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 *visitFCmpInst(FCmpInst &I);
262 Instruction *visitICmpInst(ICmpInst &I);
263 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
264 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
267 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
268 ConstantInt *DivRHS);
269 Instruction *FoldICmpAddOpCst(ICmpInst &ICI, Value *X, ConstantInt *CI,
270 ICmpInst::Predicate Pred, Value *TheAdd);
271 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
272 ICmpInst::Predicate Cond, Instruction &I);
273 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
275 Instruction *commonCastTransforms(CastInst &CI);
276 Instruction *commonIntCastTransforms(CastInst &CI);
277 Instruction *commonPointerCastTransforms(CastInst &CI);
278 Instruction *visitTrunc(TruncInst &CI);
279 Instruction *visitZExt(ZExtInst &CI);
280 Instruction *visitSExt(SExtInst &CI);
281 Instruction *visitFPTrunc(FPTruncInst &CI);
282 Instruction *visitFPExt(CastInst &CI);
283 Instruction *visitFPToUI(FPToUIInst &FI);
284 Instruction *visitFPToSI(FPToSIInst &FI);
285 Instruction *visitUIToFP(CastInst &CI);
286 Instruction *visitSIToFP(CastInst &CI);
287 Instruction *visitPtrToInt(PtrToIntInst &CI);
288 Instruction *visitIntToPtr(IntToPtrInst &CI);
289 Instruction *visitBitCast(BitCastInst &CI);
290 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
292 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
293 Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
294 Value *A, Value *B, Instruction &Outer,
295 SelectPatternFlavor SPF2, Value *C);
296 Instruction *visitSelectInst(SelectInst &SI);
297 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
298 Instruction *visitCallInst(CallInst &CI);
299 Instruction *visitInvokeInst(InvokeInst &II);
301 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
302 Instruction *visitPHINode(PHINode &PN);
303 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
304 Instruction *visitAllocaInst(AllocaInst &AI);
305 Instruction *visitFree(Instruction &FI);
306 Instruction *visitLoadInst(LoadInst &LI);
307 Instruction *visitStoreInst(StoreInst &SI);
308 Instruction *visitBranchInst(BranchInst &BI);
309 Instruction *visitSwitchInst(SwitchInst &SI);
310 Instruction *visitInsertElementInst(InsertElementInst &IE);
311 Instruction *visitExtractElementInst(ExtractElementInst &EI);
312 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
313 Instruction *visitExtractValueInst(ExtractValueInst &EV);
315 // visitInstruction - Specify what to return for unhandled instructions...
316 Instruction *visitInstruction(Instruction &I) { return 0; }
319 Instruction *visitCallSite(CallSite CS);
320 bool transformConstExprCastCall(CallSite CS);
321 Instruction *transformCallThroughTrampoline(CallSite CS);
322 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
323 bool DoXform = true);
324 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
325 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
329 // InsertNewInstBefore - insert an instruction New before instruction Old
330 // in the program. Add the new instruction to the worklist.
332 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
333 assert(New && New->getParent() == 0 &&
334 "New instruction already inserted into a basic block!");
335 BasicBlock *BB = Old.getParent();
336 BB->getInstList().insert(&Old, New); // Insert inst
341 // ReplaceInstUsesWith - This method is to be used when an instruction is
342 // found to be dead, replacable with another preexisting expression. Here
343 // we add all uses of I to the worklist, replace all uses of I with the new
344 // value, then return I, so that the inst combiner will know that I was
347 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
348 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
350 // If we are replacing the instruction with itself, this must be in a
351 // segment of unreachable code, so just clobber the instruction.
353 V = UndefValue::get(I.getType());
355 I.replaceAllUsesWith(V);
359 // EraseInstFromFunction - When dealing with an instruction that has side
360 // effects or produces a void value, we can't rely on DCE to delete the
361 // instruction. Instead, visit methods should return the value returned by
363 Instruction *EraseInstFromFunction(Instruction &I) {
364 DEBUG(errs() << "IC: ERASE " << I << '\n');
366 assert(I.use_empty() && "Cannot erase instruction that is used!");
367 // Make sure that we reprocess all operands now that we reduced their
369 if (I.getNumOperands() < 8) {
370 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
371 if (Instruction *Op = dyn_cast<Instruction>(*i))
377 return 0; // Don't do anything with FI
380 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
381 APInt &KnownOne, unsigned Depth = 0) const {
382 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
385 bool MaskedValueIsZero(Value *V, const APInt &Mask,
386 unsigned Depth = 0) const {
387 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
389 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
390 return llvm::ComputeNumSignBits(Op, TD, Depth);
395 /// SimplifyCommutative - This performs a few simplifications for
396 /// commutative operators.
397 bool SimplifyCommutative(BinaryOperator &I);
399 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
400 /// based on the demanded bits.
401 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
402 APInt& KnownZero, APInt& KnownOne,
404 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
405 APInt& KnownZero, APInt& KnownOne,
408 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
409 /// SimplifyDemandedBits knows about. See if the instruction has any
410 /// properties that allow us to simplify its operands.
411 bool SimplifyDemandedInstructionBits(Instruction &Inst);
413 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
414 APInt& UndefElts, unsigned Depth = 0);
416 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
417 // which has a PHI node as operand #0, see if we can fold the instruction
418 // into the PHI (which is only possible if all operands to the PHI are
421 // If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
422 // that would normally be unprofitable because they strongly encourage jump
424 Instruction *FoldOpIntoPhi(Instruction &I, bool AllowAggressive = false);
426 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
427 // operator and they all are only used by the PHI, PHI together their
428 // inputs, and do the operation once, to the result of the PHI.
429 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
430 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
431 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
432 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
435 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
436 ConstantInt *AndRHS, BinaryOperator &TheAnd);
438 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
439 bool isSub, Instruction &I);
440 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
441 bool isSigned, bool Inside, Instruction &IB);
442 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
443 Instruction *MatchBSwap(BinaryOperator &I);
444 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
445 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
446 Instruction *SimplifyMemSet(MemSetInst *MI);
449 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
451 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
452 unsigned CastOpc, int &NumCastsRemoved);
453 unsigned GetOrEnforceKnownAlignment(Value *V,
454 unsigned PrefAlign = 0);
457 } // end anonymous namespace
459 char InstCombiner::ID = 0;
460 static RegisterPass<InstCombiner>
461 X("instcombine", "Combine redundant instructions");
463 // getComplexity: Assign a complexity or rank value to LLVM Values...
464 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
465 static unsigned getComplexity(Value *V) {
466 if (isa<Instruction>(V)) {
467 if (BinaryOperator::isNeg(V) ||
468 BinaryOperator::isFNeg(V) ||
469 BinaryOperator::isNot(V))
473 if (isa<Argument>(V)) return 3;
474 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
477 // isOnlyUse - Return true if this instruction will be deleted if we stop using
479 static bool isOnlyUse(Value *V) {
480 return V->hasOneUse() || isa<Constant>(V);
483 // getPromotedType - Return the specified type promoted as it would be to pass
484 // though a va_arg area...
