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/LLVMContext.h"
40 #include "llvm/Pass.h"
41 #include "llvm/DerivedTypes.h"
42 #include "llvm/GlobalVariable.h"
43 #include "llvm/Operator.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/InstructionSimplify.h"
46 #include "llvm/Analysis/MemoryBuiltins.h"
47 #include "llvm/Analysis/ValueTracking.h"
48 #include "llvm/Target/TargetData.h"
49 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 #include "llvm/Transforms/Utils/Local.h"
51 #include "llvm/Support/CallSite.h"
52 #include "llvm/Support/ConstantRange.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/ErrorHandling.h"
55 #include "llvm/Support/GetElementPtrTypeIterator.h"
56 #include "llvm/Support/InstVisitor.h"
57 #include "llvm/Support/IRBuilder.h"
58 #include "llvm/Support/MathExtras.h"
59 #include "llvm/Support/PatternMatch.h"
60 #include "llvm/Support/TargetFolder.h"
61 #include "llvm/Support/raw_ostream.h"
62 #include "llvm/ADT/DenseMap.h"
63 #include "llvm/ADT/SmallVector.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/ADT/Statistic.h"
66 #include "llvm/ADT/STLExtras.h"
70 using namespace llvm::PatternMatch;
72 STATISTIC(NumCombined , "Number of insts combined");
73 STATISTIC(NumConstProp, "Number of constant folds");
74 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
75 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
76 STATISTIC(NumSunkInst , "Number of instructions sunk");
78 /// SelectPatternFlavor - We can match a variety of different patterns for
79 /// select operations.
80 enum SelectPatternFlavor {
88 /// InstCombineWorklist - This is the worklist management logic for
90 class InstCombineWorklist {
91 SmallVector<Instruction*, 256> Worklist;
92 DenseMap<Instruction*, unsigned> WorklistMap;
94 void operator=(const InstCombineWorklist&RHS); // DO NOT IMPLEMENT
95 InstCombineWorklist(const InstCombineWorklist&); // DO NOT IMPLEMENT
97 InstCombineWorklist() {}
99 bool isEmpty() const { return Worklist.empty(); }
101 /// Add - Add the specified instruction to the worklist if it isn't already
103 void Add(Instruction *I) {
104 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second) {
105 DEBUG(errs() << "IC: ADD: " << *I << '\n');
106 Worklist.push_back(I);
110 void AddValue(Value *V) {
111 if (Instruction *I = dyn_cast<Instruction>(V))
115 /// AddInitialGroup - Add the specified batch of stuff in reverse order.
116 /// which should only be done when the worklist is empty and when the group
117 /// has no duplicates.
118 void AddInitialGroup(Instruction *const *List, unsigned NumEntries) {
119 assert(Worklist.empty() && "Worklist must be empty to add initial group");
120 Worklist.reserve(NumEntries+16);
121 DEBUG(errs() << "IC: ADDING: " << NumEntries << " instrs to worklist\n");
122 for (; NumEntries; --NumEntries) {
123 Instruction *I = List[NumEntries-1];
124 WorklistMap.insert(std::make_pair(I, Worklist.size()));
125 Worklist.push_back(I);
129 // Remove - remove I from the worklist if it exists.
130 void Remove(Instruction *I) {
131 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
132 if (It == WorklistMap.end()) return; // Not in worklist.
134 // Don't bother moving everything down, just null out the slot.
135 Worklist[It->second] = 0;
137 WorklistMap.erase(It);
140 Instruction *RemoveOne() {
141 Instruction *I = Worklist.back();
143 WorklistMap.erase(I);
147 /// AddUsersToWorkList - When an instruction is simplified, add all users of
148 /// the instruction to the work lists because they might get more simplified
151 void AddUsersToWorkList(Instruction &I) {
152 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
154 Add(cast<Instruction>(*UI));
158 /// Zap - check that the worklist is empty and nuke the backing store for
159 /// the map if it is large.
161 assert(WorklistMap.empty() && "Worklist empty, but map not?");
163 // Do an explicit clear, this shrinks the map if needed.
167 } // end anonymous namespace.
171 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works
172 /// just like the normal insertion helper, but also adds any new instructions
173 /// to the instcombine worklist.
174 class InstCombineIRInserter : public IRBuilderDefaultInserter<true> {
175 InstCombineWorklist &Worklist;
177 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
179 void InsertHelper(Instruction *I, const Twine &Name,
180 BasicBlock *BB, BasicBlock::iterator InsertPt) const {
181 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
185 } // end anonymous namespace
189 class InstCombiner : public FunctionPass,
190 public InstVisitor<InstCombiner, Instruction*> {
192 bool MustPreserveLCSSA;
195 /// Worklist - All of the instructions that need to be simplified.
196 InstCombineWorklist Worklist;
198 /// Builder - This is an IRBuilder that automatically inserts new
199 /// instructions into the worklist when they are created.
200 typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
203 static char ID; // Pass identification, replacement for typeid
204 InstCombiner() : FunctionPass(&ID), TD(0), Builder(0) {}
206 LLVMContext *Context;
207 LLVMContext *getContext() const { return Context; }
210 virtual bool runOnFunction(Function &F);
212 bool DoOneIteration(Function &F, unsigned ItNum);
214 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
215 AU.addPreservedID(LCSSAID);
216 AU.setPreservesCFG();
219 TargetData *getTargetData() const { return TD; }
221 // Visitation implementation - Implement instruction combining for different
222 // instruction types. The semantics are as follows:
224 // null - No change was made
225 // I - Change was made, I is still valid, I may be dead though
226 // otherwise - Change was made, replace I with returned instruction
228 Instruction *visitAdd(BinaryOperator &I);
229 Instruction *visitFAdd(BinaryOperator &I);
230 Value *OptimizePointerDifference(Value *LHS, Value *RHS, const Type *Ty);
231 Instruction *visitSub(BinaryOperator &I);
232 Instruction *visitFSub(BinaryOperator &I);
233 Instruction *visitMul(BinaryOperator &I);
234 Instruction *visitFMul(BinaryOperator &I);
235 Instruction *visitURem(BinaryOperator &I);
236 Instruction *visitSRem(BinaryOperator &I);
237 Instruction *visitFRem(BinaryOperator &I);
238 bool SimplifyDivRemOfSelect(BinaryOperator &I);
239 Instruction *commonRemTransforms(BinaryOperator &I);
240 Instruction *commonIRemTransforms(BinaryOperator &I);
241 Instruction *commonDivTransforms(BinaryOperator &I);
242 Instruction *commonIDivTransforms(BinaryOperator &I);
243 Instruction *visitUDiv(BinaryOperator &I);
244 Instruction *visitSDiv(BinaryOperator &I);
245 Instruction *visitFDiv(BinaryOperator &I);
246 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
247 Instruction *FoldAndOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
248 Instruction *visitAnd(BinaryOperator &I);
249 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
250 Instruction *FoldOrOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
251 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
252 Value *A, Value *B, Value *C);
253 Instruction *visitOr (BinaryOperator &I);
254 Instruction *visitXor(BinaryOperator &I);
255 Instruction *visitShl(BinaryOperator &I);
256 Instruction *visitAShr(BinaryOperator &I);
257 Instruction *visitLShr(BinaryOperator &I);
258 Instruction *commonShiftTransforms(BinaryOperator &I);
259 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
261 Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
262 GlobalVariable *GV, CmpInst &ICI);
263 Instruction *visitFCmpInst(FCmpInst &I);
264 Instruction *visitICmpInst(ICmpInst &I);
265 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
266 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
269 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
270 ConstantInt *DivRHS);
271 Instruction *FoldICmpAddOpCst(ICmpInst &ICI, Value *X, ConstantInt *CI,
272 ICmpInst::Predicate Pred, Value *TheAdd);
273 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
274 ICmpInst::Predicate Cond, Instruction &I);
275 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
277 Instruction *commonCastTransforms(CastInst &CI);
278 Instruction *commonIntCastTransforms(CastInst &CI);
279 Instruction *commonPointerCastTransforms(CastInst &CI);
280 Instruction *visitTrunc(TruncInst &CI);
281 Instruction *visitZExt(ZExtInst &CI);
282 Instruction *visitSExt(SExtInst &CI);
283 Instruction *visitFPTrunc(FPTruncInst &CI);
284 Instruction *visitFPExt(CastInst &CI);
285 Instruction *visitFPToUI(FPToUIInst &FI);
286 Instruction *visitFPToSI(FPToSIInst &FI);
287 Instruction *visitUIToFP(CastInst &CI);
288 Instruction *visitSIToFP(CastInst &CI);
289 Instruction *visitPtrToInt(PtrToIntInst &CI);
290 Instruction *visitIntToPtr(IntToPtrInst &CI);
291 Instruction *visitBitCast(BitCastInst &CI);
292 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
294 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
295 Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
296 Value *A, Value *B, Instruction &Outer,
297 SelectPatternFlavor SPF2, Value *C);
298 Instruction *visitSelectInst(SelectInst &SI);
299 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
300 Instruction *visitCallInst(CallInst &CI);
301 Instruction *visitInvokeInst(InvokeInst &II);
303 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
304 Instruction *visitPHINode(PHINode &PN);
305 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
306 Instruction *visitAllocaInst(AllocaInst &AI);
307 Instruction *visitFree(Instruction &FI);
308 Instruction *visitLoadInst(LoadInst &LI);
309 Instruction *visitStoreInst(StoreInst &SI);
310 Instruction *visitBranchInst(BranchInst &BI);
311 Instruction *visitSwitchInst(SwitchInst &SI);
312 Instruction *visitInsertElementInst(InsertElementInst &IE);
313 Instruction *visitExtractElementInst(ExtractElementInst &EI);
314 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
315 Instruction *visitExtractValueInst(ExtractValueInst &EV);
317 // visitInstruction - Specify what to return for unhandled instructions...
318 Instruction *visitInstruction(Instruction &I) { return 0; }
321 Instruction *visitCallSite(CallSite CS);
322 bool transformConstExprCastCall(CallSite CS);
323 Instruction *transformCallThroughTrampoline(CallSite CS);
324 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
325 bool DoXform = true);
326 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
327 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
331 // InsertNewInstBefore - insert an instruction New before instruction Old
332 // in the program. Add the new instruction to the worklist.
334 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
335 assert(New && New->getParent() == 0 &&
336 "New instruction already inserted into a basic block!");
337 BasicBlock *BB = Old.getParent();
338 BB->getInstList().insert(&Old, New); // Insert inst
343 // ReplaceInstUsesWith - This method is to be used when an instruction is
344 // found to be dead, replacable with another preexisting expression. Here
345 // we add all uses of I to the worklist, replace all uses of I with the new
346 // value, then return I, so that the inst combiner will know that I was
349 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
350 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
352 // If we are replacing the instruction with itself, this must be in a
353 // segment of unreachable code, so just clobber the instruction.
355 V = UndefValue::get(I.getType());
357 I.replaceAllUsesWith(V);
361 // EraseInstFromFunction - When dealing with an instruction that has side
362 // effects or produces a void value, we can't rely on DCE to delete the
363 // instruction. Instead, visit methods should return the value returned by
365 Instruction *EraseInstFromFunction(Instruction &I) {
366 DEBUG(errs() << "IC: ERASE " << I << '\n');
368 assert(I.use_empty() && "Cannot erase instruction that is used!");
369 // Make sure that we reprocess all operands now that we reduced their
371 if (I.getNumOperands() < 8) {
372 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
373 if (Instruction *Op = dyn_cast<Instruction>(*i))
379 return 0; // Don't do anything with FI
382 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
383 APInt &KnownOne, unsigned Depth = 0) const {
384 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
387 bool MaskedValueIsZero(Value *V, const APInt &Mask,
388 unsigned Depth = 0) const {
389 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
391 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
392 return llvm::ComputeNumSignBits(Op, TD, Depth);
397 /// SimplifyCommutative - This performs a few simplifications for
398 /// commutative operators.
399 bool SimplifyCommutative(BinaryOperator &I);
401 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
402 /// based on the demanded bits.
403 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
404 APInt& KnownZero, APInt& KnownOne,
406 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
407 APInt& KnownZero, APInt& KnownOne,
410 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
411 /// SimplifyDemandedBits knows about. See if the instruction has any
412 /// properties that allow us to simplify its operands.
413 bool SimplifyDemandedInstructionBits(Instruction &Inst);
415 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
416 APInt& UndefElts, unsigned Depth = 0);
418 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
419 // which has a PHI node as operand #0, see if we can fold the instruction
420 // into the PHI (which is only possible if all operands to the PHI are
423 // If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
424 // that would normally be unprofitable because they strongly encourage jump
426 Instruction *FoldOpIntoPhi(Instruction &I, bool AllowAggressive = false);
428 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
429 // operator and they all are only used by the PHI, PHI together their
430 // inputs, and do the operation once, to the result of the PHI.
431 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
432 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
433 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
434 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
437 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
438 ConstantInt *AndRHS, BinaryOperator &TheAnd);
440 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
441 bool isSub, Instruction &I);
442 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
443 bool isSigned, bool Inside, Instruction &IB);
444 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
445 Instruction *MatchBSwap(BinaryOperator &I);
446 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
447 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
448 Instruction *SimplifyMemSet(MemSetInst *MI);
451 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
453 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
454 unsigned CastOpc, int &NumCastsRemoved);
455 unsigned GetOrEnforceKnownAlignment(Value *V,
456 unsigned PrefAlign = 0);
459 } // end anonymous namespace
461 char InstCombiner::ID = 0;
462 static RegisterPass<InstCombiner>
463 X("instcombine", "Combine redundant instructions");
465 // getComplexity: Assign a complexity or rank value to LLVM Values...
466 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
467 static unsigned getComplexity(Value *V) {
468 if (isa<Instruction>(V)) {
469 if (BinaryOperator::isNeg(V) ||
470 BinaryOperator::isFNeg(V) ||
471 BinaryOperator::isNot(V))
475 if (isa<Argument>(V)) return 3;
476 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
479 // isOnlyUse - Return true if this instruction will be deleted if we stop using
481 static bool isOnlyUse(Value *V) {
482 return V->hasOneUse() || isa<Constant>(V);
485 // getPromotedType - Return the specified type promoted as it would be to pass
486 // though a va_arg area...
487 static const Type *getPromotedType(const Type *Ty) {
488 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
489 if (ITy->getBitWidth() < 32)
490 return Type::getInt32Ty(Ty->getContext());
495 /// ShouldChangeType - Return true if it is desirable to convert a computation
496 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
497 /// type for example, or from a smaller to a larger illegal type.
498 static bool ShouldChangeType(const Type *From, const Type *To,
499 const TargetData *TD) {
500 assert(isa<IntegerType>(From) && isa<IntegerType>(To));
502 // If we don't have TD, we don't know if the source/dest are legal.
503 if (!TD) return false;
505 unsigned FromWidth = From->getPrimitiveSizeInBits();
506 unsigned ToWidth = To->getPrimitiveSizeInBits();
507 bool FromLegal = TD->isLegalInteger(FromWidth);
508 bool ToLegal = TD->isLegalInteger(ToWidth);
510 // If this is a legal integer from type, and the result would be an illegal
511 // type, don't do the transformation.
512 if (FromLegal && !ToLegal)
515 // Otherwise, if both are illegal, do not increase the size of the result. We
516 // do allow things like i160 -> i64, but not i64 -> i160.
517 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
523 /// getBitCastOperand - If the specified operand is a CastInst, a constant
524 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
525 /// operand value, otherwise return null.
526 static Value *getBitCastOperand(Value *V) {
527 if (Operator *O = dyn_cast<Operator>(V)) {
528 if (O->getOpcode() == Instruction::BitCast)
529 return O->getOperand(0);
530 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
531 if (GEP->hasAllZeroIndices())
532 return GEP->getPointerOperand();
537 /// This function is a wrapper around CastInst::isEliminableCastPair. It
538 /// simply extracts arguments and returns what that function returns.
539 static Instruction::CastOps
540 isEliminableCastPair(
541 const CastInst *CI, ///< The first cast instruction
542 unsigned opcode, ///< The opcode of the second cast instruction
543 const Type *DstTy, ///< The target type for the second cast instruction
544 TargetData *TD ///< The target data for pointer size
547 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
548 const Type *MidTy = CI->getType(); // B from above
550 // Get the opcodes of the two Cast instructions
551 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
552 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
554 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
556 TD ? TD->getIntPtrType(CI->getContext()) : 0);
558 // We don't want to form an inttoptr or ptrtoint that converts to an integer
559 // type that differs from the pointer size.
560 if ((Res == Instruction::IntToPtr &&
561 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
562 (Res == Instruction::PtrToInt &&
563 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
566 return Instruction::CastOps(Res);
569 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
570 /// in any code being generated. It does not require codegen if V is simple
571 /// enough or if the cast can be folded into other casts.
572 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
573 const Type *Ty, TargetData *TD) {
574 if (V->getType() == Ty || isa<Constant>(V)) return false;
576 // If this is another cast that can be eliminated, it isn't codegen either.
577 if (const CastInst *CI = dyn_cast<CastInst>(V))
578 if (isEliminableCastPair(CI, opcode, Ty, TD))
583 // SimplifyCommutative - This performs a few simplifications for commutative
586 // 1. Order operands such that they are listed from right (least complex) to
587 // left (most complex). This puts constants before unary operators before
590 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
591 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
593 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
594 bool Changed = false;
595 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
596 Changed = !I.swapOperands();
598 if (!I.isAssociative()) return Changed;
599 Instruction::BinaryOps Opcode = I.getOpcode();
600 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
601 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
602 if (isa<Constant>(I.getOperand(1))) {
603 Constant *Folded = ConstantExpr::get(I.getOpcode(),
604 cast<Constant>(I.getOperand(1)),
605 cast<Constant>(Op->getOperand(1)));
606 I.setOperand(0, Op->getOperand(0));
607 I.setOperand(1, Folded);
609 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
610 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
611 isOnlyUse(Op) && isOnlyUse(Op1)) {
612 Constant *C1 = cast<Constant>(Op->getOperand(1));
613 Constant *C2 = cast<Constant>(Op1->getOperand(1));
615 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
616 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
617 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
621 I.setOperand(0, New);
622 I.setOperand(1, Folded);
629 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
630 // if the LHS is a constant zero (which is the 'negate' form).
632 static inline Value *dyn_castNegVal(Value *V) {
633 if (BinaryOperator::isNeg(V))
634 return BinaryOperator::getNegArgument(V);
636 // Constants can be considered to be negated values if they can be folded.
637 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
638 return ConstantExpr::getNeg(C);
640 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
641 if (C->getType()->getElementType()->isInteger())
642 return ConstantExpr::getNeg(C);
647 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
648 // instruction if the LHS is a constant negative zero (which is the 'negate'
651 static inline Value *dyn_castFNegVal(Value *V) {
652 if (BinaryOperator::isFNeg(V))
653 return BinaryOperator::getFNegArgument(V);
655 // Constants can be considered to be negated values if they can be folded.
656 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
657 return ConstantExpr::getFNeg(C);
659 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
660 if (C->getType()->getElementType()->isFloatingPoint())
661 return ConstantExpr::getFNeg(C);
666 /// MatchSelectPattern - Pattern match integer [SU]MIN, [SU]MAX, and ABS idioms,
667 /// returning the kind and providing the out parameter results if we
668 /// successfully match.
669 static SelectPatternFlavor
670 MatchSelectPattern(Value *V, Value *&LHS, Value *&RHS) {
671 SelectInst *SI = dyn_cast<SelectInst>(V);
672 if (SI == 0) return SPF_UNKNOWN;
674 ICmpInst *ICI = dyn_cast<ICmpInst>(SI->getCondition());
675 if (ICI == 0) return SPF_UNKNOWN;
677 LHS = ICI->getOperand(0);
678 RHS = ICI->getOperand(1);
680 // (icmp X, Y) ? X : Y
681 if (SI->getTrueValue() == ICI->getOperand(0) &&
682 SI->getFalseValue() == ICI->getOperand(1)) {
683 switch (ICI->getPredicate()) {
684 default: return SPF_UNKNOWN; // Equality.
685 case ICmpInst::ICMP_UGT:
686 case ICmpInst::ICMP_UGE: return SPF_UMAX;
687 case ICmpInst::ICMP_SGT:
688 case ICmpInst::ICMP_SGE: return SPF_SMAX;
689 case ICmpInst::ICMP_ULT:
690 case ICmpInst::ICMP_ULE: return SPF_UMIN;
691 case ICmpInst::ICMP_SLT:
692 case ICmpInst::ICMP_SLE: return SPF_SMIN;
696 // (icmp X, Y) ? Y : X
697 if (SI->getTrueValue() == ICI->getOperand(1) &&
698 SI->getFalseValue() == ICI->getOperand(0)) {
699 switch (ICI->getPredicate()) {
700 default: return SPF_UNKNOWN; // Equality.
701 case ICmpInst::ICMP_UGT:
702 case ICmpInst::ICMP_UGE: return SPF_UMIN;
703 case ICmpInst::ICMP_SGT:
704 case ICmpInst::ICMP_SGE: return SPF_SMIN;
705 case ICmpInst::ICMP_ULT:
706 case ICmpInst::ICMP_ULE: return SPF_UMAX;
707 case ICmpInst::ICMP_SLT:
708 case ICmpInst::ICMP_SLE: return SPF_SMAX;
712 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
717 /// isFreeToInvert - Return true if the specified value is free to invert (apply
718 /// ~ to). This happens in cases where the ~ can be eliminated.
719 static inline bool isFreeToInvert(Value *V) {
721 if (BinaryOperator::isNot(V))
724 // Constants can be considered to be not'ed values.
725 if (isa<ConstantInt>(V))
728 // Compares can be inverted if they have a single use.
729 if (CmpInst *CI = dyn_cast<CmpInst>(V))
730 return CI->hasOneUse();
735 static inline Value *dyn_castNotVal(Value *V) {
736 // If this is not(not(x)) don't return that this is a not: we want the two
737 // not's to be folded first.
738 if (BinaryOperator::isNot(V)) {
739 Value *Operand = BinaryOperator::getNotArgument(V);
740 if (!isFreeToInvert(Operand))
744 // Constants can be considered to be not'ed values...
745 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
746 return ConstantInt::get(C->getType(), ~C->getValue());
752 // dyn_castFoldableMul - If this value is a multiply that can be folded into
753 // other computations (because it has a constant operand), return the
754 // non-constant operand of the multiply, and set CST to point to the multiplier.
755 // Otherwise, return null.
757 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
758 if (V->hasOneUse() && V->getType()->isInteger())
759 if (Instruction *I = dyn_cast<Instruction>(V)) {
760 if (I->getOpcode() == Instruction::Mul)
761 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
762 return I->getOperand(0);
763 if (I->getOpcode() == Instruction::Shl)
764 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
765 // The multiplier is really 1 << CST.
766 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
767 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
768 CST = ConstantInt::get(V->getType()->getContext(),
769 APInt(BitWidth, 1).shl(CSTVal));
770 return I->getOperand(0);
776 /// AddOne - Add one to a ConstantInt
777 static Constant *AddOne(Constant *C) {
778 return ConstantExpr::getAdd(C,
779 ConstantInt::get(C->getType(), 1));
781 /// SubOne - Subtract one from a ConstantInt
782 static Constant *SubOne(ConstantInt *C) {
783 return ConstantExpr::getSub(C,
784 ConstantInt::get(C->getType(), 1));
786 /// MultiplyOverflows - True if the multiply can not be expressed in an int
788 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
789 uint32_t W = C1->getBitWidth();
790 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
799 APInt MulExt = LHSExt * RHSExt;
802 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
804 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
805 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
806 return MulExt.slt(Min) || MulExt.sgt(Max);
810 /// ShrinkDemandedConstant - Check to see if the specified operand of the
811 /// specified instruction is a constant integer. If so, check to see if there
812 /// are any bits set in the constant that are not demanded. If so, shrink the
813 /// constant and return true.
814 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
816 assert(I && "No instruction?");
817 assert(OpNo < I->getNumOperands() && "Operand index too large");
819 // If the operand is not a constant integer, nothing to do.
820 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
821 if (!OpC) return false;
823 // If there are no bits set that aren't demanded, nothing to do.
824 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
825 if ((~Demanded & OpC->getValue()) == 0)
828 // This instruction is producing bits that are not demanded. Shrink the RHS.
829 Demanded &= OpC->getValue();
830 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
834 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
835 // set of known zero and one bits, compute the maximum and minimum values that
836 // could have the specified known zero and known one bits, returning them in
838 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
839 const APInt& KnownOne,
840 APInt& Min, APInt& Max) {
841 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
842 KnownZero.getBitWidth() == Min.getBitWidth() &&
843 KnownZero.getBitWidth() == Max.getBitWidth() &&
844 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
845 APInt UnknownBits = ~(KnownZero|KnownOne);
847 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
848 // bit if it is unknown.
850 Max = KnownOne|UnknownBits;
852 if (UnknownBits.isNegative()) { // Sign bit is unknown
853 Min.set(Min.getBitWidth()-1);
854 Max.clear(Max.getBitWidth()-1);
858 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
859 // a set of known zero and one bits, compute the maximum and minimum values that
860 // could have the specified known zero and known one bits, returning them in
862 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
863 const APInt &KnownOne,
864 APInt &Min, APInt &Max) {
865 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
866 KnownZero.getBitWidth() == Min.getBitWidth() &&
867 KnownZero.getBitWidth() == Max.getBitWidth() &&
868 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
869 APInt UnknownBits = ~(KnownZero|KnownOne);
871 // The minimum value is when the unknown bits are all zeros.
873 // The maximum value is when the unknown bits are all ones.
874 Max = KnownOne|UnknownBits;
877 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
878 /// SimplifyDemandedBits knows about. See if the instruction has any
879 /// properties that allow us to simplify its operands.
880 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
881 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
882 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
883 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
885 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
886 KnownZero, KnownOne, 0);
887 if (V == 0) return false;
888 if (V == &Inst) return true;
889 ReplaceInstUsesWith(Inst, V);
893 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
894 /// specified instruction operand if possible, updating it in place. It returns
895 /// true if it made any change and false otherwise.
896 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
897 APInt &KnownZero, APInt &KnownOne,
899 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
900 KnownZero, KnownOne, Depth);
901 if (NewVal == 0) return false;
907 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
908 /// value based on the demanded bits. When this function is called, it is known
909 /// that only the bits set in DemandedMask of the result of V are ever used
910 /// downstream. Consequently, depending on the mask and V, it may be possible
911 /// to replace V with a constant or one of its operands. In such cases, this
912 /// function does the replacement and returns true. In all other cases, it
913 /// returns false after analyzing the expression and setting KnownOne and known
914 /// to be one in the expression. KnownZero contains all the bits that are known
915 /// to be zero in the expression. These are provided to potentially allow the
916 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
917 /// the expression. KnownOne and KnownZero always follow the invariant that
918 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
919 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
920 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
921 /// and KnownOne must all be the same.
923 /// This returns null if it did not change anything and it permits no
924 /// simplification. This returns V itself if it did some simplification of V's
925 /// operands based on the information about what bits are demanded. This returns
926 /// some other non-null value if it found out that V is equal to another value
927 /// in the context where the specified bits are demanded, but not for all users.
928 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
929 APInt &KnownZero, APInt &KnownOne,
931 assert(V != 0 && "Null pointer of Value???");
932 assert(Depth <= 6 && "Limit Search Depth");
933 uint32_t BitWidth = DemandedMask.getBitWidth();
934 const Type *VTy = V->getType();
935 assert((TD || !isa<PointerType>(VTy)) &&
936 "SimplifyDemandedBits needs to know bit widths!");
937 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
938 (!VTy->isIntOrIntVector() ||
939 VTy->getScalarSizeInBits() == BitWidth) &&
940 KnownZero.getBitWidth() == BitWidth &&
941 KnownOne.getBitWidth() == BitWidth &&
942 "Value *V, DemandedMask, KnownZero and KnownOne "
943 "must have same BitWidth");
944 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
945 // We know all of the bits for a constant!
946 KnownOne = CI->getValue() & DemandedMask;
947 KnownZero = ~KnownOne & DemandedMask;
950 if (isa<ConstantPointerNull>(V)) {
951 // We know all of the bits for a constant!
953 KnownZero = DemandedMask;
959 if (DemandedMask == 0) { // Not demanding any bits from V.
960 if (isa<UndefValue>(V))
962 return UndefValue::get(VTy);
965 if (Depth == 6) // Limit search depth.
968 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
969 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
971 Instruction *I = dyn_cast<Instruction>(V);
973 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
974 return 0; // Only analyze instructions.
977 // If there are multiple uses of this value and we aren't at the root, then
978 // we can't do any simplifications of the operands, because DemandedMask
979 // only reflects the bits demanded by *one* of the users.
980 if (Depth != 0 && !I->hasOneUse()) {
981 // Despite the fact that we can't simplify this instruction in all User's
982 // context, we can at least compute the knownzero/knownone bits, and we can
983 // do simplifications that apply to *just* the one user if we know that
984 // this instruction has a simpler value in that context.
985 if (I->getOpcode() == Instruction::And) {
986 // If either the LHS or the RHS are Zero, the result is zero.
987 ComputeMaskedBits(I->getOperand(1), DemandedMask,
988 RHSKnownZero, RHSKnownOne, Depth+1);
989 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
990 LHSKnownZero, LHSKnownOne, Depth+1);
992 // If all of the demanded bits are known 1 on one side, return the other.
993 // These bits cannot contribute to the result of the 'and' in this
995 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
996 (DemandedMask & ~LHSKnownZero))
997 return I->getOperand(0);
998 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
999 (DemandedMask & ~RHSKnownZero))
1000 return I->getOperand(1);
1002 // If all of the demanded bits in the inputs are known zeros, return zero.
1003 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1004 return Constant::getNullValue(VTy);
1006 } else if (I->getOpcode() == Instruction::Or) {
1007 // We can simplify (X|Y) -> X or Y in the user's context if we know that
1008 // only bits from X or Y are demanded.
1010 // If either the LHS or the RHS are One, the result is One.
1011 ComputeMaskedBits(I->getOperand(1), DemandedMask,
1012 RHSKnownZero, RHSKnownOne, Depth+1);
1013 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1014 LHSKnownZero, LHSKnownOne, Depth+1);
1016 // If all of the demanded bits are known zero on one side, return the
1017 // other. These bits cannot contribute to the result of the 'or' in this
1019 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1020 (DemandedMask & ~LHSKnownOne))
1021 return I->getOperand(0);
1022 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1023 (DemandedMask & ~RHSKnownOne))
1024 return I->getOperand(1);
1026 // If all of the potentially set bits on one side are known to be set on
1027 // the other side, just use the 'other' side.
1028 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1029 (DemandedMask & (~RHSKnownZero)))
1030 return I->getOperand(0);
1031 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1032 (DemandedMask & (~LHSKnownZero)))
1033 return I->getOperand(1);
1036 // Compute the KnownZero/KnownOne bits to simplify things downstream.
1037 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
1041 // If this is the root being simplified, allow it to have multiple uses,
1042 // just set the DemandedMask to all bits so that we can try to simplify the
1043 // operands. This allows visitTruncInst (for example) to simplify the
1044 // operand of a trunc without duplicating all the logic below.
1045 if (Depth == 0 && !V->hasOneUse())
1046 DemandedMask = APInt::getAllOnesValue(BitWidth);
1048 switch (I->getOpcode()) {
1050 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1052 case Instruction::And:
1053 // If either the LHS or the RHS are Zero, the result is zero.
1054 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1055 RHSKnownZero, RHSKnownOne, Depth+1) ||
1056 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
1057 LHSKnownZero, LHSKnownOne, Depth+1))
1059 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1060 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1062 // If all of the demanded bits are known 1 on one side, return the other.
1063 // These bits cannot contribute to the result of the 'and'.
1064 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1065 (DemandedMask & ~LHSKnownZero))
1066 return I->getOperand(0);
1067 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1068 (DemandedMask & ~RHSKnownZero))
1069 return I->getOperand(1);
1071 // If all of the demanded bits in the inputs are known zeros, return zero.
1072 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1073 return Constant::getNullValue(VTy);
1075 // If the RHS is a constant, see if we can simplify it.
1076 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1079 // Output known-1 bits are only known if set in both the LHS & RHS.
1080 RHSKnownOne &= LHSKnownOne;
1081 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1082 RHSKnownZero |= LHSKnownZero;
1084 case Instruction::Or:
1085 // If either the LHS or the RHS are One, the result is One.
1086 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1087 RHSKnownZero, RHSKnownOne, Depth+1) ||
1088 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
1089 LHSKnownZero, LHSKnownOne, Depth+1))
1091 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1092 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1094 // If all of the demanded bits are known zero on one side, return the other.
1095 // These bits cannot contribute to the result of the 'or'.
1096 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1097 (DemandedMask & ~LHSKnownOne))
1098 return I->getOperand(0);
1099 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1100 (DemandedMask & ~RHSKnownOne))
1101 return I->getOperand(1);
1103 // If all of the potentially set bits on one side are known to be set on
1104 // the other side, just use the 'other' side.
1105 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1106 (DemandedMask & (~RHSKnownZero)))
1107 return I->getOperand(0);
1108 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1109 (DemandedMask & (~LHSKnownZero)))
1110 return I->getOperand(1);
1112 // If the RHS is a constant, see if we can simplify it.
1113 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1116 // Output known-0 bits are only known if clear in both the LHS & RHS.
1117 RHSKnownZero &= LHSKnownZero;
1118 // Output known-1 are known to be set if set in either the LHS | RHS.
1119 RHSKnownOne |= LHSKnownOne;
1121 case Instruction::Xor: {
1122 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1123 RHSKnownZero, RHSKnownOne, Depth+1) ||
1124 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1125 LHSKnownZero, LHSKnownOne, Depth+1))
1127 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1128 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1130 // If all of the demanded bits are known zero on one side, return the other.
1131 // These bits cannot contribute to the result of the 'xor'.
1132 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1133 return I->getOperand(0);
1134 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1135 return I->getOperand(1);
1137 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1138 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1139 (RHSKnownOne & LHSKnownOne);
1140 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1141 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1142 (RHSKnownOne & LHSKnownZero);
1144 // If all of the demanded bits are known to be zero on one side or the
1145 // other, turn this into an *inclusive* or.
1146 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1147 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1149 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1151 return InsertNewInstBefore(Or, *I);
1154 // If all of the demanded bits on one side are known, and all of the set
1155 // bits on that side are also known to be set on the other side, turn this
1156 // into an AND, as we know the bits will be cleared.
1157 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1158 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1160 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1161 Constant *AndC = Constant::getIntegerValue(VTy,
1162 ~RHSKnownOne & DemandedMask);
1164 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1165 return InsertNewInstBefore(And, *I);
1169 // If the RHS is a constant, see if we can simplify it.
1170 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1171 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1174 // If our LHS is an 'and' and if it has one use, and if any of the bits we
1175 // are flipping are known to be set, then the xor is just resetting those
1176 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
1177 // simplifying both of them.
1178 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0)))
1179 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
1180 isa<ConstantInt>(I->getOperand(1)) &&
1181 isa<ConstantInt>(LHSInst->getOperand(1)) &&
1182 (LHSKnownOne & RHSKnownOne & DemandedMask) != 0) {
1183 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
1184 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
1185 APInt NewMask = ~(LHSKnownOne & RHSKnownOne & DemandedMask);
1188 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
1189 Instruction *NewAnd =
1190 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1191 InsertNewInstBefore(NewAnd, *I);
1194 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
1195 Instruction *NewXor =
1196 BinaryOperator::CreateXor(NewAnd, XorC, "tmp");
1197 return InsertNewInstBefore(NewXor, *I);
1201 RHSKnownZero = KnownZeroOut;
1202 RHSKnownOne = KnownOneOut;
1205 case Instruction::Select:
1206 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1207 RHSKnownZero, RHSKnownOne, Depth+1) ||
1208 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1209 LHSKnownZero, LHSKnownOne, Depth+1))
1211 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1212 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1214 // If the operands are constants, see if we can simplify them.
1215 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1216 ShrinkDemandedConstant(I, 2, DemandedMask))
1219 // Only known if known in both the LHS and RHS.
1220 RHSKnownOne &= LHSKnownOne;
1221 RHSKnownZero &= LHSKnownZero;
1223 case Instruction::Trunc: {
1224 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1225 DemandedMask.zext(truncBf);
1226 RHSKnownZero.zext(truncBf);
1227 RHSKnownOne.zext(truncBf);
1228 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1229 RHSKnownZero, RHSKnownOne, Depth+1))
1231 DemandedMask.trunc(BitWidth);
1232 RHSKnownZero.trunc(BitWidth);
1233 RHSKnownOne.trunc(BitWidth);
1234 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1237 case Instruction::BitCast:
1238 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1239 return false; // vector->int or fp->int?
1241 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1242 if (const VectorType *SrcVTy =
1243 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1244 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1245 // Don't touch a bitcast between vectors of different element counts.
1248 // Don't touch a scalar-to-vector bitcast.
1250 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1251 // Don't touch a vector-to-scalar bitcast.
1254 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1255 RHSKnownZero, RHSKnownOne, Depth+1))
1257 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1259 case Instruction::ZExt: {
1260 // Compute the bits in the result that are not present in the input.
1261 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1263 DemandedMask.trunc(SrcBitWidth);
1264 RHSKnownZero.trunc(SrcBitWidth);
1265 RHSKnownOne.trunc(SrcBitWidth);
1266 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1267 RHSKnownZero, RHSKnownOne, Depth+1))
1269 DemandedMask.zext(BitWidth);
1270 RHSKnownZero.zext(BitWidth);
1271 RHSKnownOne.zext(BitWidth);
1272 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1273 // The top bits are known to be zero.
1274 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1277 case Instruction::SExt: {
1278 // Compute the bits in the result that are not present in the input.
1279 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1281 APInt InputDemandedBits = DemandedMask &
1282 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1284 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1285 // If any of the sign extended bits are demanded, we know that the sign
1287 if ((NewBits & DemandedMask) != 0)
1288 InputDemandedBits.set(SrcBitWidth-1);
1290 InputDemandedBits.trunc(SrcBitWidth);
1291 RHSKnownZero.trunc(SrcBitWidth);
1292 RHSKnownOne.trunc(SrcBitWidth);
1293 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1294 RHSKnownZero, RHSKnownOne, Depth+1))
1296 InputDemandedBits.zext(BitWidth);
1297 RHSKnownZero.zext(BitWidth);
1298 RHSKnownOne.zext(BitWidth);
1299 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1301 // If the sign bit of the input is known set or clear, then we know the
1302 // top bits of the result.
1304 // If the input sign bit is known zero, or if the NewBits are not demanded
1305 // convert this into a zero extension.
1306 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1307 // Convert to ZExt cast
1308 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1309 return InsertNewInstBefore(NewCast, *I);
1310 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1311 RHSKnownOne |= NewBits;
1315 case Instruction::Add: {
1316 // Figure out what the input bits are. If the top bits of the and result
1317 // are not demanded, then the add doesn't demand them from its input
1319 unsigned NLZ = DemandedMask.countLeadingZeros();
1321 // If there is a constant on the RHS, there are a variety of xformations
1323 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1324 // If null, this should be simplified elsewhere. Some of the xforms here
1325 // won't work if the RHS is zero.
1329 // If the top bit of the output is demanded, demand everything from the
1330 // input. Otherwise, we demand all the input bits except NLZ top bits.
1331 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1333 // Find information about known zero/one bits in the input.
1334 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1335 LHSKnownZero, LHSKnownOne, Depth+1))
1338 // If the RHS of the add has bits set that can't affect the input, reduce
1340 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1343 // Avoid excess work.
1344 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1347 // Turn it into OR if input bits are zero.
1348 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1350 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1352 return InsertNewInstBefore(Or, *I);
1355 // We can say something about the output known-zero and known-one bits,
1356 // depending on potential carries from the input constant and the
1357 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1358 // bits set and the RHS constant is 0x01001, then we know we have a known
1359 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1361 // To compute this, we first compute the potential carry bits. These are
1362 // the bits which may be modified. I'm not aware of a better way to do
1364 const APInt &RHSVal = RHS->getValue();
1365 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1367 // Now that we know which bits have carries, compute the known-1/0 sets.
1369 // Bits are known one if they are known zero in one operand and one in the
1370 // other, and there is no input carry.
1371 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1372 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1374 // Bits are known zero if they are known zero in both operands and there
1375 // is no input carry.
1376 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1378 // If the high-bits of this ADD are not demanded, then it does not demand
1379 // the high bits of its LHS or RHS.
1380 if (DemandedMask[BitWidth-1] == 0) {
1381 // Right fill the mask of bits for this ADD to demand the most
1382 // significant bit and all those below it.
1383 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1384 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1385 LHSKnownZero, LHSKnownOne, Depth+1) ||
1386 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1387 LHSKnownZero, LHSKnownOne, Depth+1))
1393 case Instruction::Sub:
1394 // If the high-bits of this SUB are not demanded, then it does not demand
1395 // the high bits of its LHS or RHS.
1396 if (DemandedMask[BitWidth-1] == 0) {
1397 // Right fill the mask of bits for this SUB to demand the most
1398 // significant bit and all those below it.
1399 uint32_t NLZ = DemandedMask.countLeadingZeros();
1400 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1401 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1402 LHSKnownZero, LHSKnownOne, Depth+1) ||
1403 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1404 LHSKnownZero, LHSKnownOne, Depth+1))
1407 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1408 // the known zeros and ones.
1409 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1411 case Instruction::Shl:
1412 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1413 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1414 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1415 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1416 RHSKnownZero, RHSKnownOne, Depth+1))
1418 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1419 RHSKnownZero <<= ShiftAmt;
1420 RHSKnownOne <<= ShiftAmt;
1421 // low bits known zero.
1423 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1426 case Instruction::LShr:
1427 // For a logical shift right
1428 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1429 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1431 // Unsigned shift right.
1432 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1433 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1434 RHSKnownZero, RHSKnownOne, Depth+1))
1436 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1437 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1438 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1440 // Compute the new bits that are at the top now.
1441 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1442 RHSKnownZero |= HighBits; // high bits known zero.
1446 case Instruction::AShr:
1447 // If this is an arithmetic shift right and only the low-bit is set, we can
1448 // always convert this into a logical shr, even if the shift amount is
1449 // variable. The low bit of the shift cannot be an input sign bit unless
1450 // the shift amount is >= the size of the datatype, which is undefined.
1451 if (DemandedMask == 1) {
1452 // Perform the logical shift right.
1453 Instruction *NewVal = BinaryOperator::CreateLShr(
1454 I->getOperand(0), I->getOperand(1), I->getName());
1455 return InsertNewInstBefore(NewVal, *I);
1458 // If the sign bit is the only bit demanded by this ashr, then there is no
1459 // need to do it, the shift doesn't change the high bit.
1460 if (DemandedMask.isSignBit())
1461 return I->getOperand(0);
1463 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1464 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1466 // Signed shift right.
1467 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1468 // If any of the "high bits" are demanded, we should set the sign bit as
1470 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1471 DemandedMaskIn.set(BitWidth-1);
1472 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1473 RHSKnownZero, RHSKnownOne, Depth+1))
1475 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1476 // Compute the new bits that are at the top now.
1477 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1478 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1479 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1481 // Handle the sign bits.
1482 APInt SignBit(APInt::getSignBit(BitWidth));
1483 // Adjust to where it is now in the mask.
1484 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1486 // If the input sign bit is known to be zero, or if none of the top bits
1487 // are demanded, turn this into an unsigned shift right.
1488 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1489 (HighBits & ~DemandedMask) == HighBits) {
1490 // Perform the logical shift right.
1491 Instruction *NewVal = BinaryOperator::CreateLShr(
1492 I->getOperand(0), SA, I->getName());
1493 return InsertNewInstBefore(NewVal, *I);
1494 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1495 RHSKnownOne |= HighBits;
1499 case Instruction::SRem:
1500 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1501 APInt RA = Rem->getValue().abs();
1502 if (RA.isPowerOf2()) {
1503 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1504 return I->getOperand(0);
1506 APInt LowBits = RA - 1;
1507 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1508 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1509 LHSKnownZero, LHSKnownOne, Depth+1))
1512 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1513 LHSKnownZero |= ~LowBits;
1515 KnownZero |= LHSKnownZero & DemandedMask;
1517 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1521 case Instruction::URem: {
1522 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1523 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1524 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1525 KnownZero2, KnownOne2, Depth+1) ||
1526 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1527 KnownZero2, KnownOne2, Depth+1))
1530 unsigned Leaders = KnownZero2.countLeadingOnes();
1531 Leaders = std::max(Leaders,
1532 KnownZero2.countLeadingOnes());
1533 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1536 case Instruction::Call:
1537 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1538 switch (II->getIntrinsicID()) {
1540 case Intrinsic::bswap: {
1541 // If the only bits demanded come from one byte of the bswap result,
1542 // just shift the input byte into position to eliminate the bswap.
1543 unsigned NLZ = DemandedMask.countLeadingZeros();
1544 unsigned NTZ = DemandedMask.countTrailingZeros();
1546 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1547 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1548 // have 14 leading zeros, round to 8.
1551 // If we need exactly one byte, we can do this transformation.
1552 if (BitWidth-NLZ-NTZ == 8) {
1553 unsigned ResultBit = NTZ;
1554 unsigned InputBit = BitWidth-NTZ-8;
1556 // Replace this with either a left or right shift to get the byte into
1558 Instruction *NewVal;
1559 if (InputBit > ResultBit)
1560 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1561 ConstantInt::get(I->getType(), InputBit-ResultBit));
1563 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1564 ConstantInt::get(I->getType(), ResultBit-InputBit));
1565 NewVal->takeName(I);
1566 return InsertNewInstBefore(NewVal, *I);
1569 // TODO: Could compute known zero/one bits based on the input.
1574 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1578 // If the client is only demanding bits that we know, return the known
1580 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1581 return Constant::getIntegerValue(VTy, RHSKnownOne);
1586 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1587 /// any number of elements. DemandedElts contains the set of elements that are
1588 /// actually used by the caller. This method analyzes which elements of the
1589 /// operand are undef and returns that information in UndefElts.
