Elt, CI);
9651 ExtractedElts[Idx] = Elt;
9654 // Insert this value into the result vector.
9655 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9657 InsertNewInstBefore(cast<Instruction>(Result), CI);
9659 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9664 case Intrinsic::stackrestore: {
9665 // If the save is right next to the restore, remove the restore. This can
9666 // happen when variable allocas are DCE'd.
9667 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9668 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9669 BasicBlock::iterator BI = SS;
9671 return EraseInstFromFunction(CI);
9675 // Scan down this block to see if there is another stack restore in the
9676 // same block without an intervening call/alloca.
9677 BasicBlock::iterator BI = II;
9678 TerminatorInst *TI = II->getParent()->getTerminator();
9679 bool CannotRemove = false;
9680 for (++BI; &*BI != TI; ++BI) {
9681 if (isa<AllocaInst>(BI)) {
9682 CannotRemove = true;
9685 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9686 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9687 // If there is a stackrestore below this one, remove this one.
9688 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9689 return EraseInstFromFunction(CI);
9690 // Otherwise, ignore the intrinsic.
9692 // If we found a non-intrinsic call, we can't remove the stack
9694 CannotRemove = true;
9700 // If the stack restore is in a return/unwind block and if there are no
9701 // allocas or calls between the restore and the return, nuke the restore.
9702 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9703 return EraseInstFromFunction(CI);
9708 return visitCallSite(II);
9711 // InvokeInst simplification
9713 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9714 return visitCallSite(&II);
9717 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9718 /// passed through the varargs area, we can eliminate the use of the cast.
9719 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9720 const CastInst * const CI,
9721 const TargetData * const TD,
9723 if (!CI->isLosslessCast())
9726 // The size of ByVal arguments is derived from the type, so we
9727 // can't change to a type with a different size. If the size were
9728 // passed explicitly we could avoid this check.
9729 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9733 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9734 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9735 if (!SrcTy->isSized() || !DstTy->isSized())
9737 if (TD->getTypePaddedSize(SrcTy) != TD->getTypePaddedSize(DstTy))
9742 // visitCallSite - Improvements for call and invoke instructions.
9744 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9745 bool Changed = false;
9747 // If the callee is a constexpr cast of a function, attempt to move the cast
9748 // to the arguments of the call/invoke.
9749 if (transformConstExprCastCall(CS)) return 0;
9751 Value *Callee = CS.getCalledValue();
9753 if (Function *CalleeF = dyn_cast<Function>(Callee))
9754 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9755 Instruction *OldCall = CS.getInstruction();
9756 // If the call and callee calling conventions don't match, this call must
9757 // be unreachable, as the call is undefined.
9758 new StoreInst(ConstantInt::getTrue(),
9759 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9761 if (!OldCall->use_empty())
9762 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9763 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9764 return EraseInstFromFunction(*OldCall);
9768 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9769 // This instruction is not reachable, just remove it. We insert a store to
9770 // undef so that we know that this code is not reachable, despite the fact
9771 // that we can't modify the CFG here.
9772 new StoreInst(ConstantInt::getTrue(),
9773 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9774 CS.getInstruction());
9776 if (!CS.getInstruction()->use_empty())
9777 CS.getInstruction()->
9778 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9780 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9781 // Don't break the CFG, insert a dummy cond branch.
9782 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9783 ConstantInt::getTrue(), II);
9785 return EraseInstFromFunction(*CS.getInstruction());
9788 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9789 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9790 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9791 return transformCallThroughTrampoline(CS);
9793 const PointerType *PTy = cast<PointerType>(Callee->getType());
9794 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9795 if (FTy->isVarArg()) {
9796 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9797 // See if we can optimize any arguments passed through the varargs area of
9799 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9800 E = CS.arg_end(); I != E; ++I, ++ix) {
9801 CastInst *CI = dyn_cast<CastInst>(*I);
9802 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9803 *I = CI->getOperand(0);
9809 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9810 // Inline asm calls cannot throw - mark them 'nounwind'.
9811 CS.setDoesNotThrow();
9815 return Changed ? CS.getInstruction() : 0;
9818 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9819 // attempt to move the cast to the arguments of the call/invoke.
9821 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9822 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9823 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9824 if (CE->getOpcode() != Instruction::BitCast ||
9825 !isa<Function>(CE->getOperand(0)))
9827 Function *Callee = cast<Function>(CE->getOperand(0));
9828 Instruction *Caller = CS.getInstruction();
9829 const AttrListPtr &CallerPAL = CS.getAttributes();
9831 // Okay, this is a cast from a function to a different type. Unless doing so
9832 // would cause a type conversion of one of our arguments, change this call to
9833 // be a direct call with arguments casted to the appropriate types.
9835 const FunctionType *FT = Callee->getFunctionType();
9836 const Type *OldRetTy = Caller->getType();
9837 const Type *NewRetTy = FT->getReturnType();
9839 if (isa<StructType>(NewRetTy))
9840 return false; // TODO: Handle multiple return values.
9842 // Check to see if we are changing the return type...
9843 if (OldRetTy != NewRetTy) {
9844 if (Callee->isDeclaration() &&
9845 // Conversion is ok if changing from one pointer type to another or from
9846 // a pointer to an integer of the same size.
9847 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9848 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9849 return false; // Cannot transform this return value.
9851 if (!Caller->use_empty() &&
9852 // void -> non-void is handled specially
9853 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9854 return false; // Cannot transform this return value.
9856 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9857 Attributes RAttrs = CallerPAL.getRetAttributes();
9858 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9859 return false; // Attribute not compatible with transformed value.
9862 // If the callsite is an invoke instruction, and the return value is used by
9863 // a PHI node in a successor, we cannot change the return type of the call
9864 // because there is no place to put the cast instruction (without breaking
9865 // the critical edge). Bail out in this case.
9866 if (!Caller->use_empty())
9867 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9868 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9870 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9871 if (PN->getParent() == II->getNormalDest() ||
9872 PN->getParent() == II->getUnwindDest())
9876 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9877 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9879 CallSite::arg_iterator AI = CS.arg_begin();
9880 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9881 const Type *ParamTy = FT->getParamType(i);
9882 const Type *ActTy = (*AI)->getType();
9884 if (!CastInst::isCastable(ActTy, ParamTy))
9885 return false; // Cannot transform this parameter value.
9887 if (CallerPAL.getParamAttributes(i + 1)
9888 & Attribute::typeIncompatible(ParamTy))
9889 return false; // Attribute not compatible with transformed value.
9891 // Converting from one pointer type to another or between a pointer and an
9892 // integer of the same size is safe even if we do not have a body.
9893 bool isConvertible = ActTy == ParamTy ||
9894 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9895 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9896 if (Callee->isDeclaration() && !isConvertible) return false;
9899 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9900 Callee->isDeclaration())
9901 return false; // Do not delete arguments unless we have a function body.
9903 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9904 !CallerPAL.isEmpty())
9905 // In this case we have more arguments than the new function type, but we
9906 // won't be dropping them. Check that these extra arguments have attributes
9907 // that are compatible with being a vararg call argument.
9908 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9909 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9911 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9912 if (PAttrs & Attribute::VarArgsIncompatible)
9916 // Okay, we decided that this is a safe thing to do: go ahead and start
9917 // inserting cast instructions as necessary...
9918 std::vector<Value*> Args;
9919 Args.reserve(NumActualArgs);
9920 SmallVector<AttributeWithIndex, 8> attrVec;
9921 attrVec.reserve(NumCommonArgs);
9923 // Get any return attributes.
9924 Attributes RAttrs = CallerPAL.getRetAttributes();
9926 // If the return value is not being used, the type may not be compatible
9927 // with the existing attributes. Wipe out any problematic attributes.
9928 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9930 // Add the new return attributes.
9932 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9934 AI = CS.arg_begin();
9935 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9936 const Type *ParamTy = FT->getParamType(i);
9937 if ((*AI)->getType() == ParamTy) {
9938 Args.push_back(*AI);
9940 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9941 false, ParamTy, false);
9942 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9943 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9946 // Add any parameter attributes.
9947 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9948 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9951 // If the function takes more arguments than the call was taking, add them
9953 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9954 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9956 // If we are removing arguments to the function, emit an obnoxious warning...
9957 if (FT->getNumParams() < NumActualArgs) {
9958 if (!FT->isVarArg()) {
9959 cerr << "WARNING: While resolving call to function '"
9960 << Callee->getName() << "' arguments were dropped!\n";
9962 // Add all of the arguments in their promoted form to the arg list...
9963 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9964 const Type *PTy = getPromotedType((*AI)->getType());
9965 if (PTy != (*AI)->getType()) {
9966 // Must promote to pass through va_arg area!
9967 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9969 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9970 InsertNewInstBefore(Cast, *Caller);
9971 Args.push_back(Cast);
9973 Args.push_back(*AI);
9976 // Add any parameter attributes.
9977 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9978 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9983 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9984 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9986 if (NewRetTy == Type::VoidTy)
9987 Caller->setName(""); // Void type should not have a name.
9989 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9992 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9993 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9994 Args.begin(), Args.end(),
9995 Caller->getName(), Caller);
9996 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9997 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9999 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10000 Caller->getName(), Caller);
10001 CallInst *CI = cast<CallInst>(Caller);
10002 if (CI->isTailCall())
10003 cast<CallInst>(NC)->setTailCall();
10004 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10005 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10008 // Insert a cast of the return type as necessary.
10010 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10011 if (NV->getType() != Type::VoidTy) {
10012 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10014 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10016 // If this is an invoke instruction, we should insert it after the first
10017 // non-phi, instruction in the normal successor block.
10018 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10019 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10020 InsertNewInstBefore(NC, *I);
10022 // Otherwise, it's a call, just insert cast right after the call instr
10023 InsertNewInstBefore(NC, *Caller);
10025 AddUsersToWorkList(*Caller);
10027 NV = UndefValue::get(Caller->getType());
10031 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10032 Caller->replaceAllUsesWith(NV);
10033 Caller->eraseFromParent();
10034 RemoveFromWorkList(Caller);
10038 // transformCallThroughTrampoline - Turn a call to a function created by the
10039 // init_trampoline intrinsic into a direct call to the underlying function.
10041 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10042 Value *Callee = CS.getCalledValue();
10043 const PointerType *PTy = cast<PointerType>(Callee->getType());
10044 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10045 const AttrListPtr &Attrs = CS.getAttributes();
10047 // If the call already has the 'nest' attribute somewhere then give up -
10048 // otherwise 'nest' would occur twice after splicing in the chain.