485 static const Type *getPromotedType(const Type *Ty) {
486 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
487 if (ITy->getBitWidth() < 32)
488 return Type::getInt32Ty(Ty->getContext());
493 /// ShouldChangeType - Return true if it is desirable to convert a computation
494 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
495 /// type for example, or from a smaller to a larger illegal type.
496 static bool ShouldChangeType(const Type *From, const Type *To,
497 const TargetData *TD) {
498 assert(isa<IntegerType>(From) && isa<IntegerType>(To));
500 // If we don't have TD, we don't know if the source/dest are legal.
501 if (!TD) return false;
503 unsigned FromWidth = From->getPrimitiveSizeInBits();
504 unsigned ToWidth = To->getPrimitiveSizeInBits();
505 bool FromLegal = TD->isLegalInteger(FromWidth);
506 bool ToLegal = TD->isLegalInteger(ToWidth);
508 // If this is a legal integer from type, and the result would be an illegal
509 // type, don't do the transformation.
510 if (FromLegal && !ToLegal)
513 // Otherwise, if both are illegal, do not increase the size of the result. We
514 // do allow things like i160 -> i64, but not i64 -> i160.
515 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
521 /// getBitCastOperand - If the specified operand is a CastInst, a constant
522 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
523 /// operand value, otherwise return null.
524 static Value *getBitCastOperand(Value *V) {
525 if (Operator *O = dyn_cast<Operator>(V)) {
526 if (O->getOpcode() == Instruction::BitCast)
527 return O->getOperand(0);
528 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
529 if (GEP->hasAllZeroIndices())
530 return GEP->getPointerOperand();
535 /// This function is a wrapper around CastInst::isEliminableCastPair. It
536 /// simply extracts arguments and returns what that function returns.
537 static Instruction::CastOps
538 isEliminableCastPair(
539 const CastInst *CI, ///< The first cast instruction
540 unsigned opcode, ///< The opcode of the second cast instruction
541 const Type *DstTy, ///< The target type for the second cast instruction
542 TargetData *TD ///< The target data for pointer size
545 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
546 const Type *MidTy = CI->getType(); // B from above
548 // Get the opcodes of the two Cast instructions
549 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
550 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
552 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
554 TD ? TD->getIntPtrType(CI->getContext()) : 0);
556 // We don't want to form an inttoptr or ptrtoint that converts to an integer
557 // type that differs from the pointer size.
558 if ((Res == Instruction::IntToPtr &&
559 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
560 (Res == Instruction::PtrToInt &&
561 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
564 return Instruction::CastOps(Res);
567 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
568 /// in any code being generated. It does not require codegen if V is simple
569 /// enough or if the cast can be folded into other casts.
570 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
571 const Type *Ty, TargetData *TD) {
572 if (V->getType() == Ty || isa<Constant>(V)) return false;
574 // If this is another cast that can be eliminated, it isn't codegen either.
575 if (const CastInst *CI = dyn_cast<CastInst>(V))
576 if (isEliminableCastPair(CI, opcode, Ty, TD))
581 // SimplifyCommutative - This performs a few simplifications for commutative
584 // 1. Order operands such that they are listed from right (least complex) to
585 // left (most complex). This puts constants before unary operators before
588 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
589 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
591 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
592 bool Changed = false;
593 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
594 Changed = !I.swapOperands();
596 if (!I.isAssociative()) return Changed;
597 Instruction::BinaryOps Opcode = I.getOpcode();
598 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
599 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
600 if (isa<Constant>(I.getOperand(1))) {
601 Constant *Folded = ConstantExpr::get(I.getOpcode(),
602 cast<Constant>(I.getOperand(1)),
603 cast<Constant>(Op->getOperand(1)));
604 I.setOperand(0, Op->getOperand(0));
605 I.setOperand(1, Folded);
607 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
608 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
609 isOnlyUse(Op) && isOnlyUse(Op1)) {
610 Constant *C1 = cast<Constant>(Op->getOperand(1));
611 Constant *C2 = cast<Constant>(Op1->getOperand(1));
613 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
614 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
615 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
619 I.setOperand(0, New);
620 I.setOperand(1, Folded);
627 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
628 // if the LHS is a constant zero (which is the 'negate' form).
630 static inline Value *dyn_castNegVal(Value *V) {
631 if (BinaryOperator::isNeg(V))
632 return BinaryOperator::getNegArgument(V);
634 // Constants can be considered to be negated values if they can be folded.
635 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
636 return ConstantExpr::getNeg(C);
638 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
639 if (C->getType()->getElementType()->isInteger())
640 return ConstantExpr::getNeg(C);
645 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
646 // instruction if the LHS is a constant negative zero (which is the 'negate'
649 static inline Value *dyn_castFNegVal(Value *V) {
650 if (BinaryOperator::isFNeg(V))
651 return BinaryOperator::getFNegArgument(V);
653 // Constants can be considered to be negated values if they can be folded.
654 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
655 return ConstantExpr::getFNeg(C);
657 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
658 if (C->getType()->getElementType()->isFloatingPoint())
659 return ConstantExpr::getFNeg(C);
664 /// MatchSelectPattern - Pattern match integer [SU]MIN, [SU]MAX, and ABS idioms,
665 /// returning the kind and providing the out parameter results if we
666 /// successfully match.
667 static SelectPatternFlavor
668 MatchSelectPattern(Value *V, Value *&LHS, Value *&RHS) {
669 SelectInst *SI = dyn_cast<SelectInst>(V);
670 if (SI == 0) return SPF_UNKNOWN;
672 ICmpInst *ICI = dyn_cast<ICmpInst>(SI->getCondition());
673 if (ICI == 0) return SPF_UNKNOWN;
675 LHS = ICI->getOperand(0);
676 RHS = ICI->getOperand(1);
678 // (icmp X, Y) ? X : Y
679 if (SI->getTrueValue() == ICI->getOperand(0) &&
680 SI->getFalseValue() == ICI->getOperand(1)) {
681 switch (ICI->getPredicate()) {
682 default: return SPF_UNKNOWN; // Equality.
683 case ICmpInst::ICMP_UGT:
684 case ICmpInst::ICMP_UGE: return SPF_UMAX;
685 case ICmpInst::ICMP_SGT:
686 case ICmpInst::ICMP_SGE: return SPF_SMAX;
687 case ICmpInst::ICMP_ULT:
688 case ICmpInst::ICMP_ULE: return SPF_UMIN;
689 case ICmpInst::ICMP_SLT:
690 case ICmpInst::ICMP_SLE: return SPF_SMIN;
694 // (icmp X, Y) ? Y : X
695 if (SI->getTrueValue() == ICI->getOperand(1) &&
696 SI->getFalseValue() == ICI->getOperand(0)) {
697 switch (ICI->getPredicate()) {
698 default: return SPF_UNKNOWN; // Equality.
699 case ICmpInst::ICMP_UGT:
700 case ICmpInst::ICMP_UGE: return SPF_UMIN;
701 case ICmpInst::ICMP_SGT:
702 case ICmpInst::ICMP_SGE: return SPF_SMIN;
703 case ICmpInst::ICMP_ULT:
704 case ICmpInst::ICMP_ULE: return SPF_UMAX;
705 case ICmpInst::ICMP_SLT:
706 case ICmpInst::ICMP_SLE: return SPF_SMAX;
710 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
715 /// isFreeToInvert - Return true if the specified value is free to invert (apply
716 /// ~ to). This happens in cases where the ~ can be eliminated.