1591 /// If the information about demanded elements can be used to simplify the
1592 /// operation, the operation is simplified, then the resultant value is
1593 /// returned. This returns null if no change was made.
1594 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1597 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1598 APInt EltMask(APInt::getAllOnesValue(VWidth));
1599 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1601 if (isa<UndefValue>(V)) {
1602 // If the entire vector is undefined, just return this info.
1603 UndefElts = EltMask;
1605 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1606 UndefElts = EltMask;
1607 return UndefValue::get(V->getType());
1611 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1612 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1613 Constant *Undef = UndefValue::get(EltTy);
1615 std::vector<Constant*> Elts;
1616 for (unsigned i = 0; i != VWidth; ++i)
1617 if (!DemandedElts[i]) { // If not demanded, set to undef.
1618 Elts.push_back(Undef);
1620 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1621 Elts.push_back(Undef);
1623 } else { // Otherwise, defined.
1624 Elts.push_back(CP->getOperand(i));
1627 // If we changed the constant, return it.
1628 Constant *NewCP = ConstantVector::get(Elts);
1629 return NewCP != CP ? NewCP : 0;
1630 } else if (isa<ConstantAggregateZero>(V)) {
1631 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1634 // Check if this is identity. If so, return 0 since we are not simplifying
1636 if (DemandedElts == ((1ULL << VWidth) -1))
1639 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1640 Constant *Zero = Constant::getNullValue(EltTy);
1641 Constant *Undef = UndefValue::get(EltTy);
1642 std::vector<Constant*> Elts;
1643 for (unsigned i = 0; i != VWidth; ++i) {
1644 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1645 Elts.push_back(Elt);
1647 UndefElts = DemandedElts ^ EltMask;
1648 return ConstantVector::get(Elts);
1651 // Limit search depth.
1655 // If multiple users are using the root value, procede with
1656 // simplification conservatively assuming that all elements
1658 if (!V->hasOneUse()) {
1659 // Quit if we find multiple users of a non-root value though.
1660 // They'll be handled when it's their turn to be visited by
1661 // the main instcombine process.
1663 // TODO: Just compute the UndefElts information recursively.
1666 // Conservatively assume that all elements are needed.
1667 DemandedElts = EltMask;
1670 Instruction *I = dyn_cast<Instruction>(V);
1671 if (!I) return 0; // Only analyze instructions.
1673 bool MadeChange = false;
1674 APInt UndefElts2(VWidth, 0);
1676 switch (I->getOpcode()) {
1679 case Instruction::InsertElement: {
1680 // If this is a variable index, we don't know which element it overwrites.
1681 // demand exactly the same input as we produce.
1682 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1684 // Note that we can't propagate undef elt info, because we don't know
1685 // which elt is getting updated.
1686 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1687 UndefElts2, Depth+1);
1688 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1692 // If this is inserting an element that isn't demanded, remove this
1694 unsigned IdxNo = Idx->getZExtValue();
1695 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1697 return I->getOperand(0);
1700 // Otherwise, the element inserted overwrites whatever was there, so the
1701 // input demanded set is simpler than the output set.
1702 APInt DemandedElts2 = DemandedElts;
1703 DemandedElts2.clear(IdxNo);
1704 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1705 UndefElts, Depth+1);
1706 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1708 // The inserted element is defined.
1709 UndefElts.clear(IdxNo);
1712 case Instruction::ShuffleVector: {
1713 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1714 uint64_t LHSVWidth =
1715 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1716 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1717 for (unsigned i = 0; i < VWidth; i++) {
1718 if (DemandedElts[i]) {
1719 unsigned MaskVal = Shuffle->getMaskValue(i);
1720 if (MaskVal != -1u) {
1721 assert(MaskVal < LHSVWidth * 2 &&
1722 "shufflevector mask index out of range!");
1723 if (MaskVal < LHSVWidth)
1724 LeftDemanded.set(MaskVal);
1726 RightDemanded.set(MaskVal - LHSVWidth);
1731 APInt UndefElts4(LHSVWidth, 0);
1732 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1733 UndefElts4, Depth+1);
1734 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1736 APInt UndefElts3(LHSVWidth, 0);
1737 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1738 UndefElts3, Depth+1);
1739 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1741 bool NewUndefElts = false;
1742 for (unsigned i = 0; i < VWidth; i++) {
1743 unsigned MaskVal = Shuffle->getMaskValue(i);
1744 if (MaskVal == -1u) {
1746 } else if (MaskVal < LHSVWidth) {
1747 if (UndefElts4[MaskVal]) {
1748 NewUndefElts = true;
1752 if (UndefElts3[MaskVal - LHSVWidth]) {
1753 NewUndefElts = true;
1760 // Add additional discovered undefs.
1761 std::vector<Constant*> Elts;
1762 for (unsigned i = 0; i < VWidth; ++i) {
1764 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1766 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1767 Shuffle->getMaskValue(i)));
1769 I->setOperand(2, ConstantVector::get(Elts));
1774 case Instruction::BitCast: {
1775 // Vector->vector casts only.
1776 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1778 unsigned InVWidth = VTy->getNumElements();
1779 APInt InputDemandedElts(InVWidth, 0);
1782 if (VWidth == InVWidth) {
1783 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1784 // elements as are demanded of us.
1786 InputDemandedElts = DemandedElts;
1787 } else if (VWidth > InVWidth) {
1791 // If there are more elements in the result than there are in the source,
1792 // then an input element is live if any of the corresponding output
1793 // elements are live.
1794 Ratio = VWidth/InVWidth;
1795 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1796 if (DemandedElts[OutIdx])
1797 InputDemandedElts.set(OutIdx/Ratio);
1803 // If there are more elements in the source than there are in the result,
1804 // then an input element is live if the corresponding output element is
1806 Ratio = InVWidth/VWidth;
1807 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1808 if (DemandedElts[InIdx/Ratio])
1809 InputDemandedElts.set(InIdx);
1812 // div/rem demand all inputs, because they don't want divide by zero.
1813 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1814 UndefElts2, Depth+1);
1816 I->setOperand(0, TmpV);
1820 UndefElts = UndefElts2;
1821 if (VWidth > InVWidth) {
1822 llvm_unreachable("Unimp");
1823 // If there are more elements in the result than there are in the source,
1824 // then an output element is undef if the corresponding input element is
1826 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1827 if (UndefElts2[OutIdx/Ratio])
1828 UndefElts.set(OutIdx);
1829 } else if (VWidth < InVWidth) {
1830 llvm_unreachable("Unimp");
1831 // If there are more elements in the source than there are in the result,
1832 // then a result element is undef if all of the corresponding input
1833 // elements are undef.
1834 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1835 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1836 if (!UndefElts2[InIdx]) // Not undef?
1837 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1841 case Instruction::And:
1842 case Instruction::Or:
1843 case Instruction::Xor:
1844 case Instruction::Add:
1845 case Instruction::Sub:
1846 case Instruction::Mul:
1847 // div/rem demand all inputs, because they don't want divide by zero.
1848 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1849 UndefElts, Depth+1);
1850 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1851 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1852 UndefElts2, Depth+1);
1853 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1855 // Output elements are undefined if both are undefined. Consider things
1856 // like undef&0. The result is known zero, not undef.
1857 UndefElts &= UndefElts2;
1860 case Instruction::Call: {
1861 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1863 switch (II->getIntrinsicID()) {
1866 // Binary vector operations that work column-wise. A dest element is a
1867 // function of the corresponding input elements from the two inputs.
1868 case Intrinsic::x86_sse_sub_ss:
1869 case Intrinsic::x86_sse_mul_ss:
1870 case Intrinsic::x86_sse_min_ss:
1871 case Intrinsic::x86_sse_max_ss:
1872 case Intrinsic::x86_sse2_sub_sd:
1873 case Intrinsic::x86_sse2_mul_sd:
1874 case Intrinsic::x86_sse2_min_sd:
1875 case Intrinsic::x86_sse2_max_sd:
1876 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1877 UndefElts, Depth+1);
1878 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1879 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1880 UndefElts2, Depth+1);
1881 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1883 // If only the low elt is demanded and this is a scalarizable intrinsic,
1884 // scalarize it now.
1885 if (DemandedElts == 1) {
1886 switch (II->getIntrinsicID()) {
1888 case Intrinsic::x86_sse_sub_ss:
1889 case Intrinsic::x86_sse_mul_ss:
1890 case Intrinsic::x86_sse2_sub_sd:
1891 case Intrinsic::x86_sse2_mul_sd:
1892 // TODO: Lower MIN/MAX/ABS/etc
1893 Value *LHS = II->getOperand(1);
1894 Value *RHS = II->getOperand(2);
1895 // Extract the element as scalars.
1896 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1897 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1898 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1899 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1901 switch (II->getIntrinsicID()) {
1902 default: llvm_unreachable("Case stmts out of sync!");
1903 case Intrinsic::x86_sse_sub_ss:
1904 case Intrinsic::x86_sse2_sub_sd:
1905 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1906 II->getName()), *II);
1908 case Intrinsic::x86_sse_mul_ss:
1909 case Intrinsic::x86_sse2_mul_sd:
1910 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1911 II->getName()), *II);
1916 InsertElementInst::Create(
1917 UndefValue::get(II->getType()), TmpV,
1918 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1919 InsertNewInstBefore(New, *II);
1924 // Output elements are undefined if both are undefined. Consider things
1925 // like undef&0. The result is known zero, not undef.
1926 UndefElts &= UndefElts2;
1932 return MadeChange ? I : 0;
1936 /// AssociativeOpt - Perform an optimization on an associative operator. This
1937 /// function is designed to check a chain of associative operators for a
1938 /// potential to apply a certain optimization. Since the optimization may be
1939 /// applicable if the expression was reassociated, this checks the chain, then
1940 /// reassociates the expression as necessary to expose the optimization
1941 /// opportunity. This makes use of a special Functor, which must define
1942 /// 'shouldApply' and 'apply' methods.
1944 template<typename Functor>
1945 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1946 unsigned Opcode = Root.getOpcode();
1947 Value *LHS = Root.getOperand(0);
1949 // Quick check, see if the immediate LHS matches...
1950 if (F.shouldApply(LHS))
1951 return F.apply(Root);
1953 // Otherwise, if the LHS is not of the same opcode as the root, return.
1954 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1955 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1956 // Should we apply this transform to the RHS?
1957 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1959 // If not to the RHS, check to see if we should apply to the LHS...
1960 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1961 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1965 // If the functor wants to apply the optimization to the RHS of LHSI,
1966 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1968 // Now all of the instructions are in the current basic block, go ahead
1969 // and perform the reassociation.
1970 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1972 // First move the selected RHS to the LHS of the root...
1973 Root.setOperand(0, LHSI->getOperand(1));
1975 // Make what used to be the LHS of the root be the user of the root...
1976 Value *ExtraOperand = TmpLHSI->getOperand(1);
1977 if (&Root == TmpLHSI) {
1978 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1981 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1982 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1983 BasicBlock::iterator ARI = &Root; ++ARI;
1984 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1987 // Now propagate the ExtraOperand down the chain of instructions until we
1989 while (TmpLHSI != LHSI) {
1990 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1991 // Move the instruction to immediately before the chain we are
1992 // constructing to avoid breaking dominance properties.
1993 NextLHSI->moveBefore(ARI);
1996 Value *NextOp = NextLHSI->getOperand(1);
1997 NextLHSI->setOperand(1, ExtraOperand);
1999 ExtraOperand = NextOp;
2002 // Now that the instructions are reassociated, have the functor perform
2003 // the transformation...
2004 return F.apply(Root);
2007 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2014 // AddRHS - Implements: X + X --> X << 1
2017 explicit AddRHS(Value *rhs) : RHS(rhs) {}
2018 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2019 Instruction *apply(BinaryOperator &Add) const {
2020 return BinaryOperator::CreateShl(Add.getOperand(0),
2021 ConstantInt::get(Add.getType(), 1));
2025 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2027 struct AddMaskingAnd {
2029 explicit AddMaskingAnd(Constant *c) : C2(c) {}
2030 bool shouldApply(Value *LHS) const {
2032 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2033 ConstantExpr::getAnd(C1, C2)->isNullValue();
2035 Instruction *apply(BinaryOperator &Add) const {
2036 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
2042 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2044 if (CastInst *CI = dyn_cast<CastInst>(&I))
2045 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
2047 // Figure out if the constant is the left or the right argument.
2048 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2049 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2051 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2053 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2054 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2057 Value *Op0 = SO, *Op1 = ConstOperand;
2059 std::swap(Op0, Op1);
2061 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2062 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
2063 SO->getName()+".op");
2064 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
2065 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
2066 SO->getName()+".cmp");
2067 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
2068 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
2069 SO->getName()+".cmp");
2070 llvm_unreachable("Unknown binary instruction type!");
2073 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2074 // constant as the other operand, try to fold the binary operator into the
2075 // select arguments. This also works for Cast instructions, which obviously do
2076 // not have a second operand.
2077 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2079 // Don't modify shared select instructions
2080 if (!SI->hasOneUse()) return 0;
2081 Value *TV = SI->getOperand(1);
2082 Value *FV = SI->getOperand(2);
2084 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2085 // Bool selects with constant operands can be folded to logical ops.
2086 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
2088 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2089 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2091 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
2098 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
2099 /// has a PHI node as operand #0, see if we can fold the instruction into the
2100 /// PHI (which is only possible if all operands to the PHI are constants).
2102 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
2103 /// that would normally be unprofitable because they strongly encourage jump
2105 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
2106 bool AllowAggressive) {
2107 AllowAggressive = false;
2108 PHINode *PN = cast<PHINode>(I.getOperand(0));
2109 unsigned NumPHIValues = PN->getNumIncomingValues();
2110 if (NumPHIValues == 0 ||
2111 // We normally only transform phis with a single use, unless we're trying
2112 // hard to make jump threading happen.
2113 (!PN->hasOneUse() && !AllowAggressive))
2117 // Check to see if all of the operands of the PHI are simple constants
2118 // (constantint/constantfp/undef). If there is one non-constant value,
2119 // remember the BB it is in. If there is more than one or if *it* is a PHI,
2120 // bail out. We don't do arbitrary constant expressions here because moving
2121 // their computation can be expensive without a cost model.
2122 BasicBlock *NonConstBB = 0;
2123 for (unsigned i = 0; i != NumPHIValues; ++i)
2124 if (!isa<Constant>(PN->getIncomingValue(i)) ||
2125 isa<ConstantExpr>(PN->getIncomingValue(i))) {
2126 if (NonConstBB) return 0; // More than one non-const value.
2127 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2128 NonConstBB = PN->getIncomingBlock(i);
2130 // If the incoming non-constant value is in I's block, we have an infinite
2132 if (NonConstBB == I.getParent())
2136 // If there is exactly one non-constant value, we can insert a copy of the
2137 // operation in that block. However, if this is a critical edge, we would be
2138 // inserting the computation one some other paths (e.g. inside a loop). Only
2139 // do this if the pred block is unconditionally branching into the phi block.
2140 if (NonConstBB != 0 && !AllowAggressive) {
2141 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2142 if (!BI || !BI->isUnconditional()) return 0;
2145 // Okay, we can do the transformation: create the new PHI node.
2146 PHINode *NewPN = PHINode::Create(I.getType(), "");
2147 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2148 InsertNewInstBefore(NewPN, *PN);
2149 NewPN->takeName(PN);
2151 // Next, add all of the operands to the PHI.
2152 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
2153 // We only currently try to fold the condition of a select when it is a phi,
2154 // not the true/false values.
2155 Value *TrueV = SI->getTrueValue();
2156 Value *FalseV = SI->getFalseValue();
2157 BasicBlock *PhiTransBB = PN->getParent();
2158 for (unsigned i = 0; i != NumPHIValues; ++i) {
2159 BasicBlock *ThisBB = PN->getIncomingBlock(i);
2160 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
2161 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
2163 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2164 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
2166 assert(PN->getIncomingBlock(i) == NonConstBB);
2167 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
2169 "phitmp", NonConstBB->getTerminator());
2170 Worklist.Add(cast<Instruction>(InV));
2172 NewPN->addIncoming(InV, ThisBB);
2174 } else if (I.getNumOperands() == 2) {
2175 Constant *C = cast<Constant>(I.getOperand(1));
2176 for (unsigned i = 0; i != NumPHIValues; ++i) {
2178 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2179 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2180 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2182 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2184 assert(PN->getIncomingBlock(i) == NonConstBB);
2185 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2186 InV = BinaryOperator::Create(BO->getOpcode(),
2187 PN->getIncomingValue(i), C, "phitmp",
2188 NonConstBB->getTerminator());
2189 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2190 InV = CmpInst::Create(CI->getOpcode(),
2192 PN->getIncomingValue(i), C, "phitmp",
2193 NonConstBB->getTerminator());
2195 llvm_unreachable("Unknown binop!");
2197 Worklist.Add(cast<Instruction>(InV));
2199 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2202 CastInst *CI = cast<CastInst>(&I);
2203 const Type *RetTy = CI->getType();
2204 for (unsigned i = 0; i != NumPHIValues; ++i) {
2206 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2207 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2209 assert(PN->getIncomingBlock(i) == NonConstBB);
2210 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2211 I.getType(), "phitmp",
2212 NonConstBB->getTerminator());
2213 Worklist.Add(cast<Instruction>(InV));
2215 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2218 return ReplaceInstUsesWith(I, NewPN);
2222 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2223 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2224 /// This basically requires proving that the add in the original type would not
2225 /// overflow to change the sign bit or have a carry out.
2226 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2227 // There are different heuristics we can use for this. Here are some simple
2230 // Add has the property that adding any two 2's complement numbers can only
2231 // have one carry bit which can change a sign. As such, if LHS and RHS each
2232 // have at least two sign bits, we know that the addition of the two values
2233 // will sign extend fine.
2234 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2238 // If one of the operands only has one non-zero bit, and if the other operand
2239 // has a known-zero bit in a more significant place than it (not including the
2240 // sign bit) the ripple may go up to and fill the zero, but won't change the
2241 // sign. For example, (X & ~4) + 1.
2249 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2250 bool Changed = SimplifyCommutative(I);
2251 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2253 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
2254 I.hasNoUnsignedWrap(), TD))
2255 return ReplaceInstUsesWith(I, V);
2258 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2259 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2260 // X + (signbit) --> X ^ signbit
2261 const APInt& Val = CI->getValue();
2262 uint32_t BitWidth = Val.getBitWidth();
2263 if (Val == APInt::getSignBit(BitWidth))
2264 return BinaryOperator::CreateXor(LHS, RHS);
2266 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2267 // (X & 254)+1 -> (X&254)|1
2268 if (SimplifyDemandedInstructionBits(I))
2271 // zext(bool) + C -> bool ? C + 1 : C
2272 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2273 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2274 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2277 if (isa<PHINode>(LHS))
2278 if (Instruction *NV = FoldOpIntoPhi(I))
2281 ConstantInt *XorRHS = 0;
2283 if (isa<ConstantInt>(RHSC) &&
2284 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2285 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2286 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2288 uint32_t Size = TySizeBits / 2;
2289 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2290 APInt CFF80Val(-C0080Val);
2292 if (TySizeBits > Size) {
2293 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2294 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2295 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2296 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2297 // This is a sign extend if the top bits are known zero.
2298 if (!MaskedValueIsZero(XorLHS,
2299 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2300 Size = 0; // Not a sign ext, but can't be any others either.
2305 C0080Val = APIntOps::lshr(C0080Val, Size);
2306 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2307 } while (Size >= 1);
2309 // FIXME: This shouldn't be necessary. When the backends can handle types
2310 // with funny bit widths then this switch statement should be removed. It
2311 // is just here to get the size of the "middle" type back up to something
2312 // that the back ends can handle.
2313 const Type *MiddleType = 0;
2316 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2317 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2318 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2321 Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext");
2322 return new SExtInst(NewTrunc, I.getType(), I.getName());
2327 if (I.getType() == Type::getInt1Ty(*Context))
2328 return BinaryOperator::CreateXor(LHS, RHS);
2331 if (I.getType()->isInteger()) {
2332 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2335 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2336 if (RHSI->getOpcode() == Instruction::Sub)
2337 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2338 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2340 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2341 if (LHSI->getOpcode() == Instruction::Sub)
2342 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2343 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2348 // -A + -B --> -(A + B)
2349 if (Value *LHSV = dyn_castNegVal(LHS)) {
2350 if (LHS->getType()->isIntOrIntVector()) {
2351 if (Value *RHSV = dyn_castNegVal(RHS)) {
2352 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
2353 return BinaryOperator::CreateNeg(NewAdd);
2357 return BinaryOperator::CreateSub(RHS, LHSV);
2361 if (!isa<Constant>(RHS))
2362 if (Value *V = dyn_castNegVal(RHS))
2363 return BinaryOperator::CreateSub(LHS, V);
2367 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2368 if (X == RHS) // X*C + X --> X * (C+1)
2369 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2371 // X*C1 + X*C2 --> X * (C1+C2)
2373 if (X == dyn_castFoldableMul(RHS, C1))
2374 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2377 // X + X*C --> X * (C+1)
2378 if (dyn_castFoldableMul(RHS, C2) == LHS)
2379 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2381 // X + ~X --> -1 since ~X = -X-1
2382 if (dyn_castNotVal(LHS) == RHS ||
2383 dyn_castNotVal(RHS) == LHS)
2384 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2387 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2388 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2389 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2392 // A+B --> A|B iff A and B have no bits set in common.
2393 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2394 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2395 APInt LHSKnownOne(IT->getBitWidth(), 0);
2396 APInt LHSKnownZero(IT->getBitWidth(), 0);
2397 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2398 if (LHSKnownZero != 0) {
2399 APInt RHSKnownOne(IT->getBitWidth(), 0);
2400 APInt RHSKnownZero(IT->getBitWidth(), 0);
2401 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2403 // No bits in common -> bitwise or.
2404 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2405 return BinaryOperator::CreateOr(LHS, RHS);
2409 // W*X + Y*Z --> W * (X+Z) iff W == Y
2410 if (I.getType()->isIntOrIntVector()) {
2411 Value *W, *X, *Y, *Z;
2412 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2413 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2417 } else if (Y == X) {
2419 } else if (X == Z) {
2426 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
2427 return BinaryOperator::CreateMul(W, NewAdd);
2432 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2434 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2435 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2437 // (X & FF00) + xx00 -> (X+xx00) & FF00
2438 if (LHS->hasOneUse() &&
2439 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2440 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2441 if (Anded == CRHS) {
2442 // See if all bits from the first bit set in the Add RHS up are included
2443 // in the mask. First, get the rightmost bit.
2444 const APInt& AddRHSV = CRHS->getValue();
2446 // Form a mask of all bits from the lowest bit added through the top.
2447 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2449 // See if the and mask includes all of these bits.
2450 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2452 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2453 // Okay, the xform is safe. Insert the new add pronto.
2454 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
2455 return BinaryOperator::CreateAnd(NewAdd, C2);
2460 // Try to fold constant add into select arguments.
2461 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2462 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2466 // add (select X 0 (sub n A)) A --> select X A n
2468 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2471 SI = dyn_cast<SelectInst>(RHS);
2474 if (SI && SI->hasOneUse()) {
2475 Value *TV = SI->getTrueValue();
2476 Value *FV = SI->getFalseValue();
2479 // Can we fold the add into the argument of the select?
2480 // We check both true and false select arguments for a matching subtract.
2481 if (match(FV, m_Zero()) &&
2482 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2483 // Fold the add into the true select value.
2484 return SelectInst::Create(SI->getCondition(), N, A);
2485 if (match(TV, m_Zero()) &&
2486 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2487 // Fold the add into the false select value.
2488 return SelectInst::Create(SI->getCondition(), A, N);
2492 // Check for (add (sext x), y), see if we can merge this into an
2493 // integer add followed by a sext.
2494 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2495 // (add (sext x), cst) --> (sext (add x, cst'))
2496 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2498 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2499 if (LHSConv->hasOneUse() &&
2500 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2501 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2502 // Insert the new, smaller add.
2503 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2505 return new SExtInst(NewAdd, I.getType());
2509 // (add (sext x), (sext y)) --> (sext (add int x, y))
2510 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2511 // Only do this if x/y have the same type, if at last one of them has a
2512 // single use (so we don't increase the number of sexts), and if the
2513 // integer add will not overflow.
2514 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2515 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2516 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2517 RHSConv->getOperand(0))) {
2518 // Insert the new integer add.
2519 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2520 RHSConv->getOperand(0), "addconv");
2521 return new SExtInst(NewAdd, I.getType());
2526 return Changed ? &I : 0;
2529 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2530 bool Changed = SimplifyCommutative(I);
2531 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2533 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2535 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2536 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2537 (I.getType())->getValueAPF()))
2538 return ReplaceInstUsesWith(I, LHS);
2541 if (isa<PHINode>(LHS))
2542 if (Instruction *NV = FoldOpIntoPhi(I))
2547 // -A + -B --> -(A + B)
2548 if (Value *LHSV = dyn_castFNegVal(LHS))
2549 return BinaryOperator::CreateFSub(RHS, LHSV);
2552 if (!isa<Constant>(RHS))
2553 if (Value *V = dyn_castFNegVal(RHS))
2554 return BinaryOperator::CreateFSub(LHS, V);
2556 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2557 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2558 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2559 return ReplaceInstUsesWith(I, LHS);
2561 // Check for (add double (sitofp x), y), see if we can merge this into an
2562 // integer add followed by a promotion.
2563 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2564 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2565 // ... if the constant fits in the integer value. This is useful for things
2566 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2567 // requires a constant pool load, and generally allows the add to be better
2569 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2571 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2572 if (LHSConv->hasOneUse() &&
2573 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2574 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2575 // Insert the new integer add.
2576 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2578 return new SIToFPInst(NewAdd, I.getType());
2582 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2583 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2584 // Only do this if x/y have the same type, if at last one of them has a
2585 // single use (so we don't increase the number of int->fp conversions),
2586 // and if the integer add will not overflow.
2587 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2588 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2589 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2590 RHSConv->getOperand(0))) {
2591 // Insert the new integer add.
2592 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2593 RHSConv->getOperand(0),"addconv");
2594 return new SIToFPInst(NewAdd, I.getType());
2599 return Changed ? &I : 0;
2603 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
2604 /// code necessary to compute the offset from the base pointer (without adding
2605 /// in the base pointer). Return the result as a signed integer of intptr size.
2606 static Value *EmitGEPOffset(User *GEP, InstCombiner &IC) {
2607 TargetData &TD = *IC.getTargetData();
2608 gep_type_iterator GTI = gep_type_begin(GEP);
2609 const Type *IntPtrTy = TD.getIntPtrType(GEP->getContext());
2610 Value *Result = Constant::getNullValue(IntPtrTy);
2612 // Build a mask for high order bits.
2613 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2614 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
2616 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
2619 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
2620 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
2621 if (OpC->isZero()) continue;
2623 // Handle a struct index, which adds its field offset to the pointer.
2624 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
2625 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
2627 Result = IC.Builder->CreateAdd(Result,
2628 ConstantInt::get(IntPtrTy, Size),
2629 GEP->getName()+".offs");
2633 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
2635 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
2636 Scale = ConstantExpr::getMul(OC, Scale);
2637 // Emit an add instruction.
2638 Result = IC.Builder->CreateAdd(Result, Scale, GEP->getName()+".offs");
2641 // Convert to correct type.
2642 if (Op->getType() != IntPtrTy)
2643 Op = IC.Builder->CreateIntCast(Op, IntPtrTy, true, Op->getName()+".c");
2645 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
2646 // We'll let instcombine(mul) convert this to a shl if possible.
2647 Op = IC.Builder->CreateMul(Op, Scale, GEP->getName()+".idx");
2650 // Emit an add instruction.
2651 Result = IC.Builder->CreateAdd(Op, Result, GEP->getName()+".offs");
2657 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
2658 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
2659 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
2660 /// be complex, and scales are involved. The above expression would also be
2661 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
2662 /// This later form is less amenable to optimization though, and we are allowed
2663 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
2665 /// If we can't emit an optimized form for this expression, this returns null.
2667 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
2669 TargetData &TD = *IC.getTargetData();
2670 gep_type_iterator GTI = gep_type_begin(GEP);
2672 // Check to see if this gep only has a single variable index. If so, and if
2673 // any constant indices are a multiple of its scale, then we can compute this
2674 // in terms of the scale of the variable index. For example, if the GEP
2675 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
2676 // because the expression will cross zero at the same point.
2677 unsigned i, e = GEP->getNumOperands();
2679 for (i = 1; i != e; ++i, ++GTI) {
2680 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
2681 // Compute the aggregate offset of constant indices.
2682 if (CI->isZero()) continue;
2684 // Handle a struct index, which adds its field offset to the pointer.
2685 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
2686 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
2688 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
2689 Offset += Size*CI->getSExtValue();
2692 // Found our variable index.
2697 // If there are no variable indices, we must have a constant offset, just
2698 // evaluate it the general way.
2699 if (i == e) return 0;
2701 Value *VariableIdx = GEP->getOperand(i);
2702 // Determine the scale factor of the variable element. For example, this is
2703 // 4 if the variable index is into an array of i32.
2704 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
2706 // Verify that there are no other variable indices. If so, emit the hard way.
2707 for (++i, ++GTI; i != e; ++i, ++GTI) {
2708 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
2711 // Compute the aggregate offset of constant indices.
2712 if (CI->isZero()) continue;
2714 // Handle a struct index, which adds its field offset to the pointer.
2715 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
2716 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
2718 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
2719 Offset += Size*CI->getSExtValue();
2723 // Okay, we know we have a single variable index, which must be a
2724 // pointer/array/vector index. If there is no offset, life is simple, return
2726 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2728 // Cast to intptrty in case a truncation occurs. If an extension is needed,
2729 // we don't need to bother extending: the extension won't affect where the
2730 // computation crosses zero.
2731 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
2732 VariableIdx = new TruncInst(VariableIdx,
2733 TD.getIntPtrType(VariableIdx->getContext()),
2734 VariableIdx->getName(), &I);
2738 // Otherwise, there is an index. The computation we will do will be modulo
2739 // the pointer size, so get it.
2740 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
2742 Offset &= PtrSizeMask;
2743 VariableScale &= PtrSizeMask;
2745 // To do this transformation, any constant index must be a multiple of the
2746 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
2747 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
2748 // multiple of the variable scale.
2749 int64_t NewOffs = Offset / (int64_t)VariableScale;
2750 if (Offset != NewOffs*(int64_t)VariableScale)
2753 // Okay, we can do this evaluation. Start by converting the index to intptr.
2754 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
2755 if (VariableIdx->getType() != IntPtrTy)
2756 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
2758 VariableIdx->getName(), &I);
2759 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
2760 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
2764 /// Optimize pointer differences into the same array into a size. Consider:
2765 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
2766 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
2768 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
2770 assert(TD && "Must have target data info for this");
2772 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
2775 GetElementPtrInst *GEP = 0;
2776 ConstantExpr *CstGEP = 0;
2778 // TODO: Could also optimize &A[i] - &A[j] -> "i-j", and "&A.foo[i] - &A.foo".
2779 // For now we require one side to be the base pointer "A" or a constant
2780 // expression derived from it.
2781 if (GetElementPtrInst *LHSGEP = dyn_cast<GetElementPtrInst>(LHS)) {
2783 if (LHSGEP->getOperand(0) == RHS) {
2786 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(RHS)) {
2787 // (gep X, ...) - (ce_gep X, ...)
2788 if (CE->getOpcode() == Instruction::GetElementPtr &&
2789 LHSGEP->getOperand(0) == CE->getOperand(0)) {
2797 if (GetElementPtrInst *RHSGEP = dyn_cast<GetElementPtrInst>(RHS)) {
2799 if (RHSGEP->getOperand(0) == LHS) {
2802 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(LHS)) {
2803 // (ce_gep X, ...) - (gep X, ...)
2804 if (CE->getOpcode() == Instruction::GetElementPtr &&
2805 RHSGEP->getOperand(0) == CE->getOperand(0)) {
2816 // Emit the offset of the GEP and an intptr_t.
2817 Value *Result = EmitGEPOffset(GEP, *this);
2819 // If we had a constant expression GEP on the other side offsetting the
2820 // pointer, subtract it from the offset we have.
2822 Value *CstOffset = EmitGEPOffset(CstGEP, *this);
2823 Result = Builder->CreateSub(Result, CstOffset);
2827 // If we have p - gep(p, ...) then we have to negate the result.
2829 Result = Builder->CreateNeg(Result, "diff.neg");
2831 return Builder->CreateIntCast(Result, Ty, true);
2835 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2836 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2838 if (Op0 == Op1) // sub X, X -> 0
2839 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2841 // If this is a 'B = x-(-A)', change to B = x+A. This preserves NSW/NUW.
2842 if (Value *V = dyn_castNegVal(Op1)) {
2843 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2844 Res->setHasNoSignedWrap(I.hasNoSignedWrap());
2845 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
2849 if (isa<UndefValue>(Op0))
2850 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2851 if (isa<UndefValue>(Op1))
2852 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2853 if (I.getType() == Type::getInt1Ty(*Context))
2854 return BinaryOperator::CreateXor(Op0, Op1);
2856 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2857 // Replace (-1 - A) with (~A).
2858 if (C->isAllOnesValue())
2859 return BinaryOperator::CreateNot(Op1);
2861 // C - ~X == X + (1+C)
2863 if (match(Op1, m_Not(m_Value(X))))
2864 return BinaryOperator::CreateAdd(X, AddOne(C));
2866 // -(X >>u 31) -> (X >>s 31)
2867 // -(X >>s 31) -> (X >>u 31)
2869 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2870 if (SI->getOpcode() == Instruction::LShr) {
2871 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2872 // Check to see if we are shifting out everything but the sign bit.
2873 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2874 SI->getType()->getPrimitiveSizeInBits()-1) {
2875 // Ok, the transformation is safe. Insert AShr.
2876 return BinaryOperator::Create(Instruction::AShr,
2877 SI->getOperand(0), CU, SI->getName());
2880 } else if (SI->getOpcode() == Instruction::AShr) {
2881 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2882 // Check to see if we are shifting out everything but the sign bit.
2883 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2884 SI->getType()->getPrimitiveSizeInBits()-1) {
2885 // Ok, the transformation is safe. Insert LShr.
2886 return BinaryOperator::CreateLShr(
2887 SI->getOperand(0), CU, SI->getName());
2894 // Try to fold constant sub into select arguments.
2895 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2896 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2899 // C - zext(bool) -> bool ? C - 1 : C
2900 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2901 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2902 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2905 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2906 if (Op1I->getOpcode() == Instruction::Add) {
2907 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2908 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2910 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2911 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2913 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2914 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2915 // C1-(X+C2) --> (C1-C2)-X
2916 return BinaryOperator::CreateSub(
2917 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2921 if (Op1I->hasOneUse()) {
2922 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2923 // is not used by anyone else...
2925 if (Op1I->getOpcode() == Instruction::Sub) {
2926 // Swap the two operands of the subexpr...
2927 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2928 Op1I->setOperand(0, IIOp1);
2929 Op1I->setOperand(1, IIOp0);
2931 // Create the new top level add instruction...
2932 return BinaryOperator::CreateAdd(Op0, Op1);
2935 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2937 if (Op1I->getOpcode() == Instruction::And &&
2938 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2939 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2941 Value *NewNot = Builder->CreateNot(OtherOp, "B.not");
2942 return BinaryOperator::CreateAnd(Op0, NewNot);
2945 // 0 - (X sdiv C) -> (X sdiv -C)
2946 if (Op1I->getOpcode() == Instruction::SDiv)
2947 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2949 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2950 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2951 ConstantExpr::getNeg(DivRHS));
2953 // X - X*C --> X * (1-C)
2954 ConstantInt *C2 = 0;
2955 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2957 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2959 return BinaryOperator::CreateMul(Op0, CP1);
2964 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2965 if (Op0I->getOpcode() == Instruction::Add) {
2966 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2967 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2968 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2969 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2970 } else if (Op0I->getOpcode() == Instruction::Sub) {
2971 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2972 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2978 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2979 if (X == Op1) // X*C - X --> X * (C-1)
2980 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2982 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2983 if (X == dyn_castFoldableMul(Op1, C2))
2984 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2987 // Optimize pointer differences into the same array into a size. Consider:
2988 // &A[10] - &A[0]: we should compile this to "10".
2990 Value *LHSOp, *RHSOp;
2991 if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
2992 match(Op1, m_PtrToInt(m_Value(RHSOp))))
2993 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
2994 return ReplaceInstUsesWith(I, Res);
2996 // trunc(p)-trunc(q) -> trunc(p-q)
2997 if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
2998 match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
2999 if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType()))
3000 return ReplaceInstUsesWith(I, Res);
3006 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
3007 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3009 // If this is a 'B = x-(-A)', change to B = x+A...
3010 if (Value *V = dyn_castFNegVal(Op1))
3011 return BinaryOperator::CreateFAdd(Op0, V);
3013 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
3014 if (Op1I->getOpcode() == Instruction::FAdd) {
3015 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
3016 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
3018 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
3019 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
3027 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
3028 /// comparison only checks the sign bit. If it only checks the sign bit, set
3029 /// TrueIfSigned if the result of the comparison is true when the input value is
3031 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
3032 bool &TrueIfSigned) {
3034 case ICmpInst::ICMP_SLT: // True if LHS s< 0
3035 TrueIfSigned = true;
3036 return RHS->isZero();
3037 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
3038 TrueIfSigned = true;
3039 return RHS->isAllOnesValue();
3040 case ICmpInst::ICMP_SGT: // True if LHS s> -1
3041 TrueIfSigned = false;
3042 return RHS->isAllOnesValue();
3043 case ICmpInst::ICMP_UGT:
3044 // True if LHS u> RHS and RHS == high-bit-mask - 1
3045 TrueIfSigned = true;
3046 return RHS->getValue() ==
3047 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
3048 case ICmpInst::ICMP_UGE:
3049 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
3050 TrueIfSigned = true;
3051 return RHS->getValue().isSignBit();
3057 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
3058 bool Changed = SimplifyCommutative(I);
3059 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3061 if (isa<UndefValue>(Op1)) // undef * X -> 0
3062 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3064 // Simplify mul instructions with a constant RHS.
3065 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3066 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1C)) {
3068 // ((X << C1)*C2) == (X * (C2 << C1))
3069 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
3070 if (SI->getOpcode() == Instruction::Shl)
3071 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
3072 return BinaryOperator::CreateMul(SI->getOperand(0),
3073 ConstantExpr::getShl(CI, ShOp));
3076 return ReplaceInstUsesWith(I, Op1C); // X * 0 == 0
3077 if (CI->equalsInt(1)) // X * 1 == X
3078 return ReplaceInstUsesWith(I, Op0);
3079 if (CI->isAllOnesValue()) // X * -1 == 0 - X
3080 return BinaryOperator::CreateNeg(Op0, I.getName());
3082 const APInt& Val = cast<ConstantInt>(CI)->getValue();
3083 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
3084 return BinaryOperator::CreateShl(Op0,
3085 ConstantInt::get(Op0->getType(), Val.logBase2()));
3087 } else if (isa<VectorType>(Op1C->getType())) {
3088 if (Op1C->isNullValue())
3089 return ReplaceInstUsesWith(I, Op1C);
3091 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
3092 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
3093 return BinaryOperator::CreateNeg(Op0, I.getName());
3095 // As above, vector X*splat(1.0) -> X in all defined cases.
3096 if (Constant *Splat = Op1V->getSplatValue()) {
3097 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
3098 if (CI->equalsInt(1))
3099 return ReplaceInstUsesWith(I, Op0);
3104 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3105 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
3106 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1C)) {
3107 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
3108 Value *Add = Builder->CreateMul(Op0I->getOperand(0), Op1C, "tmp");
3109 Value *C1C2 = Builder->CreateMul(Op1C, Op0I->getOperand(1));
3110 return BinaryOperator::CreateAdd(Add, C1C2);
3114 // Try to fold constant mul into select arguments.
3115 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3116 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3119 if (isa<PHINode>(Op0))
3120 if (Instruction *NV = FoldOpIntoPhi(I))
3124 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
3125 if (Value *Op1v = dyn_castNegVal(Op1))
3126 return BinaryOperator::CreateMul(Op0v, Op1v);
3128 // (X / Y) * Y = X - (X % Y)
3129 // (X / Y) * -Y = (X % Y) - X
3132 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
3134 (BO->getOpcode() != Instruction::UDiv &&
3135 BO->getOpcode() != Instruction::SDiv)) {
3137 BO = dyn_cast<BinaryOperator>(Op1);
3139 Value *Neg = dyn_castNegVal(Op1C);
3140 if (BO && BO->hasOneUse() &&
3141 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
3142 (BO->getOpcode() == Instruction::UDiv ||
3143 BO->getOpcode() == Instruction::SDiv)) {
3144 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
3146 // If the division is exact, X % Y is zero.
3147 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
3148 if (SDiv->isExact()) {
3150 return ReplaceInstUsesWith(I, Op0BO);
3151 return BinaryOperator::CreateNeg(Op0BO);
3155 if (BO->getOpcode() == Instruction::UDiv)
3156 Rem = Builder->CreateURem(Op0BO, Op1BO);
3158 Rem = Builder->CreateSRem(Op0BO, Op1BO);
3162 return BinaryOperator::CreateSub(Op0BO, Rem);
3163 return BinaryOperator::CreateSub(Rem, Op0BO);
3167 /// i1 mul -> i1 and.
3168 if (I.getType() == Type::getInt1Ty(*Context))
3169 return BinaryOperator::CreateAnd(Op0, Op1);
3171 // X*(1 << Y) --> X << Y
3172 // (1 << Y)*X --> X << Y
3175 if (match(Op0, m_Shl(m_One(), m_Value(Y))))
3176 return BinaryOperator::CreateShl(Op1, Y);
3177 if (match(Op1, m_Shl(m_One(), m_Value(Y))))
3178 return BinaryOperator::CreateShl(Op0, Y);
3181 // If one of the operands of the multiply is a cast from a boolean value, then
3182 // we know the bool is either zero or one, so this is a 'masking' multiply.
3183 // X * Y (where Y is 0 or 1) -> X & (0-Y)
3184 if (!isa<VectorType>(I.getType())) {
3185 // -2 is "-1 << 1" so it is all bits set except the low one.
3186 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
3188 Value *BoolCast = 0, *OtherOp = 0;
3189 if (MaskedValueIsZero(Op0, Negative2))
3190 BoolCast = Op0, OtherOp = Op1;
3191 else if (MaskedValueIsZero(Op1, Negative2))
3192 BoolCast = Op1, OtherOp = Op0;
3195 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
3197 return BinaryOperator::CreateAnd(V, OtherOp);
3201 return Changed ? &I : 0;
3204 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
3205 bool Changed = SimplifyCommutative(I);
3206 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3208 // Simplify mul instructions with a constant RHS...
3209 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3210 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1C)) {
3211 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
3212 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
3213 if (Op1F->isExactlyValue(1.0))
3214 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
3215 } else if (isa<VectorType>(Op1C->getType())) {
3216 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
3217 // As above, vector X*splat(1.0) -> X in all defined cases.
3218 if (Constant *Splat = Op1V->getSplatValue()) {
3219 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
3220 if (F->isExactlyValue(1.0))
3221 return ReplaceInstUsesWith(I, Op0);
3226 // Try to fold constant mul into select arguments.
3227 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3228 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3231 if (isa<PHINode>(Op0))
3232 if (Instruction *NV = FoldOpIntoPhi(I))
3236 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
3237 if (Value *Op1v = dyn_castFNegVal(Op1))
3238 return BinaryOperator::CreateFMul(Op0v, Op1v);
3240 return Changed ? &I : 0;
3243 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
3245 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
3246 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
3248 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
3249 int NonNullOperand = -1;
3250 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3251 if (ST->isNullValue())
3253 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
3254 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3255 if (ST->isNullValue())
3258 if (NonNullOperand == -1)
3261 Value *SelectCond = SI->getOperand(0);
3263 // Change the div/rem to use 'Y' instead of the select.
3264 I.setOperand(1, SI->getOperand(NonNullOperand));
3266 // Okay, we know we replace the operand of the div/rem with 'Y' with no
3267 // problem. However, the select, or the condition of the select may have
3268 // multiple uses. Based on our knowledge that the operand must be non-zero,
3269 // propagate the known value for the select into other uses of it, and
3270 // propagate a known value of the condition into its other users.
3272 // If the select and condition only have a single use, don't bother with this,
3274 if (SI->use_empty() && SelectCond->hasOneUse())
3277 // Scan the current block backward, looking for other uses of SI.
3278 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
3280 while (BBI != BBFront) {
3282 // If we found a call to a function, we can't assume it will return, so
3283 // information from below it cannot be propagated above it.
3284 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
3287 // Replace uses of the select or its condition with the known values.
3288 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
3291 *I = SI->getOperand(NonNullOperand);
3293 } else if (*I == SelectCond) {
3294 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
3295 ConstantInt::getFalse(*Context);
3300 // If we past the instruction, quit looking for it.
3303 if (&*BBI == SelectCond)
3306 // If we ran out of things to eliminate, break out of the loop.
3307 if (SelectCond == 0 && SI == 0)
3315 /// This function implements the transforms on div instructions that work
3316 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
3317 /// used by the visitors to those instructions.
3318 /// @brief Transforms common to all three div instructions
3319 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
3320 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3322 // undef / X -> 0 for integer.
3323 // undef / X -> undef for FP (the undef could be a snan).
3324 if (isa<UndefValue>(Op0)) {
3325 if (Op0->getType()->isFPOrFPVector())
3326 return ReplaceInstUsesWith(I, Op0);
3327 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3330 // X / undef -> undef
3331 if (isa<UndefValue>(Op1))
3332 return ReplaceInstUsesWith(I, Op1);
3337 /// This function implements the transforms common to both integer division
3338 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3339 /// division instructions.
3340 /// @brief Common integer divide transforms
3341 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3342 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3344 // (sdiv X, X) --> 1 (udiv X, X) --> 1
3346 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
3347 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
3348 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
3349 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
3352 Constant *CI = ConstantInt::get(I.getType(), 1);
3353 return ReplaceInstUsesWith(I, CI);
3356 if (Instruction *Common = commonDivTransforms(I))
3359 // Handle cases involving: [su]div X, (select Cond, Y, Z)
3360 // This does not apply for fdiv.