10049 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10052 IntrinsicInst *Tramp =
10053 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10055 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10056 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10057 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10059 const AttrListPtr &NestAttrs = NestF->getAttributes();
10060 if (!NestAttrs.isEmpty()) {
10061 unsigned NestIdx = 1;
10062 const Type *NestTy = 0;
10063 Attributes NestAttr = Attribute::None;
10065 // Look for a parameter marked with the 'nest' attribute.
10066 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10067 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10068 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10069 // Record the parameter type and any other attributes.
10071 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10076 Instruction *Caller = CS.getInstruction();
10077 std::vector<Value*> NewArgs;
10078 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10080 SmallVector<AttributeWithIndex, 8> NewAttrs;
10081 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10083 // Insert the nest argument into the call argument list, which may
10084 // mean appending it. Likewise for attributes.
10086 // Add any result attributes.
10087 if (Attributes Attr = Attrs.getRetAttributes())
10088 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10092 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10094 if (Idx == NestIdx) {
10095 // Add the chain argument and attributes.
10096 Value *NestVal = Tramp->getOperand(3);
10097 if (NestVal->getType() != NestTy)
10098 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10099 NewArgs.push_back(NestVal);
10100 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10106 // Add the original argument and attributes.
10107 NewArgs.push_back(*I);
10108 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10110 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10116 // Add any function attributes.
10117 if (Attributes Attr = Attrs.getFnAttributes())
10118 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10120 // The trampoline may have been bitcast to a bogus type (FTy).
10121 // Handle this by synthesizing a new function type, equal to FTy
10122 // with the chain parameter inserted.
10124 std::vector<const Type*> NewTypes;
10125 NewTypes.reserve(FTy->getNumParams()+1);
10127 // Insert the chain's type into the list of parameter types, which may
10128 // mean appending it.
10131 FunctionType::param_iterator I = FTy->param_begin(),
10132 E = FTy->param_end();
10135 if (Idx == NestIdx)
10136 // Add the chain's type.
10137 NewTypes.push_back(NestTy);
10142 // Add the original type.
10143 NewTypes.push_back(*I);
10149 // Replace the trampoline call with a direct call. Let the generic
10150 // code sort out any function type mismatches.
10151 FunctionType *NewFTy =
10152 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
10153 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
10154 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
10155 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10157 Instruction *NewCaller;
10158 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10159 NewCaller = InvokeInst::Create(NewCallee,
10160 II->getNormalDest(), II->getUnwindDest(),
10161 NewArgs.begin(), NewArgs.end(),
10162 Caller->getName(), Caller);
10163 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10164 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10166 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10167 Caller->getName(), Caller);
10168 if (cast<CallInst>(Caller)->isTailCall())
10169 cast<CallInst>(NewCaller)->setTailCall();
10170 cast<CallInst>(NewCaller)->
10171 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10172 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10174 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10175 Caller->replaceAllUsesWith(NewCaller);
10176 Caller->eraseFromParent();
10177 RemoveFromWorkList(Caller);
10182 // Replace the trampoline call with a direct call. Since there is no 'nest'
10183 // parameter, there is no need to adjust the argument list. Let the generic
10184 // code sort out any function type mismatches.
10185 Constant *NewCallee =
10186 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
10187 CS.setCalledFunction(NewCallee);
10188 return CS.getInstruction();
10191 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10192 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10193 /// and a single binop.
10194 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10195 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10196 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10197 unsigned Opc = FirstInst->getOpcode();
10198 Value *LHSVal = FirstInst->getOperand(0);
10199 Value *RHSVal = FirstInst->getOperand(1);
10201 const Type *LHSType = LHSVal->getType();
10202 const Type *RHSType = RHSVal->getType();
10204 // Scan to see if all operands are the same opcode, all have one use, and all
10205 // kill their operands (i.e. the operands have one use).
10206 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10207 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10208 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10209 // Verify type of the LHS matches so we don't fold cmp's of different
10210 // types or GEP's with different index types.
10211 I->getOperand(0)->getType() != LHSType ||
10212 I->getOperand(1)->getType() != RHSType)
10215 // If they are CmpInst instructions, check their predicates
10216 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10217 if (cast<CmpInst>(I)->getPredicate() !=
10218 cast<CmpInst>(FirstInst)->getPredicate())
10221 // Keep track of which operand needs a phi node.
10222 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10223 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10226 // Otherwise, this is safe to transform!
10228 Value *InLHS = FirstInst->getOperand(0);
10229 Value *InRHS = FirstInst->getOperand(1);
10230 PHINode *NewLHS = 0, *NewRHS = 0;
10232 NewLHS = PHINode::Create(LHSType,
10233 FirstInst->getOperand(0)->getName() + ".pn");
10234 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10235 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10236 InsertNewInstBefore(NewLHS, PN);
10241 NewRHS = PHINode::Create(RHSType,
10242 FirstInst->getOperand(1)->getName() + ".pn");
10243 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10244 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10245 InsertNewInstBefore(NewRHS, PN);
10249 // Add all operands to the new PHIs.
10250 if (NewLHS || NewRHS) {
10251 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10252 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10254 Value *NewInLHS = InInst->getOperand(0);
10255 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10258 Value *NewInRHS = InInst->getOperand(1);
10259 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10264 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10265 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10266 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10267 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
10271 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10272 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10274 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10275 FirstInst->op_end());
10276 // This is true if all GEP bases are allocas and if all indices into them are
10278 bool AllBasePointersAreAllocas = true;
10280 // Scan to see if all operands are the same opcode, all have one use, and all
10281 // kill their operands (i.e. the operands have one use).
10282 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10283 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10284 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10285 GEP->getNumOperands() != FirstInst->getNumOperands())
10288 // Keep track of whether or not all GEPs are of alloca pointers.
10289 if (AllBasePointersAreAllocas &&
10290 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10291 !GEP->hasAllConstantIndices()))
10292 AllBasePointersAreAllocas = false;
10294 // Compare the operand lists.
10295 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10296 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10299 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10300 // if one of the PHIs has a constant for the index. The index may be
10301 // substantially cheaper to compute for the constants, so making it a
10302 // variable index could pessimize the path. This also handles the case
10303 // for struct indices, which must always be constant.
10304 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10305 isa<ConstantInt>(GEP->getOperand(op)))
10308 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10310 FixedOperands[op] = 0; // Needs a PHI.
10314 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10315 // bother doing this transformation. At best, this will just save a bit of
10316 // offset calculation, but all the predecessors will have to materialize the
10317 // stack address into a register anyway. We'd actually rather *clone* the
10318 // load up into the predecessors so that we have a load of a gep of an alloca,
10319 // which can usually all be folded into the load.
10320 if (AllBasePointersAreAllocas)
10323 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10324 // that is variable.
10325 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10327 bool HasAnyPHIs = false;
10328 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10329 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10330 Value *FirstOp = FirstInst->getOperand(i);
10331 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10332 FirstOp->getName()+".pn");
10333 InsertNewInstBefore(NewPN, PN);
10335 NewPN->reserveOperandSpace(e);
10336 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10337 OperandPhis[i] = NewPN;
10338 FixedOperands[i] = NewPN;
10343 // Add all operands to the new PHIs.
10345 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10346 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10347 BasicBlock *InBB = PN.getIncomingBlock(i);
10349 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10350 if (PHINode *OpPhi = OperandPhis[op])
10351 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10355 Value *Base = FixedOperands[0];
10356 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10357 FixedOperands.end());
10361 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10362 /// sink the load out of the block that defines it. This means that it must be
10363 /// obvious the value of the load is not changed from the point of the load to
10364 /// the end of the block it is in.
10366 /// Finally, it is safe, but not profitable, to sink a load targetting a
10367 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10369 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10370 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10372 for (++BBI; BBI != E; ++BBI)
10373 if (BBI->mayWriteToMemory())
10376 // Check for non-address taken alloca. If not address-taken already, it isn't
10377 // profitable to do this xform.
10378 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10379 bool isAddressTaken = false;
10380 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10382 if (isa<LoadInst>(UI)) continue;
10383 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10384 // If storing TO the alloca, then the address isn't taken.
10385 if (SI->getOperand(1) == AI) continue;
10387 isAddressTaken = true;
10391 if (!isAddressTaken && AI->isStaticAlloca())
10395 // If this load is a load from a GEP with a constant offset from an alloca,
10396 // then we don't want to sink it. In its present form, it will be
10397 // load [constant stack offset]. Sinking it will cause us to have to
10398 // materialize the stack addresses in each predecessor in a register only to
10399 // do a shared load from register in the successor.
10400 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10401 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10402 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10409 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10410 // operator and they all are only used by the PHI, PHI together their
10411 // inputs, and do the operation once, to the result of the PHI.
10412 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10413 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10415 // Scan the instruction, looking for input operations that can be folded away.
10416 // If all input operands to the phi are the same instruction (e.g. a cast from
10417 // the same type or "+42") we can pull the operation through the PHI, reducing
10418 // code size and simplifying code.
10419 Constant *ConstantOp = 0;
10420 const Type *CastSrcTy = 0;
10421 bool isVolatile = false;
10422 if (isa<CastInst>(FirstInst)) {
10423 CastSrcTy = FirstInst->getOperand(0)->getType();
10424 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10425 // Can fold binop, compare or shift here if the RHS is a constant,
10426 // otherwise call FoldPHIArgBinOpIntoPHI.
10427 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10428 if (ConstantOp == 0)
10429 return FoldPHIArgBinOpIntoPHI(PN);
10430 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10431 isVolatile = LI->isVolatile();
10432 // We can't sink the load if the loaded value could be modified between the
10433 // load and the PHI.
10434 if (LI->getParent() != PN.getIncomingBlock(0) ||
10435 !isSafeAndProfitableToSinkLoad(LI))
10438 // If the PHI is of volatile loads and the load block has multiple
10439 // successors, sinking it would remove a load of the volatile value from
10440 // the path through the other successor.
10442 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10445 } else if (isa<GetElementPtrInst>(FirstInst)) {
10446 return FoldPHIArgGEPIntoPHI(PN);
10448 return 0; // Cannot fold this operation.
10451 // Check to see if all arguments are the same operation.
10452 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10453 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10454 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10455 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10458 if (I->getOperand(0)->getType() != CastSrcTy)
10459 return 0; // Cast operation must match.
10460 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10461 // We can't sink the load if the loaded value could be modified between
10462 // the load and the PHI.