717 static inline bool isFreeToInvert(Value *V) {
719 if (BinaryOperator::isNot(V))
722 // Constants can be considered to be not'ed values.
723 if (isa<ConstantInt>(V))
726 // Compares can be inverted if they have a single use.
727 if (CmpInst *CI = dyn_cast<CmpInst>(V))
728 return CI->hasOneUse();
733 static inline Value *dyn_castNotVal(Value *V) {
734 // If this is not(not(x)) don't return that this is a not: we want the two
735 // not's to be folded first.
736 if (BinaryOperator::isNot(V)) {
737 Value *Operand = BinaryOperator::getNotArgument(V);
738 if (!isFreeToInvert(Operand))
742 // Constants can be considered to be not'ed values...
743 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
744 return ConstantInt::get(C->getType(), ~C->getValue());
750 // dyn_castFoldableMul - If this value is a multiply that can be folded into
751 // other computations (because it has a constant operand), return the
752 // non-constant operand of the multiply, and set CST to point to the multiplier.
753 // Otherwise, return null.
755 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
756 if (V->hasOneUse() && V->getType()->isInteger())
757 if (Instruction *I = dyn_cast<Instruction>(V)) {
758 if (I->getOpcode() == Instruction::Mul)
759 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
760 return I->getOperand(0);
761 if (I->getOpcode() == Instruction::Shl)
762 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
763 // The multiplier is really 1 << CST.
764 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
765 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
766 CST = ConstantInt::get(V->getType()->getContext(),
767 APInt(BitWidth, 1).shl(CSTVal));
768 return I->getOperand(0);
774 /// AddOne - Add one to a ConstantInt
775 static Constant *AddOne(Constant *C) {
776 return ConstantExpr::getAdd(C,
777 ConstantInt::get(C->getType(), 1));
779 /// SubOne - Subtract one from a ConstantInt
780 static Constant *SubOne(ConstantInt *C) {
781 return ConstantExpr::getSub(C,
782 ConstantInt::get(C->getType(), 1));
784 /// MultiplyOverflows - True if the multiply can not be expressed in an int
786 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
787 uint32_t W = C1->getBitWidth();
788 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
797 APInt MulExt = LHSExt * RHSExt;
800 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
802 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
803 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
804 return MulExt.slt(Min) || MulExt.sgt(Max);
808 /// ShrinkDemandedConstant - Check to see if the specified operand of the
809 /// specified instruction is a constant integer. If so, check to see if there
810 /// are any bits set in the constant that are not demanded. If so, shrink the
811 /// constant and return true.
812 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
814 assert(I && "No instruction?");
815 assert(OpNo < I->getNumOperands() && "Operand index too large");
817 // If the operand is not a constant integer, nothing to do.
818 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
819 if (!OpC) return false;
821 // If there are no bits set that aren't demanded, nothing to do.
822 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
823 if ((~Demanded & OpC->getValue()) == 0)
826 // This instruction is producing bits that are not demanded. Shrink the RHS.
827 Demanded &= OpC->getValue();
828 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
832 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
833 // set of known zero and one bits, compute the maximum and minimum values that
834 // could have the specified known zero and known one bits, returning them in
836 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
837 const APInt& KnownOne,
838 APInt& Min, APInt& Max) {
839 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
840 KnownZero.getBitWidth() == Min.getBitWidth() &&
841 KnownZero.getBitWidth() == Max.getBitWidth() &&
842 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
843 APInt UnknownBits = ~(KnownZero|KnownOne);
845 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
846 // bit if it is unknown.
848 Max = KnownOne|UnknownBits;
850 if (UnknownBits.isNegative()) { // Sign bit is unknown
851 Min.set(Min.getBitWidth()-1);
852 Max.clear(Max.getBitWidth()-1);
856 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
857 // a set of known zero and one bits, compute the maximum and minimum values that
858 // could have the specified known zero and known one bits, returning them in
860 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
861 const APInt &KnownOne,
862 APInt &Min, APInt &Max) {
863 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
864 KnownZero.getBitWidth() == Min.getBitWidth() &&
865 KnownZero.getBitWidth() == Max.getBitWidth() &&
866 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
867 APInt UnknownBits = ~(KnownZero|KnownOne);
869 // The minimum value is when the unknown bits are all zeros.
871 // The maximum value is when the unknown bits are all ones.
872 Max = KnownOne|UnknownBits;
875 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
876 /// SimplifyDemandedBits knows about. See if the instruction has any
877 /// properties that allow us to simplify its operands.
878 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
879 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
880 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
881 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
883 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
884 KnownZero, KnownOne, 0);
885 if (V == 0) return false;
886 if (V == &Inst) return true;
887 ReplaceInstUsesWith(Inst, V);
891 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
892 /// specified instruction operand if possible, updating it in place. It returns
893 /// true if it made any change and false otherwise.
894 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
895 APInt &KnownZero, APInt &KnownOne,
897 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
898 KnownZero, KnownOne, Depth);
899 if (NewVal == 0) return false;
905 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
906 /// value based on the demanded bits. When this function is called, it is known
907 /// that only the bits set in DemandedMask of the result of V are ever used
908 /// downstream. Consequently, depending on the mask and V, it may be possible
909 /// to replace V with a constant or one of its operands. In such cases, this
910 /// function does the replacement and returns true. In all other cases, it
911 /// returns false after analyzing the expression and setting KnownOne and known
912 /// to be one in the expression. KnownZero contains all the bits that are known
913 /// to be zero in the expression. These are provided to potentially allow the
914 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
915 /// the expression. KnownOne and KnownZero always follow the invariant that
916 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
917 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
918 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
919 /// and KnownOne must all be the same.
921 /// This returns null if it did not change anything and it permits no
922 /// simplification. This returns V itself if it did some simplification of V's
923 /// operands based on the information about what bits are demanded. This returns
924 /// some other non-null value if it found out that V is equal to another value
925 /// in the context where the specified bits are demanded, but not for all users.
926 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
927 APInt &KnownZero, APInt &KnownOne,
929 assert(V != 0 && "Null pointer of Value???");
930 assert(Depth <= 6 && "Limit Search Depth");
931 uint32_t BitWidth = DemandedMask.getBitWidth();
932 const Type *VTy = V->getType();
933 assert((TD || !isa<PointerType>(VTy)) &&
934 "SimplifyDemandedBits needs to know bit widths!");
935 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
936 (!VTy->isIntOrIntVector() ||
937 VTy->getScalarSizeInBits() == BitWidth) &&
938 KnownZero.getBitWidth() == BitWidth &&
939 KnownOne.getBitWidth() == BitWidth &&
940 "Value *V, DemandedMask, KnownZero and KnownOne "
941 "must have same BitWidth");
942 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
943 // We know all of the bits for a constant!