3361 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3364 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3366 if (RHS->equalsInt(1))
3367 return ReplaceInstUsesWith(I, Op0);
3369 // (X / C1) / C2 -> X / (C1*C2)
3370 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3371 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3372 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3373 if (MultiplyOverflows(RHS, LHSRHS,
3374 I.getOpcode()==Instruction::SDiv))
3375 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3377 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3378 ConstantExpr::getMul(RHS, LHSRHS));
3381 if (!RHS->isZero()) { // avoid X udiv 0
3382 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3383 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3385 if (isa<PHINode>(Op0))
3386 if (Instruction *NV = FoldOpIntoPhi(I))
3391 // 0 / X == 0, we don't need to preserve faults!
3392 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3393 if (LHS->equalsInt(0))
3394 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3396 // It can't be division by zero, hence it must be division by one.
3397 if (I.getType() == Type::getInt1Ty(*Context))
3398 return ReplaceInstUsesWith(I, Op0);
3400 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3401 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3404 return ReplaceInstUsesWith(I, Op0);
3410 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3411 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3413 // Handle the integer div common cases
3414 if (Instruction *Common = commonIDivTransforms(I))
3417 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3418 // X udiv C^2 -> X >> C
3419 // Check to see if this is an unsigned division with an exact power of 2,
3420 // if so, convert to a right shift.
3421 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3422 return BinaryOperator::CreateLShr(Op0,
3423 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3425 // X udiv C, where C >= signbit
3426 if (C->getValue().isNegative()) {
3427 Value *IC = Builder->CreateICmpULT( Op0, C);
3428 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3429 ConstantInt::get(I.getType(), 1));
3433 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3434 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3435 if (RHSI->getOpcode() == Instruction::Shl &&
3436 isa<ConstantInt>(RHSI->getOperand(0))) {
3437 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3438 if (C1.isPowerOf2()) {
3439 Value *N = RHSI->getOperand(1);
3440 const Type *NTy = N->getType();
3441 if (uint32_t C2 = C1.logBase2())
3442 N = Builder->CreateAdd(N, ConstantInt::get(NTy, C2), "tmp");
3443 return BinaryOperator::CreateLShr(Op0, N);
3448 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3449 // where C1&C2 are powers of two.
3450 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3451 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3452 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3453 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3454 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3455 // Compute the shift amounts
3456 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3457 // Construct the "on true" case of the select
3458 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3459 Value *TSI = Builder->CreateLShr(Op0, TC, SI->getName()+".t");
3461 // Construct the "on false" case of the select
3462 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3463 Value *FSI = Builder->CreateLShr(Op0, FC, SI->getName()+".f");
3465 // construct the select instruction and return it.
3466 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3472 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3473 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3475 // Handle the integer div common cases
3476 if (Instruction *Common = commonIDivTransforms(I))
3479 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3481 if (RHS->isAllOnesValue())
3482 return BinaryOperator::CreateNeg(Op0);
3484 // sdiv X, C --> ashr X, log2(C)
3485 if (cast<SDivOperator>(&I)->isExact() &&
3486 RHS->getValue().isNonNegative() &&
3487 RHS->getValue().isPowerOf2()) {
3488 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3489 RHS->getValue().exactLogBase2());
3490 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3493 // -X/C --> X/-C provided the negation doesn't overflow.
3494 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3495 if (isa<Constant>(Sub->getOperand(0)) &&
3496 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3497 Sub->hasNoSignedWrap())
3498 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3499 ConstantExpr::getNeg(RHS));
3502 // If the sign bits of both operands are zero (i.e. we can prove they are
3503 // unsigned inputs), turn this into a udiv.
3504 if (I.getType()->isInteger()) {
3505 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3506 if (MaskedValueIsZero(Op0, Mask)) {
3507 if (MaskedValueIsZero(Op1, Mask)) {
3508 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3509 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3511 ConstantInt *ShiftedInt;
3512 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3513 ShiftedInt->getValue().isPowerOf2()) {
3514 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3515 // Safe because the only negative value (1 << Y) can take on is
3516 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3517 // the sign bit set.
3518 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3526 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3527 return commonDivTransforms(I);
3530 /// This function implements the transforms on rem instructions that work
3531 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3532 /// is used by the visitors to those instructions.
3533 /// @brief Transforms common to all three rem instructions
3534 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3535 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3537 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3538 if (I.getType()->isFPOrFPVector())
3539 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3540 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3542 if (isa<UndefValue>(Op1))
3543 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3545 // Handle cases involving: rem X, (select Cond, Y, Z)
3546 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3552 /// This function implements the transforms common to both integer remainder
3553 /// instructions (urem and srem). It is called by the visitors to those integer
3554 /// remainder instructions.
3555 /// @brief Common integer remainder transforms
3556 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3557 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3559 if (Instruction *common = commonRemTransforms(I))
3562 // 0 % X == 0 for integer, we don't need to preserve faults!
3563 if (Constant *LHS = dyn_cast<Constant>(Op0))
3564 if (LHS->isNullValue())
3565 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3567 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3568 // X % 0 == undef, we don't need to preserve faults!
3569 if (RHS->equalsInt(0))
3570 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3572 if (RHS->equalsInt(1)) // X % 1 == 0
3573 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3575 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3576 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3577 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3579 } else if (isa<PHINode>(Op0I)) {
3580 if (Instruction *NV = FoldOpIntoPhi(I))
3584 // See if we can fold away this rem instruction.
3585 if (SimplifyDemandedInstructionBits(I))
3593 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3594 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3596 if (Instruction *common = commonIRemTransforms(I))
3599 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3600 // X urem C^2 -> X and C
3601 // Check to see if this is an unsigned remainder with an exact power of 2,
3602 // if so, convert to a bitwise and.
3603 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3604 if (C->getValue().isPowerOf2())
3605 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3608 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3609 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3610 if (RHSI->getOpcode() == Instruction::Shl &&
3611 isa<ConstantInt>(RHSI->getOperand(0))) {
3612 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3613 Constant *N1 = Constant::getAllOnesValue(I.getType());
3614 Value *Add = Builder->CreateAdd(RHSI, N1, "tmp");
3615 return BinaryOperator::CreateAnd(Op0, Add);
3620 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3621 // where C1&C2 are powers of two.
3622 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3623 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3624 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3625 // STO == 0 and SFO == 0 handled above.
3626 if ((STO->getValue().isPowerOf2()) &&
3627 (SFO->getValue().isPowerOf2())) {
3628 Value *TrueAnd = Builder->CreateAnd(Op0, SubOne(STO),
3629 SI->getName()+".t");
3630 Value *FalseAnd = Builder->CreateAnd(Op0, SubOne(SFO),
3631 SI->getName()+".f");
3632 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3640 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3641 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3643 // Handle the integer rem common cases
3644 if (Instruction *Common = commonIRemTransforms(I))
3647 if (Value *RHSNeg = dyn_castNegVal(Op1))
3648 if (!isa<Constant>(RHSNeg) ||
3649 (isa<ConstantInt>(RHSNeg) &&
3650 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3652 Worklist.AddValue(I.getOperand(1));
3653 I.setOperand(1, RHSNeg);
3657 // If the sign bits of both operands are zero (i.e. we can prove they are
3658 // unsigned inputs), turn this into a urem.
3659 if (I.getType()->isInteger()) {
3660 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3661 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3662 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3663 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3667 // If it's a constant vector, flip any negative values positive.
3668 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3669 unsigned VWidth = RHSV->getNumOperands();
3671 bool hasNegative = false;
3672 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3673 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3674 if (RHS->getValue().isNegative())
3678 std::vector<Constant *> Elts(VWidth);
3679 for (unsigned i = 0; i != VWidth; ++i) {
3680 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3681 if (RHS->getValue().isNegative())
3682 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3688 Constant *NewRHSV = ConstantVector::get(Elts);
3689 if (NewRHSV != RHSV) {
3690 Worklist.AddValue(I.getOperand(1));
3691 I.setOperand(1, NewRHSV);
3700 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3701 return commonRemTransforms(I);
3704 // isOneBitSet - Return true if there is exactly one bit set in the specified
3706 static bool isOneBitSet(const ConstantInt *CI) {
3707 return CI->getValue().isPowerOf2();
3710 // isHighOnes - Return true if the constant is of the form 1+0+.
3711 // This is the same as lowones(~X).
3712 static bool isHighOnes(const ConstantInt *CI) {
3713 return (~CI->getValue() + 1).isPowerOf2();
3716 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3717 /// are carefully arranged to allow folding of expressions such as:
3719 /// (A < B) | (A > B) --> (A != B)
3721 /// Note that this is only valid if the first and second predicates have the
3722 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3724 /// Three bits are used to represent the condition, as follows:
3729 /// <=> Value Definition
3730 /// 000 0 Always false
3737 /// 111 7 Always true
3739 static unsigned getICmpCode(const ICmpInst *ICI) {
3740 switch (ICI->getPredicate()) {
3742 case ICmpInst::ICMP_UGT: return 1; // 001
3743 case ICmpInst::ICMP_SGT: return 1; // 001
3744 case ICmpInst::ICMP_EQ: return 2; // 010
3745 case ICmpInst::ICMP_UGE: return 3; // 011
3746 case ICmpInst::ICMP_SGE: return 3; // 011
3747 case ICmpInst::ICMP_ULT: return 4; // 100
3748 case ICmpInst::ICMP_SLT: return 4; // 100
3749 case ICmpInst::ICMP_NE: return 5; // 101
3750 case ICmpInst::ICMP_ULE: return 6; // 110
3751 case ICmpInst::ICMP_SLE: return 6; // 110
3754 llvm_unreachable("Invalid ICmp predicate!");
3759 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3760 /// predicate into a three bit mask. It also returns whether it is an ordered
3761 /// predicate by reference.
3762 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3765 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3766 case FCmpInst::FCMP_UNO: return 0; // 000
3767 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3768 case FCmpInst::FCMP_UGT: return 1; // 001
3769 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3770 case FCmpInst::FCMP_UEQ: return 2; // 010
3771 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3772 case FCmpInst::FCMP_UGE: return 3; // 011
3773 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3774 case FCmpInst::FCMP_ULT: return 4; // 100
3775 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3776 case FCmpInst::FCMP_UNE: return 5; // 101
3777 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3778 case FCmpInst::FCMP_ULE: return 6; // 110
3781 // Not expecting FCMP_FALSE and FCMP_TRUE;
3782 llvm_unreachable("Unexpected FCmp predicate!");
3787 /// getICmpValue - This is the complement of getICmpCode, which turns an
3788 /// opcode and two operands into either a constant true or false, or a brand
3789 /// new ICmp instruction. The sign is passed in to determine which kind
3790 /// of predicate to use in the new icmp instruction.
3791 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3792 LLVMContext *Context) {
3794 default: llvm_unreachable("Illegal ICmp code!");
3795 case 0: return ConstantInt::getFalse(*Context);
3798 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3800 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3801 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3804 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3806 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3809 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3811 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3812 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3815 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3817 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3818 case 7: return ConstantInt::getTrue(*Context);
3822 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3823 /// opcode and two operands into either a FCmp instruction. isordered is passed
3824 /// in to determine which kind of predicate to use in the new fcmp instruction.
3825 static Value *getFCmpValue(bool isordered, unsigned code,
3826 Value *LHS, Value *RHS, LLVMContext *Context) {
3828 default: llvm_unreachable("Illegal FCmp code!");
3831 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3833 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3836 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3838 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3841 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3843 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3846 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3848 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3851 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3853 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3856 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3858 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3861 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3863 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3864 case 7: return ConstantInt::getTrue(*Context);
3868 /// PredicatesFoldable - Return true if both predicates match sign or if at
3869 /// least one of them is an equality comparison (which is signless).
3870 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3871 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
3872 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
3873 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
3877 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3878 struct FoldICmpLogical {
3881 ICmpInst::Predicate pred;
3882 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3883 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3884 pred(ICI->getPredicate()) {}
3885 bool shouldApply(Value *V) const {
3886 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3887 if (PredicatesFoldable(pred, ICI->getPredicate()))
3888 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3889 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3892 Instruction *apply(Instruction &Log) const {
3893 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3894 if (ICI->getOperand(0) != LHS) {
3895 assert(ICI->getOperand(1) == LHS);
3896 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3899 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3900 unsigned LHSCode = getICmpCode(ICI);
3901 unsigned RHSCode = getICmpCode(RHSICI);
3903 switch (Log.getOpcode()) {
3904 case Instruction::And: Code = LHSCode & RHSCode; break;
3905 case Instruction::Or: Code = LHSCode | RHSCode; break;
3906 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3907 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3910 bool isSigned = RHSICI->isSigned() || ICI->isSigned();
3911 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3912 if (Instruction *I = dyn_cast<Instruction>(RV))
3914 // Otherwise, it's a constant boolean value...
3915 return IC.ReplaceInstUsesWith(Log, RV);
3918 } // end anonymous namespace
3920 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3921 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3922 // guaranteed to be a binary operator.
3923 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3925 ConstantInt *AndRHS,
3926 BinaryOperator &TheAnd) {
3927 Value *X = Op->getOperand(0);
3928 Constant *Together = 0;
3930 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3932 switch (Op->getOpcode()) {
3933 case Instruction::Xor:
3934 if (Op->hasOneUse()) {
3935 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3936 Value *And = Builder->CreateAnd(X, AndRHS);
3938 return BinaryOperator::CreateXor(And, Together);
3941 case Instruction::Or:
3942 if (Together == AndRHS) // (X | C) & C --> C
3943 return ReplaceInstUsesWith(TheAnd, AndRHS);
3945 if (Op->hasOneUse() && Together != OpRHS) {
3946 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3947 Value *Or = Builder->CreateOr(X, Together);
3949 return BinaryOperator::CreateAnd(Or, AndRHS);
3952 case Instruction::Add:
3953 if (Op->hasOneUse()) {
3954 // Adding a one to a single bit bit-field should be turned into an XOR
3955 // of the bit. First thing to check is to see if this AND is with a
3956 // single bit constant.
3957 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3959 // If there is only one bit set...
3960 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3961 // Ok, at this point, we know that we are masking the result of the
3962 // ADD down to exactly one bit. If the constant we are adding has
3963 // no bits set below this bit, then we can eliminate the ADD.
3964 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3966 // Check to see if any bits below the one bit set in AndRHSV are set.
3967 if ((AddRHS & (AndRHSV-1)) == 0) {
3968 // If not, the only thing that can effect the output of the AND is
3969 // the bit specified by AndRHSV. If that bit is set, the effect of
3970 // the XOR is to toggle the bit. If it is clear, then the ADD has
3972 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3973 TheAnd.setOperand(0, X);
3976 // Pull the XOR out of the AND.
3977 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
3978 NewAnd->takeName(Op);
3979 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3986 case Instruction::Shl: {
3987 // We know that the AND will not produce any of the bits shifted in, so if
3988 // the anded constant includes them, clear them now!
3990 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3991 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3992 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3993 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3995 if (CI->getValue() == ShlMask) {
3996 // Masking out bits that the shift already masks
3997 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3998 } else if (CI != AndRHS) { // Reducing bits set in and.
3999 TheAnd.setOperand(1, CI);
4004 case Instruction::LShr:
4006 // We know that the AND will not produce any of the bits shifted in, so if
4007 // the anded constant includes them, clear them now! This only applies to
4008 // unsigned shifts, because a signed shr may bring in set bits!
4010 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
4011 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
4012 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
4013 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
4015 if (CI->getValue() == ShrMask) {
4016 // Masking out bits that the shift already masks.
4017 return ReplaceInstUsesWith(TheAnd, Op);
4018 } else if (CI != AndRHS) {
4019 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
4024 case Instruction::AShr:
4026 // See if this is shifting in some sign extension, then masking it out
4028 if (Op->hasOneUse()) {
4029 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
4030 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
4031 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
4032 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
4033 if (C == AndRHS) { // Masking out bits shifted in.
4034 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
4035 // Make the argument unsigned.
4036 Value *ShVal = Op->getOperand(0);
4037 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
4038 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
4047 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
4048 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
4049 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
4050 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
4051 /// insert new instructions.
4052 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
4053 bool isSigned, bool Inside,
4055 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
4056 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
4057 "Lo is not <= Hi in range emission code!");
4060 if (Lo == Hi) // Trivially false.
4061 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
4063 // V >= Min && V < Hi --> V < Hi
4064 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
4065 ICmpInst::Predicate pred = (isSigned ?
4066 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
4067 return new ICmpInst(pred, V, Hi);
4070 // Emit V-Lo <u Hi-Lo
4071 Constant *NegLo = ConstantExpr::getNeg(Lo);
4072 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
4073 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
4074 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
4077 if (Lo == Hi) // Trivially true.
4078 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
4080 // V < Min || V >= Hi -> V > Hi-1
4081 Hi = SubOne(cast<ConstantInt>(Hi));
4082 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
4083 ICmpInst::Predicate pred = (isSigned ?
4084 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
4085 return new ICmpInst(pred, V, Hi);
4088 // Emit V-Lo >u Hi-1-Lo
4089 // Note that Hi has already had one subtracted from it, above.
4090 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
4091 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
4092 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
4093 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
4096 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
4097 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
4098 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
4099 // not, since all 1s are not contiguous.
4100 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
4101 const APInt& V = Val->getValue();
4102 uint32_t BitWidth = Val->getType()->getBitWidth();
4103 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
4105 // look for the first zero bit after the run of ones
4106 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
4107 // look for the first non-zero bit
4108 ME = V.getActiveBits();
4112 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
4113 /// where isSub determines whether the operator is a sub. If we can fold one of
4114 /// the following xforms:
4116 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
4117 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
4118 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
4120 /// return (A +/- B).
4122 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
4123 ConstantInt *Mask, bool isSub,
4125 Instruction *LHSI = dyn_cast<Instruction>(LHS);
4126 if (!LHSI || LHSI->getNumOperands() != 2 ||
4127 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
4129 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
4131 switch (LHSI->getOpcode()) {
4133 case Instruction::And:
4134 if (ConstantExpr::getAnd(N, Mask) == Mask) {
4135 // If the AndRHS is a power of two minus one (0+1+), this is simple.
4136 if ((Mask->getValue().countLeadingZeros() +
4137 Mask->getValue().countPopulation()) ==
4138 Mask->getValue().getBitWidth())
4141 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
4142 // part, we don't need any explicit masks to take them out of A. If that
4143 // is all N is, ignore it.
4144 uint32_t MB = 0, ME = 0;
4145 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
4146 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
4147 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
4148 if (MaskedValueIsZero(RHS, Mask))
4153 case Instruction::Or:
4154 case Instruction::Xor:
4155 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
4156 if ((Mask->getValue().countLeadingZeros() +
4157 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
4158 && ConstantExpr::getAnd(N, Mask)->isNullValue())
4164 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
4165 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
4168 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
4169 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
4170 ICmpInst *LHS, ICmpInst *RHS) {
4172 ConstantInt *LHSCst, *RHSCst;
4173 ICmpInst::Predicate LHSCC, RHSCC;
4175 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
4176 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4177 m_ConstantInt(LHSCst))) ||
4178 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4179 m_ConstantInt(RHSCst))))
4182 if (LHSCst == RHSCst && LHSCC == RHSCC) {
4183 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
4184 // where C is a power of 2
4185 if (LHSCC == ICmpInst::ICMP_ULT &&
4186 LHSCst->getValue().isPowerOf2()) {
4187 Value *NewOr = Builder->CreateOr(Val, Val2);
4188 return new ICmpInst(LHSCC, NewOr, LHSCst);
4191 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
4192 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
4193 Value *NewOr = Builder->CreateOr(Val, Val2);
4194 return new ICmpInst(LHSCC, NewOr, LHSCst);
4198 // From here on, we only handle:
4199 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
4200 if (Val != Val2) return 0;
4202 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4203 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4204 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4205 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4206 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4209 // We can't fold (ugt x, C) & (sgt x, C2).
4210 if (!PredicatesFoldable(LHSCC, RHSCC))
4213 // Ensure that the larger constant is on the RHS.
4215 if (CmpInst::isSigned(LHSCC) ||
4216 (ICmpInst::isEquality(LHSCC) &&
4217 CmpInst::isSigned(RHSCC)))
4218 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4220 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4223 std::swap(LHS, RHS);
4224 std::swap(LHSCst, RHSCst);
4225 std::swap(LHSCC, RHSCC);
4228 // At this point, we know we have have two icmp instructions
4229 // comparing a value against two constants and and'ing the result
4230 // together. Because of the above check, we know that we only have
4231 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
4232 // (from the FoldICmpLogical check above), that the two constants
4233 // are not equal and that the larger constant is on the RHS
4234 assert(LHSCst != RHSCst && "Compares not folded above?");
4237 default: llvm_unreachable("Unknown integer condition code!");
4238 case ICmpInst::ICMP_EQ:
4240 default: llvm_unreachable("Unknown integer condition code!");
4241 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
4242 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
4243 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
4244 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4245 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
4246 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
4247 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
4248 return ReplaceInstUsesWith(I, LHS);
4250 case ICmpInst::ICMP_NE:
4252 default: llvm_unreachable("Unknown integer condition code!");
4253 case ICmpInst::ICMP_ULT:
4254 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
4255 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
4256 break; // (X != 13 & X u< 15) -> no change
4257 case ICmpInst::ICMP_SLT:
4258 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
4259 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
4260 break; // (X != 13 & X s< 15) -> no change
4261 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
4262 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
4263 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
4264 return ReplaceInstUsesWith(I, RHS);
4265 case ICmpInst::ICMP_NE:
4266 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
4267 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4268 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4269 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
4270 ConstantInt::get(Add->getType(), 1));
4272 break; // (X != 13 & X != 15) -> no change
4275 case ICmpInst::ICMP_ULT:
4277 default: llvm_unreachable("Unknown integer condition code!");
4278 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
4279 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
4280 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4281 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
4283 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
4284 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
4285 return ReplaceInstUsesWith(I, LHS);
4286 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
4290 case ICmpInst::ICMP_SLT:
4292 default: llvm_unreachable("Unknown integer condition code!");
4293 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
4294 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
4295 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4296 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
4298 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
4299 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
4300 return ReplaceInstUsesWith(I, LHS);
4301 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
4305 case ICmpInst::ICMP_UGT:
4307 default: llvm_unreachable("Unknown integer condition code!");
4308 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
4309 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
4310 return ReplaceInstUsesWith(I, RHS);
4311 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
4313 case ICmpInst::ICMP_NE:
4314 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
4315 return new ICmpInst(LHSCC, Val, RHSCst);
4316 break; // (X u> 13 & X != 15) -> no change
4317 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
4318 return InsertRangeTest(Val, AddOne(LHSCst),
4319 RHSCst, false, true, I);
4320 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4324 case ICmpInst::ICMP_SGT:
4326 default: llvm_unreachable("Unknown integer condition code!");
4327 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
4328 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4329 return ReplaceInstUsesWith(I, RHS);
4330 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4332 case ICmpInst::ICMP_NE:
4333 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4334 return new ICmpInst(LHSCC, Val, RHSCst);
4335 break; // (X s> 13 & X != 15) -> no change
4336 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
4337 return InsertRangeTest(Val, AddOne(LHSCst),
4338 RHSCst, true, true, I);
4339 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4348 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
4351 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4352 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4353 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4354 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4355 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4356 // If either of the constants are nans, then the whole thing returns
4358 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4359 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4360 return new FCmpInst(FCmpInst::FCMP_ORD,
4361 LHS->getOperand(0), RHS->getOperand(0));
4364 // Handle vector zeros. This occurs because the canonical form of
4365 // "fcmp ord x,x" is "fcmp ord x, 0".
4366 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4367 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4368 return new FCmpInst(FCmpInst::FCMP_ORD,
4369 LHS->getOperand(0), RHS->getOperand(0));
4373 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4374 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4375 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4378 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4379 // Swap RHS operands to match LHS.
4380 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4381 std::swap(Op1LHS, Op1RHS);
4384 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4385 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4387 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4389 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
4390 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4391 if (Op0CC == FCmpInst::FCMP_TRUE)
4392 return ReplaceInstUsesWith(I, RHS);
4393 if (Op1CC == FCmpInst::FCMP_TRUE)
4394 return ReplaceInstUsesWith(I, LHS);
4398 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4399 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4401 std::swap(LHS, RHS);
4402 std::swap(Op0Pred, Op1Pred);
4403 std::swap(Op0Ordered, Op1Ordered);
4406 // uno && ueq -> uno && (uno || eq) -> ueq
4407 // ord && olt -> ord && (ord && lt) -> olt
4408 if (Op0Ordered == Op1Ordered)
4409 return ReplaceInstUsesWith(I, RHS);
4411 // uno && oeq -> uno && (ord && eq) -> false
4412 // uno && ord -> false
4414 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4415 // ord && ueq -> ord && (uno || eq) -> oeq
4416 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4417 Op0LHS, Op0RHS, Context));
4425 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4426 bool Changed = SimplifyCommutative(I);
4427 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4429 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
4430 return ReplaceInstUsesWith(I, V);
4432 // See if we can simplify any instructions used by the instruction whose sole
4433 // purpose is to compute bits we don't care about.
4434 if (SimplifyDemandedInstructionBits(I))
4438 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4439 const APInt &AndRHSMask = AndRHS->getValue();
4440 APInt NotAndRHS(~AndRHSMask);
4442 // Optimize a variety of ((val OP C1) & C2) combinations...
4443 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4444 Value *Op0LHS = Op0I->getOperand(0);
4445 Value *Op0RHS = Op0I->getOperand(1);
4446 switch (Op0I->getOpcode()) {
4448 case Instruction::Xor:
4449 case Instruction::Or:
4450 // If the mask is only needed on one incoming arm, push it up.
4451 if (!Op0I->hasOneUse()) break;
4453 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4454 // Not masking anything out for the LHS, move to RHS.
4455 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
4456 Op0RHS->getName()+".masked");
4457 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
4459 if (!isa<Constant>(Op0RHS) &&
4460 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4461 // Not masking anything out for the RHS, move to LHS.
4462 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
4463 Op0LHS->getName()+".masked");
4464 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
4468 case Instruction::Add:
4469 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4470 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4471 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4472 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4473 return BinaryOperator::CreateAnd(V, AndRHS);
4474 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4475 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4478 case Instruction::Sub:
4479 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4480 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4481 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4482 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4483 return BinaryOperator::CreateAnd(V, AndRHS);
4485 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4486 // has 1's for all bits that the subtraction with A might affect.
4487 if (Op0I->hasOneUse()) {
4488 uint32_t BitWidth = AndRHSMask.getBitWidth();
4489 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4490 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4492 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4493 if (!(A && A->isZero()) && // avoid infinite recursion.
4494 MaskedValueIsZero(Op0LHS, Mask)) {
4495 Value *NewNeg = Builder->CreateNeg(Op0RHS);
4496 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4501 case Instruction::Shl:
4502 case Instruction::LShr:
4503 // (1 << x) & 1 --> zext(x == 0)
4504 // (1 >> x) & 1 --> zext(x == 0)
4505 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4507 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
4508 return new ZExtInst(NewICmp, I.getType());
4513 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4514 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4516 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4517 // If this is an integer truncation or change from signed-to-unsigned, and
4518 // if the source is an and/or with immediate, transform it. This
4519 // frequently occurs for bitfield accesses.
4520 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4521 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4522 CastOp->getNumOperands() == 2)
4523 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
4524 if (CastOp->getOpcode() == Instruction::And) {
4525 // Change: and (cast (and X, C1) to T), C2
4526 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4527 // This will fold the two constants together, which may allow
4528 // other simplifications.
4529 Value *NewCast = Builder->CreateTruncOrBitCast(
4530 CastOp->getOperand(0), I.getType(),
4531 CastOp->getName()+".shrunk");
4532 // trunc_or_bitcast(C1)&C2
4533 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4534 C3 = ConstantExpr::getAnd(C3, AndRHS);
4535 return BinaryOperator::CreateAnd(NewCast, C3);
4536 } else if (CastOp->getOpcode() == Instruction::Or) {
4537 // Change: and (cast (or X, C1) to T), C2
4538 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4539 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4540 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4542 return ReplaceInstUsesWith(I, AndRHS);
4548 // Try to fold constant and into select arguments.
4549 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4550 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4552 if (isa<PHINode>(Op0))
4553 if (Instruction *NV = FoldOpIntoPhi(I))
4558 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4559 if (Value *Op0NotVal = dyn_castNotVal(Op0))
4560 if (Value *Op1NotVal = dyn_castNotVal(Op1))
4561 if (Op0->hasOneUse() && Op1->hasOneUse()) {
4562 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
4563 I.getName()+".demorgan");
4564 return BinaryOperator::CreateNot(Or);
4568 Value *A = 0, *B = 0, *C = 0, *D = 0;
4569 // (A|B) & ~(A&B) -> A^B
4570 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
4571 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
4572 ((A == C && B == D) || (A == D && B == C)))
4573 return BinaryOperator::CreateXor(A, B);
4575 // ~(A&B) & (A|B) -> A^B
4576 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
4577 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
4578 ((A == C && B == D) || (A == D && B == C)))
4579 return BinaryOperator::CreateXor(A, B);
4581 if (Op0->hasOneUse() &&
4582 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4583 if (A == Op1) { // (A^B)&A -> A&(A^B)
4584 I.swapOperands(); // Simplify below
4585 std::swap(Op0, Op1);
4586 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4587 cast<BinaryOperator>(Op0)->swapOperands();
4588 I.swapOperands(); // Simplify below
4589 std::swap(Op0, Op1);
4593 if (Op1->hasOneUse() &&
4594 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4595 if (B == Op0) { // B&(A^B) -> B&(B^A)
4596 cast<BinaryOperator>(Op1)->swapOperands();
4599 if (A == Op0) // A&(A^B) -> A & ~B
4600 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
4603 // (A&((~A)|B)) -> A&B
4604 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4605 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4606 return BinaryOperator::CreateAnd(A, Op1);
4607 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4608 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4609 return BinaryOperator::CreateAnd(A, Op0);
4612 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4613 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4614 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4617 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4618 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4622 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4623 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4624 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4625 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4626 const Type *SrcTy = Op0C->getOperand(0)->getType();
4627 if (SrcTy == Op1C->getOperand(0)->getType() &&
4628 SrcTy->isIntOrIntVector() &&
4629 // Only do this if the casts both really cause code to be generated.
4630 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4632 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4634 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
4635 Op1C->getOperand(0), I.getName());
4636 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4640 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4641 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4642 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4643 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4644 SI0->getOperand(1) == SI1->getOperand(1) &&
4645 (SI0->hasOneUse() || SI1->hasOneUse())) {
4647 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
4649 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4650 SI1->getOperand(1));
4654 // If and'ing two fcmp, try combine them into one.
4655 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4656 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4657 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4661 return Changed ? &I : 0;
4664 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4665 /// capable of providing pieces of a bswap. The subexpression provides pieces
4666 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4667 /// the expression came from the corresponding "byte swapped" byte in some other
4668 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4669 /// we know that the expression deposits the low byte of %X into the high byte
4670 /// of the bswap result and that all other bytes are zero. This expression is
4671 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4674 /// This function returns true if the match was unsuccessful and false if so.
4675 /// On entry to the function the "OverallLeftShift" is a signed integer value
4676 /// indicating the number of bytes that the subexpression is later shifted. For
4677 /// example, if the expression is later right shifted by 16 bits, the
4678 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4679 /// byte of ByteValues is actually being set.
4681 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4682 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4683 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4684 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4685 /// always in the local (OverallLeftShift) coordinate space.
4687 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4688 SmallVector<Value*, 8> &ByteValues) {
4689 if (Instruction *I = dyn_cast<Instruction>(V)) {
4690 // If this is an or instruction, it may be an inner node of the bswap.
4691 if (I->getOpcode() == Instruction::Or) {
4692 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4694 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4698 // If this is a logical shift by a constant multiple of 8, recurse with
4699 // OverallLeftShift and ByteMask adjusted.
4700 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4702 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4703 // Ensure the shift amount is defined and of a byte value.
4704 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4707 unsigned ByteShift = ShAmt >> 3;
4708 if (I->getOpcode() == Instruction::Shl) {
4709 // X << 2 -> collect(X, +2)
4710 OverallLeftShift += ByteShift;
4711 ByteMask >>= ByteShift;
4713 // X >>u 2 -> collect(X, -2)
4714 OverallLeftShift -= ByteShift;
4715 ByteMask <<= ByteShift;
4716 ByteMask &= (~0U >> (32-ByteValues.size()));
4719 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4720 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4722 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4726 // If this is a logical 'and' with a mask that clears bytes, clear the
4727 // corresponding bytes in ByteMask.
4728 if (I->getOpcode() == Instruction::And &&
4729 isa<ConstantInt>(I->getOperand(1))) {
4730 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4731 unsigned NumBytes = ByteValues.size();
4732 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4733 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4735 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4736 // If this byte is masked out by a later operation, we don't care what
4738 if ((ByteMask & (1 << i)) == 0)
4741 // If the AndMask is all zeros for this byte, clear the bit.
4742 APInt MaskB = AndMask & Byte;
4744 ByteMask &= ~(1U << i);
4748 // If the AndMask is not all ones for this byte, it's not a bytezap.
4752 // Otherwise, this byte is kept.
4755 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4760 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4761 // the input value to the bswap. Some observations: 1) if more than one byte
4762 // is demanded from this input, then it could not be successfully assembled
4763 // into a byteswap. At least one of the two bytes would not be aligned with
4764 // their ultimate destination.
4765 if (!isPowerOf2_32(ByteMask)) return true;
4766 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4768 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4769 // is demanded, it needs to go into byte 0 of the result. This means that the
4770 // byte needs to be shifted until it lands in the right byte bucket. The
4771 // shift amount depends on the position: if the byte is coming from the high
4772 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4773 // low part, it must be shifted left.
4774 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4775 if (InputByteNo < ByteValues.size()/2) {
4776 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4779 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4783 // If the destination byte value is already defined, the values are or'd
4784 // together, which isn't a bswap (unless it's an or of the same bits).
4785 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4787 ByteValues[DestByteNo] = V;
4791 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4792 /// If so, insert the new bswap intrinsic and return it.
4793 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4794 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4795 if (!ITy || ITy->getBitWidth() % 16 ||
4796 // ByteMask only allows up to 32-byte values.
4797 ITy->getBitWidth() > 32*8)
4798 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4800 /// ByteValues - For each byte of the result, we keep track of which value
4801 /// defines each byte.
4802 SmallVector<Value*, 8> ByteValues;
4803 ByteValues.resize(ITy->getBitWidth()/8);
4805 // Try to find all the pieces corresponding to the bswap.
4806 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4807 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4810 // Check to see if all of the bytes come from the same value.
4811 Value *V = ByteValues[0];
4812 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4814 // Check to make sure that all of the bytes come from the same value.
4815 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4816 if (ByteValues[i] != V)
4818 const Type *Tys[] = { ITy };
4819 Module *M = I.getParent()->getParent()->getParent();
4820 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4821 return CallInst::Create(F, V);
4824 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4825 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4826 /// we can simplify this expression to "cond ? C : D or B".
4827 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4829 LLVMContext *Context) {
4830 // If A is not a select of -1/0, this cannot match.
4832 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4835 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4836 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4837 return SelectInst::Create(Cond, C, B);
4838 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4839 return SelectInst::Create(Cond, C, B);
4840 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4841 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4842 return SelectInst::Create(Cond, C, D);
4843 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4844 return SelectInst::Create(Cond, C, D);
4848 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4849 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4850 ICmpInst *LHS, ICmpInst *RHS) {
4852 ConstantInt *LHSCst, *RHSCst;
4853 ICmpInst::Predicate LHSCC, RHSCC;
4855 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4856 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4857 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4861 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
4862 if (LHSCst == RHSCst && LHSCC == RHSCC &&
4863 LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
4864 Value *NewOr = Builder->CreateOr(Val, Val2);
4865 return new ICmpInst(LHSCC, NewOr, LHSCst);
4868 // From here on, we only handle:
4869 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4870 if (Val != Val2) return 0;
4872 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4873 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4874 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4875 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4876 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4879 // We can't fold (ugt x, C) | (sgt x, C2).
4880 if (!PredicatesFoldable(LHSCC, RHSCC))
4883 // Ensure that the larger constant is on the RHS.
4885 if (CmpInst::isSigned(LHSCC) ||
4886 (ICmpInst::isEquality(LHSCC) &&
4887 CmpInst::isSigned(RHSCC)))
4888 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4890 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4893 std::swap(LHS, RHS);
4894 std::swap(LHSCst, RHSCst);
4895 std::swap(LHSCC, RHSCC);
4898 // At this point, we know we have have two icmp instructions
4899 // comparing a value against two constants and or'ing the result
4900 // together. Because of the above check, we know that we only have
4901 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4902 // FoldICmpLogical check above), that the two constants are not
4904 assert(LHSCst != RHSCst && "Compares not folded above?");
4907 default: llvm_unreachable("Unknown integer condition code!");
4908 case ICmpInst::ICMP_EQ:
4910 default: llvm_unreachable("Unknown integer condition code!");
4911 case ICmpInst::ICMP_EQ:
4912 if (LHSCst == SubOne(RHSCst)) {
4913 // (X == 13 | X == 14) -> X-13 <u 2
4914 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4915 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4916 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4917 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4919 break; // (X == 13 | X == 15) -> no change
4920 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4921 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4923 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4924 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4925 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4926 return ReplaceInstUsesWith(I, RHS);
4929 case ICmpInst::ICMP_NE:
4931 default: llvm_unreachable("Unknown integer condition code!");
4932 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4933 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4934 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4935 return ReplaceInstUsesWith(I, LHS);
4936 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4937 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4938 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4939 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4942 case ICmpInst::ICMP_ULT:
4944 default: llvm_unreachable("Unknown integer condition code!");
4945 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4947 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4948 // If RHSCst is [us]MAXINT, it is always false. Not handling
4949 // this can cause overflow.
4950 if (RHSCst->isMaxValue(false))
4951 return ReplaceInstUsesWith(I, LHS);
4952 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4954 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4956 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4957 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4958 return ReplaceInstUsesWith(I, RHS);
4959 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4963 case ICmpInst::ICMP_SLT:
4965 default: llvm_unreachable("Unknown integer condition code!");
4966 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4968 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4969 // If RHSCst is [us]MAXINT, it is always false. Not handling
4970 // this can cause overflow.
4971 if (RHSCst->isMaxValue(true))
4972 return ReplaceInstUsesWith(I, LHS);
4973 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4975 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4977 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4978 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4979 return ReplaceInstUsesWith(I, RHS);
4980 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4984 case ICmpInst::ICMP_UGT:
4986 default: llvm_unreachable("Unknown integer condition code!");
4987 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4988 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4989 return ReplaceInstUsesWith(I, LHS);
4990 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4992 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4993 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4994 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4995 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4999 case ICmpInst::ICMP_SGT:
5001 default: llvm_unreachable("Unknown integer condition code!");
5002 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
5003 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
5004 return ReplaceInstUsesWith(I, LHS);
5005 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
5007 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
5008 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
5009 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5010 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
5018 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
5020 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
5021 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
5022 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
5023 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
5024 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
5025 // If either of the constants are nans, then the whole thing returns
5027 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
5028 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5030 // Otherwise, no need to compare the two constants, compare the
5032 return new FCmpInst(FCmpInst::FCMP_UNO,
5033 LHS->getOperand(0), RHS->getOperand(0));
5036 // Handle vector zeros. This occurs because the canonical form of
5037 // "fcmp uno x,x" is "fcmp uno x, 0".
5038 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
5039 isa<ConstantAggregateZero>(RHS->getOperand(1)))
5040 return new FCmpInst(FCmpInst::FCMP_UNO,
5041 LHS->getOperand(0), RHS->getOperand(0));
5046 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
5047 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
5048 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
5050 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
5051 // Swap RHS operands to match LHS.
5052 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
5053 std::swap(Op1LHS, Op1RHS);
5055 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
5056 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
5058 return new FCmpInst((FCmpInst::Predicate)Op0CC,
5060 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
5061 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5062 if (Op0CC == FCmpInst::FCMP_FALSE)
5063 return ReplaceInstUsesWith(I, RHS);
5064 if (Op1CC == FCmpInst::FCMP_FALSE)
5065 return ReplaceInstUsesWith(I, LHS);
5068 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
5069 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
5070 if (Op0Ordered == Op1Ordered) {
5071 // If both are ordered or unordered, return a new fcmp with
5072 // or'ed predicates.
5073 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
5074 Op0LHS, Op0RHS, Context);
5075 if (Instruction *I = dyn_cast<Instruction>(RV))
5077 // Otherwise, it's a constant boolean value...
5078 return ReplaceInstUsesWith(I, RV);
5084 /// FoldOrWithConstants - This helper function folds:
5086 /// ((A | B) & C1) | (B & C2)
5092 /// when the XOR of the two constants is "all ones" (-1).
5093 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
5094 Value *A, Value *B, Value *C) {
5095 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
5099 ConstantInt *CI2 = 0;
5100 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
5102 APInt Xor = CI1->getValue() ^ CI2->getValue();
5103 if (!Xor.isAllOnesValue()) return 0;
5105 if (V1 == A || V1 == B) {
5106 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
5107 return BinaryOperator::CreateOr(NewOp, V1);
5113 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
5114 bool Changed = SimplifyCommutative(I);
5115 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5117 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
5118 return ReplaceInstUsesWith(I, V);
5121 // See if we can simplify any instructions used by the instruction whose sole
5122 // purpose is to compute bits we don't care about.
5123 if (SimplifyDemandedInstructionBits(I))
5126 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5127 ConstantInt *C1 = 0; Value *X = 0;
5128 // (X & C1) | C2 --> (X | C2) & (C1|C2)
5129 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
5131 Value *Or = Builder->CreateOr(X, RHS);
5133 return BinaryOperator::CreateAnd(Or,
5134 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
5137 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
5138 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
5140 Value *Or = Builder->CreateOr(X, RHS);
5142 return BinaryOperator::CreateXor(Or,
5143 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
5146 // Try to fold constant and into select arguments.
5147 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5148 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5150 if (isa<PHINode>(Op0))
5151 if (Instruction *NV = FoldOpIntoPhi(I))
5155 Value *A = 0, *B = 0;
5156 ConstantInt *C1 = 0, *C2 = 0;
5158 // (A | B) | C and A | (B | C) -> bswap if possible.
5159 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
5160 if (match(Op0, m_Or(m_Value(), m_Value())) ||
5161 match(Op1, m_Or(m_Value(), m_Value())) ||
5162 (match(Op0, m_Shift(m_Value(), m_Value())) &&
5163 match(Op1, m_Shift(m_Value(), m_Value())))) {
5164 if (Instruction *BSwap = MatchBSwap(I))
5168 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
5169 if (Op0->hasOneUse() &&
5170 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
5171 MaskedValueIsZero(Op1, C1->getValue())) {
5172 Value *NOr = Builder->CreateOr(A, Op1);
5174 return BinaryOperator::CreateXor(NOr, C1);
5177 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
5178 if (Op1->hasOneUse() &&
5179 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
5180 MaskedValueIsZero(Op0, C1->getValue())) {
5181 Value *NOr = Builder->CreateOr(A, Op0);
5183 return BinaryOperator::CreateXor(NOr, C1);
5187 Value *C = 0, *D = 0;
5188 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
5189 match(Op1, m_And(m_Value(B), m_Value(D)))) {
5190 Value *V1 = 0, *V2 = 0, *V3 = 0;
5191 C1 = dyn_cast<ConstantInt>(C);
5192 C2 = dyn_cast<ConstantInt>(D);
5193 if (C1 && C2) { // (A & C1)|(B & C2)
5194 // If we have: ((V + N) & C1) | (V & C2)
5195 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
5196 // replace with V+N.
5197 if (C1->getValue() == ~C2->getValue()) {
5198 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
5199 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
5200 // Add commutes, try both ways.
5201 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
5202 return ReplaceInstUsesWith(I, A);
5203 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
5204 return ReplaceInstUsesWith(I, A);
5206 // Or commutes, try both ways.
5207 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
5208 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
5209 // Add commutes, try both ways.
5210 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
5211 return ReplaceInstUsesWith(I, B);
5212 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
5213 return ReplaceInstUsesWith(I, B);
5216 V1 = 0; V2 = 0; V3 = 0;
5219 // Check to see if we have any common things being and'ed. If so, find the
5220 // terms for V1 & (V2|V3).