10463 if (LI->isVolatile() != isVolatile ||
10464 LI->getParent() != PN.getIncomingBlock(i) ||
10465 !isSafeAndProfitableToSinkLoad(LI))
10468 // If the PHI is of volatile loads and the load block has multiple
10469 // successors, sinking it would remove a load of the volatile value from
10470 // the path through the other successor.
10472 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10475 } else if (I->getOperand(1) != ConstantOp) {
10480 // Okay, they are all the same operation. Create a new PHI node of the
10481 // correct type, and PHI together all of the LHS's of the instructions.
10482 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10483 PN.getName()+".in");
10484 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10486 Value *InVal = FirstInst->getOperand(0);
10487 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10489 // Add all operands to the new PHI.
10490 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10491 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10492 if (NewInVal != InVal)
10494 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10499 // The new PHI unions all of the same values together. This is really
10500 // common, so we handle it intelligently here for compile-time speed.
10504 InsertNewInstBefore(NewPN, PN);
10508 // Insert and return the new operation.
10509 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10510 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10511 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10512 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10513 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10514 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10515 PhiVal, ConstantOp);
10516 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10518 // If this was a volatile load that we are merging, make sure to loop through
10519 // and mark all the input loads as non-volatile. If we don't do this, we will
10520 // insert a new volatile load and the old ones will not be deletable.
10522 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10523 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10525 return new LoadInst(PhiVal, "", isVolatile);
10528 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10530 static bool DeadPHICycle(PHINode *PN,
10531 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10532 if (PN->use_empty()) return true;
10533 if (!PN->hasOneUse()) return false;
10535 // Remember this node, and if we find the cycle, return.
10536 if (!PotentiallyDeadPHIs.insert(PN))
10539 // Don't scan crazily complex things.
10540 if (PotentiallyDeadPHIs.size() == 16)
10543 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10544 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10549 /// PHIsEqualValue - Return true if this phi node is always equal to
10550 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10551 /// z = some value; x = phi (y, z); y = phi (x, z)
10552 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10553 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10554 // See if we already saw this PHI node.
10555 if (!ValueEqualPHIs.insert(PN))
10558 // Don't scan crazily complex things.
10559 if (ValueEqualPHIs.size() == 16)
10562 // Scan the operands to see if they are either phi nodes or are equal to
10564 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10565 Value *Op = PN->getIncomingValue(i);
10566 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10567 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10569 } else if (Op != NonPhiInVal)
10577 // PHINode simplification
10579 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10580 // If LCSSA is around, don't mess with Phi nodes
10581 if (MustPreserveLCSSA) return 0;
10583 if (Value *V = PN.hasConstantValue())
10584 return ReplaceInstUsesWith(PN, V);
10586 // If all PHI operands are the same operation, pull them through the PHI,
10587 // reducing code size.
10588 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10589 isa<Instruction>(PN.getIncomingValue(1)) &&
10590 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10591 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10592 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10593 // than themselves more than once.
10594 PN.getIncomingValue(0)->hasOneUse())
10595 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10598 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10599 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10600 // PHI)... break the cycle.
10601 if (PN.hasOneUse()) {
10602 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10603 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10604 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10605 PotentiallyDeadPHIs.insert(&PN);
10606 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10607 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10610 // If this phi has a single use, and if that use just computes a value for
10611 // the next iteration of a loop, delete the phi. This occurs with unused
10612 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10613 // common case here is good because the only other things that catch this
10614 // are induction variable analysis (sometimes) and ADCE, which is only run
10616 if (PHIUser->hasOneUse() &&
10617 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10618 PHIUser->use_back() == &PN) {
10619 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10623 // We sometimes end up with phi cycles that non-obviously end up being the
10624 // same value, for example:
10625 // z = some value; x = phi (y, z); y = phi (x, z)
10626 // where the phi nodes don't necessarily need to be in the same block. Do a
10627 // quick check to see if the PHI node only contains a single non-phi value, if
10628 // so, scan to see if the phi cycle is actually equal to that value.
10630 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10631 // Scan for the first non-phi operand.
10632 while (InValNo != NumOperandVals &&
10633 isa<PHINode>(PN.getIncomingValue(InValNo)))
10636 if (InValNo != NumOperandVals) {
10637 Value *NonPhiInVal = PN.getOperand(InValNo);
10639 // Scan the rest of the operands to see if there are any conflicts, if so
10640 // there is no need to recursively scan other phis.
10641 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10642 Value *OpVal = PN.getIncomingValue(InValNo);
10643 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10647 // If we scanned over all operands, then we have one unique value plus
10648 // phi values. Scan PHI nodes to see if they all merge in each other or
10650 if (InValNo == NumOperandVals) {
10651 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10652 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10653 return ReplaceInstUsesWith(PN, NonPhiInVal);
10660 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10661 Instruction *InsertPoint,
10662 InstCombiner *IC) {
10663 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10664 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10665 // We must cast correctly to the pointer type. Ensure that we
10666 // sign extend the integer value if it is smaller as this is
10667 // used for address computation.
10668 Instruction::CastOps opcode =
10669 (VTySize < PtrSize ? Instruction::SExt :
10670 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10671 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10675 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10676 Value *PtrOp = GEP.getOperand(0);
10677 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10678 // If so, eliminate the noop.
10679 if (GEP.getNumOperands() == 1)
10680 return ReplaceInstUsesWith(GEP, PtrOp);
10682 if (isa<UndefValue>(GEP.getOperand(0)))
10683 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10685 bool HasZeroPointerIndex = false;
10686 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10687 HasZeroPointerIndex = C->isNullValue();
10689 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10690 return ReplaceInstUsesWith(GEP, PtrOp);
10692 // Eliminate unneeded casts for indices.
10693 bool MadeChange = false;
10695 gep_type_iterator GTI = gep_type_begin(GEP);
10696 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10697 i != e; ++i, ++GTI) {
10698 if (isa<SequentialType>(*GTI)) {
10699 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10700 if (CI->getOpcode() == Instruction::ZExt ||
10701 CI->getOpcode() == Instruction::SExt) {
10702 const Type *SrcTy = CI->getOperand(0)->getType();
10703 // We can eliminate a cast from i32 to i64 iff the target
10704 // is a 32-bit pointer target.
10705 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10707 *i = CI->getOperand(0);
10711 // If we are using a wider index than needed for this platform, shrink it
10712 // to what we need. If narrower, sign-extend it to what we need.
10713 // If the incoming value needs a cast instruction,
10714 // insert it. This explicit cast can make subsequent optimizations more
10717 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10718 if (Constant *C = dyn_cast<Constant>(Op)) {
10719 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10722 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10727 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10728 if (Constant *C = dyn_cast<Constant>(Op)) {
10729 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10732 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10740 if (MadeChange) return &GEP;
10742 // Combine Indices - If the source pointer to this getelementptr instruction
10743 // is a getelementptr instruction, combine the indices of the two
10744 // getelementptr instructions into a single instruction.
10746 SmallVector<Value*, 8> SrcGEPOperands;
10747 if (User *Src = dyn_castGetElementPtr(PtrOp))
10748 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10750 if (!SrcGEPOperands.empty()) {
10751 // Note that if our source is a gep chain itself that we wait for that
10752 // chain to be resolved before we perform this transformation. This
10753 // avoids us creating a TON of code in some cases.
10755 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10756 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10757 return 0; // Wait until our source is folded to completion.
10759 SmallVector<Value*, 8> Indices;
10761 // Find out whether the last index in the source GEP is a sequential idx.
10762 bool EndsWithSequential = false;
10763 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10764 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10765 EndsWithSequential = !isa<StructType>(*I);
10767 // Can we combine the two pointer arithmetics offsets?
10768 if (EndsWithSequential) {
10769 // Replace: gep (gep %P, long B), long A, ...
10770 // With: T = long A+B; gep %P, T, ...
10772 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10773 if (SO1 == Constant::getNullValue(SO1->getType())) {
10775 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10778 // If they aren't the same type, convert both to an integer of the
10779 // target's pointer size.
10780 if (SO1->getType() != GO1->getType()) {
10781 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10782 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10783 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10784 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10786 unsigned PS = TD->getPointerSizeInBits();
10787 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10788 // Convert GO1 to SO1's type.
10789 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10791 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10792 // Convert SO1 to GO1's type.
10793 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10795 const Type *PT = TD->getIntPtrType();
10796 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10797 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10801 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10802 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10804 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10805 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10809 // Recycle the GEP we already have if possible.
10810 if (SrcGEPOperands.size() == 2) {
10811 GEP.setOperand(0, SrcGEPOperands[0]);
10812 GEP.setOperand(1, Sum);
10815 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10816 SrcGEPOperands.end()-1);
10817 Indices.push_back(Sum);
10818 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10820 } else if (isa<Constant>(*GEP.idx_begin()) &&
10821 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10822 SrcGEPOperands.size() != 1) {
10823 // Otherwise we can do the fold if the first index of the GEP is a zero
10824 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10825 SrcGEPOperands.end());
10826 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10829 if (!Indices.empty())
10830 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10831 Indices.end(), GEP.getName());
10833 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10834 // GEP of global variable. If all of the indices for this GEP are
10835 // constants, we can promote this to a constexpr instead of an instruction.
10837 // Scan for nonconstants...
10838 SmallVector<Constant*, 8> Indices;
10839 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10840 for (; I != E && isa<Constant>(*I); ++I)
10841 Indices.push_back(cast<Constant>(*I));
10843 if (I == E) { // If they are all constants...
10844 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10845 &Indices[0],Indices.size());
10847 // Replace all uses of the GEP with the new constexpr...
10848 return ReplaceInstUsesWith(GEP, CE);
10850 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10851 if (!isa<PointerType>(X->getType())) {
10852 // Not interesting. Source pointer must be a cast from pointer.
10853 } else if (HasZeroPointerIndex) {
10854 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10855 // into : GEP [10 x i8]* X, i32 0, ...
10857 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
10858 // into : GEP i8* X, ...
10860 // This occurs when the program declares an array extern like "int X[];"
10861 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10862 const PointerType *XTy = cast<PointerType>(X->getType());
10863 if (const ArrayType *CATy =
10864 dyn_cast<ArrayType>(CPTy->getElementType())) {
10865 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
10866 if (CATy->getElementType() == XTy->getElementType()) {
10867 // -> GEP i8* X, ...
10868 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
10869 return GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
10871 } else if (const ArrayType *XATy =
10872 dyn_cast<ArrayType>(XTy->getElementType())) {
10873 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
10874 if (CATy->getElementType() == XATy->getElementType()) {
10875 // -> GEP [10 x i8]* X, i32 0, ...