944 KnownOne = CI->getValue() & DemandedMask;
945 KnownZero = ~KnownOne & DemandedMask;
948 if (isa<ConstantPointerNull>(V)) {
949 // We know all of the bits for a constant!
951 KnownZero = DemandedMask;
957 if (DemandedMask == 0) { // Not demanding any bits from V.
958 if (isa<UndefValue>(V))
960 return UndefValue::get(VTy);
963 if (Depth == 6) // Limit search depth.
966 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
967 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
969 Instruction *I = dyn_cast<Instruction>(V);
971 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
972 return 0; // Only analyze instructions.
975 // If there are multiple uses of this value and we aren't at the root, then
976 // we can't do any simplifications of the operands, because DemandedMask
977 // only reflects the bits demanded by *one* of the users.
978 if (Depth != 0 && !I->hasOneUse()) {
979 // Despite the fact that we can't simplify this instruction in all User's
980 // context, we can at least compute the knownzero/knownone bits, and we can
981 // do simplifications that apply to *just* the one user if we know that
982 // this instruction has a simpler value in that context.
983 if (I->getOpcode() == Instruction::And) {
984 // If either the LHS or the RHS are Zero, the result is zero.
985 ComputeMaskedBits(I->getOperand(1), DemandedMask,
986 RHSKnownZero, RHSKnownOne, Depth+1);
987 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
988 LHSKnownZero, LHSKnownOne, Depth+1);
990 // If all of the demanded bits are known 1 on one side, return the other.
991 // These bits cannot contribute to the result of the 'and' in this
993 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
994 (DemandedMask & ~LHSKnownZero))
995 return I->getOperand(0);
996 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
997 (DemandedMask & ~RHSKnownZero))
998 return I->getOperand(1);
1000 // If all of the demanded bits in the inputs are known zeros, return zero.
1001 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1002 return Constant::getNullValue(VTy);
1004 } else if (I->getOpcode() == Instruction::Or) {
1005 // We can simplify (X|Y) -> X or Y in the user's context if we know that
1006 // only bits from X or Y are demanded.
1008 // If either the LHS or the RHS are One, the result is One.
1009 ComputeMaskedBits(I->getOperand(1), DemandedMask,
1010 RHSKnownZero, RHSKnownOne, Depth+1);
1011 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1012 LHSKnownZero, LHSKnownOne, Depth+1);
1014 // If all of the demanded bits are known zero on one side, return the
1015 // other. These bits cannot contribute to the result of the 'or' in this
1017 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1018 (DemandedMask & ~LHSKnownOne))
1019 return I->getOperand(0);
1020 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1021 (DemandedMask & ~RHSKnownOne))
1022 return I->getOperand(1);
1024 // If all of the potentially set bits on one side are known to be set on
1025 // the other side, just use the 'other' side.
1026 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1027 (DemandedMask & (~RHSKnownZero)))
1028 return I->getOperand(0);
1029 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1030 (DemandedMask & (~LHSKnownZero)))
1031 return I->getOperand(1);
1034 // Compute the KnownZero/KnownOne bits to simplify things downstream.
1035 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
1039 // If this is the root being simplified, allow it to have multiple uses,
1040 // just set the DemandedMask to all bits so that we can try to simplify the
1041 // operands. This allows visitTruncInst (for example) to simplify the
1042 // operand of a trunc without duplicating all the logic below.
1043 if (Depth == 0 && !V->hasOneUse())
1044 DemandedMask = APInt::getAllOnesValue(BitWidth);
1046 switch (I->getOpcode()) {
1048 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1050 case Instruction::And:
1051 // If either the LHS or the RHS are Zero, the result is zero.
1052 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1053 RHSKnownZero, RHSKnownOne, Depth+1) ||
1054 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
1055 LHSKnownZero, LHSKnownOne, Depth+1))
1057 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1058 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1060 // If all of the demanded bits are known 1 on one side, return the other.
1061 // These bits cannot contribute to the result of the 'and'.
1062 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1063 (DemandedMask & ~LHSKnownZero))
1064 return I->getOperand(0);
1065 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1066 (DemandedMask & ~RHSKnownZero))
1067 return I->getOperand(1);
1069 // If all of the demanded bits in the inputs are known zeros, return zero.
1070 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1071 return Constant::getNullValue(VTy);
1073 // If the RHS is a constant, see if we can simplify it.
1074 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1077 // Output known-1 bits are only known if set in both the LHS & RHS.
1078 RHSKnownOne &= LHSKnownOne;
1079 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1080 RHSKnownZero |= LHSKnownZero;
1082 case Instruction::Or:
1083 // If either the LHS or the RHS are One, the result is One.
1084 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1085 RHSKnownZero, RHSKnownOne, Depth+1) ||
1086 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
1087 LHSKnownZero, LHSKnownOne, Depth+1))
1089 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1090 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1092 // If all of the demanded bits are known zero on one side, return the other.
1093 // These bits cannot contribute to the result of the 'or'.
1094 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1095 (DemandedMask & ~LHSKnownOne))
1096 return I->getOperand(0);
1097 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1098 (DemandedMask & ~RHSKnownOne))
1099 return I->getOperand(1);
1101 // If all of the potentially set bits on one side are known to be set on
1102 // the other side, just use the 'other' side.
1103 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1104 (DemandedMask & (~RHSKnownZero)))
1105 return I->getOperand(0);
1106 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1107 (DemandedMask & (~LHSKnownZero)))
1108 return I->getOperand(1);
1110 // If the RHS is a constant, see if we can simplify it.
1111 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1114 // Output known-0 bits are only known if clear in both the LHS & RHS.
1115 RHSKnownZero &= LHSKnownZero;
1116 // Output known-1 are known to be set if set in either the LHS | RHS.
1117 RHSKnownOne |= LHSKnownOne;
1119 case Instruction::Xor: {
1120 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1121 RHSKnownZero, RHSKnownOne, Depth+1) ||
1122 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1123 LHSKnownZero, LHSKnownOne, Depth+1))
1125 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1126 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1128 // If all of the demanded bits are known zero on one side, return the other.
1129 // These bits cannot contribute to the result of the 'xor'.
1130 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1131 return I->getOperand(0);
1132 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1133 return I->getOperand(1);
1135 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1136 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1137 (RHSKnownOne & LHSKnownOne);
1138 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1139 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1140 (RHSKnownOne & LHSKnownZero);
1142 // If all of the demanded bits are known to be zero on one side or the
1143 // other, turn this into an *inclusive* or.
1144 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1145 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1147 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1149 return InsertNewInstBefore(Or, *I);
1152 // If all of the demanded bits on one side are known, and all of the set
1153 // bits on that side are also known to be set on the other side, turn this
1154 // into an AND, as we know the bits will be cleared.
1155 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1156 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1158 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1159 Constant *AndC = Constant::getIntegerValue(VTy,
1160 ~RHSKnownOne & DemandedMask);
1162 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1163 return InsertNewInstBefore(And, *I);
1167 // If the RHS is a constant, see if we can simplify it.
1168 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1169 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1172 // If our LHS is an 'and' and if it has one use, and if any of the bits we
1173 // are flipping are known to be set, then the xor is just resetting those
1174 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
1175 // simplifying both of them.