5221 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
5222 if (A == B) // (A & C)|(A & D) == A & (C|D)
5223 V1 = A, V2 = C, V3 = D;
5224 else if (A == D) // (A & C)|(B & A) == A & (B|C)
5225 V1 = A, V2 = B, V3 = C;
5226 else if (C == B) // (A & C)|(C & D) == C & (A|D)
5227 V1 = C, V2 = A, V3 = D;
5228 else if (C == D) // (A & C)|(B & C) == C & (A|B)
5229 V1 = C, V2 = A, V3 = B;
5232 Value *Or = Builder->CreateOr(V2, V3, "tmp");
5233 return BinaryOperator::CreateAnd(V1, Or);
5237 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
5238 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
5240 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
5242 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
5244 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
5247 // ((A&~B)|(~A&B)) -> A^B
5248 if ((match(C, m_Not(m_Specific(D))) &&
5249 match(B, m_Not(m_Specific(A)))))
5250 return BinaryOperator::CreateXor(A, D);
5251 // ((~B&A)|(~A&B)) -> A^B
5252 if ((match(A, m_Not(m_Specific(D))) &&
5253 match(B, m_Not(m_Specific(C)))))
5254 return BinaryOperator::CreateXor(C, D);
5255 // ((A&~B)|(B&~A)) -> A^B
5256 if ((match(C, m_Not(m_Specific(B))) &&
5257 match(D, m_Not(m_Specific(A)))))
5258 return BinaryOperator::CreateXor(A, B);
5259 // ((~B&A)|(B&~A)) -> A^B
5260 if ((match(A, m_Not(m_Specific(B))) &&
5261 match(D, m_Not(m_Specific(C)))))
5262 return BinaryOperator::CreateXor(C, B);
5265 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
5266 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
5267 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
5268 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
5269 SI0->getOperand(1) == SI1->getOperand(1) &&
5270 (SI0->hasOneUse() || SI1->hasOneUse())) {
5271 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
5273 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
5274 SI1->getOperand(1));
5278 // ((A|B)&1)|(B&-2) -> (A&1) | B
5279 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
5280 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
5281 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
5282 if (Ret) return Ret;
5284 // (B&-2)|((A|B)&1) -> (A&1) | B
5285 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
5286 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
5287 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
5288 if (Ret) return Ret;
5291 // (~A | ~B) == (~(A & B)) - De Morgan's Law
5292 if (Value *Op0NotVal = dyn_castNotVal(Op0))
5293 if (Value *Op1NotVal = dyn_castNotVal(Op1))
5294 if (Op0->hasOneUse() && Op1->hasOneUse()) {
5295 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
5296 I.getName()+".demorgan");
5297 return BinaryOperator::CreateNot(And);
5300 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
5301 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
5302 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5305 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
5306 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
5310 // fold (or (cast A), (cast B)) -> (cast (or A, B))
5311 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5312 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5313 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
5314 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
5315 !isa<ICmpInst>(Op1C->getOperand(0))) {
5316 const Type *SrcTy = Op0C->getOperand(0)->getType();
5317 if (SrcTy == Op1C->getOperand(0)->getType() &&
5318 SrcTy->isIntOrIntVector() &&
5319 // Only do this if the casts both really cause code to be
5321 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5323 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5325 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
5326 Op1C->getOperand(0), I.getName());
5327 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5334 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
5335 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
5336 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
5337 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5341 return Changed ? &I : 0;
5346 // XorSelf - Implements: X ^ X --> 0
5349 XorSelf(Value *rhs) : RHS(rhs) {}
5350 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5351 Instruction *apply(BinaryOperator &Xor) const {
5358 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5359 bool Changed = SimplifyCommutative(I);
5360 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5362 if (isa<UndefValue>(Op1)) {
5363 if (isa<UndefValue>(Op0))
5364 // Handle undef ^ undef -> 0 special case. This is a common
5366 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5367 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5370 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5371 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5372 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5373 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5376 // See if we can simplify any instructions used by the instruction whose sole
5377 // purpose is to compute bits we don't care about.
5378 if (SimplifyDemandedInstructionBits(I))
5380 if (isa<VectorType>(I.getType()))
5381 if (isa<ConstantAggregateZero>(Op1))
5382 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5384 // Is this a ~ operation?
5385 if (Value *NotOp = dyn_castNotVal(&I)) {
5386 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5387 if (Op0I->getOpcode() == Instruction::And ||
5388 Op0I->getOpcode() == Instruction::Or) {
5389 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5390 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5391 if (dyn_castNotVal(Op0I->getOperand(1)))
5392 Op0I->swapOperands();
5393 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5395 Builder->CreateNot(Op0I->getOperand(1),
5396 Op0I->getOperand(1)->getName()+".not");
5397 if (Op0I->getOpcode() == Instruction::And)
5398 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5399 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5402 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
5403 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
5404 if (isFreeToInvert(Op0I->getOperand(0)) &&
5405 isFreeToInvert(Op0I->getOperand(1))) {
5407 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
5409 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
5410 if (Op0I->getOpcode() == Instruction::And)
5411 return BinaryOperator::CreateOr(NotX, NotY);
5412 return BinaryOperator::CreateAnd(NotX, NotY);
5419 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5420 if (RHS->isOne() && Op0->hasOneUse()) {
5421 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5422 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5423 return new ICmpInst(ICI->getInversePredicate(),
5424 ICI->getOperand(0), ICI->getOperand(1));
5426 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5427 return new FCmpInst(FCI->getInversePredicate(),
5428 FCI->getOperand(0), FCI->getOperand(1));
5431 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5432 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5433 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5434 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5435 Instruction::CastOps Opcode = Op0C->getOpcode();
5436 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
5437 (RHS == ConstantExpr::getCast(Opcode,
5438 ConstantInt::getTrue(*Context),
5439 Op0C->getDestTy()))) {
5440 CI->setPredicate(CI->getInversePredicate());
5441 return CastInst::Create(Opcode, CI, Op0C->getType());
5447 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5448 // ~(c-X) == X-c-1 == X+(-c-1)
5449 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5450 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5451 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5452 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5453 ConstantInt::get(I.getType(), 1));
5454 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5457 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5458 if (Op0I->getOpcode() == Instruction::Add) {
5459 // ~(X-c) --> (-c-1)-X
5460 if (RHS->isAllOnesValue()) {
5461 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5462 return BinaryOperator::CreateSub(
5463 ConstantExpr::getSub(NegOp0CI,
5464 ConstantInt::get(I.getType(), 1)),
5465 Op0I->getOperand(0));
5466 } else if (RHS->getValue().isSignBit()) {
5467 // (X + C) ^ signbit -> (X + C + signbit)
5468 Constant *C = ConstantInt::get(*Context,
5469 RHS->getValue() + Op0CI->getValue());
5470 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5473 } else if (Op0I->getOpcode() == Instruction::Or) {
5474 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5475 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5476 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5477 // Anything in both C1 and C2 is known to be zero, remove it from
5479 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5480 NewRHS = ConstantExpr::getAnd(NewRHS,
5481 ConstantExpr::getNot(CommonBits));
5483 I.setOperand(0, Op0I->getOperand(0));
5484 I.setOperand(1, NewRHS);
5491 // Try to fold constant and into select arguments.
5492 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5493 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5495 if (isa<PHINode>(Op0))
5496 if (Instruction *NV = FoldOpIntoPhi(I))
5500 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5502 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5504 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5506 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5509 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5512 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5513 if (A == Op0) { // B^(B|A) == (A|B)^B
5514 Op1I->swapOperands();
5516 std::swap(Op0, Op1);
5517 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5518 I.swapOperands(); // Simplified below.
5519 std::swap(Op0, Op1);
5521 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5522 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5523 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5524 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5525 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5527 if (A == Op0) { // A^(A&B) -> A^(B&A)
5528 Op1I->swapOperands();
5531 if (B == Op0) { // A^(B&A) -> (B&A)^A
5532 I.swapOperands(); // Simplified below.
5533 std::swap(Op0, Op1);
5538 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5541 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5542 Op0I->hasOneUse()) {
5543 if (A == Op1) // (B|A)^B == (A|B)^B
5545 if (B == Op1) // (A|B)^B == A & ~B
5546 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
5547 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5548 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5549 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5550 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5551 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5553 if (A == Op1) // (A&B)^A -> (B&A)^A
5555 if (B == Op1 && // (B&A)^A == ~B & A
5556 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5557 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
5562 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5563 if (Op0I && Op1I && Op0I->isShift() &&
5564 Op0I->getOpcode() == Op1I->getOpcode() &&
5565 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5566 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5568 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
5570 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5571 Op1I->getOperand(1));
5575 Value *A, *B, *C, *D;
5576 // (A & B)^(A | B) -> A ^ B
5577 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5578 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5579 if ((A == C && B == D) || (A == D && B == C))
5580 return BinaryOperator::CreateXor(A, B);
5582 // (A | B)^(A & B) -> A ^ B
5583 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5584 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5585 if ((A == C && B == D) || (A == D && B == C))
5586 return BinaryOperator::CreateXor(A, B);
5590 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5591 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5592 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5593 // (X & Y)^(X & Y) -> (Y^Z) & X
5594 Value *X = 0, *Y = 0, *Z = 0;
5596 X = A, Y = B, Z = D;
5598 X = A, Y = B, Z = C;
5600 X = B, Y = A, Z = D;
5602 X = B, Y = A, Z = C;
5605 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
5606 return BinaryOperator::CreateAnd(NewOp, X);
5611 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5612 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5613 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5616 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5617 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5618 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5619 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5620 const Type *SrcTy = Op0C->getOperand(0)->getType();
5621 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5622 // Only do this if the casts both really cause code to be generated.
5623 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5625 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5627 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
5628 Op1C->getOperand(0), I.getName());
5629 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5634 return Changed ? &I : 0;
5637 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5638 LLVMContext *Context) {
5639 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5642 static bool HasAddOverflow(ConstantInt *Result,
5643 ConstantInt *In1, ConstantInt *In2,
5646 if (In2->getValue().isNegative())
5647 return Result->getValue().sgt(In1->getValue());
5649 return Result->getValue().slt(In1->getValue());
5651 return Result->getValue().ult(In1->getValue());
5654 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5655 /// overflowed for this type.
5656 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5657 Constant *In2, LLVMContext *Context,
5658 bool IsSigned = false) {
5659 Result = ConstantExpr::getAdd(In1, In2);
5661 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5662 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5663 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5664 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5665 ExtractElement(In1, Idx, Context),
5666 ExtractElement(In2, Idx, Context),
5673 return HasAddOverflow(cast<ConstantInt>(Result),
5674 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5678 static bool HasSubOverflow(ConstantInt *Result,
5679 ConstantInt *In1, ConstantInt *In2,
5682 if (In2->getValue().isNegative())
5683 return Result->getValue().slt(In1->getValue());
5685 return Result->getValue().sgt(In1->getValue());
5687 return Result->getValue().ugt(In1->getValue());
5690 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5691 /// overflowed for this type.
5692 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5693 Constant *In2, LLVMContext *Context,
5694 bool IsSigned = false) {
5695 Result = ConstantExpr::getSub(In1, In2);
5697 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5698 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5699 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5700 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5701 ExtractElement(In1, Idx, Context),
5702 ExtractElement(In2, Idx, Context),
5709 return HasSubOverflow(cast<ConstantInt>(Result),
5710 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5715 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5716 /// else. At this point we know that the GEP is on the LHS of the comparison.
5717 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5718 ICmpInst::Predicate Cond,
5720 // Look through bitcasts.
5721 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5722 RHS = BCI->getOperand(0);
5724 Value *PtrBase = GEPLHS->getOperand(0);
5725 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5726 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5727 // This transformation (ignoring the base and scales) is valid because we
5728 // know pointers can't overflow since the gep is inbounds. See if we can
5729 // output an optimized form.
5730 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5732 // If not, synthesize the offset the hard way.
5734 Offset = EmitGEPOffset(GEPLHS, *this);
5735 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5736 Constant::getNullValue(Offset->getType()));
5737 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5738 // If the base pointers are different, but the indices are the same, just
5739 // compare the base pointer.
5740 if (PtrBase != GEPRHS->getOperand(0)) {
5741 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5742 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5743 GEPRHS->getOperand(0)->getType();
5745 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5746 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5747 IndicesTheSame = false;
5751 // If all indices are the same, just compare the base pointers.
5753 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5754 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5756 // Otherwise, the base pointers are different and the indices are
5757 // different, bail out.
5761 // If one of the GEPs has all zero indices, recurse.
5762 bool AllZeros = true;
5763 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5764 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5765 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5770 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5771 ICmpInst::getSwappedPredicate(Cond), I);
5773 // If the other GEP has all zero indices, recurse.
5775 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5776 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5777 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5782 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5784 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5785 // If the GEPs only differ by one index, compare it.
5786 unsigned NumDifferences = 0; // Keep track of # differences.
5787 unsigned DiffOperand = 0; // The operand that differs.
5788 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5789 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5790 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5791 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5792 // Irreconcilable differences.
5796 if (NumDifferences++) break;
5801 if (NumDifferences == 0) // SAME GEP?
5802 return ReplaceInstUsesWith(I, // No comparison is needed here.
5803 ConstantInt::get(Type::getInt1Ty(*Context),
5804 ICmpInst::isTrueWhenEqual(Cond)));
5806 else if (NumDifferences == 1) {
5807 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5808 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5809 // Make sure we do a signed comparison here.
5810 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5814 // Only lower this if the icmp is the only user of the GEP or if we expect
5815 // the result to fold to a constant!
5817 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5818 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5819 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5820 Value *L = EmitGEPOffset(GEPLHS, *this);
5821 Value *R = EmitGEPOffset(GEPRHS, *this);
5822 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5828 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5830 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5833 if (!isa<ConstantFP>(RHSC)) return 0;
5834 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5836 // Get the width of the mantissa. We don't want to hack on conversions that
5837 // might lose information from the integer, e.g. "i64 -> float"
5838 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5839 if (MantissaWidth == -1) return 0; // Unknown.
5841 // Check to see that the input is converted from an integer type that is small
5842 // enough that preserves all bits. TODO: check here for "known" sign bits.
5843 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5844 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5846 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5847 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5851 // If the conversion would lose info, don't hack on this.
5852 if ((int)InputSize > MantissaWidth)
5855 // Otherwise, we can potentially simplify the comparison. We know that it
5856 // will always come through as an integer value and we know the constant is
5857 // not a NAN (it would have been previously simplified).
5858 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5860 ICmpInst::Predicate Pred;
5861 switch (I.getPredicate()) {
5862 default: llvm_unreachable("Unexpected predicate!");
5863 case FCmpInst::FCMP_UEQ:
5864 case FCmpInst::FCMP_OEQ:
5865 Pred = ICmpInst::ICMP_EQ;
5867 case FCmpInst::FCMP_UGT:
5868 case FCmpInst::FCMP_OGT:
5869 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5871 case FCmpInst::FCMP_UGE:
5872 case FCmpInst::FCMP_OGE:
5873 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5875 case FCmpInst::FCMP_ULT:
5876 case FCmpInst::FCMP_OLT:
5877 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5879 case FCmpInst::FCMP_ULE:
5880 case FCmpInst::FCMP_OLE:
5881 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5883 case FCmpInst::FCMP_UNE:
5884 case FCmpInst::FCMP_ONE:
5885 Pred = ICmpInst::ICMP_NE;
5887 case FCmpInst::FCMP_ORD:
5888 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5889 case FCmpInst::FCMP_UNO:
5890 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5893 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5895 // Now we know that the APFloat is a normal number, zero or inf.
5897 // See if the FP constant is too large for the integer. For example,
5898 // comparing an i8 to 300.0.
5899 unsigned IntWidth = IntTy->getScalarSizeInBits();
5902 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5903 // and large values.
5904 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5905 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5906 APFloat::rmNearestTiesToEven);
5907 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5908 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5909 Pred == ICmpInst::ICMP_SLE)
5910 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5911 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5914 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5915 // +INF and large values.
5916 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5917 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5918 APFloat::rmNearestTiesToEven);
5919 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5920 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5921 Pred == ICmpInst::ICMP_ULE)
5922 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5923 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5928 // See if the RHS value is < SignedMin.
5929 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5930 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5931 APFloat::rmNearestTiesToEven);
5932 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5933 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5934 Pred == ICmpInst::ICMP_SGE)
5935 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5936 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5940 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5941 // [0, UMAX], but it may still be fractional. See if it is fractional by
5942 // casting the FP value to the integer value and back, checking for equality.
5943 // Don't do this for zero, because -0.0 is not fractional.
5944 Constant *RHSInt = LHSUnsigned
5945 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5946 : ConstantExpr::getFPToSI(RHSC, IntTy);
5947 if (!RHS.isZero()) {
5948 bool Equal = LHSUnsigned
5949 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5950 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5952 // If we had a comparison against a fractional value, we have to adjust
5953 // the compare predicate and sometimes the value. RHSC is rounded towards
5954 // zero at this point.
5956 default: llvm_unreachable("Unexpected integer comparison!");
5957 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5958 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5959 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5960 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5961 case ICmpInst::ICMP_ULE:
5962 // (float)int <= 4.4 --> int <= 4
5963 // (float)int <= -4.4 --> false
5964 if (RHS.isNegative())
5965 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5967 case ICmpInst::ICMP_SLE:
5968 // (float)int <= 4.4 --> int <= 4
5969 // (float)int <= -4.4 --> int < -4
5970 if (RHS.isNegative())
5971 Pred = ICmpInst::ICMP_SLT;
5973 case ICmpInst::ICMP_ULT:
5974 // (float)int < -4.4 --> false
5975 // (float)int < 4.4 --> int <= 4
5976 if (RHS.isNegative())
5977 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5978 Pred = ICmpInst::ICMP_ULE;
5980 case ICmpInst::ICMP_SLT:
5981 // (float)int < -4.4 --> int < -4
5982 // (float)int < 4.4 --> int <= 4
5983 if (!RHS.isNegative())
5984 Pred = ICmpInst::ICMP_SLE;
5986 case ICmpInst::ICMP_UGT:
5987 // (float)int > 4.4 --> int > 4
5988 // (float)int > -4.4 --> true
5989 if (RHS.isNegative())
5990 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5992 case ICmpInst::ICMP_SGT:
5993 // (float)int > 4.4 --> int > 4
5994 // (float)int > -4.4 --> int >= -4
5995 if (RHS.isNegative())
5996 Pred = ICmpInst::ICMP_SGE;
5998 case ICmpInst::ICMP_UGE:
5999 // (float)int >= -4.4 --> true
6000 // (float)int >= 4.4 --> int > 4
6001 if (!RHS.isNegative())
6002 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6003 Pred = ICmpInst::ICMP_UGT;
6005 case ICmpInst::ICMP_SGE:
6006 // (float)int >= -4.4 --> int >= -4
6007 // (float)int >= 4.4 --> int > 4
6008 if (!RHS.isNegative())
6009 Pred = ICmpInst::ICMP_SGT;
6015 // Lower this FP comparison into an appropriate integer version of the
6017 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
6021 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
6022 /// cmp pred (load (gep GV, ...)), cmpcst
6023 /// where GV is a global variable with a constant initializer. Try to simplify
6024 /// this into one or two simpler comparisons that do not need the load. For
6025 /// example, we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into
6026 /// "icmp eq i, 3". We assume that eliminating a load is always goodness.
6027 Instruction *InstCombiner::
6028 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
6031 // There are many forms of this optimization we can handle, for now, just do
6032 // the simple index into a single-dimensional array.
6034 // Require: GEP GV, 0, i
6035 if (GEP->getNumOperands() != 3 ||
6036 !isa<ConstantInt>(GEP->getOperand(1)) ||
6037 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
6040 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
6041 if (Init == 0 || Init->getNumOperands() > 1024) return 0;
6044 // Variables for our state machines.
6046 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
6047 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
6048 // and 87 is the second (and last) index. FirstTrueElement is -1 when
6049 // undefined, otherwise set to the first true element. SecondTrueElement is
6050 // -1 when undefined, -2 when overdefined and >= 0 when that index is true.
6051 int FirstTrueElement = -1, SecondTrueElement = -1;
6053 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
6054 // form "i != 47 & i != 87". Same state transitions as for true elements.
6055 int FirstFalseElement = -1, SecondFalseElement = -1;
6057 // MagicBitvector - This is a magic bitvector where we set a bit if the
6058 // comparison is true for element 'i'. If there are 64 elements or less in
6059 // the array, this will fully represent all the comparison results.
6060 uint64_t MagicBitvector = 0;
6063 // Scan the array and see if one of our patterns matches.
6064 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
6065 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
6066 // Find out if the comparison would be true or false for the i'th element.
6067 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(),
6068 Init->getOperand(i),
6070 // If the result is undef for this element, ignore it.
6071 if (isa<UndefValue>(C)) continue;
6073 // If we can't compute the result for any of the elements, we have to give
6074 // up evaluating the entire conditional.
6075 if (!isa<ConstantInt>(C)) return 0;
6077 // Otherwise, we know if the comparison is true or false for this element,
6078 // update our state machines.
6079 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
6081 // State machine for single index comparison.
6083 // Update the TrueElement state machine.
6084 if (FirstTrueElement == -1)
6085 FirstTrueElement = i;
6086 else if (SecondTrueElement == -1)
6087 SecondTrueElement = i;
6089 SecondTrueElement = -2;
6091 // Update the FalseElement state machine.
6092 if (FirstFalseElement == -1)
6093 FirstFalseElement = i;
6094 else if (SecondFalseElement == -1)
6095 SecondFalseElement = i;
6097 SecondFalseElement = -2;
6100 // If this element is in range, update our magic bitvector.
6101 if (i < 64 && IsTrueForElt)
6102 MagicBitvector |= 1ULL << i;
6104 // If all of our states become overdefined, bail out early.
6105 if (i >= 64 && SecondTrueElement == -2 && SecondFalseElement == -2)
6109 // Now that we've scanned the entire array, emit our new comparison(s). We
6110 // order the state machines in complexity of the generated code.
6111 Value *Idx = GEP->getOperand(2);
6113 // If the comparison is only true for one or two elements, emit direct
6115 if (SecondTrueElement != -2) {
6116 // None true -> false.
6117 if (FirstTrueElement == -1)
6118 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6120 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
6122 // True for one element -> 'i == 47'.
6123 if (SecondTrueElement == -1)
6124 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
6126 // True for two elements -> 'i == 47 | i == 72'.
6127 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
6128 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
6129 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
6130 return BinaryOperator::CreateOr(C1, C2);
6133 // If the comparison is only false for one or two elements, emit direct
6135 if (SecondFalseElement != -2) {
6136 // None false -> true.
6137 if (FirstFalseElement == -1)
6138 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6140 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
6142 // False for one element -> 'i != 47'.
6143 if (SecondFalseElement == -1)
6144 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
6146 // False for two elements -> 'i != 47 & i != 72'.
6147 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
6148 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
6149 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
6150 return BinaryOperator::CreateAnd(C1, C2);
6153 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
6154 // of this load, replace it with computation that does:
6155 // ((magic_cst >> i) & 1) != 0
6156 if (Init->getNumOperands() <= 32 ||
6157 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
6159 if (Init->getNumOperands() <= 32)
6160 Ty = Type::getInt32Ty(Init->getContext());
6162 Ty = Type::getInt64Ty(Init->getContext());
6163 Value *V = Builder->CreateIntCast(Idx, Ty, false);
6164 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
6165 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
6166 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
6169 // TODO: Range check
6170 // TODO: GEP 0, i, 4
6175 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
6176 bool Changed = false;
6178 /// Orders the operands of the compare so that they are listed from most
6179 /// complex to least complex. This puts constants before unary operators,
6180 /// before binary operators.
6181 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
6186 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6188 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
6189 return ReplaceInstUsesWith(I, V);
6191 // Simplify 'fcmp pred X, X'
6193 switch (I.getPredicate()) {
6194 default: llvm_unreachable("Unknown predicate!");
6195 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
6196 case FCmpInst::FCMP_ULT: // True if unordered or less than
6197 case FCmpInst::FCMP_UGT: // True if unordered or greater than
6198 case FCmpInst::FCMP_UNE: // True if unordered or not equal
6199 // Canonicalize these to be 'fcmp uno %X, 0.0'.
6200 I.setPredicate(FCmpInst::FCMP_UNO);
6201 I.setOperand(1, Constant::getNullValue(Op0->getType()));
6204 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
6205 case FCmpInst::FCMP_OEQ: // True if ordered and equal
6206 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
6207 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
6208 // Canonicalize these to be 'fcmp ord %X, 0.0'.
6209 I.setPredicate(FCmpInst::FCMP_ORD);
6210 I.setOperand(1, Constant::getNullValue(Op0->getType()));
6215 // Handle fcmp with constant RHS
6216 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6217 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6218 switch (LHSI->getOpcode()) {
6219 case Instruction::PHI:
6220 // Only fold fcmp into the PHI if the phi and fcmp are in the same
6221 // block. If in the same block, we're encouraging jump threading. If
6222 // not, we are just pessimizing the code by making an i1 phi.
6223 if (LHSI->getParent() == I.getParent())
6224 if (Instruction *NV = FoldOpIntoPhi(I, true))
6227 case Instruction::SIToFP:
6228 case Instruction::UIToFP:
6229 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
6232 case Instruction::Select: {
6233 // If either operand of the select is a constant, we can fold the
6234 // comparison into the select arms, which will cause one to be
6235 // constant folded and the select turned into a bitwise or.
6236 Value *Op1 = 0, *Op2 = 0;
6237 if (LHSI->hasOneUse()) {
6238 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6239 // Fold the known value into the constant operand.
6240 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
6241 // Insert a new FCmp of the other select operand.
6242 Op2 = Builder->CreateFCmp(I.getPredicate(),
6243 LHSI->getOperand(2), RHSC, I.getName());
6244 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6245 // Fold the known value into the constant operand.
6246 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
6247 // Insert a new FCmp of the other select operand.
6248 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
6254 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6257 case Instruction::Load:
6258 if (GetElementPtrInst *GEP =
6259 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
6260 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
6261 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
6262 !cast<LoadInst>(LHSI)->isVolatile())
6263 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
6265 //errs() << "NOT HANDLED: " << *GV << "\n";
6266 //errs() << "\t" << *GEP << "\n";
6267 //errs() << "\t " << I << "\n\n\n";
6273 return Changed ? &I : 0;
6276 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
6277 bool Changed = false;
6279 /// Orders the operands of the compare so that they are listed from most
6280 /// complex to least complex. This puts constants before unary operators,
6281 /// before binary operators.
6282 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
6287 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6289 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
6290 return ReplaceInstUsesWith(I, V);
6292 const Type *Ty = Op0->getType();
6294 // icmp's with boolean values can always be turned into bitwise operations
6295 if (Ty == Type::getInt1Ty(*Context)) {
6296 switch (I.getPredicate()) {
6297 default: llvm_unreachable("Invalid icmp instruction!");
6298 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6299 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
6300 return BinaryOperator::CreateNot(Xor);
6302 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6303 return BinaryOperator::CreateXor(Op0, Op1);
6305 case ICmpInst::ICMP_UGT:
6306 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6308 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6309 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6310 return BinaryOperator::CreateAnd(Not, Op1);
6312 case ICmpInst::ICMP_SGT:
6313 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6315 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6316 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6317 return BinaryOperator::CreateAnd(Not, Op0);
6319 case ICmpInst::ICMP_UGE:
6320 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6322 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6323 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6324 return BinaryOperator::CreateOr(Not, Op1);
6326 case ICmpInst::ICMP_SGE:
6327 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6329 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6330 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6331 return BinaryOperator::CreateOr(Not, Op0);
6336 unsigned BitWidth = 0;
6338 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6339 else if (Ty->isIntOrIntVector())
6340 BitWidth = Ty->getScalarSizeInBits();
6342 bool isSignBit = false;
6344 // See if we are doing a comparison with a constant.
6345 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6346 Value *A = 0, *B = 0;
6348 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6349 if (I.isEquality() && CI->isZero() &&
6350 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6351 // (icmp cond A B) if cond is equality
6352 return new ICmpInst(I.getPredicate(), A, B);
6355 // If we have an icmp le or icmp ge instruction, turn it into the
6356 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6357 // them being folded in the code below. The SimplifyICmpInst code has
6358 // already handled the edge cases for us, so we just assert on them.
6359 switch (I.getPredicate()) {
6361 case ICmpInst::ICMP_ULE:
6362 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
6363 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
6365 case ICmpInst::ICMP_SLE:
6366 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
6367 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6369 case ICmpInst::ICMP_UGE:
6370 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
6371 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
6373 case ICmpInst::ICMP_SGE:
6374 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
6375 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6379 // If this comparison is a normal comparison, it demands all
6380 // bits, if it is a sign bit comparison, it only demands the sign bit.
6382 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6385 // See if we can fold the comparison based on range information we can get
6386 // by checking whether bits are known to be zero or one in the input.
6387 if (BitWidth != 0) {
6388 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6389 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6391 if (SimplifyDemandedBits(I.getOperandUse(0),
6392 isSignBit ? APInt::getSignBit(BitWidth)
6393 : APInt::getAllOnesValue(BitWidth),
6394 Op0KnownZero, Op0KnownOne, 0))
6396 if (SimplifyDemandedBits(I.getOperandUse(1),
6397 APInt::getAllOnesValue(BitWidth),
6398 Op1KnownZero, Op1KnownOne, 0))
6401 // Given the known and unknown bits, compute a range that the LHS could be
6402 // in. Compute the Min, Max and RHS values based on the known bits. For the
6403 // EQ and NE we use unsigned values.
6404 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6405 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6407 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6409 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6412 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6414 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6418 // If Min and Max are known to be the same, then SimplifyDemandedBits
6419 // figured out that the LHS is a constant. Just constant fold this now so
6420 // that code below can assume that Min != Max.
6421 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6422 return new ICmpInst(I.getPredicate(),
6423 ConstantInt::get(*Context, Op0Min), Op1);
6424 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6425 return new ICmpInst(I.getPredicate(), Op0,
6426 ConstantInt::get(*Context, Op1Min));
6428 // Based on the range information we know about the LHS, see if we can
6429 // simplify this comparison. For example, (x&4) < 8 is always true.
6430 switch (I.getPredicate()) {
6431 default: llvm_unreachable("Unknown icmp opcode!");
6432 case ICmpInst::ICMP_EQ:
6433 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6434 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6436 case ICmpInst::ICMP_NE:
6437 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6438 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6440 case ICmpInst::ICMP_ULT:
6441 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6442 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6443 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6444 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6445 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6446 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6447 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6448 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6449 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6452 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6453 if (CI->isMinValue(true))
6454 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6455 Constant::getAllOnesValue(Op0->getType()));
6458 case ICmpInst::ICMP_UGT:
6459 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6460 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6461 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6462 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6464 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6465 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6466 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6467 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6468 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6471 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6472 if (CI->isMaxValue(true))
6473 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6474 Constant::getNullValue(Op0->getType()));
6477 case ICmpInst::ICMP_SLT:
6478 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6479 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6480 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6481 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6482 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6483 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6484 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6485 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6486 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6490 case ICmpInst::ICMP_SGT:
6491 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6492 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6493 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6494 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6496 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6497 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6498 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6499 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6500 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6504 case ICmpInst::ICMP_SGE:
6505 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6506 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6507 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6508 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6509 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6511 case ICmpInst::ICMP_SLE:
6512 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6513 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6514 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6515 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6516 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6518 case ICmpInst::ICMP_UGE:
6519 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6520 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6521 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6522 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6523 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6525 case ICmpInst::ICMP_ULE:
6526 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6527 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6528 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6529 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6530 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6534 // Turn a signed comparison into an unsigned one if both operands
6535 // are known to have the same sign.
6537 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6538 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6539 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6542 // Test if the ICmpInst instruction is used exclusively by a select as
6543 // part of a minimum or maximum operation. If so, refrain from doing
6544 // any other folding. This helps out other analyses which understand
6545 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6546 // and CodeGen. And in this case, at least one of the comparison
6547 // operands has at least one user besides the compare (the select),
6548 // which would often largely negate the benefit of folding anyway.
6550 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6551 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6552 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6555 // See if we are doing a comparison between a constant and an instruction that
6556 // can be folded into the comparison.
6557 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6558 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6559 // instruction, see if that instruction also has constants so that the
6560 // instruction can be folded into the icmp
6561 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6562 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6566 // Handle icmp with constant (but not simple integer constant) RHS
6567 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6568 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6569 switch (LHSI->getOpcode()) {
6570 case Instruction::GetElementPtr:
6571 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6572 if (RHSC->isNullValue() &&
6573 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
6574 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6575 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6577 case Instruction::PHI:
6578 // Only fold icmp into the PHI if the phi and icmp are in the same
6579 // block. If in the same block, we're encouraging jump threading. If
6580 // not, we are just pessimizing the code by making an i1 phi.
6581 if (LHSI->getParent() == I.getParent())
6582 if (Instruction *NV = FoldOpIntoPhi(I, true))
6585 case Instruction::Select: {
6586 // If either operand of the select is a constant, we can fold the
6587 // comparison into the select arms, which will cause one to be
6588 // constant folded and the select turned into a bitwise or.
6589 Value *Op1 = 0, *Op2 = 0;
6590 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
6591 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6592 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
6593 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6595 // We only want to perform this transformation if it will not lead to
6596 // additional code. This is true if either both sides of the select
6597 // fold to a constant (in which case the icmp is replaced with a select
6598 // which will usually simplify) or this is the only user of the
6599 // select (in which case we are trading a select+icmp for a simpler
6601 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
6603 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
6606 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
6608 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6612 case Instruction::Call:
6613 // If we have (malloc != null), and if the malloc has a single use, we
6614 // can assume it is successful and remove the malloc.
6615 if (isMalloc(LHSI) && LHSI->hasOneUse() &&
6616 isa<ConstantPointerNull>(RHSC)) {
6617 // Need to explicitly erase malloc call here, instead of adding it to
6618 // Worklist, because it won't get DCE'd from the Worklist since
6619 // isInstructionTriviallyDead() returns false for function calls.
6620 // It is OK to replace LHSI/MallocCall with Undef because the
6621 // instruction that uses it will be erased via Worklist.
6622 if (extractMallocCall(LHSI)) {
6623 LHSI->replaceAllUsesWith(UndefValue::get(LHSI->getType()));
6624 EraseInstFromFunction(*LHSI);
6625 return ReplaceInstUsesWith(I,
6626 ConstantInt::get(Type::getInt1Ty(*Context),
6627 !I.isTrueWhenEqual()));
6629 if (CallInst* MallocCall = extractMallocCallFromBitCast(LHSI))
6630 if (MallocCall->hasOneUse()) {
6631 MallocCall->replaceAllUsesWith(
6632 UndefValue::get(MallocCall->getType()));
6633 EraseInstFromFunction(*MallocCall);
6634 Worklist.Add(LHSI); // The malloc's bitcast use.
6635 return ReplaceInstUsesWith(I,
6636 ConstantInt::get(Type::getInt1Ty(*Context),
6637 !I.isTrueWhenEqual()));
6641 case Instruction::IntToPtr:
6642 // icmp pred inttoptr(X), null -> icmp pred X, 0
6643 if (RHSC->isNullValue() && TD &&
6644 TD->getIntPtrType(RHSC->getContext()) ==
6645 LHSI->getOperand(0)->getType())
6646 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6647 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6650 case Instruction::Load:
6651 if (GetElementPtrInst *GEP =
6652 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
6653 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
6654 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
6655 !cast<LoadInst>(LHSI)->isVolatile())
6656 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
6658 //errs() << "NOT HANDLED: " << *GV << "\n";
6659 //errs() << "\t" << *GEP << "\n";
6660 //errs() << "\t " << I << "\n\n\n";
6666 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6667 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6668 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6670 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6671 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6672 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6675 // Test to see if the operands of the icmp are casted versions of other
6676 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6678 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6679 if (isa<PointerType>(Op0->getType()) &&
6680 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6681 // We keep moving the cast from the left operand over to the right
6682 // operand, where it can often be eliminated completely.
6683 Op0 = CI->getOperand(0);
6685 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6686 // so eliminate it as well.
6687 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6688 Op1 = CI2->getOperand(0);
6690 // If Op1 is a constant, we can fold the cast into the constant.
6691 if (Op0->getType() != Op1->getType()) {
6692 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6693 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6695 // Otherwise, cast the RHS right before the icmp
6696 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
6699 return new ICmpInst(I.getPredicate(), Op0, Op1);
6703 if (isa<CastInst>(Op0)) {
6704 // Handle the special case of: icmp (cast bool to X), <cst>
6705 // This comes up when you have code like
6708 // For generality, we handle any zero-extension of any operand comparison
6709 // with a constant or another cast from the same type.
6710 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
6711 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6715 // See if it's the same type of instruction on the left and right.
6716 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6717 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6718 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6719 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6720 switch (Op0I->getOpcode()) {
6722 case Instruction::Add:
6723 case Instruction::Sub:
6724 case Instruction::Xor:
6725 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6726 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6727 Op1I->getOperand(0));
6728 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6729 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6730 if (CI->getValue().isSignBit()) {
6731 ICmpInst::Predicate Pred = I.isSigned()
6732 ? I.getUnsignedPredicate()
6733 : I.getSignedPredicate();
6734 return new ICmpInst(Pred, Op0I->getOperand(0),
6735 Op1I->getOperand(0));
6738 if (CI->getValue().isMaxSignedValue()) {
6739 ICmpInst::Predicate Pred = I.isSigned()
6740 ? I.getUnsignedPredicate()
6741 : I.getSignedPredicate();
6742 Pred = I.getSwappedPredicate(Pred);
6743 return new ICmpInst(Pred, Op0I->getOperand(0),
6744 Op1I->getOperand(0));
6748 case Instruction::Mul:
6749 if (!I.isEquality())
6752 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6753 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6754 // Mask = -1 >> count-trailing-zeros(Cst).
6755 if (!CI->isZero() && !CI->isOne()) {
6756 const APInt &AP = CI->getValue();
6757 ConstantInt *Mask = ConstantInt::get(*Context,
6758 APInt::getLowBitsSet(AP.getBitWidth(),
6760 AP.countTrailingZeros()));
6761 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
6762 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
6763 return new ICmpInst(I.getPredicate(), And1, And2);
6772 // ~x < ~y --> y < x
6774 if (match(Op0, m_Not(m_Value(A))) &&
6775 match(Op1, m_Not(m_Value(B))))
6776 return new ICmpInst(I.getPredicate(), B, A);
6779 if (I.isEquality()) {
6780 Value *A, *B, *C, *D;
6782 // -x == -y --> x == y
6783 if (match(Op0, m_Neg(m_Value(A))) &&
6784 match(Op1, m_Neg(m_Value(B))))
6785 return new ICmpInst(I.getPredicate(), A, B);
6787 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6788 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6789 Value *OtherVal = A == Op1 ? B : A;
6790 return new ICmpInst(I.getPredicate(), OtherVal,
6791 Constant::getNullValue(A->getType()));
6794 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6795 // A^c1 == C^c2 --> A == C^(c1^c2)
6796 ConstantInt *C1, *C2;
6797 if (match(B, m_ConstantInt(C1)) &&
6798 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6800 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6801 Value *Xor = Builder->CreateXor(C, NC, "tmp");
6802 return new ICmpInst(I.getPredicate(), A, Xor);
6805 // A^B == A^D -> B == D
6806 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6807 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6808 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6809 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6813 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6814 (A == Op0 || B == Op0)) {
6815 // A == (A^B) -> B == 0
6816 Value *OtherVal = A == Op0 ? B : A;
6817 return new ICmpInst(I.getPredicate(), OtherVal,
6818 Constant::getNullValue(A->getType()));
6821 // (A-B) == A -> B == 0
6822 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6823 return new ICmpInst(I.getPredicate(), B,
6824 Constant::getNullValue(B->getType()));
6826 // A == (A-B) -> B == 0
6827 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6828 return new ICmpInst(I.getPredicate(), B,
6829 Constant::getNullValue(B->getType()));
6831 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6832 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6833 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6834 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6835 Value *X = 0, *Y = 0, *Z = 0;
6838 X = B; Y = D; Z = A;
6839 } else if (A == D) {
6840 X = B; Y = C; Z = A;
6841 } else if (B == C) {
6842 X = A; Y = D; Z = B;
6843 } else if (B == D) {
6844 X = A; Y = C; Z = B;
6847 if (X) { // Build (X^Y) & Z
6848 Op1 = Builder->CreateXor(X, Y, "tmp");
6849 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
6850 I.setOperand(0, Op1);
6851 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6858 Value *X; ConstantInt *Cst;
6860 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
6861 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
6864 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
6865 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
6867 return Changed ? &I : 0;
6870 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
6871 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
6872 Value *X, ConstantInt *CI,
6873 ICmpInst::Predicate Pred,
6875 // If we have X+0, exit early (simplifying logic below) and let it get folded
6876 // elsewhere. icmp X+0, X -> icmp X, X
6878 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
6879 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
6882 // (X+4) == X -> false.
6883 if (Pred == ICmpInst::ICMP_EQ)
6884 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
6886 // (X+4) != X -> true.
6887 if (Pred == ICmpInst::ICMP_NE)
6888 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
6890 // If this is an instruction (as opposed to constantexpr) get NUW/NSW info.
6891 bool isNUW = false, isNSW = false;
6892 if (BinaryOperator *Add = dyn_cast<BinaryOperator>(TheAdd)) {
6893 isNUW = Add->hasNoUnsignedWrap();
6894 isNSW = Add->hasNoSignedWrap();
6897 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
6898 // so the values can never be equal. Similiarly for all other "or equals"
6901 // (X+1) <u X --> X >u (MAXUINT-1) --> X != 255
6902 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
6903 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
6904 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
6905 // If this is an NUW add, then this is always false.
6907 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
6909 Value *R = ConstantExpr::getSub(ConstantInt::get(CI->getType(), -1ULL), CI);
6910 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
6913 // (X+1) >u X --> X <u (0-1) --> X != 255
6914 // (X+2) >u X --> X <u (0-2) --> X <u 254
6915 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
6916 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
6917 // If this is an NUW add, then this is always true.
6919 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
6920 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
6923 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
6924 ConstantInt *SMax = ConstantInt::get(X->getContext(),
6925 APInt::getSignedMaxValue(BitWidth));
6927 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
6928 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
6929 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
6930 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
6931 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
6932 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
6933 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
6934 // If this is an NSW add, then we have two cases: if the constant is
6935 // positive, then this is always false, if negative, this is always true.
6937 bool isTrue = CI->getValue().isNegative();
6938 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
6941 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
6944 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
6945 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
6946 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
6947 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
6948 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
6949 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
6951 // If this is an NSW add, then we have two cases: if the constant is
6952 // positive, then this is always true, if negative, this is always false.
6954 bool isTrue = !CI->getValue().isNegative();
6955 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
6958 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
6959 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
6960 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
6963 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6964 /// and CmpRHS are both known to be integer constants.
6965 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6966 ConstantInt *DivRHS) {
6967 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6968 const APInt &CmpRHSV = CmpRHS->getValue();
6970 // FIXME: If the operand types don't match the type of the divide
6971 // then don't attempt this transform. The code below doesn't have the
6972 // logic to deal with a signed divide and an unsigned compare (and
6973 // vice versa). This is because (x /s C1) <s C2 produces different
6974 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6975 // (x /u C1) <u C2. Simply casting the operands and result won't
6976 // work. :( The if statement below tests that condition and bails
6978 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6979 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
6981 if (DivRHS->isZero())
6982 return 0; // The ProdOV computation fails on divide by zero.
6983 if (DivIsSigned && DivRHS->isAllOnesValue())
6984 return 0; // The overflow computation also screws up here
6985 if (DivRHS->isOne())
6986 return 0; // Not worth bothering, and eliminates some funny cases
6989 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6990 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6991 // C2 (CI). By solving for X we can turn this into a range check
6992 // instead of computing a divide.
6993 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6995 // Determine if the product overflows by seeing if the product is
6996 // not equal to the divide. Make sure we do the same kind of divide
6997 // as in the LHS instruction that we're folding.
6998 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6999 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
7001 // Get the ICmp opcode
7002 ICmpInst::Predicate Pred = ICI.getPredicate();
7004 // Figure out the interval that is being checked. For example, a comparison
7005 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
7006 // Compute this interval based on the constants involved and the signedness of
7007 // the compare/divide. This computes a half-open interval, keeping track of
7008 // whether either value in the interval overflows. After analysis each
7009 // overflow variable is set to 0 if it's corresponding bound variable is valid
7010 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
7011 int LoOverflow = 0, HiOverflow = 0;
7012 Constant *LoBound = 0, *HiBound = 0;
7014 if (!DivIsSigned) { // udiv
7015 // e.g. X/5 op 3 --> [15, 20)
7017 HiOverflow = LoOverflow = ProdOV;
7019 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
7020 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
7021 if (CmpRHSV == 0) { // (X / pos) op 0
7022 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
7023 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
7025 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
7026 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
7027 HiOverflow = LoOverflow = ProdOV;
7029 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
7030 } else { // (X / pos) op neg
7031 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
7032 HiBound = AddOne(Prod);
7033 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
7035 ConstantInt* DivNeg =
7036 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
7037 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
7041 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
7042 if (CmpRHSV == 0) { // (X / neg) op 0
7043 // e.g. X/-5 op 0 --> [-4, 5)
7044 LoBound = AddOne(DivRHS);
7045 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
7046 if (HiBound == DivRHS) { // -INTMIN = INTMIN
7047 HiOverflow = 1; // [INTMIN+1, overflow)
7048 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
7050 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
7051 // e.g. X/-5 op 3 --> [-19, -14)
7052 HiBound = AddOne(Prod);
7053 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
7055 LoOverflow = AddWithOverflow(LoBound, HiBound,
7056 DivRHS, Context, true) ? -1 : 0;
7057 } else { // (X / neg) op neg
7058 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
7059 LoOverflow = HiOverflow = ProdOV;
7061 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
7064 // Dividing by a negative swaps the condition. LT <-> GT
7065 Pred = ICmpInst::getSwappedPredicate(Pred);
7068 Value *X = DivI->getOperand(0);
7070 default: llvm_unreachable("Unhandled icmp opcode!");
7071 case ICmpInst::ICMP_EQ:
7072 if (LoOverflow && HiOverflow)
7073 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7074 else if (HiOverflow)
7075 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
7076 ICmpInst::ICMP_UGE, X, LoBound);
7077 else if (LoOverflow)
7078 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
7079 ICmpInst::ICMP_ULT, X, HiBound);
7081 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
7082 case ICmpInst::ICMP_NE:
7083 if (LoOverflow && HiOverflow)
7084 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7085 else if (HiOverflow)
7086 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
7087 ICmpInst::ICMP_ULT, X, LoBound);
7088 else if (LoOverflow)
7089 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
7090 ICmpInst::ICMP_UGE, X, HiBound);
7092 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
7093 case ICmpInst::ICMP_ULT:
7094 case ICmpInst::ICMP_SLT:
7095 if (LoOverflow == +1) // Low bound is greater than input range.
7096 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7097 if (LoOverflow == -1) // Low bound is less than input range.
7098 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7099 return new ICmpInst(Pred, X, LoBound);
7100 case ICmpInst::ICMP_UGT:
7101 case ICmpInst::ICMP_SGT:
7102 if (HiOverflow == +1) // High bound greater than input range.
7103 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7104 else if (HiOverflow == -1) // High bound less than input range.