10876 // At this point, we know that the cast source type is a pointer
10877 // to an array of the same type as the destination pointer
10878 // array. Because the array type is never stepped over (there
10879 // is a leading zero) we can fold the cast into this GEP.
10880 GEP.setOperand(0, X);
10885 } else if (GEP.getNumOperands() == 2) {
10886 // Transform things like:
10887 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10888 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10889 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10890 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10891 if (isa<ArrayType>(SrcElTy) &&
10892 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10893 TD->getTypePaddedSize(ResElTy)) {
10895 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10896 Idx[1] = GEP.getOperand(1);
10897 Value *V = InsertNewInstBefore(
10898 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10899 // V and GEP are both pointer types --> BitCast
10900 return new BitCastInst(V, GEP.getType());
10903 // Transform things like:
10904 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10905 // (where tmp = 8*tmp2) into:
10906 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10908 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10909 uint64_t ArrayEltSize =
10910 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType());
10912 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10913 // allow either a mul, shift, or constant here.
10915 ConstantInt *Scale = 0;
10916 if (ArrayEltSize == 1) {
10917 NewIdx = GEP.getOperand(1);
10918 Scale = ConstantInt::get(NewIdx->getType(), 1);
10919 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10920 NewIdx = ConstantInt::get(CI->getType(), 1);
10922 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10923 if (Inst->getOpcode() == Instruction::Shl &&
10924 isa<ConstantInt>(Inst->getOperand(1))) {
10925 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10926 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10927 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10928 NewIdx = Inst->getOperand(0);
10929 } else if (Inst->getOpcode() == Instruction::Mul &&
10930 isa<ConstantInt>(Inst->getOperand(1))) {
10931 Scale = cast<ConstantInt>(Inst->getOperand(1));
10932 NewIdx = Inst->getOperand(0);
10936 // If the index will be to exactly the right offset with the scale taken
10937 // out, perform the transformation. Note, we don't know whether Scale is
10938 // signed or not. We'll use unsigned version of division/modulo
10939 // operation after making sure Scale doesn't have the sign bit set.
10940 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
10941 Scale->getZExtValue() % ArrayEltSize == 0) {
10942 Scale = ConstantInt::get(Scale->getType(),
10943 Scale->getZExtValue() / ArrayEltSize);
10944 if (Scale->getZExtValue() != 1) {
10945 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10947 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10948 NewIdx = InsertNewInstBefore(Sc, GEP);
10951 // Insert the new GEP instruction.
10953 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10955 Instruction *NewGEP =
10956 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10957 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10958 // The NewGEP must be pointer typed, so must the old one -> BitCast
10959 return new BitCastInst(NewGEP, GEP.getType());
10965 /// See if we can simplify:
10966 /// X = bitcast A to B*
10967 /// Y = gep X, <...constant indices...>
10968 /// into a gep of the original struct. This is important for SROA and alias
10969 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
10970 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
10971 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
10972 // Determine how much the GEP moves the pointer. We are guaranteed to get
10973 // a constant back from EmitGEPOffset.
10974 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
10975 int64_t Offset = OffsetV->getSExtValue();
10977 // If this GEP instruction doesn't move the pointer, just replace the GEP
10978 // with a bitcast of the real input to the dest type.
10980 // If the bitcast is of an allocation, and the allocation will be
10981 // converted to match the type of the cast, don't touch this.
10982 if (isa<AllocationInst>(BCI->getOperand(0))) {
10983 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10984 if (Instruction *I = visitBitCast(*BCI)) {
10987 BCI->getParent()->getInstList().insert(BCI, I);
10988 ReplaceInstUsesWith(*BCI, I);
10993 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10996 // Otherwise, if the offset is non-zero, we need to find out if there is a
10997 // field at Offset in 'A's type. If so, we can pull the cast through the
10999 SmallVector<Value*, 8> NewIndices;
11001 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11002 if (FindElementAtOffset(InTy, Offset, NewIndices, TD)) {
11003 Instruction *NGEP =
11004 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11006 if (NGEP->getType() == GEP.getType()) return NGEP;
11007 InsertNewInstBefore(NGEP, GEP);
11008 NGEP->takeName(&GEP);
11009 return new BitCastInst(NGEP, GEP.getType());
11017 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11018 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11019 if (AI.isArrayAllocation()) { // Check C != 1
11020 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11021 const Type *NewTy =
11022 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11023 AllocationInst *New = 0;
11025 // Create and insert the replacement instruction...
11026 if (isa<MallocInst>(AI))
11027 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11029 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11030 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11033 InsertNewInstBefore(New, AI);
11035 // Scan to the end of the allocation instructions, to skip over a block of
11036 // allocas if possible...also skip interleaved debug info
11038 BasicBlock::iterator It = New;
11039 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11041 // Now that I is pointing to the first non-allocation-inst in the block,
11042 // insert our getelementptr instruction...
11044 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
11048 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11049 New->getName()+".sub", It);
11051 // Now make everything use the getelementptr instead of the original
11053 return ReplaceInstUsesWith(AI, V);
11054 } else if (isa<UndefValue>(AI.getArraySize())) {
11055 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11059 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11060 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11061 // Note that we only do this for alloca's, because malloc should allocate
11062 // and return a unique pointer, even for a zero byte allocation.
11063 if (TD->getTypePaddedSize(AI.getAllocatedType()) == 0)
11064 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11066 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11067 if (AI.getAlignment() == 0)
11068 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11074 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11075 Value *Op = FI.getOperand(0);
11077 // free undef -> unreachable.
11078 if (isa<UndefValue>(Op)) {
11079 // Insert a new store to null because we cannot modify the CFG here.
11080 new StoreInst(ConstantInt::getTrue(),
11081 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
11082 return EraseInstFromFunction(FI);
11085 // If we have 'free null' delete the instruction. This can happen in stl code
11086 // when lots of inlining happens.
11087 if (isa<ConstantPointerNull>(Op))
11088 return EraseInstFromFunction(FI);
11090 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11091 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11092 FI.setOperand(0, CI->getOperand(0));
11096 // Change free (gep X, 0,0,0,0) into free(X)
11097 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11098 if (GEPI->hasAllZeroIndices()) {
11099 AddToWorkList(GEPI);
11100 FI.setOperand(0, GEPI->getOperand(0));
11105 // Change free(malloc) into nothing, if the malloc has a single use.
11106 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11107 if (MI->hasOneUse()) {
11108 EraseInstFromFunction(FI);
11109 return EraseInstFromFunction(*MI);
11116 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11117 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11118 const TargetData *TD) {
11119 User *CI = cast<User>(LI.getOperand(0));
11120 Value *CastOp = CI->getOperand(0);
11122 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11123 // Instead of loading constant c string, use corresponding integer value
11124 // directly if string length is small enough.
11126 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11127 unsigned len = Str.length();
11128 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11129 unsigned numBits = Ty->getPrimitiveSizeInBits();
11130 // Replace LI with immediate integer store.
11131 if ((numBits >> 3) == len + 1) {
11132 APInt StrVal(numBits, 0);
11133 APInt SingleChar(numBits, 0);
11134 if (TD->isLittleEndian()) {
11135 for (signed i = len-1; i >= 0; i--) {
11136 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11137 StrVal = (StrVal << 8) | SingleChar;
11140 for (unsigned i = 0; i < len; i++) {
11141 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11142 StrVal = (StrVal << 8) | SingleChar;
11144 // Append NULL at the end.
11146 StrVal = (StrVal << 8) | SingleChar;
11148 Value *NL = ConstantInt::get(StrVal);
11149 return IC.ReplaceInstUsesWith(LI, NL);
11154 const PointerType *DestTy = cast<PointerType>(CI->getType());
11155 const Type *DestPTy = DestTy->getElementType();
11156 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11158 // If the address spaces don't match, don't eliminate the cast.
11159 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11162 const Type *SrcPTy = SrcTy->getElementType();
11164 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11165 isa<VectorType>(DestPTy)) {
11166 // If the source is an array, the code below will not succeed. Check to
11167 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11169 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11170 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11171 if (ASrcTy->getNumElements() != 0) {
11173 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11174 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11175 SrcTy = cast<PointerType>(CastOp->getType());
11176 SrcPTy = SrcTy->getElementType();
11179 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11180 isa<VectorType>(SrcPTy)) &&
11181 // Do not allow turning this into a load of an integer, which is then
11182 // casted to a pointer, this pessimizes pointer analysis a lot.
11183 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11184 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11185 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11187 // Okay, we are casting from one integer or pointer type to another of
11188 // the same size. Instead of casting the pointer before the load, cast
11189 // the result of the loaded value.
11190 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11192 LI.isVolatile()),LI);
11193 // Now cast the result of the load.
11194 return new BitCastInst(NewLoad, LI.getType());
11201 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
11202 /// from this value cannot trap. If it is not obviously safe to load from the
11203 /// specified pointer, we do a quick local scan of the basic block containing
11204 /// ScanFrom, to determine if the address is already accessed.
11205 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
11206 // If it is an alloca it is always safe to load from.
11207 if (isa<AllocaInst>(V)) return true;
11209 // If it is a global variable it is mostly safe to load from.
11210 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
11211 // Don't try to evaluate aliases. External weak GV can be null.
11212 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
11214 // Otherwise, be a little bit agressive by scanning the local block where we
11215 // want to check to see if the pointer is already being loaded or stored
11216 // from/to. If so, the previous load or store would have already trapped,
11217 // so there is no harm doing an extra load (also, CSE will later eliminate
11218 // the load entirely).
11219 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
11224 // If we see a free or a call (which might do a free) the pointer could be
11226 if (isa<FreeInst>(BBI) ||
11227 (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)))
11230 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11231 if (LI->getOperand(0) == V) return true;
11232 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
11233 if (SI->getOperand(1) == V) return true;
11240 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11241 Value *Op = LI.getOperand(0);
11243 // Attempt to improve the alignment.
11244 unsigned KnownAlign =
11245 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11247 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11248 LI.getAlignment()))
11249 LI.setAlignment(KnownAlign);
11251 // load (cast X) --> cast (load X) iff safe
11252 if (isa<CastInst>(Op))
11253 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11256 // None of the following transforms are legal for volatile loads.
11257 if (LI.isVolatile()) return 0;
11259 // Do really simple store-to-load forwarding and load CSE, to catch cases
11260 // where there are several consequtive memory accesses to the same location,
11261 // separated by a few arithmetic operations.