1176 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0)))
1177 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
1178 isa<ConstantInt>(I->getOperand(1)) &&
1179 isa<ConstantInt>(LHSInst->getOperand(1)) &&
1180 (LHSKnownOne & RHSKnownOne & DemandedMask) != 0) {
1181 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
1182 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
1183 APInt NewMask = ~(LHSKnownOne & RHSKnownOne & DemandedMask);
1186 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
1187 Instruction *NewAnd =
1188 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1189 InsertNewInstBefore(NewAnd, *I);
1192 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
1193 Instruction *NewXor =
1194 BinaryOperator::CreateXor(NewAnd, XorC, "tmp");
1195 return InsertNewInstBefore(NewXor, *I);
1199 RHSKnownZero = KnownZeroOut;
1200 RHSKnownOne = KnownOneOut;
1203 case Instruction::Select:
1204 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1205 RHSKnownZero, RHSKnownOne, Depth+1) ||
1206 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1207 LHSKnownZero, LHSKnownOne, Depth+1))
1209 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1210 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1212 // If the operands are constants, see if we can simplify them.
1213 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1214 ShrinkDemandedConstant(I, 2, DemandedMask))
1217 // Only known if known in both the LHS and RHS.
1218 RHSKnownOne &= LHSKnownOne;
1219 RHSKnownZero &= LHSKnownZero;
1221 case Instruction::Trunc: {
1222 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1223 DemandedMask.zext(truncBf);
1224 RHSKnownZero.zext(truncBf);
1225 RHSKnownOne.zext(truncBf);
1226 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1227 RHSKnownZero, RHSKnownOne, Depth+1))
1229 DemandedMask.trunc(BitWidth);
1230 RHSKnownZero.trunc(BitWidth);
1231 RHSKnownOne.trunc(BitWidth);
1232 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1235 case Instruction::BitCast:
1236 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1237 return false; // vector->int or fp->int?
1239 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1240 if (const VectorType *SrcVTy =
1241 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1242 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1243 // Don't touch a bitcast between vectors of different element counts.
1246 // Don't touch a scalar-to-vector bitcast.
1248 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1249 // Don't touch a vector-to-scalar bitcast.
1252 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1253 RHSKnownZero, RHSKnownOne, Depth+1))
1255 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1257 case Instruction::ZExt: {
1258 // Compute the bits in the result that are not present in the input.
1259 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1261 DemandedMask.trunc(SrcBitWidth);
1262 RHSKnownZero.trunc(SrcBitWidth);
1263 RHSKnownOne.trunc(SrcBitWidth);
1264 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1265 RHSKnownZero, RHSKnownOne, Depth+1))
1267 DemandedMask.zext(BitWidth);
1268 RHSKnownZero.zext(BitWidth);
1269 RHSKnownOne.zext(BitWidth);
1270 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1271 // The top bits are known to be zero.
1272 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1275 case Instruction::SExt: {
1276 // Compute the bits in the result that are not present in the input.
1277 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1279 APInt InputDemandedBits = DemandedMask &
1280 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1282 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1283 // If any of the sign extended bits are demanded, we know that the sign
1285 if ((NewBits & DemandedMask) != 0)
1286 InputDemandedBits.set(SrcBitWidth-1);
1288 InputDemandedBits.trunc(SrcBitWidth);
1289 RHSKnownZero.trunc(SrcBitWidth);
1290 RHSKnownOne.trunc(SrcBitWidth);
1291 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1292 RHSKnownZero, RHSKnownOne, Depth+1))
1294 InputDemandedBits.zext(BitWidth);
1295 RHSKnownZero.zext(BitWidth);
1296 RHSKnownOne.zext(BitWidth);
1297 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1299 // If the sign bit of the input is known set or clear, then we know the
1300 // top bits of the result.
1302 // If the input sign bit is known zero, or if the NewBits are not demanded
1303 // convert this into a zero extension.
1304 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1305 // Convert to ZExt cast
1306 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1307 return InsertNewInstBefore(NewCast, *I);
1308 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1309 RHSKnownOne |= NewBits;
1313 case Instruction::Add: {
1314 // Figure out what the input bits are. If the top bits of the and result
1315 // are not demanded, then the add doesn't demand them from its input
1317 unsigned NLZ = DemandedMask.countLeadingZeros();
1319 // If there is a constant on the RHS, there are a variety of xformations
1321 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1322 // If null, this should be simplified elsewhere. Some of the xforms here
1323 // won't work if the RHS is zero.
1327 // If the top bit of the output is demanded, demand everything from the
1328 // input. Otherwise, we demand all the input bits except NLZ top bits.
1329 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1331 // Find information about known zero/one bits in the input.
1332 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1333 LHSKnownZero, LHSKnownOne, Depth+1))
1336 // If the RHS of the add has bits set that can't affect the input, reduce
1338 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1341 // Avoid excess work.
1342 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1345 // Turn it into OR if input bits are zero.
1346 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1348 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1350 return InsertNewInstBefore(Or, *I);
1353 // We can say something about the output known-zero and known-one bits,
1354 // depending on potential carries from the input constant and the
1355 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1356 // bits set and the RHS constant is 0x01001, then we know we have a known
1357 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1359 // To compute this, we first compute the potential carry bits. These are
1360 // the bits which may be modified. I'm not aware of a better way to do
1362 const APInt &RHSVal = RHS->getValue();
1363 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1365 // Now that we know which bits have carries, compute the known-1/0 sets.
1367 // Bits are known one if they are known zero in one operand and one in the
1368 // other, and there is no input carry.
1369 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1370 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1372 // Bits are known zero if they are known zero in both operands and there
1373 // is no input carry.
1374 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1376 // If the high-bits of this ADD are not demanded, then it does not demand
1377 // the high bits of its LHS or RHS.
1378 if (DemandedMask[BitWidth-1] == 0) {
1379 // Right fill the mask of bits for this ADD to demand the most
1380 // significant bit and all those below it.
1381 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1382 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1383 LHSKnownZero, LHSKnownOne, Depth+1) ||
1384 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1385 LHSKnownZero, LHSKnownOne, Depth+1))
1391 case Instruction::Sub:
1392 // If the high-bits of this SUB are not demanded, then it does not demand
1393 // the high bits of its LHS or RHS.
1394 if (DemandedMask[BitWidth-1] == 0) {
1395 // Right fill the mask of bits for this SUB to demand the most
1396 // significant bit and all those below it.
1397 uint32_t NLZ = DemandedMask.countLeadingZeros();
1398 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1399 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1400 LHSKnownZero, LHSKnownOne, Depth+1) ||
1401 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1402 LHSKnownZero, LHSKnownOne, Depth+1))
1405 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1406 // the known zeros and ones.
1407 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1409 case Instruction::Shl:
1410 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1411 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1412 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1413 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1414 RHSKnownZero, RHSKnownOne, Depth+1))
1416 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1417 RHSKnownZero <<= ShiftAmt;
1418 RHSKnownOne <<= ShiftAmt;
1419 // low bits known zero.