7105 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7106 if (Pred == ICmpInst::ICMP_UGT)
7107 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
7109 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
7114 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
7116 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
7119 const APInt &RHSV = RHS->getValue();
7121 switch (LHSI->getOpcode()) {
7122 case Instruction::Trunc:
7123 if (ICI.isEquality() && LHSI->hasOneUse()) {
7124 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
7125 // of the high bits truncated out of x are known.
7126 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
7127 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
7128 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
7129 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
7130 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
7132 // If all the high bits are known, we can do this xform.
7133 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
7134 // Pull in the high bits from known-ones set.
7135 APInt NewRHS(RHS->getValue());
7136 NewRHS.zext(SrcBits);
7138 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
7139 ConstantInt::get(*Context, NewRHS));
7144 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
7145 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
7146 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
7148 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
7149 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
7150 Value *CompareVal = LHSI->getOperand(0);
7152 // If the sign bit of the XorCST is not set, there is no change to
7153 // the operation, just stop using the Xor.
7154 if (!XorCST->getValue().isNegative()) {
7155 ICI.setOperand(0, CompareVal);
7160 // Was the old condition true if the operand is positive?
7161 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
7163 // If so, the new one isn't.
7164 isTrueIfPositive ^= true;
7166 if (isTrueIfPositive)
7167 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
7170 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
7174 if (LHSI->hasOneUse()) {
7175 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
7176 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
7177 const APInt &SignBit = XorCST->getValue();
7178 ICmpInst::Predicate Pred = ICI.isSigned()
7179 ? ICI.getUnsignedPredicate()
7180 : ICI.getSignedPredicate();
7181 return new ICmpInst(Pred, LHSI->getOperand(0),
7182 ConstantInt::get(*Context, RHSV ^ SignBit));
7185 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
7186 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
7187 const APInt &NotSignBit = XorCST->getValue();
7188 ICmpInst::Predicate Pred = ICI.isSigned()
7189 ? ICI.getUnsignedPredicate()
7190 : ICI.getSignedPredicate();
7191 Pred = ICI.getSwappedPredicate(Pred);
7192 return new ICmpInst(Pred, LHSI->getOperand(0),
7193 ConstantInt::get(*Context, RHSV ^ NotSignBit));
7198 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
7199 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
7200 LHSI->getOperand(0)->hasOneUse()) {
7201 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
7203 // If the LHS is an AND of a truncating cast, we can widen the
7204 // and/compare to be the input width without changing the value
7205 // produced, eliminating a cast.
7206 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
7207 // We can do this transformation if either the AND constant does not
7208 // have its sign bit set or if it is an equality comparison.
7209 // Extending a relational comparison when we're checking the sign
7210 // bit would not work.
7211 if (Cast->hasOneUse() &&
7212 (ICI.isEquality() ||
7213 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
7215 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
7216 APInt NewCST = AndCST->getValue();
7217 NewCST.zext(BitWidth);
7219 NewCI.zext(BitWidth);
7221 Builder->CreateAnd(Cast->getOperand(0),
7222 ConstantInt::get(*Context, NewCST), LHSI->getName());
7223 return new ICmpInst(ICI.getPredicate(), NewAnd,
7224 ConstantInt::get(*Context, NewCI));
7228 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
7229 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
7230 // happens a LOT in code produced by the C front-end, for bitfield
7232 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
7233 if (Shift && !Shift->isShift())
7237 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
7238 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
7239 const Type *AndTy = AndCST->getType(); // Type of the and.
7241 // We can fold this as long as we can't shift unknown bits
7242 // into the mask. This can only happen with signed shift
7243 // rights, as they sign-extend.
7245 bool CanFold = Shift->isLogicalShift();
7247 // To test for the bad case of the signed shr, see if any
7248 // of the bits shifted in could be tested after the mask.
7249 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
7250 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
7252 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
7253 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
7254 AndCST->getValue()) == 0)
7260 if (Shift->getOpcode() == Instruction::Shl)
7261 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
7263 NewCst = ConstantExpr::getShl(RHS, ShAmt);
7265 // Check to see if we are shifting out any of the bits being
7267 if (ConstantExpr::get(Shift->getOpcode(),
7268 NewCst, ShAmt) != RHS) {
7269 // If we shifted bits out, the fold is not going to work out.
7270 // As a special case, check to see if this means that the
7271 // result is always true or false now.
7272 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7273 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7274 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7275 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7277 ICI.setOperand(1, NewCst);
7278 Constant *NewAndCST;
7279 if (Shift->getOpcode() == Instruction::Shl)
7280 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
7282 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
7283 LHSI->setOperand(1, NewAndCST);
7284 LHSI->setOperand(0, Shift->getOperand(0));
7285 Worklist.Add(Shift); // Shift is dead.
7291 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
7292 // preferable because it allows the C<<Y expression to be hoisted out
7293 // of a loop if Y is invariant and X is not.
7294 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
7295 ICI.isEquality() && !Shift->isArithmeticShift() &&
7296 !isa<Constant>(Shift->getOperand(0))) {
7299 if (Shift->getOpcode() == Instruction::LShr) {
7300 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
7302 // Insert a logical shift.
7303 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
7306 // Compute X & (C << Y).
7308 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
7310 ICI.setOperand(0, NewAnd);
7316 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
7317 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7320 uint32_t TypeBits = RHSV.getBitWidth();
7322 // Check that the shift amount is in range. If not, don't perform
7323 // undefined shifts. When the shift is visited it will be
7325 if (ShAmt->uge(TypeBits))
7328 if (ICI.isEquality()) {
7329 // If we are comparing against bits always shifted out, the
7330 // comparison cannot succeed.
7332 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
7334 if (Comp != RHS) {// Comparing against a bit that we know is zero.
7335 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7336 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
7337 return ReplaceInstUsesWith(ICI, Cst);
7340 if (LHSI->hasOneUse()) {
7341 // Otherwise strength reduce the shift into an and.
7342 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
7344 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
7345 TypeBits-ShAmtVal));
7348 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
7349 return new ICmpInst(ICI.getPredicate(), And,
7350 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
7354 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
7355 bool TrueIfSigned = false;
7356 if (LHSI->hasOneUse() &&
7357 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
7358 // (X << 31) <s 0 --> (X&1) != 0
7359 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
7360 (TypeBits-ShAmt->getZExtValue()-1));
7362 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
7363 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
7364 And, Constant::getNullValue(And->getType()));
7369 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
7370 case Instruction::AShr: {
7371 // Only handle equality comparisons of shift-by-constant.
7372 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7373 if (!ShAmt || !ICI.isEquality()) break;
7375 // Check that the shift amount is in range. If not, don't perform
7376 // undefined shifts. When the shift is visited it will be
7378 uint32_t TypeBits = RHSV.getBitWidth();
7379 if (ShAmt->uge(TypeBits))
7382 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
7384 // If we are comparing against bits always shifted out, the
7385 // comparison cannot succeed.
7386 APInt Comp = RHSV << ShAmtVal;
7387 if (LHSI->getOpcode() == Instruction::LShr)
7388 Comp = Comp.lshr(ShAmtVal);
7390 Comp = Comp.ashr(ShAmtVal);
7392 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
7393 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7394 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
7395 return ReplaceInstUsesWith(ICI, Cst);
7398 // Otherwise, check to see if the bits shifted out are known to be zero.
7399 // If so, we can compare against the unshifted value:
7400 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
7401 if (LHSI->hasOneUse() &&
7402 MaskedValueIsZero(LHSI->getOperand(0),
7403 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
7404 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
7405 ConstantExpr::getShl(RHS, ShAmt));
7408 if (LHSI->hasOneUse()) {
7409 // Otherwise strength reduce the shift into an and.
7410 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7411 Constant *Mask = ConstantInt::get(*Context, Val);
7413 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
7414 Mask, LHSI->getName()+".mask");
7415 return new ICmpInst(ICI.getPredicate(), And,
7416 ConstantExpr::getShl(RHS, ShAmt));
7421 case Instruction::SDiv:
7422 case Instruction::UDiv:
7423 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7424 // Fold this div into the comparison, producing a range check.
7425 // Determine, based on the divide type, what the range is being
7426 // checked. If there is an overflow on the low or high side, remember
7427 // it, otherwise compute the range [low, hi) bounding the new value.
7428 // See: InsertRangeTest above for the kinds of replacements possible.
7429 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7430 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7435 case Instruction::Add:
7436 // Fold: icmp pred (add X, C1), C2
7437 if (!ICI.isEquality()) {
7438 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7440 const APInt &LHSV = LHSC->getValue();
7442 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7445 if (ICI.isSigned()) {
7446 if (CR.getLower().isSignBit()) {
7447 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7448 ConstantInt::get(*Context, CR.getUpper()));
7449 } else if (CR.getUpper().isSignBit()) {
7450 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7451 ConstantInt::get(*Context, CR.getLower()));
7454 if (CR.getLower().isMinValue()) {
7455 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7456 ConstantInt::get(*Context, CR.getUpper()));
7457 } else if (CR.getUpper().isMinValue()) {
7458 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7459 ConstantInt::get(*Context, CR.getLower()));
7466 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7467 if (ICI.isEquality()) {
7468 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7470 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7471 // the second operand is a constant, simplify a bit.
7472 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7473 switch (BO->getOpcode()) {
7474 case Instruction::SRem:
7475 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7476 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7477 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7478 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7480 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
7482 return new ICmpInst(ICI.getPredicate(), NewRem,
7483 Constant::getNullValue(BO->getType()));
7487 case Instruction::Add:
7488 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7489 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7490 if (BO->hasOneUse())
7491 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7492 ConstantExpr::getSub(RHS, BOp1C));
7493 } else if (RHSV == 0) {
7494 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7495 // efficiently invertible, or if the add has just this one use.
7496 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7498 if (Value *NegVal = dyn_castNegVal(BOp1))
7499 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
7500 else if (Value *NegVal = dyn_castNegVal(BOp0))
7501 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
7502 else if (BO->hasOneUse()) {
7503 Value *Neg = Builder->CreateNeg(BOp1);
7505 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
7509 case Instruction::Xor:
7510 // For the xor case, we can xor two constants together, eliminating
7511 // the explicit xor.
7512 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7513 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7514 ConstantExpr::getXor(RHS, BOC));
7517 case Instruction::Sub:
7518 // Replace (([sub|xor] A, B) != 0) with (A != B)
7520 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7524 case Instruction::Or:
7525 // If bits are being or'd in that are not present in the constant we
7526 // are comparing against, then the comparison could never succeed!
7527 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7528 Constant *NotCI = ConstantExpr::getNot(RHS);
7529 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7530 return ReplaceInstUsesWith(ICI,
7531 ConstantInt::get(Type::getInt1Ty(*Context),
7536 case Instruction::And:
7537 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7538 // If bits are being compared against that are and'd out, then the
7539 // comparison can never succeed!
7540 if ((RHSV & ~BOC->getValue()) != 0)
7541 return ReplaceInstUsesWith(ICI,
7542 ConstantInt::get(Type::getInt1Ty(*Context),
7545 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7546 if (RHS == BOC && RHSV.isPowerOf2())
7547 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7548 ICmpInst::ICMP_NE, LHSI,
7549 Constant::getNullValue(RHS->getType()));
7551 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7552 if (BOC->getValue().isSignBit()) {
7553 Value *X = BO->getOperand(0);
7554 Constant *Zero = Constant::getNullValue(X->getType());
7555 ICmpInst::Predicate pred = isICMP_NE ?
7556 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7557 return new ICmpInst(pred, X, Zero);
7560 // ((X & ~7) == 0) --> X < 8
7561 if (RHSV == 0 && isHighOnes(BOC)) {
7562 Value *X = BO->getOperand(0);
7563 Constant *NegX = ConstantExpr::getNeg(BOC);
7564 ICmpInst::Predicate pred = isICMP_NE ?
7565 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7566 return new ICmpInst(pred, X, NegX);
7571 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7572 // Handle icmp {eq|ne} <intrinsic>, intcst.
7573 if (II->getIntrinsicID() == Intrinsic::bswap) {
7575 ICI.setOperand(0, II->getOperand(1));
7576 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7584 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7585 /// We only handle extending casts so far.
7587 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7588 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7589 Value *LHSCIOp = LHSCI->getOperand(0);
7590 const Type *SrcTy = LHSCIOp->getType();
7591 const Type *DestTy = LHSCI->getType();
7594 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7595 // integer type is the same size as the pointer type.
7596 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7597 TD->getPointerSizeInBits() ==
7598 cast<IntegerType>(DestTy)->getBitWidth()) {
7600 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7601 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7602 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7603 RHSOp = RHSC->getOperand(0);
7604 // If the pointer types don't match, insert a bitcast.
7605 if (LHSCIOp->getType() != RHSOp->getType())
7606 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
7610 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7613 // The code below only handles extension cast instructions, so far.
7615 if (LHSCI->getOpcode() != Instruction::ZExt &&
7616 LHSCI->getOpcode() != Instruction::SExt)
7619 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7620 bool isSignedCmp = ICI.isSigned();
7622 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7623 // Not an extension from the same type?
7624 RHSCIOp = CI->getOperand(0);
7625 if (RHSCIOp->getType() != LHSCIOp->getType())
7628 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7629 // and the other is a zext), then we can't handle this.
7630 if (CI->getOpcode() != LHSCI->getOpcode())
7633 // Deal with equality cases early.
7634 if (ICI.isEquality())
7635 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7637 // A signed comparison of sign extended values simplifies into a
7638 // signed comparison.
7639 if (isSignedCmp && isSignedExt)
7640 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7642 // The other three cases all fold into an unsigned comparison.
7643 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7646 // If we aren't dealing with a constant on the RHS, exit early
7647 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7651 // Compute the constant that would happen if we truncated to SrcTy then
7652 // reextended to DestTy.
7653 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7654 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7657 // If the re-extended constant didn't change...
7659 // Deal with equality cases early.
7660 if (ICI.isEquality())
7661 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7663 // A signed comparison of sign extended values simplifies into a
7664 // signed comparison.
7665 if (isSignedExt && isSignedCmp)
7666 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7668 // The other three cases all fold into an unsigned comparison.
7669 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
7672 // The re-extended constant changed so the constant cannot be represented
7673 // in the shorter type. Consequently, we cannot emit a simple comparison.
7675 // First, handle some easy cases. We know the result cannot be equal at this
7676 // point so handle the ICI.isEquality() cases
7677 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7678 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7679 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7680 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7682 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7683 // should have been folded away previously and not enter in here.
7686 // We're performing a signed comparison.
7687 if (cast<ConstantInt>(CI)->getValue().isNegative())
7688 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7690 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7692 // We're performing an unsigned comparison.
7694 // We're performing an unsigned comp with a sign extended value.
7695 // This is true if the input is >= 0. [aka >s -1]
7696 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7697 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
7699 // Unsigned extend & unsigned compare -> always true.
7700 Result = ConstantInt::getTrue(*Context);
7704 // Finally, return the value computed.
7705 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7706 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7707 return ReplaceInstUsesWith(ICI, Result);
7709 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7710 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7711 "ICmp should be folded!");
7712 if (Constant *CI = dyn_cast<Constant>(Result))
7713 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7714 return BinaryOperator::CreateNot(Result);
7717 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7718 return commonShiftTransforms(I);
7721 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7722 return commonShiftTransforms(I);
7725 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7726 if (Instruction *R = commonShiftTransforms(I))
7729 Value *Op0 = I.getOperand(0);
7731 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7732 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7733 if (CSI->isAllOnesValue())
7734 return ReplaceInstUsesWith(I, CSI);
7736 // See if we can turn a signed shr into an unsigned shr.
7737 if (MaskedValueIsZero(Op0,
7738 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7739 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7741 // Arithmetic shifting an all-sign-bit value is a no-op.
7742 unsigned NumSignBits = ComputeNumSignBits(Op0);
7743 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7744 return ReplaceInstUsesWith(I, Op0);
7749 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7750 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7751 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7753 // shl X, 0 == X and shr X, 0 == X
7754 // shl 0, X == 0 and shr 0, X == 0
7755 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7756 Op0 == Constant::getNullValue(Op0->getType()))
7757 return ReplaceInstUsesWith(I, Op0);
7759 if (isa<UndefValue>(Op0)) {
7760 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7761 return ReplaceInstUsesWith(I, Op0);
7762 else // undef << X -> 0, undef >>u X -> 0
7763 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7765 if (isa<UndefValue>(Op1)) {
7766 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7767 return ReplaceInstUsesWith(I, Op0);
7768 else // X << undef, X >>u undef -> 0
7769 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7772 // See if we can fold away this shift.
7773 if (SimplifyDemandedInstructionBits(I))
7776 // Try to fold constant and into select arguments.
7777 if (isa<Constant>(Op0))
7778 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7779 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7782 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7783 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7788 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7789 BinaryOperator &I) {
7790 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7792 // See if we can simplify any instructions used by the instruction whose sole
7793 // purpose is to compute bits we don't care about.
7794 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7796 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7799 if (Op1->uge(TypeBits)) {
7800 if (I.getOpcode() != Instruction::AShr)
7801 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7803 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7808 // ((X*C1) << C2) == (X * (C1 << C2))
7809 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7810 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7811 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7812 return BinaryOperator::CreateMul(BO->getOperand(0),
7813 ConstantExpr::getShl(BOOp, Op1));
7815 // Try to fold constant and into select arguments.
7816 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7817 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7819 if (isa<PHINode>(Op0))
7820 if (Instruction *NV = FoldOpIntoPhi(I))
7823 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7824 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7825 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7826 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7827 // place. Don't try to do this transformation in this case. Also, we
7828 // require that the input operand is a shift-by-constant so that we have
7829 // confidence that the shifts will get folded together. We could do this
7830 // xform in more cases, but it is unlikely to be profitable.
7831 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7832 isa<ConstantInt>(TrOp->getOperand(1))) {
7833 // Okay, we'll do this xform. Make the shift of shift.
7834 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7835 // (shift2 (shift1 & 0x00FF), c2)
7836 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
7838 // For logical shifts, the truncation has the effect of making the high
7839 // part of the register be zeros. Emulate this by inserting an AND to
7840 // clear the top bits as needed. This 'and' will usually be zapped by
7841 // other xforms later if dead.
7842 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7843 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7844 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7846 // The mask we constructed says what the trunc would do if occurring
7847 // between the shifts. We want to know the effect *after* the second
7848 // shift. We know that it is a logical shift by a constant, so adjust the
7849 // mask as appropriate.
7850 if (I.getOpcode() == Instruction::Shl)
7851 MaskV <<= Op1->getZExtValue();
7853 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7854 MaskV = MaskV.lshr(Op1->getZExtValue());
7858 Value *And = Builder->CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7861 // Return the value truncated to the interesting size.
7862 return new TruncInst(And, I.getType());
7866 if (Op0->hasOneUse()) {
7867 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7868 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7871 switch (Op0BO->getOpcode()) {
7873 case Instruction::Add:
7874 case Instruction::And:
7875 case Instruction::Or:
7876 case Instruction::Xor: {
7877 // These operators commute.
7878 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7879 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7880 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7881 m_Specific(Op1)))) {
7882 Value *YS = // (Y << C)
7883 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
7885 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
7886 Op0BO->getOperand(1)->getName());
7887 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7888 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7889 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7892 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7893 Value *Op0BOOp1 = Op0BO->getOperand(1);
7894 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7896 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7897 m_ConstantInt(CC))) &&
7898 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7899 Value *YS = // (Y << C)
7900 Builder->CreateShl(Op0BO->getOperand(0), Op1,
7903 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7904 V1->getName()+".mask");
7905 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7910 case Instruction::Sub: {
7911 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7912 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7913 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7914 m_Specific(Op1)))) {
7915 Value *YS = // (Y << C)
7916 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7918 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
7919 Op0BO->getOperand(0)->getName());
7920 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7921 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7922 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7925 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7926 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7927 match(Op0BO->getOperand(0),
7928 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7929 m_ConstantInt(CC))) && V2 == Op1 &&
7930 cast<BinaryOperator>(Op0BO->getOperand(0))
7931 ->getOperand(0)->hasOneUse()) {
7932 Value *YS = // (Y << C)
7933 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7935 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7936 V1->getName()+".mask");
7938 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7946 // If the operand is an bitwise operator with a constant RHS, and the
7947 // shift is the only use, we can pull it out of the shift.
7948 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7949 bool isValid = true; // Valid only for And, Or, Xor
7950 bool highBitSet = false; // Transform if high bit of constant set?
7952 switch (Op0BO->getOpcode()) {
7953 default: isValid = false; break; // Do not perform transform!
7954 case Instruction::Add:
7955 isValid = isLeftShift;
7957 case Instruction::Or:
7958 case Instruction::Xor:
7961 case Instruction::And:
7966 // If this is a signed shift right, and the high bit is modified
7967 // by the logical operation, do not perform the transformation.
7968 // The highBitSet boolean indicates the value of the high bit of
7969 // the constant which would cause it to be modified for this
7972 if (isValid && I.getOpcode() == Instruction::AShr)
7973 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7976 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7979 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
7980 NewShift->takeName(Op0BO);
7982 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7989 // Find out if this is a shift of a shift by a constant.
7990 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7991 if (ShiftOp && !ShiftOp->isShift())
7994 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7995 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7996 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7997 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7998 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7999 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
8000 Value *X = ShiftOp->getOperand(0);
8002 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
8004 const IntegerType *Ty = cast<IntegerType>(I.getType());
8006 // Check for (X << c1) << c2 and (X >> c1) >> c2
8007 if (I.getOpcode() == ShiftOp->getOpcode()) {
8008 // If this is oversized composite shift, then unsigned shifts get 0, ashr
8010 if (AmtSum >= TypeBits) {
8011 if (I.getOpcode() != Instruction::AShr)
8012 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
8013 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
8016 return BinaryOperator::Create(I.getOpcode(), X,
8017 ConstantInt::get(Ty, AmtSum));
8020 if (ShiftOp->getOpcode() == Instruction::LShr &&
8021 I.getOpcode() == Instruction::AShr) {
8022 if (AmtSum >= TypeBits)
8023 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
8025 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
8026 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
8029 if (ShiftOp->getOpcode() == Instruction::AShr &&
8030 I.getOpcode() == Instruction::LShr) {
8031 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
8032 if (AmtSum >= TypeBits)
8033 AmtSum = TypeBits-1;
8035 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
8037 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
8038 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
8041 // Okay, if we get here, one shift must be left, and the other shift must be
8042 // right. See if the amounts are equal.
8043 if (ShiftAmt1 == ShiftAmt2) {
8044 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
8045 if (I.getOpcode() == Instruction::Shl) {
8046 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
8047 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
8049 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
8050 if (I.getOpcode() == Instruction::LShr) {
8051 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
8052 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
8054 // We can simplify ((X << C) >>s C) into a trunc + sext.
8055 // NOTE: we could do this for any C, but that would make 'unusual' integer
8056 // types. For now, just stick to ones well-supported by the code
8058 const Type *SExtType = 0;
8059 switch (Ty->getBitWidth() - ShiftAmt1) {
8066 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
8071 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
8072 // Otherwise, we can't handle it yet.
8073 } else if (ShiftAmt1 < ShiftAmt2) {
8074 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
8076 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
8077 if (I.getOpcode() == Instruction::Shl) {
8078 assert(ShiftOp->getOpcode() == Instruction::LShr ||
8079 ShiftOp->getOpcode() == Instruction::AShr);
8080 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
8082 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
8083 return BinaryOperator::CreateAnd(Shift,
8084 ConstantInt::get(*Context, Mask));
8087 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
8088 if (I.getOpcode() == Instruction::LShr) {
8089 assert(ShiftOp->getOpcode() == Instruction::Shl);
8090 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
8092 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
8093 return BinaryOperator::CreateAnd(Shift,
8094 ConstantInt::get(*Context, Mask));
8097 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
8099 assert(ShiftAmt2 < ShiftAmt1);
8100 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
8102 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
8103 if (I.getOpcode() == Instruction::Shl) {
8104 assert(ShiftOp->getOpcode() == Instruction::LShr ||
8105 ShiftOp->getOpcode() == Instruction::AShr);
8106 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
8107 ConstantInt::get(Ty, ShiftDiff));
8109 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
8110 return BinaryOperator::CreateAnd(Shift,
8111 ConstantInt::get(*Context, Mask));
8114 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
8115 if (I.getOpcode() == Instruction::LShr) {
8116 assert(ShiftOp->getOpcode() == Instruction::Shl);
8117 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
8119 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
8120 return BinaryOperator::CreateAnd(Shift,
8121 ConstantInt::get(*Context, Mask));
8124 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
8131 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
8132 /// expression. If so, decompose it, returning some value X, such that Val is
8135 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
8136 int &Offset, LLVMContext *Context) {
8137 assert(Val->getType() == Type::getInt32Ty(*Context) &&
8138 "Unexpected allocation size type!");
8139 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
8140 Offset = CI->getZExtValue();
8142 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
8143 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
8144 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8145 if (I->getOpcode() == Instruction::Shl) {
8146 // This is a value scaled by '1 << the shift amt'.
8147 Scale = 1U << RHS->getZExtValue();
8149 return I->getOperand(0);
8150 } else if (I->getOpcode() == Instruction::Mul) {
8151 // This value is scaled by 'RHS'.
8152 Scale = RHS->getZExtValue();
8154 return I->getOperand(0);
8155 } else if (I->getOpcode() == Instruction::Add) {
8156 // We have X+C. Check to see if we really have (X*C2)+C1,
8157 // where C1 is divisible by C2.
8160 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
8162 Offset += RHS->getZExtValue();
8169 // Otherwise, we can't look past this.
8176 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
8177 /// try to eliminate the cast by moving the type information into the alloc.
8178 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
8180 const PointerType *PTy = cast<PointerType>(CI.getType());
8182 BuilderTy AllocaBuilder(*Builder);
8183 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
8185 // Remove any uses of AI that are dead.
8186 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
8188 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
8189 Instruction *User = cast<Instruction>(*UI++);
8190 if (isInstructionTriviallyDead(User)) {
8191 while (UI != E && *UI == User)
8192 ++UI; // If this instruction uses AI more than once, don't break UI.
8195 DEBUG(errs() << "IC: DCE: " << *User << '\n');
8196 EraseInstFromFunction(*User);
8200 // This requires TargetData to get the alloca alignment and size information.
8203 // Get the type really allocated and the type casted to.
8204 const Type *AllocElTy = AI.getAllocatedType();
8205 const Type *CastElTy = PTy->getElementType();
8206 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
8208 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
8209 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
8210 if (CastElTyAlign < AllocElTyAlign) return 0;
8212 // If the allocation has multiple uses, only promote it if we are strictly
8213 // increasing the alignment of the resultant allocation. If we keep it the
8214 // same, we open the door to infinite loops of various kinds. (A reference
8215 // from a dbg.declare doesn't count as a use for this purpose.)
8216 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
8217 CastElTyAlign == AllocElTyAlign) return 0;
8219 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
8220 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
8221 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
8223 // See if we can satisfy the modulus by pulling a scale out of the array
8225 unsigned ArraySizeScale;
8227 Value *NumElements = // See if the array size is a decomposable linear expr.
8228 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
8229 ArrayOffset, Context);
8231 // If we can now satisfy the modulus, by using a non-1 scale, we really can
8233 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
8234 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
8236 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
8241 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
8242 // Insert before the alloca, not before the cast.
8243 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
8246 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
8247 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
8248 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
8251 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
8252 New->setAlignment(AI.getAlignment());
8255 // If the allocation has one real use plus a dbg.declare, just remove the
8257 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
8258 EraseInstFromFunction(*DI);
8260 // If the allocation has multiple real uses, insert a cast and change all
8261 // things that used it to use the new cast. This will also hack on CI, but it
8263 else if (!AI.hasOneUse()) {
8264 // New is the allocation instruction, pointer typed. AI is the original
8265 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
8266 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
8267 AI.replaceAllUsesWith(NewCast);
8269 return ReplaceInstUsesWith(CI, New);
8272 /// CanEvaluateInDifferentType - Return true if we can take the specified value
8273 /// and return it as type Ty without inserting any new casts and without
8274 /// changing the computed value. This is used by code that tries to decide
8275 /// whether promoting or shrinking integer operations to wider or smaller types
8276 /// will allow us to eliminate a truncate or extend.
8278 /// This is a truncation operation if Ty is smaller than V->getType(), or an
8279 /// extension operation if Ty is larger.
8281 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
8282 /// should return true if trunc(V) can be computed by computing V in the smaller
8283 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
8284 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
8285 /// efficiently truncated.
8287 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
8288 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
8289 /// the final result.
8290 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
8292 int &NumCastsRemoved){
8293 // We can always evaluate constants in another type.
8294 if (isa<Constant>(V))
8297 Instruction *I = dyn_cast<Instruction>(V);
8298 if (!I) return false;
8300 const Type *OrigTy = V->getType();
8302 // If this is an extension or truncate, we can often eliminate it.
8303 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
8304 // If this is a cast from the destination type, we can trivially eliminate
8305 // it, and this will remove a cast overall.
8306 if (I->getOperand(0)->getType() == Ty) {
8307 // If the first operand is itself a cast, and is eliminable, do not count
8308 // this as an eliminable cast. We would prefer to eliminate those two
8310 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
8316 // We can't extend or shrink something that has multiple uses: doing so would
8317 // require duplicating the instruction in general, which isn't profitable.
8318 if (!I->hasOneUse()) return false;
8320 unsigned Opc = I->getOpcode();
8322 case Instruction::Add:
8323 case Instruction::Sub:
8324 case Instruction::Mul:
8325 case Instruction::And:
8326 case Instruction::Or:
8327 case Instruction::Xor:
8328 // These operators can all arbitrarily be extended or truncated.
8329 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8331 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
8334 case Instruction::UDiv:
8335 case Instruction::URem: {
8336 // UDiv and URem can be truncated if all the truncated bits are zero.
8337 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8338 uint32_t BitWidth = Ty->getScalarSizeInBits();
8339 if (BitWidth < OrigBitWidth) {
8340 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
8341 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
8342 MaskedValueIsZero(I->getOperand(1), Mask)) {
8343 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8345 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
8351 case Instruction::Shl:
8352 // If we are truncating the result of this SHL, and if it's a shift of a
8353 // constant amount, we can always perform a SHL in a smaller type.
8354 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8355 uint32_t BitWidth = Ty->getScalarSizeInBits();
8356 if (BitWidth < OrigTy->getScalarSizeInBits() &&
8357 CI->getLimitedValue(BitWidth) < BitWidth)
8358 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8362 case Instruction::LShr:
8363 // If this is a truncate of a logical shr, we can truncate it to a smaller
8364 // lshr iff we know that the bits we would otherwise be shifting in are
8366 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8367 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8368 uint32_t BitWidth = Ty->getScalarSizeInBits();
8369 if (BitWidth < OrigBitWidth &&
8370 MaskedValueIsZero(I->getOperand(0),
8371 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8372 CI->getLimitedValue(BitWidth) < BitWidth) {
8373 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8378 case Instruction::ZExt:
8379 case Instruction::SExt:
8380 case Instruction::Trunc:
8381 // If this is the same kind of case as our original (e.g. zext+zext), we
8382 // can safely replace it. Note that replacing it does not reduce the number
8383 // of casts in the input.
8387 // sext (zext ty1), ty2 -> zext ty2
8388 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8391 case Instruction::Select: {
8392 SelectInst *SI = cast<SelectInst>(I);
8393 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8395 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8398 case Instruction::PHI: {
8399 // We can change a phi if we can change all operands.
8400 PHINode *PN = cast<PHINode>(I);
8401 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8402 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8408 // TODO: Can handle more cases here.
8415 /// EvaluateInDifferentType - Given an expression that
8416 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8417 /// evaluate the expression.
8418 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8420 if (Constant *C = dyn_cast<Constant>(V))
8421 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
8423 // Otherwise, it must be an instruction.
8424 Instruction *I = cast<Instruction>(V);
8425 Instruction *Res = 0;
8426 unsigned Opc = I->getOpcode();
8428 case Instruction::Add:
8429 case Instruction::Sub:
8430 case Instruction::Mul:
8431 case Instruction::And:
8432 case Instruction::Or:
8433 case Instruction::Xor:
8434 case Instruction::AShr:
8435 case Instruction::LShr:
8436 case Instruction::Shl:
8437 case Instruction::UDiv:
8438 case Instruction::URem: {
8439 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8440 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8441 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8444 case Instruction::Trunc:
8445 case Instruction::ZExt:
8446 case Instruction::SExt:
8447 // If the source type of the cast is the type we're trying for then we can
8448 // just return the source. There's no need to insert it because it is not
8450 if (I->getOperand(0)->getType() == Ty)
8451 return I->getOperand(0);
8453 // Otherwise, must be the same type of cast, so just reinsert a new one.
8454 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty);
8456 case Instruction::Select: {
8457 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8458 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8459 Res = SelectInst::Create(I->getOperand(0), True, False);
8462 case Instruction::PHI: {
8463 PHINode *OPN = cast<PHINode>(I);
8464 PHINode *NPN = PHINode::Create(Ty);
8465 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8466 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8467 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8473 // TODO: Can handle more cases here.
8474 llvm_unreachable("Unreachable!");
8479 return InsertNewInstBefore(Res, *I);
8482 /// @brief Implement the transforms common to all CastInst visitors.
8483 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8484 Value *Src = CI.getOperand(0);
8486 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8487 // eliminate it now.
8488 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8489 if (Instruction::CastOps opc =
8490 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8491 // The first cast (CSrc) is eliminable so we need to fix up or replace
8492 // the second cast (CI). CSrc will then have a good chance of being dead.
8493 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8497 // If we are casting a select then fold the cast into the select
8498 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8499 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8502 // If we are casting a PHI then fold the cast into the PHI
8503 if (isa<PHINode>(Src)) {
8504 // We don't do this if this would create a PHI node with an illegal type if
8505 // it is currently legal.
8506 if (!isa<IntegerType>(Src->getType()) ||
8507 !isa<IntegerType>(CI.getType()) ||
8508 ShouldChangeType(CI.getType(), Src->getType(), TD))
8509 if (Instruction *NV = FoldOpIntoPhi(CI))
8516 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8517 /// or not there is a sequence of GEP indices into the type that will land us at
8518 /// the specified offset. If so, fill them into NewIndices and return the
8519 /// resultant element type, otherwise return null.
8520 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8521 SmallVectorImpl<Value*> &NewIndices,
8522 const TargetData *TD,
8523 LLVMContext *Context) {
8525 if (!Ty->isSized()) return 0;
8527 // Start with the index over the outer type. Note that the type size
8528 // might be zero (even if the offset isn't zero) if the indexed type
8529 // is something like [0 x {int, int}]
8530 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8531 int64_t FirstIdx = 0;
8532 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8533 FirstIdx = Offset/TySize;
8534 Offset -= FirstIdx*TySize;
8536 // Handle hosts where % returns negative instead of values [0..TySize).
8540 assert(Offset >= 0);
8542 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8545 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8547 // Index into the types. If we fail, set OrigBase to null.
8549 // Indexing into tail padding between struct/array elements.
8550 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8553 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8554 const StructLayout *SL = TD->getStructLayout(STy);
8555 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8556 "Offset must stay within the indexed type");
8558 unsigned Elt = SL->getElementContainingOffset(Offset);
8559 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8561 Offset -= SL->getElementOffset(Elt);
8562 Ty = STy->getElementType(Elt);
8563 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8564 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8565 assert(EltSize && "Cannot index into a zero-sized array");
8566 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8568 Ty = AT->getElementType();
8570 // Otherwise, we can't index into the middle of this atomic type, bail.
8578 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8579 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8580 Value *Src = CI.getOperand(0);
8582 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8583 // If casting the result of a getelementptr instruction with no offset, turn
8584 // this into a cast of the original pointer!
8585 if (GEP->hasAllZeroIndices()) {
8586 // Changing the cast operand is usually not a good idea but it is safe
8587 // here because the pointer operand is being replaced with another
8588 // pointer operand so the opcode doesn't need to change.
8590 CI.setOperand(0, GEP->getOperand(0));
8594 // If the GEP has a single use, and the base pointer is a bitcast, and the
8595 // GEP computes a constant offset, see if we can convert these three
8596 // instructions into fewer. This typically happens with unions and other
8597 // non-type-safe code.
8598 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8599 if (GEP->hasAllConstantIndices()) {
8600 // We are guaranteed to get a constant from EmitGEPOffset.
8601 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, *this));
8602 int64_t Offset = OffsetV->getSExtValue();
8604 // Get the base pointer input of the bitcast, and the type it points to.
8605 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8606 const Type *GEPIdxTy =
8607 cast<PointerType>(OrigBase->getType())->getElementType();
8608 SmallVector<Value*, 8> NewIndices;
8609 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8610 // If we were able to index down into an element, create the GEP
8611 // and bitcast the result. This eliminates one bitcast, potentially
8613 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
8614 Builder->CreateInBoundsGEP(OrigBase,
8615 NewIndices.begin(), NewIndices.end()) :
8616 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
8617 NGEP->takeName(GEP);
8619 if (isa<BitCastInst>(CI))
8620 return new BitCastInst(NGEP, CI.getType());
8621 assert(isa<PtrToIntInst>(CI));
8622 return new PtrToIntInst(NGEP, CI.getType());
8628 return commonCastTransforms(CI);
8631 /// commonIntCastTransforms - This function implements the common transforms
8632 /// for trunc, zext, and sext.
8633 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8634 if (Instruction *Result = commonCastTransforms(CI))
8637 Value *Src = CI.getOperand(0);
8638 const Type *SrcTy = Src->getType();
8639 const Type *DestTy = CI.getType();
8640 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8641 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8643 // See if we can simplify any instructions used by the LHS whose sole
8644 // purpose is to compute bits we don't care about.
8645 if (SimplifyDemandedInstructionBits(CI))
8648 // If the source isn't an instruction or has more than one use then we
8649 // can't do anything more.
8650 Instruction *SrcI = dyn_cast<Instruction>(Src);
8651 if (!SrcI || !Src->hasOneUse())
8654 // Attempt to propagate the cast into the instruction for int->int casts.
8655 int NumCastsRemoved = 0;
8656 // Only do this if the dest type is a simple type, don't convert the
8657 // expression tree to something weird like i93 unless the source is also
8659 if ((isa<VectorType>(DestTy) ||
8660 ShouldChangeType(SrcI->getType(), DestTy, TD)) &&
8661 CanEvaluateInDifferentType(SrcI, DestTy,
8662 CI.getOpcode(), NumCastsRemoved)) {
8663 // If this cast is a truncate, evaluting in a different type always
8664 // eliminates the cast, so it is always a win. If this is a zero-extension,
8665 // we need to do an AND to maintain the clear top-part of the computation,
8666 // so we require that the input have eliminated at least one cast. If this
8667 // is a sign extension, we insert two new casts (to do the extension) so we
8668 // require that two casts have been eliminated.
8669 bool DoXForm = false;
8670 bool JustReplace = false;
8671 switch (CI.getOpcode()) {
8673 // All the others use floating point so we shouldn't actually
8674 // get here because of the check above.
8675 llvm_unreachable("Unknown cast type");
8676 case Instruction::Trunc:
8679 case Instruction::ZExt: {
8680 DoXForm = NumCastsRemoved >= 1;
8682 if (!DoXForm && 0) {
8683 // If it's unnecessary to issue an AND to clear the high bits, it's
8684 // always profitable to do this xform.
8685 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8686 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8687 if (MaskedValueIsZero(TryRes, Mask))
8688 return ReplaceInstUsesWith(CI, TryRes);
8690 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8691 if (TryI->use_empty())
8692 EraseInstFromFunction(*TryI);
8696 case Instruction::SExt: {
8697 DoXForm = NumCastsRemoved >= 2;
8698 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8699 // If we do not have to emit the truncate + sext pair, then it's always
8700 // profitable to do this xform.
8702 // It's not safe to eliminate the trunc + sext pair if one of the
8703 // eliminated cast is a truncate. e.g.
8704 // t2 = trunc i32 t1 to i16
8705 // t3 = sext i16 t2 to i32
8708 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8709 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8710 if (NumSignBits > (DestBitSize - SrcBitSize))
8711 return ReplaceInstUsesWith(CI, TryRes);
8713 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8714 if (TryI->use_empty())
8715 EraseInstFromFunction(*TryI);
8722 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8723 " to avoid cast: " << CI);
8724 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8725 CI.getOpcode() == Instruction::SExt);
8727 // Just replace this cast with the result.
8728 return ReplaceInstUsesWith(CI, Res);
8730 assert(Res->getType() == DestTy);
8731 switch (CI.getOpcode()) {
8732 default: llvm_unreachable("Unknown cast type!");
8733 case Instruction::Trunc:
8734 // Just replace this cast with the result.
8735 return ReplaceInstUsesWith(CI, Res);
8736 case Instruction::ZExt: {
8737 assert(SrcBitSize < DestBitSize && "Not a zext?");
8739 // If the high bits are already zero, just replace this cast with the
8741 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8742 if (MaskedValueIsZero(Res, Mask))
8743 return ReplaceInstUsesWith(CI, Res);
8745 // We need to emit an AND to clear the high bits.
8746 Constant *C = ConstantInt::get(*Context,
8747 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8748 return BinaryOperator::CreateAnd(Res, C);
8750 case Instruction::SExt: {
8751 // If the high bits are already filled with sign bit, just replace this
8752 // cast with the result.
8753 unsigned NumSignBits = ComputeNumSignBits(Res);
8754 if (NumSignBits > (DestBitSize - SrcBitSize))
8755 return ReplaceInstUsesWith(CI, Res);
8757 // We need to emit a cast to truncate, then a cast to sext.
8758 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
8764 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8765 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8767 switch (SrcI->getOpcode()) {
8768 case Instruction::Add:
8769 case Instruction::Mul:
8770 case Instruction::And:
8771 case Instruction::Or:
8772 case Instruction::Xor:
8773 // If we are discarding information, rewrite.
8774 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8775 // Don't insert two casts unless at least one can be eliminated.
8776 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8777 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8778 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8779 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8780 return BinaryOperator::Create(
8781 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8785 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8786 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8787 SrcI->getOpcode() == Instruction::Xor &&
8788 Op1 == ConstantInt::getTrue(*Context) &&
8789 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8790 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName());
8791 return BinaryOperator::CreateXor(New,
8792 ConstantInt::get(CI.getType(), 1));
8796 case Instruction::Shl: {
8797 // Canonicalize trunc inside shl, if we can.
8798 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8799 if (CI && DestBitSize < SrcBitSize &&
8800 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8801 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8802 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8803 return BinaryOperator::CreateShl(Op0c, Op1c);
8811 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8812 if (Instruction *Result = commonIntCastTransforms(CI))
8815 Value *Src = CI.getOperand(0);
8816 const Type *Ty = CI.getType();
8817 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8818 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8820 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8821 if (DestBitWidth == 1) {
8822 Constant *One = ConstantInt::get(Src->getType(), 1);
8823 Src = Builder->CreateAnd(Src, One, "tmp");
8824 Value *Zero = Constant::getNullValue(Src->getType());
8825 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8828 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8829 ConstantInt *ShAmtV = 0;
8831 if (Src->hasOneUse() &&
8832 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8833 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8835 // Get a mask for the bits shifting in.
8836 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8837 if (MaskedValueIsZero(ShiftOp, Mask)) {
8838 if (ShAmt >= DestBitWidth) // All zeros.
8839 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8841 // Okay, we can shrink this. Truncate the input, then return a new
8843 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName());
8844 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8845 return BinaryOperator::CreateLShr(V1, V2);
8852 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8853 /// in order to eliminate the icmp.
8854 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8856 // If we are just checking for a icmp eq of a single bit and zext'ing it
8857 // to an integer, then shift the bit to the appropriate place and then
8858 // cast to integer to avoid the comparison.
8859 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8860 const APInt &Op1CV = Op1C->getValue();
8862 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8863 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8864 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8865 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8866 if (!DoXform) return ICI;
8868 Value *In = ICI->getOperand(0);
8869 Value *Sh = ConstantInt::get(In->getType(),
8870 In->getType()->getScalarSizeInBits()-1);
8871 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
8872 if (In->getType() != CI.getType())
8873 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
8875 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8876 Constant *One = ConstantInt::get(In->getType(), 1);
8877 In = Builder->CreateXor(In, One, In->getName()+".not");
8880 return ReplaceInstUsesWith(CI, In);
8885 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8886 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8887 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8888 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8889 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8890 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8891 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8892 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8893 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8894 // This only works for EQ and NE
8895 ICI->isEquality()) {
8896 // If Op1C some other power of two, convert:
8897 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8898 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8899 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8900 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8902 APInt KnownZeroMask(~KnownZero);
8903 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8904 if (!DoXform) return ICI;
8906 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8907 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8908 // (X&4) == 2 --> false
8909 // (X&4) != 2 --> true
8910 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8911 Res = ConstantExpr::getZExt(Res, CI.getType());
8912 return ReplaceInstUsesWith(CI, Res);
8915 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8916 Value *In = ICI->getOperand(0);
8918 // Perform a logical shr by shiftamt.
8919 // Insert the shift to put the result in the low bit.
8920 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
8921 In->getName()+".lobit");
8924 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8925 Constant *One = ConstantInt::get(In->getType(), 1);
8926 In = Builder->CreateXor(In, One, "tmp");
8929 if (CI.getType() == In->getType())
8930 return ReplaceInstUsesWith(CI, In);
8932 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8937 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
8938 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
8939 // may lead to additional simplifications.
8940 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
8941 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
8942 uint32_t BitWidth = ITy->getBitWidth();
8943 Value *LHS = ICI->getOperand(0);
8944 Value *RHS = ICI->getOperand(1);
8946 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
8947 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
8948 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8949 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
8950 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
8952 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
8953 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
8954 APInt UnknownBit = ~KnownBits;
8955 if (UnknownBit.countPopulation() == 1) {
8956 if (!DoXform) return ICI;
8958 Value *Result = Builder->CreateXor(LHS, RHS);
8960 // Mask off any bits that are set and won't be shifted away.
8961 if (KnownOneLHS.uge(UnknownBit))
8962 Result = Builder->CreateAnd(Result,
8963 ConstantInt::get(ITy, UnknownBit));
8965 // Shift the bit we're testing down to the lsb.