11262 BasicBlock::iterator BBI = &LI;
11263 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11264 return ReplaceInstUsesWith(LI, AvailableVal);
11266 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11267 const Value *GEPI0 = GEPI->getOperand(0);
11268 // TODO: Consider a target hook for valid address spaces for this xform.
11269 if (isa<ConstantPointerNull>(GEPI0) &&
11270 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11271 // Insert a new store to null instruction before the load to indicate
11272 // that this code is not reachable. We do this instead of inserting
11273 // an unreachable instruction directly because we cannot modify the
11275 new StoreInst(UndefValue::get(LI.getType()),
11276 Constant::getNullValue(Op->getType()), &LI);
11277 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11281 if (Constant *C = dyn_cast<Constant>(Op)) {
11282 // load null/undef -> undef
11283 // TODO: Consider a target hook for valid address spaces for this xform.
11284 if (isa<UndefValue>(C) || (C->isNullValue() &&
11285 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11286 // Insert a new store to null instruction before the load to indicate that
11287 // this code is not reachable. We do this instead of inserting an
11288 // unreachable instruction directly because we cannot modify the CFG.
11289 new StoreInst(UndefValue::get(LI.getType()),
11290 Constant::getNullValue(Op->getType()), &LI);
11291 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11294 // Instcombine load (constant global) into the value loaded.
11295 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11296 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11297 return ReplaceInstUsesWith(LI, GV->getInitializer());
11299 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11300 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11301 if (CE->getOpcode() == Instruction::GetElementPtr) {
11302 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11303 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11305 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
11306 return ReplaceInstUsesWith(LI, V);
11307 if (CE->getOperand(0)->isNullValue()) {
11308 // Insert a new store to null instruction before the load to indicate
11309 // that this code is not reachable. We do this instead of inserting
11310 // an unreachable instruction directly because we cannot modify the
11312 new StoreInst(UndefValue::get(LI.getType()),
11313 Constant::getNullValue(Op->getType()), &LI);
11314 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11317 } else if (CE->isCast()) {
11318 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11324 // If this load comes from anywhere in a constant global, and if the global
11325 // is all undef or zero, we know what it loads.
11326 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11327 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11328 if (GV->getInitializer()->isNullValue())
11329 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11330 else if (isa<UndefValue>(GV->getInitializer()))
11331 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11335 if (Op->hasOneUse()) {
11336 // Change select and PHI nodes to select values instead of addresses: this
11337 // helps alias analysis out a lot, allows many others simplifications, and
11338 // exposes redundancy in the code.
11340 // Note that we cannot do the transformation unless we know that the
11341 // introduced loads cannot trap! Something like this is valid as long as
11342 // the condition is always false: load (select bool %C, int* null, int* %G),
11343 // but it would not be valid if we transformed it to load from null
11344 // unconditionally.
11346 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11347 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11348 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11349 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11350 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11351 SI->getOperand(1)->getName()+".val"), LI);
11352 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11353 SI->getOperand(2)->getName()+".val"), LI);
11354 return SelectInst::Create(SI->getCondition(), V1, V2);
11357 // load (select (cond, null, P)) -> load P
11358 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11359 if (C->isNullValue()) {
11360 LI.setOperand(0, SI->getOperand(2));
11364 // load (select (cond, P, null)) -> load P
11365 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11366 if (C->isNullValue()) {
11367 LI.setOperand(0, SI->getOperand(1));
11375 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11376 /// when possible. This makes it generally easy to do alias analysis and/or
11377 /// SROA/mem2reg of the memory object.
11378 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11379 User *CI = cast<User>(SI.getOperand(1));
11380 Value *CastOp = CI->getOperand(0);
11382 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11383 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11384 if (SrcTy == 0) return 0;
11386 const Type *SrcPTy = SrcTy->getElementType();
11388 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11391 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11392 /// to its first element. This allows us to handle things like:
11393 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11394 /// on 32-bit hosts.
11395 SmallVector<Value*, 4> NewGEPIndices;
11397 // If the source is an array, the code below will not succeed. Check to
11398 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11400 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11401 // Index through pointer.
11402 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
11403 NewGEPIndices.push_back(Zero);
11406 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11407 if (!STy->getNumElements()) /* Struct can be empty {} */
11409 NewGEPIndices.push_back(Zero);
11410 SrcPTy = STy->getElementType(0);
11411 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11412 NewGEPIndices.push_back(Zero);
11413 SrcPTy = ATy->getElementType();
11419 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11422 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11425 // If the pointers point into different address spaces or if they point to
11426 // values with different sizes, we can't do the transformation.
11427 if (SrcTy->getAddressSpace() !=
11428 cast<PointerType>(CI->getType())->getAddressSpace() ||
11429 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11430 IC.getTargetData().getTypeSizeInBits(DestPTy))
11433 // Okay, we are casting from one integer or pointer type to another of
11434 // the same size. Instead of casting the pointer before
11435 // the store, cast the value to be stored.
11437 Value *SIOp0 = SI.getOperand(0);
11438 Instruction::CastOps opcode = Instruction::BitCast;
11439 const Type* CastSrcTy = SIOp0->getType();
11440 const Type* CastDstTy = SrcPTy;
11441 if (isa<PointerType>(CastDstTy)) {
11442 if (CastSrcTy->isInteger())
11443 opcode = Instruction::IntToPtr;
11444 } else if (isa<IntegerType>(CastDstTy)) {
11445 if (isa<PointerType>(SIOp0->getType()))
11446 opcode = Instruction::PtrToInt;
11449 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11450 // emit a GEP to index into its first field.
11451 if (!NewGEPIndices.empty()) {
11452 if (Constant *C = dyn_cast<Constant>(CastOp))
11453 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11454 NewGEPIndices.size());
11456 CastOp = IC.InsertNewInstBefore(
11457 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11458 NewGEPIndices.end()), SI);
11461 if (Constant *C = dyn_cast<Constant>(SIOp0))
11462 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11464 NewCast = IC.InsertNewInstBefore(
11465 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11467 return new StoreInst(NewCast, CastOp);
11470 /// equivalentAddressValues - Test if A and B will obviously have the same
11471 /// value. This includes recognizing that %t0 and %t1 will have the same
11472 /// value in code like this:
11473 /// %t0 = getelementptr \@a, 0, 3
11474 /// store i32 0, i32* %t0
11475 /// %t1 = getelementptr \@a, 0, 3
11476 /// %t2 = load i32* %t1
11478 static bool equivalentAddressValues(Value *A, Value *B) {
11479 // Test if the values are trivially equivalent.
11480 if (A == B) return true;
11482 // Test if the values come form identical arithmetic instructions.
11483 if (isa<BinaryOperator>(A) ||
11484 isa<CastInst>(A) ||
11486 isa<GetElementPtrInst>(A))
11487 if (Instruction *BI = dyn_cast<Instruction>(B))
11488 if (cast<Instruction>(A)->isIdenticalTo(BI))
11491 // Otherwise they may not be equivalent.
11495 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11496 // return the llvm.dbg.declare.
11497 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11498 if (!V->hasNUses(2))
11500 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11502 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11504 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11505 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11512 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11513 Value *Val = SI.getOperand(0);
11514 Value *Ptr = SI.getOperand(1);
11516 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11517 EraseInstFromFunction(SI);
11522 // If the RHS is an alloca with a single use, zapify the store, making the
11524 // If the RHS is an alloca with a two uses, the other one being a
11525 // llvm.dbg.declare, zapify the store and the declare, making the
11526 // alloca dead. We must do this to prevent declare's from affecting
11528 if (!SI.isVolatile()) {
11529 if (Ptr->hasOneUse()) {
11530 if (isa<AllocaInst>(Ptr)) {
11531 EraseInstFromFunction(SI);
11535 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11536 if (isa<AllocaInst>(GEP->getOperand(0))) {
11537 if (GEP->getOperand(0)->hasOneUse()) {
11538 EraseInstFromFunction(SI);
11542 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11543 EraseInstFromFunction(*DI);
11544 EraseInstFromFunction(SI);
11551 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11552 EraseInstFromFunction(*DI);
11553 EraseInstFromFunction(SI);
11559 // Attempt to improve the alignment.
11560 unsigned KnownAlign =
11561 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11563 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11564 SI.getAlignment()))
11565 SI.setAlignment(KnownAlign);
11567 // Do really simple DSE, to catch cases where there are several consecutive
11568 // stores to the same location, separated by a few arithmetic operations. This
11569 // situation often occurs with bitfield accesses.
11570 BasicBlock::iterator BBI = &SI;
11571 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11574 // Don't count debug info directives, lest they affect codegen,
11575 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11576 // It is necessary for correctness to skip those that feed into a
11577 // llvm.dbg.declare, as these are not present when debugging is off.
11578 if (isa<DbgInfoIntrinsic>(BBI) ||
11579 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11584 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11585 // Prev store isn't volatile, and stores to the same location?
11586 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11587 SI.getOperand(1))) {
11590 EraseInstFromFunction(*PrevSI);
11596 // If this is a load, we have to stop. However, if the loaded value is from
11597 // the pointer we're loading and is producing the pointer we're storing,
11598 // then *this* store is dead (X = load P; store X -> P).
11599 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11600 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11601 !SI.isVolatile()) {
11602 EraseInstFromFunction(SI);
11606 // Otherwise, this is a load from some other location. Stores before it
11607 // may not be dead.
11611 // Don't skip over loads or things that can modify memory.
11612 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11617 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11619 // store X, null -> turns into 'unreachable' in SimplifyCFG
11620 if (isa<ConstantPointerNull>(Ptr)) {
11621 if (!isa<UndefValue>(Val)) {
11622 SI.setOperand(0, UndefValue::get(Val->getType()));
11623 if (Instruction *U = dyn_cast<Instruction>(Val))
11624 AddToWorkList(U); // Dropped a use.
11627 return 0; // Do not modify these!
11630 // store undef, Ptr -> noop
11631 if (isa<UndefValue>(Val)) {
11632 EraseInstFromFunction(SI);
11637 // If the pointer destination is a cast, see if we can fold the cast into the
11639 if (isa<CastInst>(Ptr))
11640 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11642 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11644 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11648 // If this store is the last instruction in the basic block (possibly
11649 // excepting debug info instructions and the pointer bitcasts that feed
11650 // into them), and if the block ends with an unconditional branch, try
11651 // to move it to the successor block.