1421 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1424 case Instruction::LShr:
1425 // For a logical shift right
1426 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1427 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1429 // Unsigned shift right.
1430 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1431 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1432 RHSKnownZero, RHSKnownOne, Depth+1))
1434 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1435 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1436 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1438 // Compute the new bits that are at the top now.
1439 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1440 RHSKnownZero |= HighBits; // high bits known zero.
1444 case Instruction::AShr:
1445 // If this is an arithmetic shift right and only the low-bit is set, we can
1446 // always convert this into a logical shr, even if the shift amount is
1447 // variable. The low bit of the shift cannot be an input sign bit unless
1448 // the shift amount is >= the size of the datatype, which is undefined.
1449 if (DemandedMask == 1) {
1450 // Perform the logical shift right.
1451 Instruction *NewVal = BinaryOperator::CreateLShr(
1452 I->getOperand(0), I->getOperand(1), I->getName());
1453 return InsertNewInstBefore(NewVal, *I);
1456 // If the sign bit is the only bit demanded by this ashr, then there is no
1457 // need to do it, the shift doesn't change the high bit.
1458 if (DemandedMask.isSignBit())
1459 return I->getOperand(0);
1461 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1462 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1464 // Signed shift right.
1465 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1466 // If any of the "high bits" are demanded, we should set the sign bit as
1468 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1469 DemandedMaskIn.set(BitWidth-1);
1470 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1471 RHSKnownZero, RHSKnownOne, Depth+1))
1473 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1474 // Compute the new bits that are at the top now.
1475 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1476 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1477 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1479 // Handle the sign bits.
1480 APInt SignBit(APInt::getSignBit(BitWidth));
1481 // Adjust to where it is now in the mask.
1482 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1484 // If the input sign bit is known to be zero, or if none of the top bits
1485 // are demanded, turn this into an unsigned shift right.
1486 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1487 (HighBits & ~DemandedMask) == HighBits) {
1488 // Perform the logical shift right.
1489 Instruction *NewVal = BinaryOperator::CreateLShr(
1490 I->getOperand(0), SA, I->getName());
1491 return InsertNewInstBefore(NewVal, *I);
1492 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1493 RHSKnownOne |= HighBits;
1497 case Instruction::SRem:
1498 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1499 APInt RA = Rem->getValue().abs();
1500 if (RA.isPowerOf2()) {
1501 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1502 return I->getOperand(0);
1504 APInt LowBits = RA - 1;
1505 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1506 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1507 LHSKnownZero, LHSKnownOne, Depth+1))
1510 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1511 LHSKnownZero |= ~LowBits;
1513 KnownZero |= LHSKnownZero & DemandedMask;
1515 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1519 case Instruction::URem: {
1520 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1521 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1522 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1523 KnownZero2, KnownOne2, Depth+1) ||
1524 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1525 KnownZero2, KnownOne2, Depth+1))
1528 unsigned Leaders = KnownZero2.countLeadingOnes();
1529 Leaders = std::max(Leaders,
1530 KnownZero2.countLeadingOnes());
1531 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1534 case Instruction::Call:
1535 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1536 switch (II->getIntrinsicID()) {
1538 case Intrinsic::bswap: {
1539 // If the only bits demanded come from one byte of the bswap result,
1540 // just shift the input byte into position to eliminate the bswap.
1541 unsigned NLZ = DemandedMask.countLeadingZeros();
1542 unsigned NTZ = DemandedMask.countTrailingZeros();
1544 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1545 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1546 // have 14 leading zeros, round to 8.
1549 // If we need exactly one byte, we can do this transformation.
1550 if (BitWidth-NLZ-NTZ == 8) {
1551 unsigned ResultBit = NTZ;
1552 unsigned InputBit = BitWidth-NTZ-8;
1554 // Replace this with either a left or right shift to get the byte into
1556 Instruction *NewVal;
1557 if (InputBit > ResultBit)
1558 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1559 ConstantInt::get(I->getType(), InputBit-ResultBit));
1561 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1562 ConstantInt::get(I->getType(), ResultBit-InputBit));
1563 NewVal->takeName(I);
1564 return InsertNewInstBefore(NewVal, *I);
1567 // TODO: Could compute known zero/one bits based on the input.
1572 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1576 // If the client is only demanding bits that we know, return the known
1578 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1579 return Constant::getIntegerValue(VTy, RHSKnownOne);
1584 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1585 /// any number of elements. DemandedElts contains the set of elements that are
1586 /// actually used by the caller. This method analyzes which elements of the
1587 /// operand are undef and returns that information in UndefElts.
1589 /// If the information about demanded elements can be used to simplify the
1590 /// operation, the operation is simplified, then the resultant value is
1591 /// returned. This returns null if no change was made.
1592 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1595 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1596 APInt EltMask(APInt::getAllOnesValue(VWidth));
1597 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1599 if (isa<UndefValue>(V)) {
1600 // If the entire vector is undefined, just return this info.
1601 UndefElts = EltMask;
1603 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1604 UndefElts = EltMask;
1605 return UndefValue::get(V->getType());
1609 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1610 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1611 Constant *Undef = UndefValue::get(EltTy);
1613 std::vector<Constant*> Elts;
1614 for (unsigned i = 0; i != VWidth; ++i)
1615 if (!DemandedElts[i]) { // If not demanded, set to undef.
1616 Elts.push_back(Undef);
1618 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1619 Elts.push_back(Undef);
1621 } else { // Otherwise, defined.
1622 Elts.push_back(CP->getOperand(i));
1625 // If we changed the constant, return it.
1626 Constant *NewCP = ConstantVector::get(Elts);
1627 return NewCP != CP ? NewCP : 0;
1628 } else if (isa<ConstantAggregateZero>(V)) {
1629 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1632 // Check if this is identity. If so, return 0 since we are not simplifying
1634 if (DemandedElts == ((1ULL << VWidth) -1))
1637 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1638 Constant *Zero = Constant::getNullValue(EltTy);
1639 Constant *Undef = UndefValue::get(EltTy);
1640 std::vector<Constant*> Elts;
1641 for (unsigned i = 0; i != VWidth; ++i) {
1642 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1643 Elts.push_back(Elt);
1645 UndefElts = DemandedElts ^ EltMask;
1646 return ConstantVector::get(Elts);
1649 // Limit search depth.
1653 // If multiple users are using the root value, procede with
1654 // simplification conservatively assuming that all elements
1656 if (!V->hasOneUse()) {
1657 // Quit if we find multiple users of a non-root value though.
1658 // They'll be handled when it's their turn to be visited by
1659 // the main instcombine process.
1661 // TODO: Just compute the UndefElts information recursively.
1664 // Conservatively assume that all elements are needed.
1665 DemandedElts = EltMask;
1668 Instruction *I = dyn_cast<Instruction>(V);
1669 if (!I) return 0; // Only analyze instructions.
1671 bool MadeChange = false;
1672 APInt UndefElts2(VWidth, 0);
1674 switch (I->getOpcode()) {
1677 case Instruction::InsertElement: {
1678 // If this is a variable index, we don't know which element it overwrites.
1679 // demand exactly the same input as we produce.