8966 Result = Builder->CreateLShr(
8967 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
8969 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8970 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
8971 Result->takeName(ICI);
8972 return ReplaceInstUsesWith(CI, Result);
8981 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8982 // If one of the common conversion will work ..
8983 if (Instruction *Result = commonIntCastTransforms(CI))
8986 Value *Src = CI.getOperand(0);
8988 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8989 // types and if the sizes are just right we can convert this into a logical
8990 // 'and' which will be much cheaper than the pair of casts.
8991 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8992 // Get the sizes of the types involved. We know that the intermediate type
8993 // will be smaller than A or C, but don't know the relation between A and C.
8994 Value *A = CSrc->getOperand(0);
8995 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8996 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8997 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8998 // If we're actually extending zero bits, then if
8999 // SrcSize < DstSize: zext(a & mask)
9000 // SrcSize == DstSize: a & mask
9001 // SrcSize > DstSize: trunc(a) & mask
9002 if (SrcSize < DstSize) {
9003 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
9004 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
9005 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
9006 return new ZExtInst(And, CI.getType());
9009 if (SrcSize == DstSize) {
9010 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
9011 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
9014 if (SrcSize > DstSize) {
9015 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
9016 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
9017 return BinaryOperator::CreateAnd(Trunc,
9018 ConstantInt::get(Trunc->getType(),
9023 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
9024 return transformZExtICmp(ICI, CI);
9026 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
9027 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
9028 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
9029 // of the (zext icmp) will be transformed.
9030 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
9031 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
9032 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
9033 (transformZExtICmp(LHS, CI, false) ||
9034 transformZExtICmp(RHS, CI, false))) {
9035 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
9036 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
9037 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
9041 // zext(trunc(t) & C) -> (t & zext(C)).
9042 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
9043 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
9044 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
9045 Value *TI0 = TI->getOperand(0);
9046 if (TI0->getType() == CI.getType())
9048 BinaryOperator::CreateAnd(TI0,
9049 ConstantExpr::getZExt(C, CI.getType()));
9052 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
9053 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
9054 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
9055 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
9056 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
9057 And->getOperand(1) == C)
9058 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
9059 Value *TI0 = TI->getOperand(0);
9060 if (TI0->getType() == CI.getType()) {
9061 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
9062 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
9063 return BinaryOperator::CreateXor(NewAnd, ZC);
9070 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
9071 if (Instruction *I = commonIntCastTransforms(CI))
9074 Value *Src = CI.getOperand(0);
9076 // Canonicalize sign-extend from i1 to a select.
9077 if (Src->getType() == Type::getInt1Ty(*Context))
9078 return SelectInst::Create(Src,
9079 Constant::getAllOnesValue(CI.getType()),
9080 Constant::getNullValue(CI.getType()));
9082 // See if the value being truncated is already sign extended. If so, just
9083 // eliminate the trunc/sext pair.
9084 if (Operator::getOpcode(Src) == Instruction::Trunc) {
9085 Value *Op = cast<User>(Src)->getOperand(0);
9086 unsigned OpBits = Op->getType()->getScalarSizeInBits();
9087 unsigned MidBits = Src->getType()->getScalarSizeInBits();
9088 unsigned DestBits = CI.getType()->getScalarSizeInBits();
9089 unsigned NumSignBits = ComputeNumSignBits(Op);
9091 if (OpBits == DestBits) {
9092 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
9093 // bits, it is already ready.
9094 if (NumSignBits > DestBits-MidBits)
9095 return ReplaceInstUsesWith(CI, Op);
9096 } else if (OpBits < DestBits) {
9097 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
9098 // bits, just sext from i32.
9099 if (NumSignBits > OpBits-MidBits)
9100 return new SExtInst(Op, CI.getType(), "tmp");
9102 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
9103 // bits, just truncate to i32.
9104 if (NumSignBits > OpBits-MidBits)
9105 return new TruncInst(Op, CI.getType(), "tmp");
9109 // If the input is a shl/ashr pair of a same constant, then this is a sign
9110 // extension from a smaller value. If we could trust arbitrary bitwidth
9111 // integers, we could turn this into a truncate to the smaller bit and then
9112 // use a sext for the whole extension. Since we don't, look deeper and check
9113 // for a truncate. If the source and dest are the same type, eliminate the
9114 // trunc and extend and just do shifts. For example, turn:
9115 // %a = trunc i32 %i to i8
9116 // %b = shl i8 %a, 6
9117 // %c = ashr i8 %b, 6
9118 // %d = sext i8 %c to i32
9120 // %a = shl i32 %i, 30
9121 // %d = ashr i32 %a, 30
9123 ConstantInt *BA = 0, *CA = 0;
9124 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
9125 m_ConstantInt(CA))) &&
9126 BA == CA && isa<TruncInst>(A)) {
9127 Value *I = cast<TruncInst>(A)->getOperand(0);
9128 if (I->getType() == CI.getType()) {
9129 unsigned MidSize = Src->getType()->getScalarSizeInBits();
9130 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
9131 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
9132 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
9133 I = Builder->CreateShl(I, ShAmtV, CI.getName());
9134 return BinaryOperator::CreateAShr(I, ShAmtV);
9141 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
9142 /// in the specified FP type without changing its value.
9143 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
9144 LLVMContext *Context) {
9146 APFloat F = CFP->getValueAPF();
9147 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
9149 return ConstantFP::get(*Context, F);
9153 /// LookThroughFPExtensions - If this is an fp extension instruction, look
9154 /// through it until we get the source value.
9155 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
9156 if (Instruction *I = dyn_cast<Instruction>(V))
9157 if (I->getOpcode() == Instruction::FPExt)
9158 return LookThroughFPExtensions(I->getOperand(0), Context);
9160 // If this value is a constant, return the constant in the smallest FP type
9161 // that can accurately represent it. This allows us to turn
9162 // (float)((double)X+2.0) into x+2.0f.
9163 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
9164 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
9165 return V; // No constant folding of this.
9166 // See if the value can be truncated to float and then reextended.
9167 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
9169 if (CFP->getType() == Type::getDoubleTy(*Context))
9170 return V; // Won't shrink.
9171 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
9173 // Don't try to shrink to various long double types.
9179 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
9180 if (Instruction *I = commonCastTransforms(CI))
9183 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
9184 // smaller than the destination type, we can eliminate the truncate by doing
9185 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
9186 // many builtins (sqrt, etc).
9187 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
9188 if (OpI && OpI->hasOneUse()) {
9189 switch (OpI->getOpcode()) {
9191 case Instruction::FAdd:
9192 case Instruction::FSub:
9193 case Instruction::FMul:
9194 case Instruction::FDiv:
9195 case Instruction::FRem:
9196 const Type *SrcTy = OpI->getType();
9197 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
9198 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
9199 if (LHSTrunc->getType() != SrcTy &&
9200 RHSTrunc->getType() != SrcTy) {
9201 unsigned DstSize = CI.getType()->getScalarSizeInBits();
9202 // If the source types were both smaller than the destination type of
9203 // the cast, do this xform.
9204 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
9205 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
9206 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
9207 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
9208 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
9217 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
9218 return commonCastTransforms(CI);
9221 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
9222 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
9224 return commonCastTransforms(FI);
9226 // fptoui(uitofp(X)) --> X
9227 // fptoui(sitofp(X)) --> X
9228 // This is safe if the intermediate type has enough bits in its mantissa to
9229 // accurately represent all values of X. For example, do not do this with
9230 // i64->float->i64. This is also safe for sitofp case, because any negative
9231 // 'X' value would cause an undefined result for the fptoui.
9232 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
9233 OpI->getOperand(0)->getType() == FI.getType() &&
9234 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
9235 OpI->getType()->getFPMantissaWidth())
9236 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
9238 return commonCastTransforms(FI);
9241 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
9242 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
9244 return commonCastTransforms(FI);
9246 // fptosi(sitofp(X)) --> X
9247 // fptosi(uitofp(X)) --> X
9248 // This is safe if the intermediate type has enough bits in its mantissa to
9249 // accurately represent all values of X. For example, do not do this with
9250 // i64->float->i64. This is also safe for sitofp case, because any negative
9251 // 'X' value would cause an undefined result for the fptoui.
9252 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
9253 OpI->getOperand(0)->getType() == FI.getType() &&
9254 (int)FI.getType()->getScalarSizeInBits() <=
9255 OpI->getType()->getFPMantissaWidth())
9256 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
9258 return commonCastTransforms(FI);
9261 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
9262 return commonCastTransforms(CI);
9265 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
9266 return commonCastTransforms(CI);
9269 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
9270 // If the destination integer type is smaller than the intptr_t type for
9271 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
9272 // trunc to be exposed to other transforms. Don't do this for extending
9273 // ptrtoint's, because we don't know if the target sign or zero extends its
9276 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
9277 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
9278 TD->getIntPtrType(CI.getContext()),
9280 return new TruncInst(P, CI.getType());
9283 return commonPointerCastTransforms(CI);
9286 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
9287 // If the source integer type is larger than the intptr_t type for
9288 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
9289 // allows the trunc to be exposed to other transforms. Don't do this for
9290 // extending inttoptr's, because we don't know if the target sign or zero
9291 // extends to pointers.
9292 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
9293 TD->getPointerSizeInBits()) {
9294 Value *P = Builder->CreateTrunc(CI.getOperand(0),
9295 TD->getIntPtrType(CI.getContext()), "tmp");
9296 return new IntToPtrInst(P, CI.getType());
9299 if (Instruction *I = commonCastTransforms(CI))
9305 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
9306 // If the operands are integer typed then apply the integer transforms,
9307 // otherwise just apply the common ones.
9308 Value *Src = CI.getOperand(0);
9309 const Type *SrcTy = Src->getType();
9310 const Type *DestTy = CI.getType();
9312 if (isa<PointerType>(SrcTy)) {
9313 if (Instruction *I = commonPointerCastTransforms(CI))
9316 if (Instruction *Result = commonCastTransforms(CI))
9321 // Get rid of casts from one type to the same type. These are useless and can
9322 // be replaced by the operand.
9323 if (DestTy == Src->getType())
9324 return ReplaceInstUsesWith(CI, Src);
9326 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
9327 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
9328 const Type *DstElTy = DstPTy->getElementType();
9329 const Type *SrcElTy = SrcPTy->getElementType();
9331 // If the address spaces don't match, don't eliminate the bitcast, which is
9332 // required for changing types.
9333 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
9336 // If we are casting a alloca to a pointer to a type of the same
9337 // size, rewrite the allocation instruction to allocate the "right" type.
9338 // There is no need to modify malloc calls because it is their bitcast that
9339 // needs to be cleaned up.
9340 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
9341 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
9344 // If the source and destination are pointers, and this cast is equivalent
9345 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
9346 // This can enhance SROA and other transforms that want type-safe pointers.
9347 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
9348 unsigned NumZeros = 0;
9349 while (SrcElTy != DstElTy &&
9350 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
9351 SrcElTy->getNumContainedTypes() /* not "{}" */) {
9352 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
9356 // If we found a path from the src to dest, create the getelementptr now.
9357 if (SrcElTy == DstElTy) {
9358 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
9359 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(), "",
9360 ((Instruction*) NULL));
9364 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
9365 if (DestVTy->getNumElements() == 1) {
9366 if (!isa<VectorType>(SrcTy)) {
9367 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
9368 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
9369 Constant::getNullValue(Type::getInt32Ty(*Context)));
9371 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
9375 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
9376 if (SrcVTy->getNumElements() == 1) {
9377 if (!isa<VectorType>(DestTy)) {
9379 Builder->CreateExtractElement(Src,
9380 Constant::getNullValue(Type::getInt32Ty(*Context)));
9381 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
9386 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9387 if (SVI->hasOneUse()) {
9388 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9389 // a bitconvert to a vector with the same # elts.
9390 if (isa<VectorType>(DestTy) &&
9391 cast<VectorType>(DestTy)->getNumElements() ==
9392 SVI->getType()->getNumElements() &&
9393 SVI->getType()->getNumElements() ==
9394 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9396 // If either of the operands is a cast from CI.getType(), then
9397 // evaluating the shuffle in the casted destination's type will allow
9398 // us to eliminate at least one cast.
9399 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9400 Tmp->getOperand(0)->getType() == DestTy) ||
9401 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9402 Tmp->getOperand(0)->getType() == DestTy)) {
9403 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
9404 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
9405 // Return a new shuffle vector. Use the same element ID's, as we
9406 // know the vector types match #elts.
9407 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9415 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9417 /// %D = select %cond, %C, %A
9419 /// %C = select %cond, %B, 0
9422 /// Assuming that the specified instruction is an operand to the select, return
9423 /// a bitmask indicating which operands of this instruction are foldable if they
9424 /// equal the other incoming value of the select.
9426 static unsigned GetSelectFoldableOperands(Instruction *I) {
9427 switch (I->getOpcode()) {
9428 case Instruction::Add:
9429 case Instruction::Mul:
9430 case Instruction::And:
9431 case Instruction::Or:
9432 case Instruction::Xor:
9433 return 3; // Can fold through either operand.
9434 case Instruction::Sub: // Can only fold on the amount subtracted.
9435 case Instruction::Shl: // Can only fold on the shift amount.
9436 case Instruction::LShr:
9437 case Instruction::AShr:
9440 return 0; // Cannot fold
9444 /// GetSelectFoldableConstant - For the same transformation as the previous
9445 /// function, return the identity constant that goes into the select.
9446 static Constant *GetSelectFoldableConstant(Instruction *I,
9447 LLVMContext *Context) {
9448 switch (I->getOpcode()) {
9449 default: llvm_unreachable("This cannot happen!");
9450 case Instruction::Add:
9451 case Instruction::Sub:
9452 case Instruction::Or:
9453 case Instruction::Xor:
9454 case Instruction::Shl:
9455 case Instruction::LShr:
9456 case Instruction::AShr:
9457 return Constant::getNullValue(I->getType());
9458 case Instruction::And:
9459 return Constant::getAllOnesValue(I->getType());
9460 case Instruction::Mul:
9461 return ConstantInt::get(I->getType(), 1);
9465 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9466 /// have the same opcode and only one use each. Try to simplify this.
9467 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9469 if (TI->getNumOperands() == 1) {
9470 // If this is a non-volatile load or a cast from the same type,
9473 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9476 return 0; // unknown unary op.
9479 // Fold this by inserting a select from the input values.
9480 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9481 FI->getOperand(0), SI.getName()+".v");
9482 InsertNewInstBefore(NewSI, SI);
9483 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9487 // Only handle binary operators here.
9488 if (!isa<BinaryOperator>(TI))
9491 // Figure out if the operations have any operands in common.
9492 Value *MatchOp, *OtherOpT, *OtherOpF;
9494 if (TI->getOperand(0) == FI->getOperand(0)) {
9495 MatchOp = TI->getOperand(0);
9496 OtherOpT = TI->getOperand(1);
9497 OtherOpF = FI->getOperand(1);
9498 MatchIsOpZero = true;
9499 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9500 MatchOp = TI->getOperand(1);
9501 OtherOpT = TI->getOperand(0);
9502 OtherOpF = FI->getOperand(0);
9503 MatchIsOpZero = false;
9504 } else if (!TI->isCommutative()) {
9506 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9507 MatchOp = TI->getOperand(0);
9508 OtherOpT = TI->getOperand(1);
9509 OtherOpF = FI->getOperand(0);
9510 MatchIsOpZero = true;
9511 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9512 MatchOp = TI->getOperand(1);
9513 OtherOpT = TI->getOperand(0);
9514 OtherOpF = FI->getOperand(1);
9515 MatchIsOpZero = true;
9520 // If we reach here, they do have operations in common.
9521 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9522 OtherOpF, SI.getName()+".v");
9523 InsertNewInstBefore(NewSI, SI);
9525 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9527 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9529 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9531 llvm_unreachable("Shouldn't get here");
9535 static bool isSelect01(Constant *C1, Constant *C2) {
9536 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9539 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9542 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9545 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9546 /// facilitate further optimization.
9547 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9549 // See the comment above GetSelectFoldableOperands for a description of the
9550 // transformation we are doing here.
9551 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9552 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9553 !isa<Constant>(FalseVal)) {
9554 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9555 unsigned OpToFold = 0;
9556 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9558 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9563 Constant *C = GetSelectFoldableConstant(TVI, Context);
9564 Value *OOp = TVI->getOperand(2-OpToFold);
9565 // Avoid creating select between 2 constants unless it's selecting
9567 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9568 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9569 InsertNewInstBefore(NewSel, SI);
9570 NewSel->takeName(TVI);
9571 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9572 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9573 llvm_unreachable("Unknown instruction!!");
9580 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9581 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9582 !isa<Constant>(TrueVal)) {
9583 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9584 unsigned OpToFold = 0;
9585 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9587 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9592 Constant *C = GetSelectFoldableConstant(FVI, Context);
9593 Value *OOp = FVI->getOperand(2-OpToFold);
9594 // Avoid creating select between 2 constants unless it's selecting
9596 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9597 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9598 InsertNewInstBefore(NewSel, SI);
9599 NewSel->takeName(FVI);
9600 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9601 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9602 llvm_unreachable("Unknown instruction!!");
9612 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9613 /// ICmpInst as its first operand.
9615 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9617 bool Changed = false;
9618 ICmpInst::Predicate Pred = ICI->getPredicate();
9619 Value *CmpLHS = ICI->getOperand(0);
9620 Value *CmpRHS = ICI->getOperand(1);
9621 Value *TrueVal = SI.getTrueValue();
9622 Value *FalseVal = SI.getFalseValue();
9624 // Check cases where the comparison is with a constant that
9625 // can be adjusted to fit the min/max idiom. We may edit ICI in
9626 // place here, so make sure the select is the only user.
9627 if (ICI->hasOneUse())
9628 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9631 case ICmpInst::ICMP_ULT:
9632 case ICmpInst::ICMP_SLT: {
9633 // X < MIN ? T : F --> F
9634 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9635 return ReplaceInstUsesWith(SI, FalseVal);
9636 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9637 Constant *AdjustedRHS = SubOne(CI);
9638 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9639 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9640 Pred = ICmpInst::getSwappedPredicate(Pred);
9641 CmpRHS = AdjustedRHS;
9642 std::swap(FalseVal, TrueVal);
9643 ICI->setPredicate(Pred);
9644 ICI->setOperand(1, CmpRHS);
9645 SI.setOperand(1, TrueVal);
9646 SI.setOperand(2, FalseVal);
9651 case ICmpInst::ICMP_UGT:
9652 case ICmpInst::ICMP_SGT: {
9653 // X > MAX ? T : F --> F
9654 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9655 return ReplaceInstUsesWith(SI, FalseVal);
9656 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9657 Constant *AdjustedRHS = AddOne(CI);
9658 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9659 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9660 Pred = ICmpInst::getSwappedPredicate(Pred);
9661 CmpRHS = AdjustedRHS;
9662 std::swap(FalseVal, TrueVal);
9663 ICI->setPredicate(Pred);
9664 ICI->setOperand(1, CmpRHS);
9665 SI.setOperand(1, TrueVal);
9666 SI.setOperand(2, FalseVal);
9673 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9674 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9675 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9676 if (match(TrueVal, m_ConstantInt<-1>()) &&
9677 match(FalseVal, m_ConstantInt<0>()))
9678 Pred = ICI->getPredicate();
9679 else if (match(TrueVal, m_ConstantInt<0>()) &&
9680 match(FalseVal, m_ConstantInt<-1>()))
9681 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9683 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9684 // If we are just checking for a icmp eq of a single bit and zext'ing it
9685 // to an integer, then shift the bit to the appropriate place and then
9686 // cast to integer to avoid the comparison.
9687 const APInt &Op1CV = CI->getValue();
9689 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9690 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9691 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9692 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9693 Value *In = ICI->getOperand(0);
9694 Value *Sh = ConstantInt::get(In->getType(),
9695 In->getType()->getScalarSizeInBits()-1);
9696 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9697 In->getName()+".lobit"),
9699 if (In->getType() != SI.getType())
9700 In = CastInst::CreateIntegerCast(In, SI.getType(),
9701 true/*SExt*/, "tmp", ICI);
9703 if (Pred == ICmpInst::ICMP_SGT)
9704 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9705 In->getName()+".not"), *ICI);
9707 return ReplaceInstUsesWith(SI, In);
9712 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9713 // Transform (X == Y) ? X : Y -> Y
9714 if (Pred == ICmpInst::ICMP_EQ)
9715 return ReplaceInstUsesWith(SI, FalseVal);
9716 // Transform (X != Y) ? X : Y -> X
9717 if (Pred == ICmpInst::ICMP_NE)
9718 return ReplaceInstUsesWith(SI, TrueVal);
9719 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9721 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9722 // Transform (X == Y) ? Y : X -> X
9723 if (Pred == ICmpInst::ICMP_EQ)
9724 return ReplaceInstUsesWith(SI, FalseVal);
9725 // Transform (X != Y) ? Y : X -> Y
9726 if (Pred == ICmpInst::ICMP_NE)
9727 return ReplaceInstUsesWith(SI, TrueVal);
9728 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9730 return Changed ? &SI : 0;
9734 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a
9735 /// PHI node (but the two may be in different blocks). See if the true/false
9736 /// values (V) are live in all of the predecessor blocks of the PHI. For
9737 /// example, cases like this cannot be mapped:
9739 /// X = phi [ C1, BB1], [C2, BB2]
9741 /// Z = select X, Y, 0
9743 /// because Y is not live in BB1/BB2.
9745 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V,
9746 const SelectInst &SI) {
9747 // If the value is a non-instruction value like a constant or argument, it
9748 // can always be mapped.
9749 const Instruction *I = dyn_cast<Instruction>(V);
9750 if (I == 0) return true;
9752 // If V is a PHI node defined in the same block as the condition PHI, we can
9753 // map the arguments.
9754 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
9756 if (const PHINode *VP = dyn_cast<PHINode>(I))
9757 if (VP->getParent() == CondPHI->getParent())
9760 // Otherwise, if the PHI and select are defined in the same block and if V is
9761 // defined in a different block, then we can transform it.
9762 if (SI.getParent() == CondPHI->getParent() &&
9763 I->getParent() != CondPHI->getParent())
9766 // Otherwise we have a 'hard' case and we can't tell without doing more
9767 // detailed dominator based analysis, punt.
9771 /// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form:
9772 /// SPF2(SPF1(A, B), C)
9773 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner,
9774 SelectPatternFlavor SPF1,
9777 SelectPatternFlavor SPF2, Value *C) {
9778 if (C == A || C == B) {
9779 // MAX(MAX(A, B), B) -> MAX(A, B)
9780 // MIN(MIN(a, b), a) -> MIN(a, b)
9782 return ReplaceInstUsesWith(Outer, Inner);
9784 // MAX(MIN(a, b), a) -> a
9785 // MIN(MAX(a, b), a) -> a
9786 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
9787 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
9788 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
9789 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
9790 return ReplaceInstUsesWith(Outer, C);
9793 // TODO: MIN(MIN(A, 23), 97)
9800 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9801 Value *CondVal = SI.getCondition();
9802 Value *TrueVal = SI.getTrueValue();
9803 Value *FalseVal = SI.getFalseValue();
9805 // select true, X, Y -> X
9806 // select false, X, Y -> Y
9807 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9808 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9810 // select C, X, X -> X
9811 if (TrueVal == FalseVal)
9812 return ReplaceInstUsesWith(SI, TrueVal);
9814 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9815 return ReplaceInstUsesWith(SI, FalseVal);
9816 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9817 return ReplaceInstUsesWith(SI, TrueVal);
9818 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9819 if (isa<Constant>(TrueVal))
9820 return ReplaceInstUsesWith(SI, TrueVal);
9822 return ReplaceInstUsesWith(SI, FalseVal);
9825 if (SI.getType() == Type::getInt1Ty(*Context)) {
9826 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9827 if (C->getZExtValue()) {
9828 // Change: A = select B, true, C --> A = or B, C
9829 return BinaryOperator::CreateOr(CondVal, FalseVal);
9831 // Change: A = select B, false, C --> A = and !B, C
9833 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9834 "not."+CondVal->getName()), SI);
9835 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9837 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9838 if (C->getZExtValue() == false) {
9839 // Change: A = select B, C, false --> A = and B, C
9840 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9842 // Change: A = select B, C, true --> A = or !B, C
9844 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9845 "not."+CondVal->getName()), SI);
9846 return BinaryOperator::CreateOr(NotCond, TrueVal);
9850 // select a, b, a -> a&b
9851 // select a, a, b -> a|b
9852 if (CondVal == TrueVal)
9853 return BinaryOperator::CreateOr(CondVal, FalseVal);
9854 else if (CondVal == FalseVal)
9855 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9858 // Selecting between two integer constants?
9859 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9860 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9861 // select C, 1, 0 -> zext C to int
9862 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9863 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9864 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9865 // select C, 0, 1 -> zext !C to int
9867 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9868 "not."+CondVal->getName()), SI);
9869 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9872 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9873 // If one of the constants is zero (we know they can't both be) and we
9874 // have an icmp instruction with zero, and we have an 'and' with the
9875 // non-constant value, eliminate this whole mess. This corresponds to
9876 // cases like this: ((X & 27) ? 27 : 0)
9877 if (TrueValC->isZero() || FalseValC->isZero())
9878 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9879 cast<Constant>(IC->getOperand(1))->isNullValue())
9880 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9881 if (ICA->getOpcode() == Instruction::And &&
9882 isa<ConstantInt>(ICA->getOperand(1)) &&
9883 (ICA->getOperand(1) == TrueValC ||
9884 ICA->getOperand(1) == FalseValC) &&
9885 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9886 // Okay, now we know that everything is set up, we just don't
9887 // know whether we have a icmp_ne or icmp_eq and whether the
9888 // true or false val is the zero.
9889 bool ShouldNotVal = !TrueValC->isZero();
9890 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9893 V = InsertNewInstBefore(BinaryOperator::Create(
9894 Instruction::Xor, V, ICA->getOperand(1)), SI);
9895 return ReplaceInstUsesWith(SI, V);
9900 // See if we are selecting two values based on a comparison of the two values.
9901 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9902 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9903 // Transform (X == Y) ? X : Y -> Y
9904 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9905 // This is not safe in general for floating point:
9906 // consider X== -0, Y== +0.
9907 // It becomes safe if either operand is a nonzero constant.
9908 ConstantFP *CFPt, *CFPf;
9909 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9910 !CFPt->getValueAPF().isZero()) ||
9911 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9912 !CFPf->getValueAPF().isZero()))
9913 return ReplaceInstUsesWith(SI, FalseVal);
9915 // Transform (X != Y) ? X : Y -> X
9916 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9917 return ReplaceInstUsesWith(SI, TrueVal);
9918 // NOTE: if we wanted to, this is where to detect MIN/MAX
9920 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9921 // Transform (X == Y) ? Y : X -> X
9922 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9923 // This is not safe in general for floating point:
9924 // consider X== -0, Y== +0.
9925 // It becomes safe if either operand is a nonzero constant.
9926 ConstantFP *CFPt, *CFPf;
9927 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9928 !CFPt->getValueAPF().isZero()) ||
9929 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9930 !CFPf->getValueAPF().isZero()))
9931 return ReplaceInstUsesWith(SI, FalseVal);
9933 // Transform (X != Y) ? Y : X -> Y
9934 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9935 return ReplaceInstUsesWith(SI, TrueVal);
9936 // NOTE: if we wanted to, this is where to detect MIN/MAX
9938 // NOTE: if we wanted to, this is where to detect ABS
9941 // See if we are selecting two values based on a comparison of the two values.
9942 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9943 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9946 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9947 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9948 if (TI->hasOneUse() && FI->hasOneUse()) {
9949 Instruction *AddOp = 0, *SubOp = 0;
9951 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9952 if (TI->getOpcode() == FI->getOpcode())
9953 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9956 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9957 // even legal for FP.
9958 if ((TI->getOpcode() == Instruction::Sub &&
9959 FI->getOpcode() == Instruction::Add) ||
9960 (TI->getOpcode() == Instruction::FSub &&
9961 FI->getOpcode() == Instruction::FAdd)) {
9962 AddOp = FI; SubOp = TI;
9963 } else if ((FI->getOpcode() == Instruction::Sub &&
9964 TI->getOpcode() == Instruction::Add) ||
9965 (FI->getOpcode() == Instruction::FSub &&
9966 TI->getOpcode() == Instruction::FAdd)) {
9967 AddOp = TI; SubOp = FI;
9971 Value *OtherAddOp = 0;
9972 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9973 OtherAddOp = AddOp->getOperand(1);
9974 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9975 OtherAddOp = AddOp->getOperand(0);
9979 // So at this point we know we have (Y -> OtherAddOp):
9980 // select C, (add X, Y), (sub X, Z)
9981 Value *NegVal; // Compute -Z
9982 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9983 NegVal = ConstantExpr::getNeg(C);
9985 NegVal = InsertNewInstBefore(
9986 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9990 Value *NewTrueOp = OtherAddOp;
9991 Value *NewFalseOp = NegVal;
9993 std::swap(NewTrueOp, NewFalseOp);
9994 Instruction *NewSel =
9995 SelectInst::Create(CondVal, NewTrueOp,
9996 NewFalseOp, SI.getName() + ".p");
9998 NewSel = InsertNewInstBefore(NewSel, SI);
9999 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
10004 // See if we can fold the select into one of our operands.
10005 if (SI.getType()->isInteger()) {
10006 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal))
10009 // MAX(MAX(a, b), a) -> MAX(a, b)
10010 // MIN(MIN(a, b), a) -> MIN(a, b)
10011 // MAX(MIN(a, b), a) -> a
10012 // MIN(MAX(a, b), a) -> a
10013 Value *LHS, *RHS, *LHS2, *RHS2;
10014 if (SelectPatternFlavor SPF = MatchSelectPattern(&SI, LHS, RHS)) {
10015 if (SelectPatternFlavor SPF2 = MatchSelectPattern(LHS, LHS2, RHS2))
10016 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2,
10019 if (SelectPatternFlavor SPF2 = MatchSelectPattern(RHS, LHS2, RHS2))
10020 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2,
10026 // ABS(-X) -> ABS(X)
10027 // ABS(ABS(X)) -> ABS(X)
10030 // See if we can fold the select into a phi node if the condition is a select.
10031 if (isa<PHINode>(SI.getCondition()))
10032 // The true/false values have to be live in the PHI predecessor's blocks.
10033 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
10034 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
10035 if (Instruction *NV = FoldOpIntoPhi(SI))
10038 if (BinaryOperator::isNot(CondVal)) {
10039 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
10040 SI.setOperand(1, FalseVal);
10041 SI.setOperand(2, TrueVal);
10048 /// EnforceKnownAlignment - If the specified pointer points to an object that
10049 /// we control, modify the object's alignment to PrefAlign. This isn't
10050 /// often possible though. If alignment is important, a more reliable approach
10051 /// is to simply align all global variables and allocation instructions to
10052 /// their preferred alignment from the beginning.
10054 static unsigned EnforceKnownAlignment(Value *V,
10055 unsigned Align, unsigned PrefAlign) {
10057 User *U = dyn_cast<User>(V);
10058 if (!U) return Align;
10060 switch (Operator::getOpcode(U)) {
10062 case Instruction::BitCast:
10063 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
10064 case Instruction::GetElementPtr: {
10065 // If all indexes are zero, it is just the alignment of the base pointer.
10066 bool AllZeroOperands = true;
10067 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
10068 if (!isa<Constant>(*i) ||
10069 !cast<Constant>(*i)->isNullValue()) {
10070 AllZeroOperands = false;
10074 if (AllZeroOperands) {
10075 // Treat this like a bitcast.
10076 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
10082 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
10083 // If there is a large requested alignment and we can, bump up the alignment
10085 if (!GV->isDeclaration()) {
10086 if (GV->getAlignment() >= PrefAlign)
10087 Align = GV->getAlignment();
10089 GV->setAlignment(PrefAlign);
10093 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
10094 // If there is a requested alignment and if this is an alloca, round up.
10095 if (AI->getAlignment() >= PrefAlign)
10096 Align = AI->getAlignment();
10098 AI->setAlignment(PrefAlign);
10106 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
10107 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
10108 /// and it is more than the alignment of the ultimate object, see if we can
10109 /// increase the alignment of the ultimate object, making this check succeed.
10110 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
10111 unsigned PrefAlign) {
10112 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
10113 sizeof(PrefAlign) * CHAR_BIT;
10114 APInt Mask = APInt::getAllOnesValue(BitWidth);
10115 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
10116 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
10117 unsigned TrailZ = KnownZero.countTrailingOnes();
10118 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
10120 if (PrefAlign > Align)
10121 Align = EnforceKnownAlignment(V, Align, PrefAlign);
10123 // We don't need to make any adjustment.
10127 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
10128 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
10129 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
10130 unsigned MinAlign = std::min(DstAlign, SrcAlign);
10131 unsigned CopyAlign = MI->getAlignment();
10133 if (CopyAlign < MinAlign) {
10134 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
10139 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
10141 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
10142 if (MemOpLength == 0) return 0;
10144 // Source and destination pointer types are always "i8*" for intrinsic. See
10145 // if the size is something we can handle with a single primitive load/store.
10146 // A single load+store correctly handles overlapping memory in the memmove
10148 unsigned Size = MemOpLength->getZExtValue();
10149 if (Size == 0) return MI; // Delete this mem transfer.
10151 if (Size > 8 || (Size&(Size-1)))
10152 return 0; // If not 1/2/4/8 bytes, exit.
10154 // Use an integer load+store unless we can find something better.
10156 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
10158 // Memcpy forces the use of i8* for the source and destination. That means
10159 // that if you're using memcpy to move one double around, you'll get a cast
10160 // from double* to i8*. We'd much rather use a double load+store rather than
10161 // an i64 load+store, here because this improves the odds that the source or
10162 // dest address will be promotable. See if we can find a better type than the
10163 // integer datatype.
10164 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
10165 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
10166 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
10167 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
10168 // down through these levels if so.
10169 while (!SrcETy->isSingleValueType()) {
10170 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
10171 if (STy->getNumElements() == 1)
10172 SrcETy = STy->getElementType(0);
10175 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
10176 if (ATy->getNumElements() == 1)
10177 SrcETy = ATy->getElementType();
10184 if (SrcETy->isSingleValueType())
10185 NewPtrTy = PointerType::getUnqual(SrcETy);
10190 // If the memcpy/memmove provides better alignment info than we can
10192 SrcAlign = std::max(SrcAlign, CopyAlign);
10193 DstAlign = std::max(DstAlign, CopyAlign);
10195 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy);
10196 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy);
10197 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
10198 InsertNewInstBefore(L, *MI);
10199 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
10201 // Set the size of the copy to 0, it will be deleted on the next iteration.
10202 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
10206 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
10207 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
10208 if (MI->getAlignment() < Alignment) {
10209 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
10210 Alignment, false));
10214 // Extract the length and alignment and fill if they are constant.
10215 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
10216 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
10217 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
10219 uint64_t Len = LenC->getZExtValue();
10220 Alignment = MI->getAlignment();
10222 // If the length is zero, this is a no-op
10223 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
10225 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
10226 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
10227 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
10229 Value *Dest = MI->getDest();
10230 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
10232 // Alignment 0 is identity for alignment 1 for memset, but not store.
10233 if (Alignment == 0) Alignment = 1;
10235 // Extract the fill value and store.
10236 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
10237 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
10238 Dest, false, Alignment), *MI);
10240 // Set the size of the copy to 0, it will be deleted on the next iteration.
10241 MI->setLength(Constant::getNullValue(LenC->getType()));
10249 /// visitCallInst - CallInst simplification. This mostly only handles folding
10250 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
10251 /// the heavy lifting.
10253 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
10254 if (isFreeCall(&CI))
10255 return visitFree(CI);
10257 // If the caller function is nounwind, mark the call as nounwind, even if the
10259 if (CI.getParent()->getParent()->doesNotThrow() &&
10260 !CI.doesNotThrow()) {
10261 CI.setDoesNotThrow();
10265 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
10266 if (!II) return visitCallSite(&CI);
10268 // Intrinsics cannot occur in an invoke, so handle them here instead of in
10270 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
10271 bool Changed = false;
10273 // memmove/cpy/set of zero bytes is a noop.
10274 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
10275 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
10277 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
10278 if (CI->getZExtValue() == 1) {
10279 // Replace the instruction with just byte operations. We would
10280 // transform other cases to loads/stores, but we don't know if
10281 // alignment is sufficient.
10285 // If we have a memmove and the source operation is a constant global,
10286 // then the source and dest pointers can't alias, so we can change this
10287 // into a call to memcpy.
10288 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
10289 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
10290 if (GVSrc->isConstant()) {
10291 Module *M = CI.getParent()->getParent()->getParent();
10292 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
10293 const Type *Tys[1];
10294 Tys[0] = CI.getOperand(3)->getType();
10296 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
10301 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
10302 // memmove(x,x,size) -> noop.
10303 if (MTI->getSource() == MTI->getDest())
10304 return EraseInstFromFunction(CI);
10307 // If we can determine a pointer alignment that is bigger than currently
10308 // set, update the alignment.
10309 if (isa<MemTransferInst>(MI)) {
10310 if (Instruction *I = SimplifyMemTransfer(MI))
10312 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
10313 if (Instruction *I = SimplifyMemSet(MSI))
10317 if (Changed) return II;
10320 switch (II->getIntrinsicID()) {
10322 case Intrinsic::bswap:
10323 // bswap(bswap(x)) -> x
10324 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
10325 if (Operand->getIntrinsicID() == Intrinsic::bswap)
10326 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
10328 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
10329 if (TruncInst *TI = dyn_cast<TruncInst>(II->getOperand(1))) {
10330 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
10331 if (Operand->getIntrinsicID() == Intrinsic::bswap) {
10332 unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
10333 TI->getType()->getPrimitiveSizeInBits();
10334 Value *CV = ConstantInt::get(Operand->getType(), C);
10335 Value *V = Builder->CreateLShr(Operand->getOperand(1), CV);
10336 return new TruncInst(V, TI->getType());
10341 case Intrinsic::powi:
10342 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getOperand(2))) {
10343 // powi(x, 0) -> 1.0
10344 if (Power->isZero())
10345 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
10347 if (Power->isOne())
10348 return ReplaceInstUsesWith(CI, II->getOperand(1));
10349 // powi(x, -1) -> 1/x
10350 if (Power->isAllOnesValue())
10351 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
10352 II->getOperand(1));
10356 case Intrinsic::uadd_with_overflow: {
10357 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
10358 const IntegerType *IT = cast<IntegerType>(II->getOperand(1)->getType());
10359 uint32_t BitWidth = IT->getBitWidth();
10360 APInt Mask = APInt::getSignBit(BitWidth);
10361 APInt LHSKnownZero(BitWidth, 0);
10362 APInt LHSKnownOne(BitWidth, 0);
10363 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
10364 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
10365 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
10367 if (LHSKnownNegative || LHSKnownPositive) {
10368 APInt RHSKnownZero(BitWidth, 0);
10369 APInt RHSKnownOne(BitWidth, 0);
10370 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
10371 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
10372 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
10373 if (LHSKnownNegative && RHSKnownNegative) {
10374 // The sign bit is set in both cases: this MUST overflow.
10375 // Create a simple add instruction, and insert it into the struct.
10376 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
10379 UndefValue::get(LHS->getType()), ConstantInt::getTrue(*Context)
10381 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10382 return InsertValueInst::Create(Struct, Add, 0);
10385 if (LHSKnownPositive && RHSKnownPositive) {
10386 // The sign bit is clear in both cases: this CANNOT overflow.
10387 // Create a simple add instruction, and insert it into the struct.
10388 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
10391 UndefValue::get(LHS->getType()), ConstantInt::getFalse(*Context)
10393 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10394 return InsertValueInst::Create(Struct, Add, 0);
10398 // FALL THROUGH uadd into sadd
10399 case Intrinsic::sadd_with_overflow:
10400 // Canonicalize constants into the RHS.
10401 if (isa<Constant>(II->getOperand(1)) &&
10402 !isa<Constant>(II->getOperand(2))) {
10403 Value *LHS = II->getOperand(1);
10404 II->setOperand(1, II->getOperand(2));
10405 II->setOperand(2, LHS);
10409 // X + undef -> undef
10410 if (isa<UndefValue>(II->getOperand(2)))
10411 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
10413 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) {
10414 // X + 0 -> {X, false}
10415 if (RHS->isZero()) {
10417 UndefValue::get(II->getOperand(0)->getType()),
10418 ConstantInt::getFalse(*Context)
10420 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10421 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
10425 case Intrinsic::usub_with_overflow:
10426 case Intrinsic::ssub_with_overflow:
10427 // undef - X -> undef
10428 // X - undef -> undef
10429 if (isa<UndefValue>(II->getOperand(1)) ||
10430 isa<UndefValue>(II->getOperand(2)))
10431 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
10433 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) {
10434 // X - 0 -> {X, false}
10435 if (RHS->isZero()) {
10437 UndefValue::get(II->getOperand(1)->getType()),
10438 ConstantInt::getFalse(*Context)
10440 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10441 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
10445 case Intrinsic::umul_with_overflow:
10446 case Intrinsic::smul_with_overflow:
10447 // Canonicalize constants into the RHS.
10448 if (isa<Constant>(II->getOperand(1)) &&
10449 !isa<Constant>(II->getOperand(2))) {
10450 Value *LHS = II->getOperand(1);
10451 II->setOperand(1, II->getOperand(2));
10452 II->setOperand(2, LHS);
10456 // X * undef -> undef
10457 if (isa<UndefValue>(II->getOperand(2)))
10458 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
10460 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getOperand(2))) {
10461 // X*0 -> {0, false}
10462 if (RHSI->isZero())
10463 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
10465 // X * 1 -> {X, false}
10466 if (RHSI->equalsInt(1)) {
10468 UndefValue::get(II->getOperand(1)->getType()),
10469 ConstantInt::getFalse(*Context)
10471 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10472 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
10476 case Intrinsic::ppc_altivec_lvx:
10477 case Intrinsic::ppc_altivec_lvxl:
10478 case Intrinsic::x86_sse_loadu_ps:
10479 case Intrinsic::x86_sse2_loadu_pd:
10480 case Intrinsic::x86_sse2_loadu_dq:
10481 // Turn PPC lvx -> load if the pointer is known aligned.
10482 // Turn X86 loadups -> load if the pointer is known aligned.
10483 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
10484 Value *Ptr = Builder->CreateBitCast(II->getOperand(1),
10485 PointerType::getUnqual(II->getType()));
10486 return new LoadInst(Ptr);
10489 case Intrinsic::ppc_altivec_stvx:
10490 case Intrinsic::ppc_altivec_stvxl:
10491 // Turn stvx -> store if the pointer is known aligned.
10492 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
10493 const Type *OpPtrTy =
10494 PointerType::getUnqual(II->getOperand(1)->getType());
10495 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy);
10496 return new StoreInst(II->getOperand(1), Ptr);
10499 case Intrinsic::x86_sse_storeu_ps:
10500 case Intrinsic::x86_sse2_storeu_pd:
10501 case Intrinsic::x86_sse2_storeu_dq:
10502 // Turn X86 storeu -> store if the pointer is known aligned.
10503 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
10504 const Type *OpPtrTy =
10505 PointerType::getUnqual(II->getOperand(2)->getType());
10506 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
10507 return new StoreInst(II->getOperand(2), Ptr);
10511 case Intrinsic::x86_sse_cvttss2si: {
10512 // These intrinsics only demands the 0th element of its input vector. If
10513 // we can simplify the input based on that, do so now.
10515 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
10516 APInt DemandedElts(VWidth, 1);
10517 APInt UndefElts(VWidth, 0);
10518 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
10520 II->setOperand(1, V);
10526 case Intrinsic::ppc_altivec_vperm:
10527 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
10528 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
10529 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
10531 // Check that all of the elements are integer constants or undefs.
10532 bool AllEltsOk = true;
10533 for (unsigned i = 0; i != 16; ++i) {
10534 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
10535 !isa<UndefValue>(Mask->getOperand(i))) {
10542 // Cast the input vectors to byte vectors.
10543 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
10544 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType());
10545 Value *Result = UndefValue::get(Op0->getType());
10547 // Only extract each element once.
10548 Value *ExtractedElts[32];
10549 memset(ExtractedElts, 0, sizeof(ExtractedElts));
10551 for (unsigned i = 0; i != 16; ++i) {
10552 if (isa<UndefValue>(Mask->getOperand(i)))
10554 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
10555 Idx &= 31; // Match the hardware behavior.
10557 if (ExtractedElts[Idx] == 0) {
10558 ExtractedElts[Idx] =
10559 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
10560 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false),
10564 // Insert this value into the result vector.
10565 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
10566 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
10569 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
10574 case Intrinsic::stackrestore: {
10575 // If the save is right next to the restore, remove the restore. This can
10576 // happen when variable allocas are DCE'd.
10577 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
10578 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
10579 BasicBlock::iterator BI = SS;
10581 return EraseInstFromFunction(CI);
10585 // Scan down this block to see if there is another stack restore in the
10586 // same block without an intervening call/alloca.
10587 BasicBlock::iterator BI = II;
10588 TerminatorInst *TI = II->getParent()->getTerminator();
10589 bool CannotRemove = false;
10590 for (++BI; &*BI != TI; ++BI) {
10591 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
10592 CannotRemove = true;
10595 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
10596 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
10597 // If there is a stackrestore below this one, remove this one.
10598 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10599 return EraseInstFromFunction(CI);
10600 // Otherwise, ignore the intrinsic.
10602 // If we found a non-intrinsic call, we can't remove the stack
10604 CannotRemove = true;
10610 // If the stack restore is in a return/unwind block and if there are no
10611 // allocas or calls between the restore and the return, nuke the restore.
10612 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10613 return EraseInstFromFunction(CI);
10618 return visitCallSite(II);
10621 // InvokeInst simplification
10623 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10624 return visitCallSite(&II);
10627 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10628 /// passed through the varargs area, we can eliminate the use of the cast.
10629 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10630 const CastInst * const CI,
10631 const TargetData * const TD,
10633 if (!CI->isLosslessCast())
10636 // The size of ByVal arguments is derived from the type, so we
10637 // can't change to a type with a different size. If the size were
10638 // passed explicitly we could avoid this check.