11655 } while (isa<DbgInfoIntrinsic>(BBI) ||
11656 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11657 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11658 if (BI->isUnconditional())
11659 if (SimplifyStoreAtEndOfBlock(SI))
11660 return 0; // xform done!
11665 /// SimplifyStoreAtEndOfBlock - Turn things like:
11666 /// if () { *P = v1; } else { *P = v2 }
11667 /// into a phi node with a store in the successor.
11669 /// Simplify things like:
11670 /// *P = v1; if () { *P = v2; }
11671 /// into a phi node with a store in the successor.
11673 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11674 BasicBlock *StoreBB = SI.getParent();
11676 // Check to see if the successor block has exactly two incoming edges. If
11677 // so, see if the other predecessor contains a store to the same location.
11678 // if so, insert a PHI node (if needed) and move the stores down.
11679 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11681 // Determine whether Dest has exactly two predecessors and, if so, compute
11682 // the other predecessor.
11683 pred_iterator PI = pred_begin(DestBB);
11684 BasicBlock *OtherBB = 0;
11685 if (*PI != StoreBB)
11688 if (PI == pred_end(DestBB))
11691 if (*PI != StoreBB) {
11696 if (++PI != pred_end(DestBB))
11699 // Bail out if all the relevant blocks aren't distinct (this can happen,
11700 // for example, if SI is in an infinite loop)
11701 if (StoreBB == DestBB || OtherBB == DestBB)
11704 // Verify that the other block ends in a branch and is not otherwise empty.
11705 BasicBlock::iterator BBI = OtherBB->getTerminator();
11706 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11707 if (!OtherBr || BBI == OtherBB->begin())
11710 // If the other block ends in an unconditional branch, check for the 'if then
11711 // else' case. there is an instruction before the branch.
11712 StoreInst *OtherStore = 0;
11713 if (OtherBr->isUnconditional()) {
11715 // Skip over debugging info.
11716 while (isa<DbgInfoIntrinsic>(BBI) ||
11717 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11718 if (BBI==OtherBB->begin())
11722 // If this isn't a store, or isn't a store to the same location, bail out.
11723 OtherStore = dyn_cast<StoreInst>(BBI);
11724 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11727 // Otherwise, the other block ended with a conditional branch. If one of the
11728 // destinations is StoreBB, then we have the if/then case.
11729 if (OtherBr->getSuccessor(0) != StoreBB &&
11730 OtherBr->getSuccessor(1) != StoreBB)
11733 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11734 // if/then triangle. See if there is a store to the same ptr as SI that
11735 // lives in OtherBB.
11737 // Check to see if we find the matching store.
11738 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11739 if (OtherStore->getOperand(1) != SI.getOperand(1))
11743 // If we find something that may be using or overwriting the stored
11744 // value, or if we run out of instructions, we can't do the xform.
11745 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11746 BBI == OtherBB->begin())
11750 // In order to eliminate the store in OtherBr, we have to
11751 // make sure nothing reads or overwrites the stored value in
11753 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11754 // FIXME: This should really be AA driven.
11755 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11760 // Insert a PHI node now if we need it.
11761 Value *MergedVal = OtherStore->getOperand(0);
11762 if (MergedVal != SI.getOperand(0)) {
11763 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11764 PN->reserveOperandSpace(2);
11765 PN->addIncoming(SI.getOperand(0), SI.getParent());
11766 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11767 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11770 // Advance to a place where it is safe to insert the new store and
11772 BBI = DestBB->getFirstNonPHI();
11773 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11774 OtherStore->isVolatile()), *BBI);
11776 // Nuke the old stores.
11777 EraseInstFromFunction(SI);
11778 EraseInstFromFunction(*OtherStore);
11784 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11785 // Change br (not X), label True, label False to: br X, label False, True
11787 BasicBlock *TrueDest;
11788 BasicBlock *FalseDest;
11789 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11790 !isa<Constant>(X)) {
11791 // Swap Destinations and condition...
11792 BI.setCondition(X);
11793 BI.setSuccessor(0, FalseDest);
11794 BI.setSuccessor(1, TrueDest);
11798 // Cannonicalize fcmp_one -> fcmp_oeq
11799 FCmpInst::Predicate FPred; Value *Y;
11800 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11801 TrueDest, FalseDest)))
11802 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11803 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11804 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11805 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11806 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11807 NewSCC->takeName(I);
11808 // Swap Destinations and condition...
11809 BI.setCondition(NewSCC);
11810 BI.setSuccessor(0, FalseDest);
11811 BI.setSuccessor(1, TrueDest);
11812 RemoveFromWorkList(I);
11813 I->eraseFromParent();
11814 AddToWorkList(NewSCC);
11818 // Cannonicalize icmp_ne -> icmp_eq
11819 ICmpInst::Predicate IPred;
11820 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11821 TrueDest, FalseDest)))
11822 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11823 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11824 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11825 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11826 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11827 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11828 NewSCC->takeName(I);
11829 // Swap Destinations and condition...
11830 BI.setCondition(NewSCC);
11831 BI.setSuccessor(0, FalseDest);
11832 BI.setSuccessor(1, TrueDest);
11833 RemoveFromWorkList(I);
11834 I->eraseFromParent();;
11835 AddToWorkList(NewSCC);
11842 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11843 Value *Cond = SI.getCondition();
11844 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11845 if (I->getOpcode() == Instruction::Add)
11846 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11847 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11848 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11849 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11851 SI.setOperand(0, I->getOperand(0));
11859 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11860 Value *Agg = EV.getAggregateOperand();
11862 if (!EV.hasIndices())
11863 return ReplaceInstUsesWith(EV, Agg);
11865 if (Constant *C = dyn_cast<Constant>(Agg)) {
11866 if (isa<UndefValue>(C))
11867 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11869 if (isa<ConstantAggregateZero>(C))
11870 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11872 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11873 // Extract the element indexed by the first index out of the constant
11874 Value *V = C->getOperand(*EV.idx_begin());
11875 if (EV.getNumIndices() > 1)
11876 // Extract the remaining indices out of the constant indexed by the
11878 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11880 return ReplaceInstUsesWith(EV, V);
11882 return 0; // Can't handle other constants
11884 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11885 // We're extracting from an insertvalue instruction, compare the indices
11886 const unsigned *exti, *exte, *insi, *inse;
11887 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11888 exte = EV.idx_end(), inse = IV->idx_end();
11889 exti != exte && insi != inse;
11891 if (*insi != *exti)
11892 // The insert and extract both reference distinctly different elements.
11893 // This means the extract is not influenced by the insert, and we can
11894 // replace the aggregate operand of the extract with the aggregate
11895 // operand of the insert. i.e., replace
11896 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11897 // %E = extractvalue { i32, { i32 } } %I, 0
11899 // %E = extractvalue { i32, { i32 } } %A, 0
11900 return ExtractValueInst::Create(IV->getAggregateOperand(),
11901 EV.idx_begin(), EV.idx_end());
11903 if (exti == exte && insi == inse)
11904 // Both iterators are at the end: Index lists are identical. Replace
11905 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11906 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11908 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11909 if (exti == exte) {
11910 // The extract list is a prefix of the insert list. i.e. replace
11911 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11912 // %E = extractvalue { i32, { i32 } } %I, 1
11914 // %X = extractvalue { i32, { i32 } } %A, 1
11915 // %E = insertvalue { i32 } %X, i32 42, 0
11916 // by switching the order of the insert and extract (though the
11917 // insertvalue should be left in, since it may have other uses).
11918 Value *NewEV = InsertNewInstBefore(
11919 ExtractValueInst::Create(IV->getAggregateOperand(),
11920 EV.idx_begin(), EV.idx_end()),
11922 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11926 // The insert list is a prefix of the extract list
11927 // We can simply remove the common indices from the extract and make it
11928 // operate on the inserted value instead of the insertvalue result.
11930 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11931 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11933 // %E extractvalue { i32 } { i32 42 }, 0
11934 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11937 // Can't simplify extracts from other values. Note that nested extracts are
11938 // already simplified implicitely by the above (extract ( extract (insert) )
11939 // will be translated into extract ( insert ( extract ) ) first and then just
11940 // the value inserted, if appropriate).
11944 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11945 /// is to leave as a vector operation.
11946 static bool CheapToScalarize(Value *V, bool isConstant) {
11947 if (isa<ConstantAggregateZero>(V))
11949 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11950 if (isConstant) return true;
11951 // If all elts are the same, we can extract.
11952 Constant *Op0 = C->getOperand(0);
11953 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11954 if (C->getOperand(i) != Op0)
11958 Instruction *I = dyn_cast<Instruction>(V);
11959 if (!I) return false;
11961 // Insert element gets simplified to the inserted element or is deleted if
11962 // this is constant idx extract element and its a constant idx insertelt.
11963 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11964 isa<ConstantInt>(I->getOperand(2)))
11966 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11968 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11969 if (BO->hasOneUse() &&
11970 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11971 CheapToScalarize(BO->getOperand(1), isConstant)))
11973 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11974 if (CI->hasOneUse() &&
11975 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11976 CheapToScalarize(CI->getOperand(1), isConstant)))
11982 /// Read and decode a shufflevector mask.
11984 /// It turns undef elements into values that are larger than the number of
11985 /// elements in the input.
11986 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11987 unsigned NElts = SVI->getType()->getNumElements();
11988 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11989 return std::vector<unsigned>(NElts, 0);
11990 if (isa<UndefValue>(SVI->getOperand(2)))
11991 return std::vector<unsigned>(NElts, 2*NElts);
11993 std::vector<unsigned> Result;
11994 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11995 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11996 if (isa<UndefValue>(*i))
11997 Result.push_back(NElts*2); // undef -> 8
11999 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12003 /// FindScalarElement - Given a vector and an element number, see if the scalar
12004 /// value is already around as a register, for example if it were inserted then
12005 /// extracted from the vector.
12006 static Value *FindScalarElement(Value *V, unsigned EltNo) {
12007 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12008 const VectorType *PTy = cast<VectorType>(V->getType());
12009 unsigned Width = PTy->getNumElements();
12010 if (EltNo >= Width) // Out of range access.
12011 return UndefValue::get(PTy->getElementType());
12013 if (isa<UndefValue>(V))
12014 return UndefValue::get(PTy->getElementType());
12015 else if (isa<ConstantAggregateZero>(V))
12016 return Constant::getNullValue(PTy->getElementType());
12017 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12018 return CP->getOperand(EltNo);
12019 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12020 // If this is an insert to a variable element, we don't know what it is.