1680 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1682 // Note that we can't propagate undef elt info, because we don't know
1683 // which elt is getting updated.
1684 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1685 UndefElts2, Depth+1);
1686 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1690 // If this is inserting an element that isn't demanded, remove this
1692 unsigned IdxNo = Idx->getZExtValue();
1693 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1695 return I->getOperand(0);
1698 // Otherwise, the element inserted overwrites whatever was there, so the
1699 // input demanded set is simpler than the output set.
1700 APInt DemandedElts2 = DemandedElts;
1701 DemandedElts2.clear(IdxNo);
1702 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1703 UndefElts, Depth+1);
1704 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1706 // The inserted element is defined.
1707 UndefElts.clear(IdxNo);
1710 case Instruction::ShuffleVector: {
1711 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1712 uint64_t LHSVWidth =
1713 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1714 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1715 for (unsigned i = 0; i < VWidth; i++) {
1716 if (DemandedElts[i]) {
1717 unsigned MaskVal = Shuffle->getMaskValue(i);
1718 if (MaskVal != -1u) {
1719 assert(MaskVal < LHSVWidth * 2 &&
1720 "shufflevector mask index out of range!");
1721 if (MaskVal < LHSVWidth)
1722 LeftDemanded.set(MaskVal);
1724 RightDemanded.set(MaskVal - LHSVWidth);
1729 APInt UndefElts4(LHSVWidth, 0);
1730 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1731 UndefElts4, Depth+1);
1732 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1734 APInt UndefElts3(LHSVWidth, 0);
1735 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1736 UndefElts3, Depth+1);
1737 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1739 bool NewUndefElts = false;
1740 for (unsigned i = 0; i < VWidth; i++) {
1741 unsigned MaskVal = Shuffle->getMaskValue(i);
1742 if (MaskVal == -1u) {
1744 } else if (MaskVal < LHSVWidth) {
1745 if (UndefElts4[MaskVal]) {
1746 NewUndefElts = true;
1750 if (UndefElts3[MaskVal - LHSVWidth]) {
1751 NewUndefElts = true;
1758 // Add additional discovered undefs.
1759 std::vector<Constant*> Elts;
1760 for (unsigned i = 0; i < VWidth; ++i) {
1762 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1764 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1765 Shuffle->getMaskValue(i)));
1767 I->setOperand(2, ConstantVector::get(Elts));
1772 case Instruction::BitCast: {
1773 // Vector->vector casts only.
1774 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1776 unsigned InVWidth = VTy->getNumElements();
1777 APInt InputDemandedElts(InVWidth, 0);
1780 if (VWidth == InVWidth) {
1781 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1782 // elements as are demanded of us.
1784 InputDemandedElts = DemandedElts;
1785 } else if (VWidth > InVWidth) {
1789 // If there are more elements in the result than there are in the source,
1790 // then an input element is live if any of the corresponding output
1791 // elements are live.
1792 Ratio = VWidth/InVWidth;
1793 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1794 if (DemandedElts[OutIdx])
1795 InputDemandedElts.set(OutIdx/Ratio);
1801 // If there are more elements in the source than there are in the result,
1802 // then an input element is live if the corresponding output element is
1804 Ratio = InVWidth/VWidth;
1805 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1806 if (DemandedElts[InIdx/Ratio])
1807 InputDemandedElts.set(InIdx);
1810 // div/rem demand all inputs, because they don't want divide by zero.
1811 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1812 UndefElts2, Depth+1);
1814 I->setOperand(0, TmpV);
1818 UndefElts = UndefElts2;
1819 if (VWidth > InVWidth) {
1820 llvm_unreachable("Unimp");
1821 // If there are more elements in the result than there are in the source,
1822 // then an output element is undef if the corresponding input element is
1824 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1825 if (UndefElts2[OutIdx/Ratio])
1826 UndefElts.set(OutIdx);
1827 } else if (VWidth < InVWidth) {
1828 llvm_unreachable("Unimp");
1829 // If there are more elements in the source than there are in the result,
1830 // then a result element is undef if all of the corresponding input
1831 // elements are undef.
1832 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1833 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1834 if (!UndefElts2[InIdx]) // Not undef?
1835 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1839 case Instruction::And:
1840 case Instruction::Or:
1841 case Instruction::Xor:
1842 case Instruction::Add:
1843 case Instruction::Sub:
1844 case Instruction::Mul:
1845 // div/rem demand all inputs, because they don't want divide by zero.
1846 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1847 UndefElts, Depth+1);
1848 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1849 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1850 UndefElts2, Depth+1);
1851 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1853 // Output elements are undefined if both are undefined. Consider things
1854 // like undef&0. The result is known zero, not undef.
1855 UndefElts &= UndefElts2;
1858 case Instruction::Call: {
1859 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1861 switch (II->getIntrinsicID()) {
1864 // Binary vector operations that work column-wise. A dest element is a
1865 // function of the corresponding input elements from the two inputs.
1866 case Intrinsic::x86_sse_sub_ss:
1867 case Intrinsic::x86_sse_mul_ss:
1868 case Intrinsic::x86_sse_min_ss:
1869 case Intrinsic::x86_sse_max_ss:
1870 case Intrinsic::x86_sse2_sub_sd:
1871 case Intrinsic::x86_sse2_mul_sd:
1872 case Intrinsic::x86_sse2_min_sd:
1873 case Intrinsic::x86_sse2_max_sd:
1874 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1875 UndefElts, Depth+1);
1876 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1877 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1878 UndefElts2, Depth+1);
1879 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1881 // If only the low elt is demanded and this is a scalarizable intrinsic,
1882 // scalarize it now.
1883 if (DemandedElts == 1) {
1884 switch (II->getIntrinsicID()) {
1886 case Intrinsic::x86_sse_sub_ss:
1887 case Intrinsic::x86_sse_mul_ss:
1888 case Intrinsic::x86_sse2_sub_sd:
1889 case Intrinsic::x86_sse2_mul_sd:
1890 // TODO: Lower MIN/MAX/ABS/etc
1891 Value *LHS = II->getOperand(1);
1892 Value *RHS = II->getOperand(2);
1893 // Extract the element as scalars.
1894 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1895 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1896 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1897 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1899 switch (II->getIntrinsicID()) {
1900 default: llvm_unreachable("Case stmts out of sync!");
1901 case Intrinsic::x86_sse_sub_ss:
1902 case Intrinsic::x86_sse2_sub_sd:
1903 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1904 II->getName()), *II);
1906 case Intrinsic::x86_sse_mul_ss:
1907 case Intrinsic::x86_sse2_mul_sd:
1908 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1909 II->getName()), *II);
1914 InsertElementInst::Create(
1915 UndefValue::get(II->getType()), TmpV,
1916 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1917 InsertNewInstBefore(New, *II);
1922 // Output elements are undefined if both are undefined. Consider things
1923 // like undef&0. The result is known zero, not undef.
1924 UndefElts &= UndefElts2;
1930 return MadeChange ? I : 0;
1934 /// AssociativeOpt - Perform an optimization on an associative operator. This
1935 /// function is designed to check a chain of associative operators for a
1936 /// potential to apply a certain optimization. Since the optimization may be
1937 /// applicable if the expression was reassociated, this checks the chain, then
1938 /// reassociates the expression as necessary to expose the optimization
1939 /// opportunity. This makes use of a special Functor, which must define
1940 /// 'shouldApply' and 'apply' methods.