10639 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10642 const Type* SrcTy =
10643 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10644 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10645 if (!SrcTy->isSized() || !DstTy->isSized())
10647 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10652 // visitCallSite - Improvements for call and invoke instructions.
10654 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10655 bool Changed = false;
10657 // If the callee is a constexpr cast of a function, attempt to move the cast
10658 // to the arguments of the call/invoke.
10659 if (transformConstExprCastCall(CS)) return 0;
10661 Value *Callee = CS.getCalledValue();
10663 if (Function *CalleeF = dyn_cast<Function>(Callee))
10664 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10665 Instruction *OldCall = CS.getInstruction();
10666 // If the call and callee calling conventions don't match, this call must
10667 // be unreachable, as the call is undefined.
10668 new StoreInst(ConstantInt::getTrue(*Context),
10669 UndefValue::get(Type::getInt1PtrTy(*Context)),
10671 // If OldCall dues not return void then replaceAllUsesWith undef.
10672 // This allows ValueHandlers and custom metadata to adjust itself.
10673 if (!OldCall->getType()->isVoidTy())
10674 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10675 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10676 return EraseInstFromFunction(*OldCall);
10680 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10681 // This instruction is not reachable, just remove it. We insert a store to
10682 // undef so that we know that this code is not reachable, despite the fact
10683 // that we can't modify the CFG here.
10684 new StoreInst(ConstantInt::getTrue(*Context),
10685 UndefValue::get(Type::getInt1PtrTy(*Context)),
10686 CS.getInstruction());
10688 // If CS dues not return void then replaceAllUsesWith undef.
10689 // This allows ValueHandlers and custom metadata to adjust itself.
10690 if (!CS.getInstruction()->getType()->isVoidTy())
10691 CS.getInstruction()->
10692 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10694 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10695 // Don't break the CFG, insert a dummy cond branch.
10696 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10697 ConstantInt::getTrue(*Context), II);
10699 return EraseInstFromFunction(*CS.getInstruction());
10702 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10703 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10704 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10705 return transformCallThroughTrampoline(CS);
10707 const PointerType *PTy = cast<PointerType>(Callee->getType());
10708 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10709 if (FTy->isVarArg()) {
10710 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10711 // See if we can optimize any arguments passed through the varargs area of
10713 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10714 E = CS.arg_end(); I != E; ++I, ++ix) {
10715 CastInst *CI = dyn_cast<CastInst>(*I);
10716 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10717 *I = CI->getOperand(0);
10723 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10724 // Inline asm calls cannot throw - mark them 'nounwind'.
10725 CS.setDoesNotThrow();
10729 return Changed ? CS.getInstruction() : 0;
10732 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10733 // attempt to move the cast to the arguments of the call/invoke.
10735 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10736 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10737 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10738 if (CE->getOpcode() != Instruction::BitCast ||
10739 !isa<Function>(CE->getOperand(0)))
10741 Function *Callee = cast<Function>(CE->getOperand(0));
10742 Instruction *Caller = CS.getInstruction();
10743 const AttrListPtr &CallerPAL = CS.getAttributes();
10745 // Okay, this is a cast from a function to a different type. Unless doing so
10746 // would cause a type conversion of one of our arguments, change this call to
10747 // be a direct call with arguments casted to the appropriate types.
10749 const FunctionType *FT = Callee->getFunctionType();
10750 const Type *OldRetTy = Caller->getType();
10751 const Type *NewRetTy = FT->getReturnType();
10753 if (isa<StructType>(NewRetTy))
10754 return false; // TODO: Handle multiple return values.
10756 // Check to see if we are changing the return type...
10757 if (OldRetTy != NewRetTy) {
10758 if (Callee->isDeclaration() &&
10759 // Conversion is ok if changing from one pointer type to another or from
10760 // a pointer to an integer of the same size.
10761 !((isa<PointerType>(OldRetTy) || !TD ||
10762 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
10763 (isa<PointerType>(NewRetTy) || !TD ||
10764 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
10765 return false; // Cannot transform this return value.
10767 if (!Caller->use_empty() &&
10768 // void -> non-void is handled specially
10769 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
10770 return false; // Cannot transform this return value.
10772 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10773 Attributes RAttrs = CallerPAL.getRetAttributes();
10774 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10775 return false; // Attribute not compatible with transformed value.
10778 // If the callsite is an invoke instruction, and the return value is used by
10779 // a PHI node in a successor, we cannot change the return type of the call
10780 // because there is no place to put the cast instruction (without breaking
10781 // the critical edge). Bail out in this case.
10782 if (!Caller->use_empty())
10783 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10784 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10786 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10787 if (PN->getParent() == II->getNormalDest() ||
10788 PN->getParent() == II->getUnwindDest())
10792 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10793 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10795 CallSite::arg_iterator AI = CS.arg_begin();
10796 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10797 const Type *ParamTy = FT->getParamType(i);
10798 const Type *ActTy = (*AI)->getType();
10800 if (!CastInst::isCastable(ActTy, ParamTy))
10801 return false; // Cannot transform this parameter value.
10803 if (CallerPAL.getParamAttributes(i + 1)
10804 & Attribute::typeIncompatible(ParamTy))
10805 return false; // Attribute not compatible with transformed value.
10807 // Converting from one pointer type to another or between a pointer and an
10808 // integer of the same size is safe even if we do not have a body.
10809 bool isConvertible = ActTy == ParamTy ||
10810 (TD && ((isa<PointerType>(ParamTy) ||
10811 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10812 (isa<PointerType>(ActTy) ||
10813 ActTy == TD->getIntPtrType(Caller->getContext()))));
10814 if (Callee->isDeclaration() && !isConvertible) return false;
10817 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10818 Callee->isDeclaration())
10819 return false; // Do not delete arguments unless we have a function body.
10821 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10822 !CallerPAL.isEmpty())
10823 // In this case we have more arguments than the new function type, but we
10824 // won't be dropping them. Check that these extra arguments have attributes
10825 // that are compatible with being a vararg call argument.
10826 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10827 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10829 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10830 if (PAttrs & Attribute::VarArgsIncompatible)
10834 // Okay, we decided that this is a safe thing to do: go ahead and start
10835 // inserting cast instructions as necessary...
10836 std::vector<Value*> Args;
10837 Args.reserve(NumActualArgs);
10838 SmallVector<AttributeWithIndex, 8> attrVec;
10839 attrVec.reserve(NumCommonArgs);
10841 // Get any return attributes.
10842 Attributes RAttrs = CallerPAL.getRetAttributes();
10844 // If the return value is not being used, the type may not be compatible
10845 // with the existing attributes. Wipe out any problematic attributes.
10846 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10848 // Add the new return attributes.
10850 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10852 AI = CS.arg_begin();
10853 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10854 const Type *ParamTy = FT->getParamType(i);
10855 if ((*AI)->getType() == ParamTy) {
10856 Args.push_back(*AI);
10858 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10859 false, ParamTy, false);
10860 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
10863 // Add any parameter attributes.
10864 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10865 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10868 // If the function takes more arguments than the call was taking, add them
10870 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10871 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10873 // If we are removing arguments to the function, emit an obnoxious warning.
10874 if (FT->getNumParams() < NumActualArgs) {
10875 if (!FT->isVarArg()) {
10876 errs() << "WARNING: While resolving call to function '"
10877 << Callee->getName() << "' arguments were dropped!\n";
10879 // Add all of the arguments in their promoted form to the arg list.
10880 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10881 const Type *PTy = getPromotedType((*AI)->getType());
10882 if (PTy != (*AI)->getType()) {
10883 // Must promote to pass through va_arg area!
10884 Instruction::CastOps opcode =
10885 CastInst::getCastOpcode(*AI, false, PTy, false);
10886 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
10888 Args.push_back(*AI);
10891 // Add any parameter attributes.
10892 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10893 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10898 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10899 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10901 if (NewRetTy->isVoidTy())
10902 Caller->setName(""); // Void type should not have a name.
10904 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10908 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10909 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10910 Args.begin(), Args.end(),
10911 Caller->getName(), Caller);
10912 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10913 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10915 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10916 Caller->getName(), Caller);
10917 CallInst *CI = cast<CallInst>(Caller);
10918 if (CI->isTailCall())
10919 cast<CallInst>(NC)->setTailCall();
10920 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10921 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10924 // Insert a cast of the return type as necessary.
10926 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10927 if (!NV->getType()->isVoidTy()) {
10928 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10930 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10932 // If this is an invoke instruction, we should insert it after the first
10933 // non-phi, instruction in the normal successor block.
10934 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10935 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10936 InsertNewInstBefore(NC, *I);
10938 // Otherwise, it's a call, just insert cast right after the call instr
10939 InsertNewInstBefore(NC, *Caller);
10941 Worklist.AddUsersToWorkList(*Caller);
10943 NV = UndefValue::get(Caller->getType());
10948 if (!Caller->use_empty())
10949 Caller->replaceAllUsesWith(NV);
10951 EraseInstFromFunction(*Caller);
10955 // transformCallThroughTrampoline - Turn a call to a function created by the
10956 // init_trampoline intrinsic into a direct call to the underlying function.
10958 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10959 Value *Callee = CS.getCalledValue();
10960 const PointerType *PTy = cast<PointerType>(Callee->getType());
10961 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10962 const AttrListPtr &Attrs = CS.getAttributes();
10964 // If the call already has the 'nest' attribute somewhere then give up -
10965 // otherwise 'nest' would occur twice after splicing in the chain.
10966 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10969 IntrinsicInst *Tramp =
10970 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10972 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10973 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10974 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10976 const AttrListPtr &NestAttrs = NestF->getAttributes();
10977 if (!NestAttrs.isEmpty()) {
10978 unsigned NestIdx = 1;
10979 const Type *NestTy = 0;
10980 Attributes NestAttr = Attribute::None;
10982 // Look for a parameter marked with the 'nest' attribute.
10983 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10984 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10985 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10986 // Record the parameter type and any other attributes.
10988 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10993 Instruction *Caller = CS.getInstruction();
10994 std::vector<Value*> NewArgs;
10995 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10997 SmallVector<AttributeWithIndex, 8> NewAttrs;
10998 NewAttrs.reserve(Attrs.getNumSlots() + 1);
11000 // Insert the nest argument into the call argument list, which may
11001 // mean appending it. Likewise for attributes.
11003 // Add any result attributes.
11004 if (Attributes Attr = Attrs.getRetAttributes())
11005 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
11009 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
11011 if (Idx == NestIdx) {
11012 // Add the chain argument and attributes.
11013 Value *NestVal = Tramp->getOperand(3);
11014 if (NestVal->getType() != NestTy)
11015 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
11016 NewArgs.push_back(NestVal);
11017 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
11023 // Add the original argument and attributes.
11024 NewArgs.push_back(*I);
11025 if (Attributes Attr = Attrs.getParamAttributes(Idx))
11027 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
11033 // Add any function attributes.
11034 if (Attributes Attr = Attrs.getFnAttributes())
11035 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
11037 // The trampoline may have been bitcast to a bogus type (FTy).
11038 // Handle this by synthesizing a new function type, equal to FTy
11039 // with the chain parameter inserted.
11041 std::vector<const Type*> NewTypes;
11042 NewTypes.reserve(FTy->getNumParams()+1);
11044 // Insert the chain's type into the list of parameter types, which may
11045 // mean appending it.
11048 FunctionType::param_iterator I = FTy->param_begin(),
11049 E = FTy->param_end();
11052 if (Idx == NestIdx)
11053 // Add the chain's type.
11054 NewTypes.push_back(NestTy);
11059 // Add the original type.
11060 NewTypes.push_back(*I);
11066 // Replace the trampoline call with a direct call. Let the generic
11067 // code sort out any function type mismatches.
11068 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
11070 Constant *NewCallee =
11071 NestF->getType() == PointerType::getUnqual(NewFTy) ?
11072 NestF : ConstantExpr::getBitCast(NestF,
11073 PointerType::getUnqual(NewFTy));
11074 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
11077 Instruction *NewCaller;
11078 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
11079 NewCaller = InvokeInst::Create(NewCallee,
11080 II->getNormalDest(), II->getUnwindDest(),
11081 NewArgs.begin(), NewArgs.end(),
11082 Caller->getName(), Caller);
11083 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
11084 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
11086 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
11087 Caller->getName(), Caller);
11088 if (cast<CallInst>(Caller)->isTailCall())
11089 cast<CallInst>(NewCaller)->setTailCall();
11090 cast<CallInst>(NewCaller)->
11091 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
11092 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
11094 if (!Caller->getType()->isVoidTy())
11095 Caller->replaceAllUsesWith(NewCaller);
11096 Caller->eraseFromParent();
11097 Worklist.Remove(Caller);
11102 // Replace the trampoline call with a direct call. Since there is no 'nest'
11103 // parameter, there is no need to adjust the argument list. Let the generic
11104 // code sort out any function type mismatches.
11105 Constant *NewCallee =
11106 NestF->getType() == PTy ? NestF :
11107 ConstantExpr::getBitCast(NestF, PTy);
11108 CS.setCalledFunction(NewCallee);
11109 return CS.getInstruction();
11112 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
11113 /// and if a/b/c and the add's all have a single use, turn this into a phi
11114 /// and a single binop.
11115 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
11116 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
11117 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
11118 unsigned Opc = FirstInst->getOpcode();
11119 Value *LHSVal = FirstInst->getOperand(0);
11120 Value *RHSVal = FirstInst->getOperand(1);
11122 const Type *LHSType = LHSVal->getType();
11123 const Type *RHSType = RHSVal->getType();
11125 // Scan to see if all operands are the same opcode, and all have one use.
11126 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
11127 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
11128 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
11129 // Verify type of the LHS matches so we don't fold cmp's of different
11130 // types or GEP's with different index types.
11131 I->getOperand(0)->getType() != LHSType ||
11132 I->getOperand(1)->getType() != RHSType)
11135 // If they are CmpInst instructions, check their predicates
11136 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
11137 if (cast<CmpInst>(I)->getPredicate() !=
11138 cast<CmpInst>(FirstInst)->getPredicate())
11141 // Keep track of which operand needs a phi node.
11142 if (I->getOperand(0) != LHSVal) LHSVal = 0;
11143 if (I->getOperand(1) != RHSVal) RHSVal = 0;
11146 // If both LHS and RHS would need a PHI, don't do this transformation,
11147 // because it would increase the number of PHIs entering the block,
11148 // which leads to higher register pressure. This is especially
11149 // bad when the PHIs are in the header of a loop.
11150 if (!LHSVal && !RHSVal)
11153 // Otherwise, this is safe to transform!
11155 Value *InLHS = FirstInst->getOperand(0);
11156 Value *InRHS = FirstInst->getOperand(1);
11157 PHINode *NewLHS = 0, *NewRHS = 0;
11159 NewLHS = PHINode::Create(LHSType,
11160 FirstInst->getOperand(0)->getName() + ".pn");
11161 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
11162 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
11163 InsertNewInstBefore(NewLHS, PN);
11168 NewRHS = PHINode::Create(RHSType,
11169 FirstInst->getOperand(1)->getName() + ".pn");
11170 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
11171 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
11172 InsertNewInstBefore(NewRHS, PN);
11176 // Add all operands to the new PHIs.
11177 if (NewLHS || NewRHS) {
11178 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11179 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
11181 Value *NewInLHS = InInst->getOperand(0);
11182 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
11185 Value *NewInRHS = InInst->getOperand(1);
11186 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
11191 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
11192 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
11193 CmpInst *CIOp = cast<CmpInst>(FirstInst);
11194 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
11198 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
11199 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
11201 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
11202 FirstInst->op_end());
11203 // This is true if all GEP bases are allocas and if all indices into them are
11205 bool AllBasePointersAreAllocas = true;
11207 // We don't want to replace this phi if the replacement would require
11208 // more than one phi, which leads to higher register pressure. This is
11209 // especially bad when the PHIs are in the header of a loop.
11210 bool NeededPhi = false;
11212 // Scan to see if all operands are the same opcode, and all have one use.
11213 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
11214 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
11215 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
11216 GEP->getNumOperands() != FirstInst->getNumOperands())
11219 // Keep track of whether or not all GEPs are of alloca pointers.
11220 if (AllBasePointersAreAllocas &&
11221 (!isa<AllocaInst>(GEP->getOperand(0)) ||
11222 !GEP->hasAllConstantIndices()))
11223 AllBasePointersAreAllocas = false;
11225 // Compare the operand lists.
11226 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
11227 if (FirstInst->getOperand(op) == GEP->getOperand(op))
11230 // Don't merge two GEPs when two operands differ (introducing phi nodes)
11231 // if one of the PHIs has a constant for the index. The index may be
11232 // substantially cheaper to compute for the constants, so making it a
11233 // variable index could pessimize the path. This also handles the case
11234 // for struct indices, which must always be constant.
11235 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
11236 isa<ConstantInt>(GEP->getOperand(op)))
11239 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
11242 // If we already needed a PHI for an earlier operand, and another operand
11243 // also requires a PHI, we'd be introducing more PHIs than we're
11244 // eliminating, which increases register pressure on entry to the PHI's
11249 FixedOperands[op] = 0; // Needs a PHI.
11254 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
11255 // bother doing this transformation. At best, this will just save a bit of
11256 // offset calculation, but all the predecessors will have to materialize the
11257 // stack address into a register anyway. We'd actually rather *clone* the
11258 // load up into the predecessors so that we have a load of a gep of an alloca,
11259 // which can usually all be folded into the load.
11260 if (AllBasePointersAreAllocas)
11263 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
11264 // that is variable.
11265 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
11267 bool HasAnyPHIs = false;
11268 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
11269 if (FixedOperands[i]) continue; // operand doesn't need a phi.
11270 Value *FirstOp = FirstInst->getOperand(i);
11271 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
11272 FirstOp->getName()+".pn");
11273 InsertNewInstBefore(NewPN, PN);
11275 NewPN->reserveOperandSpace(e);
11276 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
11277 OperandPhis[i] = NewPN;
11278 FixedOperands[i] = NewPN;
11283 // Add all operands to the new PHIs.
11285 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11286 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
11287 BasicBlock *InBB = PN.getIncomingBlock(i);
11289 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
11290 if (PHINode *OpPhi = OperandPhis[op])
11291 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
11295 Value *Base = FixedOperands[0];
11296 return cast<GEPOperator>(FirstInst)->isInBounds() ?
11297 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
11298 FixedOperands.end()) :
11299 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
11300 FixedOperands.end());
11304 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
11305 /// sink the load out of the block that defines it. This means that it must be
11306 /// obvious the value of the load is not changed from the point of the load to
11307 /// the end of the block it is in.
11309 /// Finally, it is safe, but not profitable, to sink a load targetting a
11310 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
11312 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
11313 BasicBlock::iterator BBI = L, E = L->getParent()->end();
11315 for (++BBI; BBI != E; ++BBI)
11316 if (BBI->mayWriteToMemory())
11319 // Check for non-address taken alloca. If not address-taken already, it isn't
11320 // profitable to do this xform.
11321 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
11322 bool isAddressTaken = false;
11323 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
11325 if (isa<LoadInst>(UI)) continue;
11326 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
11327 // If storing TO the alloca, then the address isn't taken.
11328 if (SI->getOperand(1) == AI) continue;
11330 isAddressTaken = true;
11334 if (!isAddressTaken && AI->isStaticAlloca())
11338 // If this load is a load from a GEP with a constant offset from an alloca,
11339 // then we don't want to sink it. In its present form, it will be
11340 // load [constant stack offset]. Sinking it will cause us to have to
11341 // materialize the stack addresses in each predecessor in a register only to
11342 // do a shared load from register in the successor.
11343 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
11344 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
11345 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
11351 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
11352 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
11354 // When processing loads, we need to propagate two bits of information to the
11355 // sunk load: whether it is volatile, and what its alignment is. We currently
11356 // don't sink loads when some have their alignment specified and some don't.
11357 // visitLoadInst will propagate an alignment onto the load when TD is around,
11358 // and if TD isn't around, we can't handle the mixed case.
11359 bool isVolatile = FirstLI->isVolatile();
11360 unsigned LoadAlignment = FirstLI->getAlignment();
11362 // We can't sink the load if the loaded value could be modified between the
11363 // load and the PHI.
11364 if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
11365 !isSafeAndProfitableToSinkLoad(FirstLI))
11368 // If the PHI is of volatile loads and the load block has multiple
11369 // successors, sinking it would remove a load of the volatile value from
11370 // the path through the other successor.
11372 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
11375 // Check to see if all arguments are the same operation.
11376 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11377 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
11378 if (!LI || !LI->hasOneUse())
11381 // We can't sink the load if the loaded value could be modified between
11382 // the load and the PHI.
11383 if (LI->isVolatile() != isVolatile ||
11384 LI->getParent() != PN.getIncomingBlock(i) ||
11385 !isSafeAndProfitableToSinkLoad(LI))
11388 // If some of the loads have an alignment specified but not all of them,
11389 // we can't do the transformation.
11390 if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
11393 LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
11395 // If the PHI is of volatile loads and the load block has multiple
11396 // successors, sinking it would remove a load of the volatile value from
11397 // the path through the other successor.
11399 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
11403 // Okay, they are all the same operation. Create a new PHI node of the
11404 // correct type, and PHI together all of the LHS's of the instructions.
11405 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
11406 PN.getName()+".in");
11407 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
11409 Value *InVal = FirstLI->getOperand(0);
11410 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
11412 // Add all operands to the new PHI.
11413 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11414 Value *NewInVal = cast<LoadInst>(PN.getIncomingValue(i))->getOperand(0);
11415 if (NewInVal != InVal)
11417 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
11422 // The new PHI unions all of the same values together. This is really
11423 // common, so we handle it intelligently here for compile-time speed.
11427 InsertNewInstBefore(NewPN, PN);
11431 // If this was a volatile load that we are merging, make sure to loop through
11432 // and mark all the input loads as non-volatile. If we don't do this, we will
11433 // insert a new volatile load and the old ones will not be deletable.
11435 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
11436 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
11438 return new LoadInst(PhiVal, "", isVolatile, LoadAlignment);
11443 /// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
11444 /// operator and they all are only used by the PHI, PHI together their
11445 /// inputs, and do the operation once, to the result of the PHI.
11446 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
11447 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
11449 if (isa<GetElementPtrInst>(FirstInst))
11450 return FoldPHIArgGEPIntoPHI(PN);
11451 if (isa<LoadInst>(FirstInst))
11452 return FoldPHIArgLoadIntoPHI(PN);
11454 // Scan the instruction, looking for input operations that can be folded away.
11455 // If all input operands to the phi are the same instruction (e.g. a cast from
11456 // the same type or "+42") we can pull the operation through the PHI, reducing
11457 // code size and simplifying code.
11458 Constant *ConstantOp = 0;
11459 const Type *CastSrcTy = 0;
11461 if (isa<CastInst>(FirstInst)) {
11462 CastSrcTy = FirstInst->getOperand(0)->getType();
11464 // Be careful about transforming integer PHIs. We don't want to pessimize
11465 // the code by turning an i32 into an i1293.
11466 if (isa<IntegerType>(PN.getType()) && isa<IntegerType>(CastSrcTy)) {
11467 if (!ShouldChangeType(PN.getType(), CastSrcTy, TD))
11470 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
11471 // Can fold binop, compare or shift here if the RHS is a constant,
11472 // otherwise call FoldPHIArgBinOpIntoPHI.
11473 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
11474 if (ConstantOp == 0)
11475 return FoldPHIArgBinOpIntoPHI(PN);
11477 return 0; // Cannot fold this operation.
11480 // Check to see if all arguments are the same operation.
11481 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11482 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
11483 if (I == 0 || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
11486 if (I->getOperand(0)->getType() != CastSrcTy)
11487 return 0; // Cast operation must match.
11488 } else if (I->getOperand(1) != ConstantOp) {
11493 // Okay, they are all the same operation. Create a new PHI node of the
11494 // correct type, and PHI together all of the LHS's of the instructions.
11495 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
11496 PN.getName()+".in");
11497 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
11499 Value *InVal = FirstInst->getOperand(0);
11500 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
11502 // Add all operands to the new PHI.
11503 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11504 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
11505 if (NewInVal != InVal)
11507 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
11512 // The new PHI unions all of the same values together. This is really
11513 // common, so we handle it intelligently here for compile-time speed.
11517 InsertNewInstBefore(NewPN, PN);
11521 // Insert and return the new operation.
11522 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst))
11523 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
11525 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
11526 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
11528 CmpInst *CIOp = cast<CmpInst>(FirstInst);
11529 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
11530 PhiVal, ConstantOp);
11533 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
11535 static bool DeadPHICycle(PHINode *PN,
11536 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
11537 if (PN->use_empty()) return true;
11538 if (!PN->hasOneUse()) return false;
11540 // Remember this node, and if we find the cycle, return.
11541 if (!PotentiallyDeadPHIs.insert(PN))
11544 // Don't scan crazily complex things.
11545 if (PotentiallyDeadPHIs.size() == 16)
11548 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
11549 return DeadPHICycle(PU, PotentiallyDeadPHIs);
11554 /// PHIsEqualValue - Return true if this phi node is always equal to
11555 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
11556 /// z = some value; x = phi (y, z); y = phi (x, z)
11557 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
11558 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
11559 // See if we already saw this PHI node.
11560 if (!ValueEqualPHIs.insert(PN))
11563 // Don't scan crazily complex things.
11564 if (ValueEqualPHIs.size() == 16)
11567 // Scan the operands to see if they are either phi nodes or are equal to
11569 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
11570 Value *Op = PN->getIncomingValue(i);
11571 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
11572 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
11574 } else if (Op != NonPhiInVal)
11583 struct PHIUsageRecord {
11584 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)
11585 unsigned Shift; // The amount shifted.
11586 Instruction *Inst; // The trunc instruction.
11588 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
11589 : PHIId(pn), Shift(Sh), Inst(User) {}
11591 bool operator<(const PHIUsageRecord &RHS) const {
11592 if (PHIId < RHS.PHIId) return true;
11593 if (PHIId > RHS.PHIId) return false;
11594 if (Shift < RHS.Shift) return true;
11595 if (Shift > RHS.Shift) return false;
11596 return Inst->getType()->getPrimitiveSizeInBits() <
11597 RHS.Inst->getType()->getPrimitiveSizeInBits();
11601 struct LoweredPHIRecord {
11602 PHINode *PN; // The PHI that was lowered.
11603 unsigned Shift; // The amount shifted.
11604 unsigned Width; // The width extracted.
11606 LoweredPHIRecord(PHINode *pn, unsigned Sh, const Type *Ty)
11607 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
11609 // Ctor form used by DenseMap.
11610 LoweredPHIRecord(PHINode *pn, unsigned Sh)
11611 : PN(pn), Shift(Sh), Width(0) {}
11617 struct DenseMapInfo<LoweredPHIRecord> {
11618 static inline LoweredPHIRecord getEmptyKey() {
11619 return LoweredPHIRecord(0, 0);
11621 static inline LoweredPHIRecord getTombstoneKey() {
11622 return LoweredPHIRecord(0, 1);
11624 static unsigned getHashValue(const LoweredPHIRecord &Val) {
11625 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
11628 static bool isEqual(const LoweredPHIRecord &LHS,
11629 const LoweredPHIRecord &RHS) {
11630 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
11631 LHS.Width == RHS.Width;
11635 struct isPodLike<LoweredPHIRecord> { static const bool value = true; };
11639 /// SliceUpIllegalIntegerPHI - This is an integer PHI and we know that it has an
11640 /// illegal type: see if it is only used by trunc or trunc(lshr) operations. If
11641 /// so, we split the PHI into the various pieces being extracted. This sort of
11642 /// thing is introduced when SROA promotes an aggregate to large integer values.
11644 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
11645 /// inttoptr. We should produce new PHIs in the right type.
11647 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
11648 // PHIUsers - Keep track of all of the truncated values extracted from a set
11649 // of PHIs, along with their offset. These are the things we want to rewrite.
11650 SmallVector<PHIUsageRecord, 16> PHIUsers;
11652 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
11653 // nodes which are extracted from. PHIsToSlice is a set we use to avoid
11654 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
11655 // check the uses of (to ensure they are all extracts).
11656 SmallVector<PHINode*, 8> PHIsToSlice;
11657 SmallPtrSet<PHINode*, 8> PHIsInspected;
11659 PHIsToSlice.push_back(&FirstPhi);
11660 PHIsInspected.insert(&FirstPhi);
11662 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
11663 PHINode *PN = PHIsToSlice[PHIId];
11665 // Scan the input list of the PHI. If any input is an invoke, and if the
11666 // input is defined in the predecessor, then we won't be split the critical
11667 // edge which is required to insert a truncate. Because of this, we have to
11669 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
11670 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
11671 if (II == 0) continue;
11672 if (II->getParent() != PN->getIncomingBlock(i))
11675 // If we have a phi, and if it's directly in the predecessor, then we have
11676 // a critical edge where we need to put the truncate. Since we can't
11677 // split the edge in instcombine, we have to bail out.
11682 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
11684 Instruction *User = cast<Instruction>(*UI);
11686 // If the user is a PHI, inspect its uses recursively.
11687 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
11688 if (PHIsInspected.insert(UserPN))
11689 PHIsToSlice.push_back(UserPN);
11693 // Truncates are always ok.
11694 if (isa<TruncInst>(User)) {
11695 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, User));
11699 // Otherwise it must be a lshr which can only be used by one trunc.
11700 if (User->getOpcode() != Instruction::LShr ||
11701 !User->hasOneUse() || !isa<TruncInst>(User->use_back()) ||
11702 !isa<ConstantInt>(User->getOperand(1)))
11705 unsigned Shift = cast<ConstantInt>(User->getOperand(1))->getZExtValue();
11706 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, User->use_back()));
11710 // If we have no users, they must be all self uses, just nuke the PHI.
11711 if (PHIUsers.empty())
11712 return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
11714 // If this phi node is transformable, create new PHIs for all the pieces
11715 // extracted out of it. First, sort the users by their offset and size.
11716 array_pod_sort(PHIUsers.begin(), PHIUsers.end());
11718 DEBUG(errs() << "SLICING UP PHI: " << FirstPhi << '\n';
11719 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
11720 errs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] <<'\n';
11723 // PredValues - This is a temporary used when rewriting PHI nodes. It is
11724 // hoisted out here to avoid construction/destruction thrashing.
11725 DenseMap<BasicBlock*, Value*> PredValues;
11727 // ExtractedVals - Each new PHI we introduce is saved here so we don't
11728 // introduce redundant PHIs.
11729 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
11731 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
11732 unsigned PHIId = PHIUsers[UserI].PHIId;
11733 PHINode *PN = PHIsToSlice[PHIId];
11734 unsigned Offset = PHIUsers[UserI].Shift;
11735 const Type *Ty = PHIUsers[UserI].Inst->getType();
11739 // If we've already lowered a user like this, reuse the previously lowered
11741 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == 0) {
11743 // Otherwise, Create the new PHI node for this user.
11744 EltPHI = PHINode::Create(Ty, PN->getName()+".off"+Twine(Offset), PN);
11745 assert(EltPHI->getType() != PN->getType() &&
11746 "Truncate didn't shrink phi?");
11748 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
11749 BasicBlock *Pred = PN->getIncomingBlock(i);
11750 Value *&PredVal = PredValues[Pred];
11752 // If we already have a value for this predecessor, reuse it.
11754 EltPHI->addIncoming(PredVal, Pred);
11758 // Handle the PHI self-reuse case.
11759 Value *InVal = PN->getIncomingValue(i);
11762 EltPHI->addIncoming(PredVal, Pred);
11766 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
11767 // If the incoming value was a PHI, and if it was one of the PHIs we
11768 // already rewrote it, just use the lowered value.
11769 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
11771 EltPHI->addIncoming(PredVal, Pred);
11776 // Otherwise, do an extract in the predecessor.
11777 Builder->SetInsertPoint(Pred, Pred->getTerminator());
11778 Value *Res = InVal;
11780 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
11781 Offset), "extract");
11782 Res = Builder->CreateTrunc(Res, Ty, "extract.t");
11784 EltPHI->addIncoming(Res, Pred);
11786 // If the incoming value was a PHI, and if it was one of the PHIs we are
11787 // rewriting, we will ultimately delete the code we inserted. This
11788 // means we need to revisit that PHI to make sure we extract out the
11790 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
11791 if (PHIsInspected.count(OldInVal)) {
11792 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
11793 OldInVal)-PHIsToSlice.begin();
11794 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
11795 cast<Instruction>(Res)));
11799 PredValues.clear();
11801 DEBUG(errs() << " Made element PHI for offset " << Offset << ": "
11802 << *EltPHI << '\n');
11803 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
11806 // Replace the use of this piece with the PHI node.
11807 ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
11810 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
11812 Value *Undef = UndefValue::get(FirstPhi.getType());
11813 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
11814 ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
11815 return ReplaceInstUsesWith(FirstPhi, Undef);
11818 // PHINode simplification
11820 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
11821 // If LCSSA is around, don't mess with Phi nodes
11822 if (MustPreserveLCSSA) return 0;
11824 if (Value *V = PN.hasConstantValue())
11825 return ReplaceInstUsesWith(PN, V);
11827 // If all PHI operands are the same operation, pull them through the PHI,
11828 // reducing code size.
11829 if (isa<Instruction>(PN.getIncomingValue(0)) &&
11830 isa<Instruction>(PN.getIncomingValue(1)) &&
11831 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
11832 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
11833 // FIXME: The hasOneUse check will fail for PHIs that use the value more
11834 // than themselves more than once.
11835 PN.getIncomingValue(0)->hasOneUse())
11836 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
11839 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
11840 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
11841 // PHI)... break the cycle.
11842 if (PN.hasOneUse()) {
11843 Instruction *PHIUser = cast<Instruction>(PN.use_back());
11844 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
11845 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
11846 PotentiallyDeadPHIs.insert(&PN);
11847 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
11848 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
11851 // If this phi has a single use, and if that use just computes a value for
11852 // the next iteration of a loop, delete the phi. This occurs with unused
11853 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
11854 // common case here is good because the only other things that catch this
11855 // are induction variable analysis (sometimes) and ADCE, which is only run
11857 if (PHIUser->hasOneUse() &&
11858 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
11859 PHIUser->use_back() == &PN) {
11860 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
11864 // We sometimes end up with phi cycles that non-obviously end up being the
11865 // same value, for example:
11866 // z = some value; x = phi (y, z); y = phi (x, z)
11867 // where the phi nodes don't necessarily need to be in the same block. Do a
11868 // quick check to see if the PHI node only contains a single non-phi value, if
11869 // so, scan to see if the phi cycle is actually equal to that value.
11871 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
11872 // Scan for the first non-phi operand.
11873 while (InValNo != NumOperandVals &&
11874 isa<PHINode>(PN.getIncomingValue(InValNo)))
11877 if (InValNo != NumOperandVals) {
11878 Value *NonPhiInVal = PN.getOperand(InValNo);
11880 // Scan the rest of the operands to see if there are any conflicts, if so
11881 // there is no need to recursively scan other phis.
11882 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
11883 Value *OpVal = PN.getIncomingValue(InValNo);
11884 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
11888 // If we scanned over all operands, then we have one unique value plus
11889 // phi values. Scan PHI nodes to see if they all merge in each other or
11891 if (InValNo == NumOperandVals) {
11892 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
11893 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
11894 return ReplaceInstUsesWith(PN, NonPhiInVal);
11899 // If there are multiple PHIs, sort their operands so that they all list
11900 // the blocks in the same order. This will help identical PHIs be eliminated
11901 // by other passes. Other passes shouldn't depend on this for correctness
11903 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
11904 if (&PN != FirstPN)
11905 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
11906 BasicBlock *BBA = PN.getIncomingBlock(i);
11907 BasicBlock *BBB = FirstPN->getIncomingBlock(i);
11909 Value *VA = PN.getIncomingValue(i);
11910 unsigned j = PN.getBasicBlockIndex(BBB);
11911 Value *VB = PN.getIncomingValue(j);
11912 PN.setIncomingBlock(i, BBB);
11913 PN.setIncomingValue(i, VB);
11914 PN.setIncomingBlock(j, BBA);
11915 PN.setIncomingValue(j, VA);
11916 // NOTE: Instcombine normally would want us to "return &PN" if we
11917 // modified any of the operands of an instruction. However, since we
11918 // aren't adding or removing uses (just rearranging them) we don't do
11919 // this in this case.
11923 // If this is an integer PHI and we know that it has an illegal type, see if
11924 // it is only used by trunc or trunc(lshr) operations. If so, we split the
11925 // PHI into the various pieces being extracted. This sort of thing is
11926 // introduced when SROA promotes an aggregate to a single large integer type.
11927 if (isa<IntegerType>(PN.getType()) && TD &&
11928 !TD->isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
11929 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
11935 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11936 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
11938 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
11939 return ReplaceInstUsesWith(GEP, V);
11941 Value *PtrOp = GEP.getOperand(0);
11943 if (isa<UndefValue>(GEP.getOperand(0)))
11944 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
11946 // Eliminate unneeded casts for indices.
11948 bool MadeChange = false;
11949 unsigned PtrSize = TD->getPointerSizeInBits();
11951 gep_type_iterator GTI = gep_type_begin(GEP);
11952 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
11953 I != E; ++I, ++GTI) {
11954 if (!isa<SequentialType>(*GTI)) continue;
11956 // If we are using a wider index than needed for this platform, shrink it
11957 // to what we need. If narrower, sign-extend it to what we need. This
11958 // explicit cast can make subsequent optimizations more obvious.
11959 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
11960 if (OpBits == PtrSize)
11963 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
11966 if (MadeChange) return &GEP;
11969 // Combine Indices - If the source pointer to this getelementptr instruction
11970 // is a getelementptr instruction, combine the indices of the two
11971 // getelementptr instructions into a single instruction.
11973 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
11974 // Note that if our source is a gep chain itself that we wait for that
11975 // chain to be resolved before we perform this transformation. This
11976 // avoids us creating a TON of code in some cases.
11978 if (GetElementPtrInst *SrcGEP =
11979 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
11980 if (SrcGEP->getNumOperands() == 2)
11981 return 0; // Wait until our source is folded to completion.
11983 SmallVector<Value*, 8> Indices;
11985 // Find out whether the last index in the source GEP is a sequential idx.
11986 bool EndsWithSequential = false;
11987 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
11989 EndsWithSequential = !isa<StructType>(*I);
11991 // Can we combine the two pointer arithmetics offsets?
11992 if (EndsWithSequential) {
11993 // Replace: gep (gep %P, long B), long A, ...
11994 // With: T = long A+B; gep %P, T, ...
11997 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
11998 Value *GO1 = GEP.getOperand(1);
11999 if (SO1 == Constant::getNullValue(SO1->getType())) {
12001 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
12004 // If they aren't the same type, then the input hasn't been processed
12005 // by the loop above yet (which canonicalizes sequential index types to
12006 // intptr_t). Just avoid transforming this until the input has been
12008 if (SO1->getType() != GO1->getType())
12010 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
12013 // Update the GEP in place if possible.
12014 if (Src->getNumOperands() == 2) {
12015 GEP.setOperand(0, Src->getOperand(0));
12016 GEP.setOperand(1, Sum);
12019 Indices.append(Src->op_begin()+1, Src->op_end()-1);
12020 Indices.push_back(Sum);
12021 Indices.append(GEP.op_begin()+2, GEP.op_end());
12022 } else if (isa<Constant>(*GEP.idx_begin()) &&
12023 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
12024 Src->getNumOperands() != 1) {
12025 // Otherwise we can do the fold if the first index of the GEP is a zero
12026 Indices.append(Src->op_begin()+1, Src->op_end());
12027 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
12030 if (!Indices.empty())
12031 return (cast<GEPOperator>(&GEP)->isInBounds() &&
12032 Src->isInBounds()) ?
12033 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
12034 Indices.end(), GEP.getName()) :
12035 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
12036 Indices.end(), GEP.getName());
12039 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
12040 if (Value *X = getBitCastOperand(PtrOp)) {
12041 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
12043 // If the input bitcast is actually "bitcast(bitcast(x))", then we don't
12044 // want to change the gep until the bitcasts are eliminated.
12045 if (getBitCastOperand(X)) {
12046 Worklist.AddValue(PtrOp);
12050 bool HasZeroPointerIndex = false;
12051 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
12052 HasZeroPointerIndex = C->isZero();
12054 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
12055 // into : GEP [10 x i8]* X, i32 0, ...
12057 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
12058 // into : GEP i8* X, ...
12060 // This occurs when the program declares an array extern like "int X[];"
12061 if (HasZeroPointerIndex) {
12062 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
12063 const PointerType *XTy = cast<PointerType>(X->getType());
12064 if (const ArrayType *CATy =
12065 dyn_cast<ArrayType>(CPTy->getElementType())) {
12066 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
12067 if (CATy->getElementType() == XTy->getElementType()) {
12068 // -> GEP i8* X, ...
12069 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
12070 return cast<GEPOperator>(&GEP)->isInBounds() ?
12071 GetElementPtrInst::CreateInBounds(X, Indices.begin(), Indices.end(),
12073 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
12077 if (const ArrayType *XATy = dyn_cast<ArrayType>(XTy->getElementType())){
12078 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
12079 if (CATy->getElementType() == XATy->getElementType()) {
12080 // -> GEP [10 x i8]* X, i32 0, ...
12081 // At this point, we know that the cast source type is a pointer
12082 // to an array of the same type as the destination pointer
12083 // array. Because the array type is never stepped over (there
12084 // is a leading zero) we can fold the cast into this GEP.
12085 GEP.setOperand(0, X);
12090 } else if (GEP.getNumOperands() == 2) {
12091 // Transform things like:
12092 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
12093 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
12094 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
12095 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
12096 if (TD && isa<ArrayType>(SrcElTy) &&
12097 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
12098 TD->getTypeAllocSize(ResElTy)) {
12100 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
12101 Idx[1] = GEP.getOperand(1);
12102 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
12103 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
12104 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
12105 // V and GEP are both pointer types --> BitCast
12106 return new BitCastInst(NewGEP, GEP.getType());
12109 // Transform things like:
12110 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
12111 // (where tmp = 8*tmp2) into:
12112 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
12114 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
12115 uint64_t ArrayEltSize =
12116 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
12118 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
12119 // allow either a mul, shift, or constant here.
12121 ConstantInt *Scale = 0;
12122 if (ArrayEltSize == 1) {
12123 NewIdx = GEP.getOperand(1);
12124 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
12125 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
12126 NewIdx = ConstantInt::get(CI->getType(), 1);
12128 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
12129 if (Inst->getOpcode() == Instruction::Shl &&
12130 isa<ConstantInt>(Inst->getOperand(1))) {
12131 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
12132 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
12133 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
12135 NewIdx = Inst->getOperand(0);
12136 } else if (Inst->getOpcode() == Instruction::Mul &&
12137 isa<ConstantInt>(Inst->getOperand(1))) {
12138 Scale = cast<ConstantInt>(Inst->getOperand(1));
12139 NewIdx = Inst->getOperand(0);
12143 // If the index will be to exactly the right offset with the scale taken
12144 // out, perform the transformation. Note, we don't know whether Scale is
12145 // signed or not. We'll use unsigned version of division/modulo
12146 // operation after making sure Scale doesn't have the sign bit set.
12147 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
12148 Scale->getZExtValue() % ArrayEltSize == 0) {
12149 Scale = ConstantInt::get(Scale->getType(),
12150 Scale->getZExtValue() / ArrayEltSize);
12151 if (Scale->getZExtValue() != 1) {
12152 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
12154 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
12157 // Insert the new GEP instruction.
12159 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
12161 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
12162 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
12163 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
12164 // The NewGEP must be pointer typed, so must the old one -> BitCast
12165 return new BitCastInst(NewGEP, GEP.getType());
12171 /// See if we can simplify:
12172 /// X = bitcast A* to B*
12173 /// Y = gep X, <...constant indices...>
12174 /// into a gep of the original struct. This is important for SROA and alias
12175 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
12176 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
12178 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
12179 // Determine how much the GEP moves the pointer. We are guaranteed to get
12180 // a constant back from EmitGEPOffset.
12181 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP, *this));
12182 int64_t Offset = OffsetV->getSExtValue();
12184 // If this GEP instruction doesn't move the pointer, just replace the GEP
12185 // with a bitcast of the real input to the dest type.
12187 // If the bitcast is of an allocation, and the allocation will be
12188 // converted to match the type of the cast, don't touch this.
12189 if (isa<AllocaInst>(BCI->getOperand(0)) ||
12190 isMalloc(BCI->getOperand(0))) {
12191 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
12192 if (Instruction *I = visitBitCast(*BCI)) {
12195 BCI->getParent()->getInstList().insert(BCI, I);
12196 ReplaceInstUsesWith(*BCI, I);
12201 return new BitCastInst(BCI->getOperand(0), GEP.getType());
12204 // Otherwise, if the offset is non-zero, we need to find out if there is a
12205 // field at Offset in 'A's type. If so, we can pull the cast through the
12207 SmallVector<Value*, 8> NewIndices;
12209 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
12210 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
12211 Value *NGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
12212 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
12213 NewIndices.end()) :
12214 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
12217 if (NGEP->getType() == GEP.getType())
12218 return ReplaceInstUsesWith(GEP, NGEP);
12219 NGEP->takeName(&GEP);
12220 return new BitCastInst(NGEP, GEP.getType());
12228 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
12229 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
12230 if (AI.isArrayAllocation()) { // Check C != 1
12231 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
12232 const Type *NewTy =
12233 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
12234 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
12235 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
12236 New->setAlignment(AI.getAlignment());
12238 // Scan to the end of the allocation instructions, to skip over a block of
12239 // allocas if possible...also skip interleaved debug info
12241 BasicBlock::iterator It = New;
12242 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
12244 // Now that I is pointing to the first non-allocation-inst in the block,
12245 // insert our getelementptr instruction...
12247 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
12251 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
12252 New->getName()+".sub", It);
12254 // Now make everything use the getelementptr instead of the original
12256 return ReplaceInstUsesWith(AI, V);
12257 } else if (isa<UndefValue>(AI.getArraySize())) {
12258 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
12262 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
12263 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
12264 // Note that we only do this for alloca's, because malloc should allocate
12265 // and return a unique pointer, even for a zero byte allocation.