12021 if (!isa<ConstantInt>(III->getOperand(2)))
12023 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12025 // If this is an insert to the element we are looking for, return the
12027 if (EltNo == IIElt)
12028 return III->getOperand(1);
12030 // Otherwise, the insertelement doesn't modify the value, recurse on its
12032 return FindScalarElement(III->getOperand(0), EltNo);
12033 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12034 unsigned LHSWidth =
12035 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12036 unsigned InEl = getShuffleMask(SVI)[EltNo];
12037 if (InEl < LHSWidth)
12038 return FindScalarElement(SVI->getOperand(0), InEl);
12039 else if (InEl < LHSWidth*2)
12040 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
12042 return UndefValue::get(PTy->getElementType());
12045 // Otherwise, we don't know.
12049 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12050 // If vector val is undef, replace extract with scalar undef.
12051 if (isa<UndefValue>(EI.getOperand(0)))
12052 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12054 // If vector val is constant 0, replace extract with scalar 0.
12055 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12056 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12058 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12059 // If vector val is constant with all elements the same, replace EI with
12060 // that element. When the elements are not identical, we cannot replace yet
12061 // (we do that below, but only when the index is constant).
12062 Constant *op0 = C->getOperand(0);
12063 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12064 if (C->getOperand(i) != op0) {
12069 return ReplaceInstUsesWith(EI, op0);
12072 // If extracting a specified index from the vector, see if we can recursively
12073 // find a previously computed scalar that was inserted into the vector.
12074 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12075 unsigned IndexVal = IdxC->getZExtValue();
12076 unsigned VectorWidth =
12077 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12079 // If this is extracting an invalid index, turn this into undef, to avoid
12080 // crashing the code below.
12081 if (IndexVal >= VectorWidth)
12082 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12084 // This instruction only demands the single element from the input vector.
12085 // If the input vector has a single use, simplify it based on this use
12087 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12088 APInt UndefElts(VectorWidth, 0);
12089 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12090 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12091 DemandedMask, UndefElts)) {
12092 EI.setOperand(0, V);
12097 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
12098 return ReplaceInstUsesWith(EI, Elt);
12100 // If the this extractelement is directly using a bitcast from a vector of
12101 // the same number of elements, see if we can find the source element from
12102 // it. In this case, we will end up needing to bitcast the scalars.
12103 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12104 if (const VectorType *VT =
12105 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12106 if (VT->getNumElements() == VectorWidth)
12107 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
12108 return new BitCastInst(Elt, EI.getType());
12112 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12113 if (I->hasOneUse()) {
12114 // Push extractelement into predecessor operation if legal and
12115 // profitable to do so
12116 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12117 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12118 if (CheapToScalarize(BO, isConstantElt)) {
12119 ExtractElementInst *newEI0 =
12120 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
12121 EI.getName()+".lhs");
12122 ExtractElementInst *newEI1 =
12123 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
12124 EI.getName()+".rhs");
12125 InsertNewInstBefore(newEI0, EI);
12126 InsertNewInstBefore(newEI1, EI);
12127 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12129 } else if (isa<LoadInst>(I)) {
12131 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12132 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12133 PointerType::get(EI.getType(), AS),EI);
12134 GetElementPtrInst *GEP =
12135 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12136 InsertNewInstBefore(GEP, EI);
12137 return new LoadInst(GEP);
12140 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12141 // Extracting the inserted element?
12142 if (IE->getOperand(2) == EI.getOperand(1))
12143 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12144 // If the inserted and extracted elements are constants, they must not
12145 // be the same value, extract from the pre-inserted value instead.
12146 if (isa<Constant>(IE->getOperand(2)) &&
12147 isa<Constant>(EI.getOperand(1))) {
12148 AddUsesToWorkList(EI);
12149 EI.setOperand(0, IE->getOperand(0));
12152 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12153 // If this is extracting an element from a shufflevector, figure out where
12154 // it came from and extract from the appropriate input element instead.
12155 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12156 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12158 unsigned LHSWidth =
12159 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12161 if (SrcIdx < LHSWidth)
12162 Src = SVI->getOperand(0);
12163 else if (SrcIdx < LHSWidth*2) {
12164 SrcIdx -= LHSWidth;
12165 Src = SVI->getOperand(1);
12167 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12169 return new ExtractElementInst(Src, SrcIdx);
12176 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12177 /// elements from either LHS or RHS, return the shuffle mask and true.
12178 /// Otherwise, return false.
12179 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12180 std::vector<Constant*> &Mask) {
12181 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12182 "Invalid CollectSingleShuffleElements");
12183 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12185 if (isa<UndefValue>(V)) {
12186 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12188 } else if (V == LHS) {
12189 for (unsigned i = 0; i != NumElts; ++i)
12190 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12192 } else if (V == RHS) {
12193 for (unsigned i = 0; i != NumElts; ++i)
12194 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
12196 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12197 // If this is an insert of an extract from some other vector, include it.
12198 Value *VecOp = IEI->getOperand(0);
12199 Value *ScalarOp = IEI->getOperand(1);
12200 Value *IdxOp = IEI->getOperand(2);
12202 if (!isa<ConstantInt>(IdxOp))
12204 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12206 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12207 // Okay, we can handle this if the vector we are insertinting into is
12208 // transitively ok.
12209 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
12210 // If so, update the mask to reflect the inserted undef.
12211 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
12214 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12215 if (isa<ConstantInt>(EI->getOperand(1)) &&
12216 EI->getOperand(0)->getType() == V->getType()) {
12217 unsigned ExtractedIdx =
12218 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12220 // This must be extracting from either LHS or RHS.
12221 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12222 // Okay, we can handle this if the vector we are insertinting into is
12223 // transitively ok.
12224 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
12225 // If so, update the mask to reflect the inserted value.
12226 if (EI->getOperand(0) == LHS) {
12227 Mask[InsertedIdx % NumElts] =
12228 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12230 assert(EI->getOperand(0) == RHS);
12231 Mask[InsertedIdx % NumElts] =
12232 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
12241 // TODO: Handle shufflevector here!
12246 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12247 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12248 /// that computes V and the LHS value of the shuffle.
12249 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12251 assert(isa<VectorType>(V->getType()) &&
12252 (RHS == 0 || V->getType() == RHS->getType()) &&
12253 "Invalid shuffle!");
12254 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12256 if (isa<UndefValue>(V)) {
12257 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12259 } else if (isa<ConstantAggregateZero>(V)) {
12260 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
12262 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12263 // If this is an insert of an extract from some other vector, include it.
12264 Value *VecOp = IEI->getOperand(0);
12265 Value *ScalarOp = IEI->getOperand(1);
12266 Value *IdxOp = IEI->getOperand(2);
12268 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12269 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12270 EI->getOperand(0)->getType() == V->getType()) {
12271 unsigned ExtractedIdx =
12272 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12273 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12275 // Either the extracted from or inserted into vector must be RHSVec,
12276 // otherwise we'd end up with a shuffle of three inputs.
12277 if (EI->getOperand(0) == RHS || RHS == 0) {
12278 RHS = EI->getOperand(0);
12279 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
12280 Mask[InsertedIdx % NumElts] =
12281 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
12285 if (VecOp == RHS) {
12286 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
12287 // Everything but the extracted element is replaced with the RHS.
12288 for (unsigned i = 0; i != NumElts; ++i) {
12289 if (i != InsertedIdx)
12290 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
12295 // If this insertelement is a chain that comes from exactly these two
12296 // vectors, return the vector and the effective shuffle.
12297 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
12298 return EI->getOperand(0);
12303 // TODO: Handle shufflevector here!
12305 // Otherwise, can't do anything fancy. Return an identity vector.
12306 for (unsigned i = 0; i != NumElts; ++i)
12307 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12311 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12312 Value *VecOp = IE.getOperand(0);
12313 Value *ScalarOp = IE.getOperand(1);
12314 Value *IdxOp = IE.getOperand(2);
12316 // Inserting an undef or into an undefined place, remove this.
12317 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12318 ReplaceInstUsesWith(IE, VecOp);
12320 // If the inserted element was extracted from some other vector, and if the
12321 // indexes are constant, try to turn this into a shufflevector operation.
12322 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12323 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12324 EI->getOperand(0)->getType() == IE.getType()) {
12325 unsigned NumVectorElts = IE.getType()->getNumElements();
12326 unsigned ExtractedIdx =
12327 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12328 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12330 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12331 return ReplaceInstUsesWith(IE, VecOp);
12333 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12334 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12336 // If we are extracting a value from a vector, then inserting it right
12337 // back into the same place, just use the input vector.
12338 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12339 return ReplaceInstUsesWith(IE, VecOp);
12341 // We could theoretically do this for ANY input. However, doing so could
12342 // turn chains of insertelement instructions into a chain of shufflevector
12343 // instructions, and right now we do not merge shufflevectors. As such,
12344 // only do this in a situation where it is clear that there is benefit.
12345 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12346 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12347 // the values of VecOp, except then one read from EIOp0.
12348 // Build a new shuffle mask.
12349 std::vector<Constant*> Mask;
12350 if (isa<UndefValue>(VecOp))
12351 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
12353 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12354 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
12357 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12358 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12359 ConstantVector::get(Mask));
12362 // If this insertelement isn't used by some other insertelement, turn it
12363 // (and any insertelements it points to), into one big shuffle.
12364 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12365 std::vector<Constant*> Mask;
12367 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
12368 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12369 // We now have a shuffle of LHS, RHS, Mask.
12370 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
12379 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12380 Value *LHS = SVI.getOperand(0);
12381 Value *RHS = SVI.getOperand(1);
12382 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12384 bool MadeChange = false;
12386 // Undefined shuffle mask -> undefined value.
12387 if (isa<UndefValue>(SVI.getOperand(2)))
12388 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12390 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12392 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12395 APInt UndefElts(VWidth, 0);
12396 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12397 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12398 LHS = SVI.getOperand(0);
12399 RHS = SVI.getOperand(1);
12403 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12404 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12405 if (LHS == RHS || isa<UndefValue>(LHS)) {
12406 if (isa<UndefValue>(LHS) && LHS == RHS) {
12407 // shuffle(undef,undef,mask) -> undef.
12408 return ReplaceInstUsesWith(SVI, LHS);
12411 // Remap any references to RHS to use LHS.