12266 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
12267 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
12269 // If the alignment is 0 (unspecified), assign it the preferred alignment.
12270 if (AI.getAlignment() == 0)
12271 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
12277 Instruction *InstCombiner::visitFree(Instruction &FI) {
12278 Value *Op = FI.getOperand(1);
12280 // free undef -> unreachable.
12281 if (isa<UndefValue>(Op)) {
12282 // Insert a new store to null because we cannot modify the CFG here.
12283 new StoreInst(ConstantInt::getTrue(*Context),
12284 UndefValue::get(Type::getInt1PtrTy(*Context)), &FI);
12285 return EraseInstFromFunction(FI);
12288 // If we have 'free null' delete the instruction. This can happen in stl code
12289 // when lots of inlining happens.
12290 if (isa<ConstantPointerNull>(Op))
12291 return EraseInstFromFunction(FI);
12293 // If we have a malloc call whose only use is a free call, delete both.
12294 if (isMalloc(Op)) {
12295 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
12296 if (Op->hasOneUse() && CI->hasOneUse()) {
12297 EraseInstFromFunction(FI);
12298 EraseInstFromFunction(*CI);
12299 return EraseInstFromFunction(*cast<Instruction>(Op));
12302 // Op is a call to malloc
12303 if (Op->hasOneUse()) {
12304 EraseInstFromFunction(FI);
12305 return EraseInstFromFunction(*cast<Instruction>(Op));
12313 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
12314 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
12315 const TargetData *TD) {
12316 User *CI = cast<User>(LI.getOperand(0));
12317 Value *CastOp = CI->getOperand(0);
12318 LLVMContext *Context = IC.getContext();
12320 const PointerType *DestTy = cast<PointerType>(CI->getType());
12321 const Type *DestPTy = DestTy->getElementType();
12322 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
12324 // If the address spaces don't match, don't eliminate the cast.
12325 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
12328 const Type *SrcPTy = SrcTy->getElementType();
12330 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
12331 isa<VectorType>(DestPTy)) {
12332 // If the source is an array, the code below will not succeed. Check to
12333 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
12335 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
12336 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
12337 if (ASrcTy->getNumElements() != 0) {
12339 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
12341 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
12342 SrcTy = cast<PointerType>(CastOp->getType());
12343 SrcPTy = SrcTy->getElementType();
12346 if (IC.getTargetData() &&
12347 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
12348 isa<VectorType>(SrcPTy)) &&
12349 // Do not allow turning this into a load of an integer, which is then
12350 // casted to a pointer, this pessimizes pointer analysis a lot.
12351 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
12352 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
12353 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
12355 // Okay, we are casting from one integer or pointer type to another of
12356 // the same size. Instead of casting the pointer before the load, cast
12357 // the result of the loaded value.
12359 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
12360 // Now cast the result of the load.
12361 return new BitCastInst(NewLoad, LI.getType());
12368 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
12369 Value *Op = LI.getOperand(0);
12371 // Attempt to improve the alignment.
12373 unsigned KnownAlign =
12374 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
12376 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
12377 LI.getAlignment()))
12378 LI.setAlignment(KnownAlign);
12381 // load (cast X) --> cast (load X) iff safe.
12382 if (isa<CastInst>(Op))
12383 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
12386 // None of the following transforms are legal for volatile loads.
12387 if (LI.isVolatile()) return 0;
12389 // Do really simple store-to-load forwarding and load CSE, to catch cases
12390 // where there are several consequtive memory accesses to the same location,
12391 // separated by a few arithmetic operations.
12392 BasicBlock::iterator BBI = &LI;
12393 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
12394 return ReplaceInstUsesWith(LI, AvailableVal);
12396 // load(gep null, ...) -> unreachable
12397 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
12398 const Value *GEPI0 = GEPI->getOperand(0);
12399 // TODO: Consider a target hook for valid address spaces for this xform.
12400 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
12401 // Insert a new store to null instruction before the load to indicate
12402 // that this code is not reachable. We do this instead of inserting
12403 // an unreachable instruction directly because we cannot modify the
12405 new StoreInst(UndefValue::get(LI.getType()),
12406 Constant::getNullValue(Op->getType()), &LI);
12407 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
12411 // load null/undef -> unreachable
12412 // TODO: Consider a target hook for valid address spaces for this xform.
12413 if (isa<UndefValue>(Op) ||
12414 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
12415 // Insert a new store to null instruction before the load to indicate that
12416 // this code is not reachable. We do this instead of inserting an
12417 // unreachable instruction directly because we cannot modify the CFG.
12418 new StoreInst(UndefValue::get(LI.getType()),
12419 Constant::getNullValue(Op->getType()), &LI);
12420 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
12423 // Instcombine load (constantexpr_cast global) -> cast (load global)
12424 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
12426 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
12429 if (Op->hasOneUse()) {
12430 // Change select and PHI nodes to select values instead of addresses: this
12431 // helps alias analysis out a lot, allows many others simplifications, and
12432 // exposes redundancy in the code.
12434 // Note that we cannot do the transformation unless we know that the
12435 // introduced loads cannot trap! Something like this is valid as long as
12436 // the condition is always false: load (select bool %C, int* null, int* %G),
12437 // but it would not be valid if we transformed it to load from null
12438 // unconditionally.
12440 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
12441 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
12442 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
12443 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
12444 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
12445 SI->getOperand(1)->getName()+".val");
12446 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
12447 SI->getOperand(2)->getName()+".val");
12448 return SelectInst::Create(SI->getCondition(), V1, V2);
12451 // load (select (cond, null, P)) -> load P
12452 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
12453 if (C->isNullValue()) {
12454 LI.setOperand(0, SI->getOperand(2));
12458 // load (select (cond, P, null)) -> load P
12459 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
12460 if (C->isNullValue()) {
12461 LI.setOperand(0, SI->getOperand(1));
12469 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
12470 /// when possible. This makes it generally easy to do alias analysis and/or
12471 /// SROA/mem2reg of the memory object.
12472 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
12473 User *CI = cast<User>(SI.getOperand(1));
12474 Value *CastOp = CI->getOperand(0);
12476 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
12477 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
12478 if (SrcTy == 0) return 0;
12480 const Type *SrcPTy = SrcTy->getElementType();
12482 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
12485 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
12486 /// to its first element. This allows us to handle things like:
12487 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
12488 /// on 32-bit hosts.
12489 SmallVector<Value*, 4> NewGEPIndices;
12491 // If the source is an array, the code below will not succeed. Check to
12492 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
12494 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
12495 // Index through pointer.
12496 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
12497 NewGEPIndices.push_back(Zero);
12500 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
12501 if (!STy->getNumElements()) /* Struct can be empty {} */
12503 NewGEPIndices.push_back(Zero);
12504 SrcPTy = STy->getElementType(0);
12505 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
12506 NewGEPIndices.push_back(Zero);
12507 SrcPTy = ATy->getElementType();
12513 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
12516 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
12519 // If the pointers point into different address spaces or if they point to
12520 // values with different sizes, we can't do the transformation.
12521 if (!IC.getTargetData() ||
12522 SrcTy->getAddressSpace() !=
12523 cast<PointerType>(CI->getType())->getAddressSpace() ||
12524 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
12525 IC.getTargetData()->getTypeSizeInBits(DestPTy))
12528 // Okay, we are casting from one integer or pointer type to another of
12529 // the same size. Instead of casting the pointer before
12530 // the store, cast the value to be stored.
12532 Value *SIOp0 = SI.getOperand(0);
12533 Instruction::CastOps opcode = Instruction::BitCast;
12534 const Type* CastSrcTy = SIOp0->getType();
12535 const Type* CastDstTy = SrcPTy;
12536 if (isa<PointerType>(CastDstTy)) {
12537 if (CastSrcTy->isInteger())
12538 opcode = Instruction::IntToPtr;
12539 } else if (isa<IntegerType>(CastDstTy)) {
12540 if (isa<PointerType>(SIOp0->getType()))
12541 opcode = Instruction::PtrToInt;
12544 // SIOp0 is a pointer to aggregate and this is a store to the first field,
12545 // emit a GEP to index into its first field.
12546 if (!NewGEPIndices.empty())
12547 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
12548 NewGEPIndices.end());
12550 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
12551 SIOp0->getName()+".c");
12552 return new StoreInst(NewCast, CastOp);
12555 /// equivalentAddressValues - Test if A and B will obviously have the same
12556 /// value. This includes recognizing that %t0 and %t1 will have the same
12557 /// value in code like this:
12558 /// %t0 = getelementptr \@a, 0, 3
12559 /// store i32 0, i32* %t0
12560 /// %t1 = getelementptr \@a, 0, 3
12561 /// %t2 = load i32* %t1
12563 static bool equivalentAddressValues(Value *A, Value *B) {
12564 // Test if the values are trivially equivalent.
12565 if (A == B) return true;
12567 // Test if the values come form identical arithmetic instructions.
12568 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
12569 // its only used to compare two uses within the same basic block, which
12570 // means that they'll always either have the same value or one of them
12571 // will have an undefined value.
12572 if (isa<BinaryOperator>(A) ||
12573 isa<CastInst>(A) ||
12575 isa<GetElementPtrInst>(A))
12576 if (Instruction *BI = dyn_cast<Instruction>(B))
12577 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
12580 // Otherwise they may not be equivalent.
12584 // If this instruction has two uses, one of which is a llvm.dbg.declare,
12585 // return the llvm.dbg.declare.
12586 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
12587 if (!V->hasNUses(2))
12589 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
12591 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
12593 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
12594 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
12601 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
12602 Value *Val = SI.getOperand(0);
12603 Value *Ptr = SI.getOperand(1);
12605 // If the RHS is an alloca with a single use, zapify the store, making the
12607 // If the RHS is an alloca with a two uses, the other one being a
12608 // llvm.dbg.declare, zapify the store and the declare, making the
12609 // alloca dead. We must do this to prevent declare's from affecting
12611 if (!SI.isVolatile()) {
12612 if (Ptr->hasOneUse()) {
12613 if (isa<AllocaInst>(Ptr)) {
12614 EraseInstFromFunction(SI);
12618 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
12619 if (isa<AllocaInst>(GEP->getOperand(0))) {
12620 if (GEP->getOperand(0)->hasOneUse()) {
12621 EraseInstFromFunction(SI);
12625 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
12626 EraseInstFromFunction(*DI);
12627 EraseInstFromFunction(SI);
12634 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
12635 EraseInstFromFunction(*DI);
12636 EraseInstFromFunction(SI);
12642 // Attempt to improve the alignment.
12644 unsigned KnownAlign =
12645 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
12647 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
12648 SI.getAlignment()))
12649 SI.setAlignment(KnownAlign);
12652 // Do really simple DSE, to catch cases where there are several consecutive
12653 // stores to the same location, separated by a few arithmetic operations. This
12654 // situation often occurs with bitfield accesses.
12655 BasicBlock::iterator BBI = &SI;
12656 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
12659 // Don't count debug info directives, lest they affect codegen,
12660 // and we skip pointer-to-pointer bitcasts, which are NOPs.
12661 // It is necessary for correctness to skip those that feed into a
12662 // llvm.dbg.declare, as these are not present when debugging is off.
12663 if (isa<DbgInfoIntrinsic>(BBI) ||
12664 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12669 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
12670 // Prev store isn't volatile, and stores to the same location?
12671 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
12672 SI.getOperand(1))) {
12675 EraseInstFromFunction(*PrevSI);
12681 // If this is a load, we have to stop. However, if the loaded value is from
12682 // the pointer we're loading and is producing the pointer we're storing,
12683 // then *this* store is dead (X = load P; store X -> P).
12684 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
12685 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
12686 !SI.isVolatile()) {
12687 EraseInstFromFunction(SI);
12691 // Otherwise, this is a load from some other location. Stores before it
12692 // may not be dead.
12696 // Don't skip over loads or things that can modify memory.
12697 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
12702 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
12704 // store X, null -> turns into 'unreachable' in SimplifyCFG
12705 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
12706 if (!isa<UndefValue>(Val)) {
12707 SI.setOperand(0, UndefValue::get(Val->getType()));
12708 if (Instruction *U = dyn_cast<Instruction>(Val))
12709 Worklist.Add(U); // Dropped a use.
12712 return 0; // Do not modify these!
12715 // store undef, Ptr -> noop
12716 if (isa<UndefValue>(Val)) {
12717 EraseInstFromFunction(SI);
12722 // If the pointer destination is a cast, see if we can fold the cast into the
12724 if (isa<CastInst>(Ptr))
12725 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
12727 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
12729 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
12733 // If this store is the last instruction in the basic block (possibly
12734 // excepting debug info instructions and the pointer bitcasts that feed
12735 // into them), and if the block ends with an unconditional branch, try
12736 // to move it to the successor block.
12740 } while (isa<DbgInfoIntrinsic>(BBI) ||
12741 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
12742 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
12743 if (BI->isUnconditional())
12744 if (SimplifyStoreAtEndOfBlock(SI))
12745 return 0; // xform done!
12750 /// SimplifyStoreAtEndOfBlock - Turn things like:
12751 /// if () { *P = v1; } else { *P = v2 }
12752 /// into a phi node with a store in the successor.
12754 /// Simplify things like:
12755 /// *P = v1; if () { *P = v2; }
12756 /// into a phi node with a store in the successor.
12758 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
12759 BasicBlock *StoreBB = SI.getParent();
12761 // Check to see if the successor block has exactly two incoming edges. If
12762 // so, see if the other predecessor contains a store to the same location.
12763 // if so, insert a PHI node (if needed) and move the stores down.
12764 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
12766 // Determine whether Dest has exactly two predecessors and, if so, compute
12767 // the other predecessor.
12768 pred_iterator PI = pred_begin(DestBB);
12769 BasicBlock *OtherBB = 0;
12770 if (*PI != StoreBB)
12773 if (PI == pred_end(DestBB))
12776 if (*PI != StoreBB) {
12781 if (++PI != pred_end(DestBB))
12784 // Bail out if all the relevant blocks aren't distinct (this can happen,
12785 // for example, if SI is in an infinite loop)
12786 if (StoreBB == DestBB || OtherBB == DestBB)
12789 // Verify that the other block ends in a branch and is not otherwise empty.
12790 BasicBlock::iterator BBI = OtherBB->getTerminator();
12791 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12792 if (!OtherBr || BBI == OtherBB->begin())
12795 // If the other block ends in an unconditional branch, check for the 'if then
12796 // else' case. there is an instruction before the branch.
12797 StoreInst *OtherStore = 0;
12798 if (OtherBr->isUnconditional()) {
12800 // Skip over debugging info.
12801 while (isa<DbgInfoIntrinsic>(BBI) ||
12802 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12803 if (BBI==OtherBB->begin())
12807 // If this isn't a store, isn't a store to the same location, or if the
12808 // alignments differ, bail out.
12809 OtherStore = dyn_cast<StoreInst>(BBI);
12810 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
12811 OtherStore->getAlignment() != SI.getAlignment())
12814 // Otherwise, the other block ended with a conditional branch. If one of the
12815 // destinations is StoreBB, then we have the if/then case.
12816 if (OtherBr->getSuccessor(0) != StoreBB &&
12817 OtherBr->getSuccessor(1) != StoreBB)
12820 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12821 // if/then triangle. See if there is a store to the same ptr as SI that
12822 // lives in OtherBB.
12824 // Check to see if we find the matching store.
12825 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12826 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
12827 OtherStore->getAlignment() != SI.getAlignment())
12831 // If we find something that may be using or overwriting the stored
12832 // value, or if we run out of instructions, we can't do the xform.
12833 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12834 BBI == OtherBB->begin())
12838 // In order to eliminate the store in OtherBr, we have to
12839 // make sure nothing reads or overwrites the stored value in
12841 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12842 // FIXME: This should really be AA driven.
12843 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12848 // Insert a PHI node now if we need it.
12849 Value *MergedVal = OtherStore->getOperand(0);
12850 if (MergedVal != SI.getOperand(0)) {
12851 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12852 PN->reserveOperandSpace(2);
12853 PN->addIncoming(SI.getOperand(0), SI.getParent());
12854 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12855 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12858 // Advance to a place where it is safe to insert the new store and
12860 BBI = DestBB->getFirstNonPHI();
12861 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12862 OtherStore->isVolatile(),
12863 SI.getAlignment()), *BBI);
12865 // Nuke the old stores.
12866 EraseInstFromFunction(SI);
12867 EraseInstFromFunction(*OtherStore);
12873 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12874 // Change br (not X), label True, label False to: br X, label False, True
12876 BasicBlock *TrueDest;
12877 BasicBlock *FalseDest;
12878 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
12879 !isa<Constant>(X)) {
12880 // Swap Destinations and condition...
12881 BI.setCondition(X);
12882 BI.setSuccessor(0, FalseDest);
12883 BI.setSuccessor(1, TrueDest);
12887 // Cannonicalize fcmp_one -> fcmp_oeq
12888 FCmpInst::Predicate FPred; Value *Y;
12889 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12890 TrueDest, FalseDest)) &&
12891 BI.getCondition()->hasOneUse())
12892 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12893 FPred == FCmpInst::FCMP_OGE) {
12894 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
12895 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
12897 // Swap Destinations and condition.
12898 BI.setSuccessor(0, FalseDest);
12899 BI.setSuccessor(1, TrueDest);
12900 Worklist.Add(Cond);
12904 // Cannonicalize icmp_ne -> icmp_eq
12905 ICmpInst::Predicate IPred;
12906 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12907 TrueDest, FalseDest)) &&
12908 BI.getCondition()->hasOneUse())
12909 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12910 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12911 IPred == ICmpInst::ICMP_SGE) {
12912 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
12913 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
12914 // Swap Destinations and condition.
12915 BI.setSuccessor(0, FalseDest);
12916 BI.setSuccessor(1, TrueDest);
12917 Worklist.Add(Cond);
12924 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12925 Value *Cond = SI.getCondition();
12926 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12927 if (I->getOpcode() == Instruction::Add)
12928 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12929 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12930 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12932 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12934 SI.setOperand(0, I->getOperand(0));
12942 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12943 Value *Agg = EV.getAggregateOperand();
12945 if (!EV.hasIndices())
12946 return ReplaceInstUsesWith(EV, Agg);
12948 if (Constant *C = dyn_cast<Constant>(Agg)) {
12949 if (isa<UndefValue>(C))
12950 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12952 if (isa<ConstantAggregateZero>(C))
12953 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12955 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12956 // Extract the element indexed by the first index out of the constant
12957 Value *V = C->getOperand(*EV.idx_begin());
12958 if (EV.getNumIndices() > 1)
12959 // Extract the remaining indices out of the constant indexed by the
12961 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12963 return ReplaceInstUsesWith(EV, V);
12965 return 0; // Can't handle other constants
12967 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12968 // We're extracting from an insertvalue instruction, compare the indices
12969 const unsigned *exti, *exte, *insi, *inse;
12970 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12971 exte = EV.idx_end(), inse = IV->idx_end();
12972 exti != exte && insi != inse;
12974 if (*insi != *exti)
12975 // The insert and extract both reference distinctly different elements.
12976 // This means the extract is not influenced by the insert, and we can
12977 // replace the aggregate operand of the extract with the aggregate
12978 // operand of the insert. i.e., replace
12979 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12980 // %E = extractvalue { i32, { i32 } } %I, 0
12982 // %E = extractvalue { i32, { i32 } } %A, 0
12983 return ExtractValueInst::Create(IV->getAggregateOperand(),
12984 EV.idx_begin(), EV.idx_end());
12986 if (exti == exte && insi == inse)
12987 // Both iterators are at the end: Index lists are identical. Replace
12988 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12989 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12991 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12992 if (exti == exte) {
12993 // The extract list is a prefix of the insert list. i.e. replace
12994 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12995 // %E = extractvalue { i32, { i32 } } %I, 1
12997 // %X = extractvalue { i32, { i32 } } %A, 1
12998 // %E = insertvalue { i32 } %X, i32 42, 0
12999 // by switching the order of the insert and extract (though the
13000 // insertvalue should be left in, since it may have other uses).
13001 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
13002 EV.idx_begin(), EV.idx_end());
13003 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
13007 // The insert list is a prefix of the extract list
13008 // We can simply remove the common indices from the extract and make it
13009 // operate on the inserted value instead of the insertvalue result.
13011 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
13012 // %E = extractvalue { i32, { i32 } } %I, 1, 0
13014 // %E extractvalue { i32 } { i32 42 }, 0
13015 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
13018 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
13019 // We're extracting from an intrinsic, see if we're the only user, which
13020 // allows us to simplify multiple result intrinsics to simpler things that
13021 // just get one value..
13022 if (II->hasOneUse()) {
13023 // Check if we're grabbing the overflow bit or the result of a 'with
13024 // overflow' intrinsic. If it's the latter we can remove the intrinsic
13025 // and replace it with a traditional binary instruction.
13026 switch (II->getIntrinsicID()) {
13027 case Intrinsic::uadd_with_overflow:
13028 case Intrinsic::sadd_with_overflow:
13029 if (*EV.idx_begin() == 0) { // Normal result.
13030 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
13031 II->replaceAllUsesWith(UndefValue::get(II->getType()));
13032 EraseInstFromFunction(*II);
13033 return BinaryOperator::CreateAdd(LHS, RHS);
13036 case Intrinsic::usub_with_overflow:
13037 case Intrinsic::ssub_with_overflow:
13038 if (*EV.idx_begin() == 0) { // Normal result.
13039 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
13040 II->replaceAllUsesWith(UndefValue::get(II->getType()));
13041 EraseInstFromFunction(*II);
13042 return BinaryOperator::CreateSub(LHS, RHS);
13045 case Intrinsic::umul_with_overflow:
13046 case Intrinsic::smul_with_overflow:
13047 if (*EV.idx_begin() == 0) { // Normal result.
13048 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
13049 II->replaceAllUsesWith(UndefValue::get(II->getType()));
13050 EraseInstFromFunction(*II);
13051 return BinaryOperator::CreateMul(LHS, RHS);
13059 // Can't simplify extracts from other values. Note that nested extracts are
13060 // already simplified implicitely by the above (extract ( extract (insert) )
13061 // will be translated into extract ( insert ( extract ) ) first and then just
13062 // the value inserted, if appropriate).
13066 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
13067 /// is to leave as a vector operation.
13068 static bool CheapToScalarize(Value *V, bool isConstant) {
13069 if (isa<ConstantAggregateZero>(V))
13071 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
13072 if (isConstant) return true;
13073 // If all elts are the same, we can extract.
13074 Constant *Op0 = C->getOperand(0);
13075 for (unsigned i = 1; i < C->getNumOperands(); ++i)
13076 if (C->getOperand(i) != Op0)
13080 Instruction *I = dyn_cast<Instruction>(V);
13081 if (!I) return false;
13083 // Insert element gets simplified to the inserted element or is deleted if
13084 // this is constant idx extract element and its a constant idx insertelt.
13085 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
13086 isa<ConstantInt>(I->getOperand(2)))
13088 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
13090 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
13091 if (BO->hasOneUse() &&
13092 (CheapToScalarize(BO->getOperand(0), isConstant) ||
13093 CheapToScalarize(BO->getOperand(1), isConstant)))
13095 if (CmpInst *CI = dyn_cast<CmpInst>(I))
13096 if (CI->hasOneUse() &&
13097 (CheapToScalarize(CI->getOperand(0), isConstant) ||
13098 CheapToScalarize(CI->getOperand(1), isConstant)))
13104 /// Read and decode a shufflevector mask.
13106 /// It turns undef elements into values that are larger than the number of
13107 /// elements in the input.
13108 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
13109 unsigned NElts = SVI->getType()->getNumElements();
13110 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
13111 return std::vector<unsigned>(NElts, 0);
13112 if (isa<UndefValue>(SVI->getOperand(2)))
13113 return std::vector<unsigned>(NElts, 2*NElts);
13115 std::vector<unsigned> Result;
13116 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
13117 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
13118 if (isa<UndefValue>(*i))
13119 Result.push_back(NElts*2); // undef -> 8
13121 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
13125 /// FindScalarElement - Given a vector and an element number, see if the scalar
13126 /// value is already around as a register, for example if it were inserted then
13127 /// extracted from the vector.
13128 static Value *FindScalarElement(Value *V, unsigned EltNo,
13129 LLVMContext *Context) {
13130 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
13131 const VectorType *PTy = cast<VectorType>(V->getType());
13132 unsigned Width = PTy->getNumElements();
13133 if (EltNo >= Width) // Out of range access.
13134 return UndefValue::get(PTy->getElementType());
13136 if (isa<UndefValue>(V))
13137 return UndefValue::get(PTy->getElementType());
13138 else if (isa<ConstantAggregateZero>(V))
13139 return Constant::getNullValue(PTy->getElementType());
13140 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
13141 return CP->getOperand(EltNo);
13142 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
13143 // If this is an insert to a variable element, we don't know what it is.
13144 if (!isa<ConstantInt>(III->getOperand(2)))
13146 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
13148 // If this is an insert to the element we are looking for, return the
13150 if (EltNo == IIElt)
13151 return III->getOperand(1);
13153 // Otherwise, the insertelement doesn't modify the value, recurse on its
13155 return FindScalarElement(III->getOperand(0), EltNo, Context);
13156 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
13157 unsigned LHSWidth =
13158 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
13159 unsigned InEl = getShuffleMask(SVI)[EltNo];
13160 if (InEl < LHSWidth)
13161 return FindScalarElement(SVI->getOperand(0), InEl, Context);
13162 else if (InEl < LHSWidth*2)
13163 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
13165 return UndefValue::get(PTy->getElementType());
13168 // Otherwise, we don't know.
13172 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
13173 // If vector val is undef, replace extract with scalar undef.
13174 if (isa<UndefValue>(EI.getOperand(0)))
13175 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
13177 // If vector val is constant 0, replace extract with scalar 0.
13178 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
13179 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
13181 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
13182 // If vector val is constant with all elements the same, replace EI with
13183 // that element. When the elements are not identical, we cannot replace yet
13184 // (we do that below, but only when the index is constant).
13185 Constant *op0 = C->getOperand(0);
13186 for (unsigned i = 1; i != C->getNumOperands(); ++i)
13187 if (C->getOperand(i) != op0) {
13192 return ReplaceInstUsesWith(EI, op0);
13195 // If extracting a specified index from the vector, see if we can recursively
13196 // find a previously computed scalar that was inserted into the vector.
13197 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
13198 unsigned IndexVal = IdxC->getZExtValue();
13199 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
13201 // If this is extracting an invalid index, turn this into undef, to avoid
13202 // crashing the code below.
13203 if (IndexVal >= VectorWidth)
13204 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
13206 // This instruction only demands the single element from the input vector.
13207 // If the input vector has a single use, simplify it based on this use
13209 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
13210 APInt UndefElts(VectorWidth, 0);
13211 APInt DemandedMask(VectorWidth, 1 << IndexVal);
13212 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
13213 DemandedMask, UndefElts)) {
13214 EI.setOperand(0, V);
13219 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
13220 return ReplaceInstUsesWith(EI, Elt);
13222 // If the this extractelement is directly using a bitcast from a vector of
13223 // the same number of elements, see if we can find the source element from
13224 // it. In this case, we will end up needing to bitcast the scalars.
13225 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
13226 if (const VectorType *VT =
13227 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
13228 if (VT->getNumElements() == VectorWidth)
13229 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
13230 IndexVal, Context))
13231 return new BitCastInst(Elt, EI.getType());
13235 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
13236 // Push extractelement into predecessor operation if legal and
13237 // profitable to do so
13238 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
13239 if (I->hasOneUse() &&
13240 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
13242 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
13243 EI.getName()+".lhs");
13245 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
13246 EI.getName()+".rhs");
13247 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
13249 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
13250 // Extracting the inserted element?
13251 if (IE->getOperand(2) == EI.getOperand(1))
13252 return ReplaceInstUsesWith(EI, IE->getOperand(1));
13253 // If the inserted and extracted elements are constants, they must not
13254 // be the same value, extract from the pre-inserted value instead.
13255 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
13256 Worklist.AddValue(EI.getOperand(0));
13257 EI.setOperand(0, IE->getOperand(0));
13260 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
13261 // If this is extracting an element from a shufflevector, figure out where
13262 // it came from and extract from the appropriate input element instead.
13263 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
13264 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
13266 unsigned LHSWidth =
13267 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
13269 if (SrcIdx < LHSWidth)
13270 Src = SVI->getOperand(0);
13271 else if (SrcIdx < LHSWidth*2) {
13272 SrcIdx -= LHSWidth;
13273 Src = SVI->getOperand(1);
13275 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
13277 return ExtractElementInst::Create(Src,
13278 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx,
13282 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
13287 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
13288 /// elements from either LHS or RHS, return the shuffle mask and true.
13289 /// Otherwise, return false.
13290 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
13291 std::vector<Constant*> &Mask,
13292 LLVMContext *Context) {
13293 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
13294 "Invalid CollectSingleShuffleElements");
13295 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
13297 if (isa<UndefValue>(V)) {
13298 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
13300 } else if (V == LHS) {
13301 for (unsigned i = 0; i != NumElts; ++i)
13302 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
13304 } else if (V == RHS) {
13305 for (unsigned i = 0; i != NumElts; ++i)
13306 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
13308 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
13309 // If this is an insert of an extract from some other vector, include it.
13310 Value *VecOp = IEI->getOperand(0);
13311 Value *ScalarOp = IEI->getOperand(1);
13312 Value *IdxOp = IEI->getOperand(2);
13314 if (!isa<ConstantInt>(IdxOp))
13316 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
13318 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
13319 // Okay, we can handle this if the vector we are insertinting into is
13320 // transitively ok.
13321 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
13322 // If so, update the mask to reflect the inserted undef.
13323 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
13326 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
13327 if (isa<ConstantInt>(EI->getOperand(1)) &&
13328 EI->getOperand(0)->getType() == V->getType()) {
13329 unsigned ExtractedIdx =
13330 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
13332 // This must be extracting from either LHS or RHS.
13333 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
13334 // Okay, we can handle this if the vector we are insertinting into is
13335 // transitively ok.
13336 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
13337 // If so, update the mask to reflect the inserted value.
13338 if (EI->getOperand(0) == LHS) {
13339 Mask[InsertedIdx % NumElts] =
13340 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
13342 assert(EI->getOperand(0) == RHS);
13343 Mask[InsertedIdx % NumElts] =
13344 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
13353 // TODO: Handle shufflevector here!
13358 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
13359 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
13360 /// that computes V and the LHS value of the shuffle.
13361 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
13362 Value *&RHS, LLVMContext *Context) {
13363 assert(isa<VectorType>(V->getType()) &&
13364 (RHS == 0 || V->getType() == RHS->getType()) &&
13365 "Invalid shuffle!");
13366 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
13368 if (isa<UndefValue>(V)) {
13369 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
13371 } else if (isa<ConstantAggregateZero>(V)) {
13372 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
13374 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
13375 // If this is an insert of an extract from some other vector, include it.
13376 Value *VecOp = IEI->getOperand(0);
13377 Value *ScalarOp = IEI->getOperand(1);
13378 Value *IdxOp = IEI->getOperand(2);
13380 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
13381 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
13382 EI->getOperand(0)->getType() == V->getType()) {
13383 unsigned ExtractedIdx =
13384 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
13385 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
13387 // Either the extracted from or inserted into vector must be RHSVec,
13388 // otherwise we'd end up with a shuffle of three inputs.
13389 if (EI->getOperand(0) == RHS || RHS == 0) {
13390 RHS = EI->getOperand(0);
13391 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
13392 Mask[InsertedIdx % NumElts] =
13393 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
13397 if (VecOp == RHS) {
13398 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
13400 // Everything but the extracted element is replaced with the RHS.
13401 for (unsigned i = 0; i != NumElts; ++i) {
13402 if (i != InsertedIdx)
13403 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
13408 // If this insertelement is a chain that comes from exactly these two
13409 // vectors, return the vector and the effective shuffle.
13410 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
13412 return EI->getOperand(0);
13417 // TODO: Handle shufflevector here!
13419 // Otherwise, can't do anything fancy. Return an identity vector.
13420 for (unsigned i = 0; i != NumElts; ++i)
13421 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
13425 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
13426 Value *VecOp = IE.getOperand(0);
13427 Value *ScalarOp = IE.getOperand(1);
13428 Value *IdxOp = IE.getOperand(2);
13430 // Inserting an undef or into an undefined place, remove this.
13431 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
13432 ReplaceInstUsesWith(IE, VecOp);
13434 // If the inserted element was extracted from some other vector, and if the
13435 // indexes are constant, try to turn this into a shufflevector operation.
13436 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
13437 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
13438 EI->getOperand(0)->getType() == IE.getType()) {
13439 unsigned NumVectorElts = IE.getType()->getNumElements();
13440 unsigned ExtractedIdx =
13441 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
13442 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
13444 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
13445 return ReplaceInstUsesWith(IE, VecOp);
13447 if (InsertedIdx >= NumVectorElts) // Out of range insert.
13448 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
13450 // If we are extracting a value from a vector, then inserting it right
13451 // back into the same place, just use the input vector.
13452 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
13453 return ReplaceInstUsesWith(IE, VecOp);
13455 // If this insertelement isn't used by some other insertelement, turn it
13456 // (and any insertelements it points to), into one big shuffle.
13457 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
13458 std::vector<Constant*> Mask;
13460 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
13461 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
13462 // We now have a shuffle of LHS, RHS, Mask.
13463 return new ShuffleVectorInst(LHS, RHS,
13464 ConstantVector::get(Mask));
13469 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
13470 APInt UndefElts(VWidth, 0);
13471 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
13472 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
13479 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
13480 Value *LHS = SVI.getOperand(0);
13481 Value *RHS = SVI.getOperand(1);
13482 std::vector<unsigned> Mask = getShuffleMask(&SVI);
13484 bool MadeChange = false;
13486 // Undefined shuffle mask -> undefined value.
13487 if (isa<UndefValue>(SVI.getOperand(2)))
13488 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
13490 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
13492 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
13495 APInt UndefElts(VWidth, 0);
13496 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
13497 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
13498 LHS = SVI.getOperand(0);
13499 RHS = SVI.getOperand(1);
13503 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
13504 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
13505 if (LHS == RHS || isa<UndefValue>(LHS)) {
13506 if (isa<UndefValue>(LHS) && LHS == RHS) {
13507 // shuffle(undef,undef,mask) -> undef.
13508 return ReplaceInstUsesWith(SVI, LHS);
13511 // Remap any references to RHS to use LHS.
13512 std::vector<Constant*> Elts;
13513 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
13514 if (Mask[i] >= 2*e)
13515 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
13517 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
13518 (Mask[i] < e && isa<UndefValue>(LHS))) {
13519 Mask[i] = 2*e; // Turn into undef.
13520 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
13522 Mask[i] = Mask[i] % e; // Force to LHS.
13523 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
13527 SVI.setOperand(0, SVI.getOperand(1));
13528 SVI.setOperand(1, UndefValue::get(RHS->getType()));
13529 SVI.setOperand(2, ConstantVector::get(Elts));
13530 LHS = SVI.getOperand(0);
13531 RHS = SVI.getOperand(1);
13535 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
13536 bool isLHSID = true, isRHSID = true;
13538 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
13539 if (Mask[i] >= e*2) continue; // Ignore undef values.
13540 // Is this an identity shuffle of the LHS value?
13541 isLHSID &= (Mask[i] == i);
13543 // Is this an identity shuffle of the RHS value?
13544 isRHSID &= (Mask[i]-e == i);
13547 // Eliminate identity shuffles.
13548 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
13549 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
13551 // If the LHS is a shufflevector itself, see if we can combine it with this
13552 // one without producing an unusual shuffle. Here we are really conservative:
13553 // we are absolutely afraid of producing a shuffle mask not in the input
13554 // program, because the code gen may not be smart enough to turn a merged
13555 // shuffle into two specific shuffles: it may produce worse code. As such,
13556 // we only merge two shuffles if the result is one of the two input shuffle
13557 // masks. In this case, merging the shuffles just removes one instruction,
13558 // which we know is safe. This is good for things like turning:
13559 // (splat(splat)) -> splat.
13560 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
13561 if (isa<UndefValue>(RHS)) {
13562 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
13564 if (LHSMask.size() == Mask.size()) {
13565 std::vector<unsigned> NewMask;
13566 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
13568 NewMask.push_back(2*e);
13570 NewMask.push_back(LHSMask[Mask[i]]);
13572 // If the result mask is equal to the src shuffle or this
13573 // shuffle mask, do the replacement.
13574 if (NewMask == LHSMask || NewMask == Mask) {
13575 unsigned LHSInNElts =
13576 cast<VectorType>(LHSSVI->getOperand(0)->getType())->
13578 std::vector<Constant*> Elts;
13579 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
13580 if (NewMask[i] >= LHSInNElts*2) {
13581 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
13583 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
13587 return new ShuffleVectorInst(LHSSVI->getOperand(0),
13588 LHSSVI->getOperand(1),
13589 ConstantVector::get(Elts));
13595 return MadeChange ? &SVI : 0;
13601 /// TryToSinkInstruction - Try to move the specified instruction from its
13602 /// current block into the beginning of DestBlock, which can only happen if it's
13603 /// safe to move the instruction past all of the instructions between it and the
13604 /// end of its block.
13605 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
13606 assert(I->hasOneUse() && "Invariants didn't hold!");
13608 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
13609 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
13612 // Do not sink alloca instructions out of the entry block.
13613 if (isa<AllocaInst>(I) && I->getParent() ==
13614 &DestBlock->getParent()->getEntryBlock())
13617 // We can only sink load instructions if there is nothing between the load and
13618 // the end of block that could change the value.
13619 if (I->mayReadFromMemory()) {
13620 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
13622 if (Scan->mayWriteToMemory())
13626 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
13628 CopyPrecedingStopPoint(I, InsertPos);
13629 I->moveBefore(InsertPos);
13635 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
13636 /// all reachable code to the worklist.
13638 /// This has a couple of tricks to make the code faster and more powerful. In
13639 /// particular, we constant fold and DCE instructions as we go, to avoid adding
13640 /// them to the worklist (this significantly speeds up instcombine on code where
13641 /// many instructions are dead or constant). Additionally, if we find a branch
13642 /// whose condition is a known constant, we only visit the reachable successors.
13644 static bool AddReachableCodeToWorklist(BasicBlock *BB,
13645 SmallPtrSet<BasicBlock*, 64> &Visited,
13647 const TargetData *TD) {
13648 bool MadeIRChange = false;
13649 SmallVector<BasicBlock*, 256> Worklist;
13650 Worklist.push_back(BB);
13652 std::vector<Instruction*> InstrsForInstCombineWorklist;
13653 InstrsForInstCombineWorklist.reserve(128);
13655 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
13657 while (!Worklist.empty()) {
13658 BB = Worklist.back();
13659 Worklist.pop_back();
13661 // We have now visited this block! If we've already been here, ignore it.
13662 if (!Visited.insert(BB)) continue;
13664 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
13665 Instruction *Inst = BBI++;
13667 // DCE instruction if trivially dead.
13668 if (isInstructionTriviallyDead(Inst)) {
13670 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
13671 Inst->eraseFromParent();
13675 // ConstantProp instruction if trivially constant.
13676 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
13677 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
13678 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
13680 Inst->replaceAllUsesWith(C);
13682 Inst->eraseFromParent();
13689 // See if we can constant fold its operands.
13690 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
13692 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
13693 if (CE == 0) continue;
13695 // If we already folded this constant, don't try again.
13696 if (!FoldedConstants.insert(CE))
13699 Constant *NewC = ConstantFoldConstantExpression(CE, TD);
13700 if (NewC && NewC != CE) {
13702 MadeIRChange = true;
13708 InstrsForInstCombineWorklist.push_back(Inst);
13711 // Recursively visit successors. If this is a branch or switch on a
13712 // constant, only visit the reachable successor.
13713 TerminatorInst *TI = BB->getTerminator();
13714 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
13715 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
13716 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
13717 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
13718 Worklist.push_back(ReachableBB);
13721 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
13722 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
13723 // See if this is an explicit destination.
13724 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
13725 if (SI->getCaseValue(i) == Cond) {
13726 BasicBlock *ReachableBB = SI->getSuccessor(i);
13727 Worklist.push_back(ReachableBB);
13731 // Otherwise it is the default destination.
13732 Worklist.push_back(SI->getSuccessor(0));
13737 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
13738 Worklist.push_back(TI->getSuccessor(i));
13741 // Once we've found all of the instructions to add to instcombine's worklist,
13742 // add them in reverse order. This way instcombine will visit from the top
13743 // of the function down. This jives well with the way that it adds all uses
13744 // of instructions to the worklist after doing a transformation, thus avoiding
13745 // some N^2 behavior in pathological cases.
13746 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
13747 InstrsForInstCombineWorklist.size());
13749 return MadeIRChange;
13752 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
13753 MadeIRChange = false;
13755 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
13756 << F.getNameStr() << "\n");
13759 // Do a depth-first traversal of the function, populate the worklist with
13760 // the reachable instructions. Ignore blocks that are not reachable. Keep
13761 // track of which blocks we visit.
13762 SmallPtrSet<BasicBlock*, 64> Visited;
13763 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
13765 // Do a quick scan over the function. If we find any blocks that are
13766 // unreachable, remove any instructions inside of them. This prevents
13767 // the instcombine code from having to deal with some bad special cases.
13768 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
13769 if (!Visited.count(BB)) {
13770 Instruction *Term = BB->getTerminator();
13771 while (Term != BB->begin()) { // Remove instrs bottom-up
13772 BasicBlock::iterator I = Term; --I;
13774 DEBUG(errs() << "IC: DCE: " << *I << '\n');
13775 // A debug intrinsic shouldn't force another iteration if we weren't
13776 // going to do one without it.
13777 if (!isa<DbgInfoIntrinsic>(I)) {
13779 MadeIRChange = true;
13782 // If I is not void type then replaceAllUsesWith undef.
13783 // This allows ValueHandlers and custom metadata to adjust itself.
13784 if (!I->getType()->isVoidTy())
13785 I->replaceAllUsesWith(UndefValue::get(I->getType()));
13786 I->eraseFromParent();
13791 while (!Worklist.isEmpty()) {
13792 Instruction *I = Worklist.RemoveOne();
13793 if (I == 0) continue; // skip null values.
13795 // Check to see if we can DCE the instruction.
13796 if (isInstructionTriviallyDead(I)) {
13797 DEBUG(errs() << "IC: DCE: " << *I << '\n');
13798 EraseInstFromFunction(*I);
13800 MadeIRChange = true;
13804 // Instruction isn't dead, see if we can constant propagate it.
13805 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
13806 if (Constant *C = ConstantFoldInstruction(I, TD)) {
13807 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
13809 // Add operands to the worklist.
13810 ReplaceInstUsesWith(*I, C);
13812 EraseInstFromFunction(*I);
13813 MadeIRChange = true;
13817 // See if we can trivially sink this instruction to a successor basic block.
13818 if (I->hasOneUse()) {
13819 BasicBlock *BB = I->getParent();
13820 Instruction *UserInst = cast<Instruction>(I->use_back());
13821 BasicBlock *UserParent;
13823 // Get the block the use occurs in.
13824 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
13825 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
13827 UserParent = UserInst->getParent();
13829 if (UserParent != BB) {
13830 bool UserIsSuccessor = false;
13831 // See if the user is one of our successors.
13832 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13833 if (*SI == UserParent) {
13834 UserIsSuccessor = true;
13838 // If the user is one of our immediate successors, and if that successor
13839 // only has us as a predecessors (we'd have to split the critical edge
13840 // otherwise), we can keep going.
13841 if (UserIsSuccessor && UserParent->getSinglePredecessor())
13842 // Okay, the CFG is simple enough, try to sink this instruction.
13843 MadeIRChange |= TryToSinkInstruction(I, UserParent);
13847 // Now that we have an instruction, try combining it to simplify it.
13848 Builder->SetInsertPoint(I->getParent(), I);
13853 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
13854 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
13856 if (Instruction *Result = visit(*I)) {
13858 // Should we replace the old instruction with a new one?
13860 DEBUG(errs() << "IC: Old = " << *I << '\n'
13861 << " New = " << *Result << '\n');
13863 // Everything uses the new instruction now.
13864 I->replaceAllUsesWith(Result);
13866 // Push the new instruction and any users onto the worklist.
13867 Worklist.Add(Result);
13868 Worklist.AddUsersToWorkList(*Result);
13870 // Move the name to the new instruction first.
13871 Result->takeName(I);
13873 // Insert the new instruction into the basic block...
13874 BasicBlock *InstParent = I->getParent();
13875 BasicBlock::iterator InsertPos = I;
13877 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13878 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13881 InstParent->getInstList().insert(InsertPos, Result);
13883 EraseInstFromFunction(*I);
13886 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
13887 << " New = " << *I << '\n');
13890 // If the instruction was modified, it's possible that it is now dead.
13891 // if so, remove it.
13892 if (isInstructionTriviallyDead(I)) {
13893 EraseInstFromFunction(*I);
13896 Worklist.AddUsersToWorkList(*I);
13899 MadeIRChange = true;
13904 return MadeIRChange;
13908 bool InstCombiner::runOnFunction(Function &F) {
13909 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13910 Context = &F.getContext();
13911 TD = getAnalysisIfAvailable<TargetData>();
13914 /// Builder - This is an IRBuilder that automatically inserts new
13915 /// instructions into the worklist when they are created.
13916 IRBuilder<true, TargetFolder, InstCombineIRInserter>
13917 TheBuilder(F.getContext(), TargetFolder(TD),
13918 InstCombineIRInserter(Worklist));
13919 Builder = &TheBuilder;
13921 bool EverMadeChange = false;
13923 // Iterate while there is work to do.
13924 unsigned Iteration = 0;
13925 while (DoOneIteration(F, Iteration++))
13926 EverMadeChange = true;
13929 return EverMadeChange;
13932 FunctionPass *llvm::createInstructionCombiningPass() {
13933 return new InstCombiner();