12412 std::vector<Constant*> Elts;
12413 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12414 if (Mask[i] >= 2*e)
12415 Elts.push_back(UndefValue::get(Type::Int32Ty));
12417 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12418 (Mask[i] < e && isa<UndefValue>(LHS))) {
12419 Mask[i] = 2*e; // Turn into undef.
12420 Elts.push_back(UndefValue::get(Type::Int32Ty));
12422 Mask[i] = Mask[i] % e; // Force to LHS.
12423 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12427 SVI.setOperand(0, SVI.getOperand(1));
12428 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12429 SVI.setOperand(2, ConstantVector::get(Elts));
12430 LHS = SVI.getOperand(0);
12431 RHS = SVI.getOperand(1);
12435 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12436 bool isLHSID = true, isRHSID = true;
12438 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12439 if (Mask[i] >= e*2) continue; // Ignore undef values.
12440 // Is this an identity shuffle of the LHS value?
12441 isLHSID &= (Mask[i] == i);
12443 // Is this an identity shuffle of the RHS value?
12444 isRHSID &= (Mask[i]-e == i);
12447 // Eliminate identity shuffles.
12448 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12449 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12451 // If the LHS is a shufflevector itself, see if we can combine it with this
12452 // one without producing an unusual shuffle. Here we are really conservative:
12453 // we are absolutely afraid of producing a shuffle mask not in the input
12454 // program, because the code gen may not be smart enough to turn a merged
12455 // shuffle into two specific shuffles: it may produce worse code. As such,
12456 // we only merge two shuffles if the result is one of the two input shuffle
12457 // masks. In this case, merging the shuffles just removes one instruction,
12458 // which we know is safe. This is good for things like turning:
12459 // (splat(splat)) -> splat.
12460 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12461 if (isa<UndefValue>(RHS)) {
12462 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12464 std::vector<unsigned> NewMask;
12465 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12466 if (Mask[i] >= 2*e)
12467 NewMask.push_back(2*e);
12469 NewMask.push_back(LHSMask[Mask[i]]);
12471 // If the result mask is equal to the src shuffle or this shuffle mask, do
12472 // the replacement.
12473 if (NewMask == LHSMask || NewMask == Mask) {
12474 unsigned LHSInNElts =
12475 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12476 std::vector<Constant*> Elts;
12477 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12478 if (NewMask[i] >= LHSInNElts*2) {
12479 Elts.push_back(UndefValue::get(Type::Int32Ty));
12481 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12484 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12485 LHSSVI->getOperand(1),
12486 ConstantVector::get(Elts));
12491 return MadeChange ? &SVI : 0;
12497 /// TryToSinkInstruction - Try to move the specified instruction from its
12498 /// current block into the beginning of DestBlock, which can only happen if it's
12499 /// safe to move the instruction past all of the instructions between it and the
12500 /// end of its block.
12501 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12502 assert(I->hasOneUse() && "Invariants didn't hold!");
12504 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12505 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
12508 // Do not sink alloca instructions out of the entry block.
12509 if (isa<AllocaInst>(I) && I->getParent() ==
12510 &DestBlock->getParent()->getEntryBlock())
12513 // We can only sink load instructions if there is nothing between the load and
12514 // the end of block that could change the value.
12515 if (I->mayReadFromMemory()) {
12516 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12518 if (Scan->mayWriteToMemory())
12522 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12524 CopyPrecedingStopPoint(I, InsertPos);
12525 I->moveBefore(InsertPos);
12531 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12532 /// all reachable code to the worklist.
12534 /// This has a couple of tricks to make the code faster and more powerful. In
12535 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12536 /// them to the worklist (this significantly speeds up instcombine on code where
12537 /// many instructions are dead or constant). Additionally, if we find a branch
12538 /// whose condition is a known constant, we only visit the reachable successors.
12540 static void AddReachableCodeToWorklist(BasicBlock *BB,
12541 SmallPtrSet<BasicBlock*, 64> &Visited,
12543 const TargetData *TD) {
12544 SmallVector<BasicBlock*, 256> Worklist;
12545 Worklist.push_back(BB);
12547 while (!Worklist.empty()) {
12548 BB = Worklist.back();
12549 Worklist.pop_back();
12551 // We have now visited this block! If we've already been here, ignore it.
12552 if (!Visited.insert(BB)) continue;
12554 DbgInfoIntrinsic *DBI_Prev = NULL;
12555 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12556 Instruction *Inst = BBI++;
12558 // DCE instruction if trivially dead.
12559 if (isInstructionTriviallyDead(Inst)) {
12561 DOUT << "IC: DCE: " << *Inst;
12562 Inst->eraseFromParent();
12566 // ConstantProp instruction if trivially constant.
12567 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12568 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12569 Inst->replaceAllUsesWith(C);
12571 Inst->eraseFromParent();
12575 // If there are two consecutive llvm.dbg.stoppoint calls then
12576 // it is likely that the optimizer deleted code in between these
12578 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12581 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12582 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12583 IC.RemoveFromWorkList(DBI_Prev);
12584 DBI_Prev->eraseFromParent();
12586 DBI_Prev = DBI_Next;
12591 IC.AddToWorkList(Inst);
12594 // Recursively visit successors. If this is a branch or switch on a
12595 // constant, only visit the reachable successor.
12596 TerminatorInst *TI = BB->getTerminator();
12597 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12598 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12599 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12600 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12601 Worklist.push_back(ReachableBB);
12604 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12605 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12606 // See if this is an explicit destination.
12607 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12608 if (SI->getCaseValue(i) == Cond) {
12609 BasicBlock *ReachableBB = SI->getSuccessor(i);
12610 Worklist.push_back(ReachableBB);
12614 // Otherwise it is the default destination.
12615 Worklist.push_back(SI->getSuccessor(0));
12620 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12621 Worklist.push_back(TI->getSuccessor(i));
12625 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12626 bool Changed = false;
12627 TD = &getAnalysis<TargetData>();
12629 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12630 << F.getNameStr() << "\n");
12633 // Do a depth-first traversal of the function, populate the worklist with
12634 // the reachable instructions. Ignore blocks that are not reachable. Keep
12635 // track of which blocks we visit.
12636 SmallPtrSet<BasicBlock*, 64> Visited;
12637 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12639 // Do a quick scan over the function. If we find any blocks that are
12640 // unreachable, remove any instructions inside of them. This prevents
12641 // the instcombine code from having to deal with some bad special cases.
12642 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12643 if (!Visited.count(BB)) {
12644 Instruction *Term = BB->getTerminator();
12645 while (Term != BB->begin()) { // Remove instrs bottom-up
12646 BasicBlock::iterator I = Term; --I;
12648 DOUT << "IC: DCE: " << *I;
12649 // A debug intrinsic shouldn't force another iteration if we weren't
12650 // going to do one without it.
12651 if (!isa<DbgInfoIntrinsic>(I)) {
12655 if (!I->use_empty())
12656 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12657 I->eraseFromParent();
12662 while (!Worklist.empty()) {
12663 Instruction *I = RemoveOneFromWorkList();
12664 if (I == 0) continue; // skip null values.
12666 // Check to see if we can DCE the instruction.
12667 if (isInstructionTriviallyDead(I)) {
12668 // Add operands to the worklist.
12669 if (I->getNumOperands() < 4)
12670 AddUsesToWorkList(*I);
12673 DOUT << "IC: DCE: " << *I;
12675 I->eraseFromParent();
12676 RemoveFromWorkList(I);
12681 // Instruction isn't dead, see if we can constant propagate it.
12682 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12683 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12685 // Add operands to the worklist.
12686 AddUsesToWorkList(*I);
12687 ReplaceInstUsesWith(*I, C);
12690 I->eraseFromParent();
12691 RemoveFromWorkList(I);
12696 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12697 // See if we can constant fold its operands.
12698 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
12699 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
12700 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12707 // See if we can trivially sink this instruction to a successor basic block.
12708 if (I->hasOneUse()) {
12709 BasicBlock *BB = I->getParent();
12710 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12711 if (UserParent != BB) {
12712 bool UserIsSuccessor = false;
12713 // See if the user is one of our successors.
12714 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12715 if (*SI == UserParent) {
12716 UserIsSuccessor = true;
12720 // If the user is one of our immediate successors, and if that successor
12721 // only has us as a predecessors (we'd have to split the critical edge
12722 // otherwise), we can keep going.
12723 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12724 next(pred_begin(UserParent)) == pred_end(UserParent))
12725 // Okay, the CFG is simple enough, try to sink this instruction.
12726 Changed |= TryToSinkInstruction(I, UserParent);
12730 // Now that we have an instruction, try combining it to simplify it...
12734 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12735 if (Instruction *Result = visit(*I)) {
12737 // Should we replace the old instruction with a new one?
12739 DOUT << "IC: Old = " << *I
12740 << " New = " << *Result;
12742 // Everything uses the new instruction now.
12743 I->replaceAllUsesWith(Result);
12745 // Push the new instruction and any users onto the worklist.
12746 AddToWorkList(Result);
12747 AddUsersToWorkList(*Result);
12749 // Move the name to the new instruction first.
12750 Result->takeName(I);
12752 // Insert the new instruction into the basic block...
12753 BasicBlock *InstParent = I->getParent();
12754 BasicBlock::iterator InsertPos = I;
12756 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12757 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12760 InstParent->getInstList().insert(InsertPos, Result);
12762 // Make sure that we reprocess all operands now that we reduced their
12764 AddUsesToWorkList(*I);
12766 // Instructions can end up on the worklist more than once. Make sure
12767 // we do not process an instruction that has been deleted.
12768 RemoveFromWorkList(I);
12770 // Erase the old instruction.
12771 InstParent->getInstList().erase(I);
12774 DOUT << "IC: Mod = " << OrigI
12775 << " New = " << *I;
12778 // If the instruction was modified, it's possible that it is now dead.
12779 // if so, remove it.
12780 if (isInstructionTriviallyDead(I)) {
12781 // Make sure we process all operands now that we are reducing their
12783 AddUsesToWorkList(*I);
12785 // Instructions may end up in the worklist more than once. Erase all
12786 // occurrences of this instruction.
12787 RemoveFromWorkList(I);
12788 I->eraseFromParent();
12791 AddUsersToWorkList(*I);
12798 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12800 // Do an explicit clear, this shrinks the map if needed.
12801 WorklistMap.clear();
12806 bool InstCombiner::runOnFunction(Function &F) {
12807 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12809 bool EverMadeChange = false;
12811 // Iterate while there is work to do.
12812 unsigned Iteration = 0;
12813 while (DoOneIteration(F, Iteration++))
12814 EverMadeChange = true;
12815 return EverMadeChange;
12818 FunctionPass *llvm::createInstructionCombiningPass() {
12819 return new InstCombiner();