1942 template<typename Functor>
1943 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1944 unsigned Opcode = Root.getOpcode();
1945 Value *LHS = Root.getOperand(0);
1947 // Quick check, see if the immediate LHS matches...
1948 if (F.shouldApply(LHS))
1949 return F.apply(Root);
1951 // Otherwise, if the LHS is not of the same opcode as the root, return.
1952 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1953 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1954 // Should we apply this transform to the RHS?
1955 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1957 // If not to the RHS, check to see if we should apply to the LHS...
1958 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1959 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1963 // If the functor wants to apply the optimization to the RHS of LHSI,
1964 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1966 // Now all of the instructions are in the current basic block, go ahead
1967 // and perform the reassociation.
1968 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1970 // First move the selected RHS to the LHS of the root...
1971 Root.setOperand(0, LHSI->getOperand(1));
1973 // Make what used to be the LHS of the root be the user of the root...
1974 Value *ExtraOperand = TmpLHSI->getOperand(1);
1975 if (&Root == TmpLHSI) {
1976 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1979 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1980 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1981 BasicBlock::iterator ARI = &Root; ++ARI;
1982 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1985 // Now propagate the ExtraOperand down the chain of instructions until we
1987 while (TmpLHSI != LHSI) {
1988 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1989 // Move the instruction to immediately before the chain we are
1990 // constructing to avoid breaking dominance properties.
1991 NextLHSI->moveBefore(ARI);
1994 Value *NextOp = NextLHSI->getOperand(1);
1995 NextLHSI->setOperand(1, ExtraOperand);
1997 ExtraOperand = NextOp;
2000 // Now that the instructions are reassociated, have the functor perform
2001 // the transformation...
2002 return F.apply(Root);
2005 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2012 // AddRHS - Implements: X + X --> X << 1
2015 explicit AddRHS(Value *rhs) : RHS(rhs) {}
2016 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2017 Instruction *apply(BinaryOperator &Add) const {
2018 return BinaryOperator::CreateShl(Add.getOperand(0),
2019 ConstantInt::get(Add.getType(), 1));
2023 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2025 struct AddMaskingAnd {
2027 explicit AddMaskingAnd(Constant *c) : C2(c) {}
2028 bool shouldApply(Value *LHS) const {
2030 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2031 ConstantExpr::getAnd(C1, C2)->isNullValue();
2033 Instruction *apply(BinaryOperator &Add) const {
2034 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
2040 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2042 if (CastInst *CI = dyn_cast<CastInst>(&I))
2043 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
2045 // Figure out if the constant is the left or the right argument.
2046 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2047 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2049 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2051 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2052 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2055 Value *Op0 = SO, *Op1 = ConstOperand;
2057 std::swap(Op0, Op1);
2059 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2060 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
2061 SO->getName()+".op");
2062 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
2063 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
2064 SO->getName()+".cmp");
2065 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
2066 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
2067 SO->getName()+".cmp");
2068 llvm_unreachable("Unknown binary instruction type!");
2071 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2072 // constant as the other operand, try to fold the binary operator into the
2073 // select arguments. This also works for Cast instructions, which obviously do
2074 // not have a second operand.
2075 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2077 // Don't modify shared select instructions
2078 if (!SI->hasOneUse()) return 0;
2079 Value *TV = SI->getOperand(1);
2080 Value *FV = SI->getOperand(2);
2082 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2083 // Bool selects with constant operands can be folded to logical ops.
2084 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
2086 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2087 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2089 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
2096 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
2097 /// has a PHI node as operand #0, see if we can fold the instruction into the
2098 /// PHI (which is only possible if all operands to the PHI are constants).
2100 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
2101 /// that would normally be unprofitable because they strongly encourage jump
2103 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
2104 bool AllowAggressive) {
2105 AllowAggressive = false;
2106 PHINode *PN = cast<PHINode>(I.getOperand(0));
2107 unsigned NumPHIValues = PN->getNumIncomingValues();
2108 if (NumPHIValues == 0 ||
2109 // We normally only transform phis with a single use, unless we're trying
2110 // hard to make jump threading happen.
2111 (!PN->hasOneUse() && !AllowAggressive))
2115 // Check to see if all of the operands of the PHI are simple constants
2116 // (constantint/constantfp/undef). If there is one non-constant value,
2117 // remember the BB it is in. If there is more than one or if *it* is a PHI,
2118 // bail out. We don't do arbitrary constant expressions here because moving
2119 // their computation can be expensive without a cost model.
2120 BasicBlock *NonConstBB = 0;
2121 for (unsigned i = 0; i != NumPHIValues; ++i)
2122 if (!isa<Constant>(PN->getIncomingValue(i)) ||
2123 isa<ConstantExpr>(PN->getIncomingValue(i))) {
2124 if (NonConstBB) return 0; // More than one non-const value.
2125 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2126 NonConstBB = PN->getIncomingBlock(i);
2128 // If the incoming non-constant value is in I's block, we have an infinite
2130 if (NonConstBB == I.getParent())
2134 // If there is exactly one non-constant value, we can insert a copy of the
2135 // operation in that block. However, if this is a critical edge, we would be
2136 // inserting the computation one some other paths (e.g. inside a loop). Only
2137 // do this if the pred block is unconditionally branching into the phi block.
2138 if (NonConstBB != 0 && !AllowAggressive) {
2139 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2140 if (!BI || !BI->isUnconditional()) return 0;
2143 // Okay, we can do the transformation: create the new PHI node.
2144 PHINode *NewPN = PHINode::Create(I.getType(), "");
2145 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2146 InsertNewInstBefore(NewPN, *PN);
2147 NewPN->takeName(PN);
2149 // Next, add all of the operands to the PHI.
2150 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
2151 // We only currently try to fold the condition of a select when it is a phi,
2152 // not the true/false values.
2153 Value *TrueV = SI->getTrueValue();
2154 Value *FalseV = SI->getFalseValue();
2155 BasicBlock *PhiTransBB = PN->getParent();
2156 for (unsigned i = 0; i != NumPHIValues; ++i) {
2157 BasicBlock *ThisBB = PN->getIncomingBlock(i);
2158 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
2159 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
2161 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2162 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
2164 assert(PN->getIncomingBlock(i) == NonConstBB);
2165 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
2167 "phitmp", NonConstBB->getTerminator());
2168 Worklist.Add(cast<Instruction>(InV));
2170 NewPN->addIncoming(InV, ThisBB);
2172 } else if (I.getNumOperands() == 2) {
2173 Constant *C = cast<Constant>(I.getOperand(1));
2174 for (unsigned i = 0; i != NumPHIValues; ++i) {
2176 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2177 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2178 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2180 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2182 assert(PN->getIncomingBlock(i) == NonConstBB);
2183 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2184 InV = BinaryOperator::Create(BO->getOpcode(),
2185 PN->getIncomingValue(i), C, "phitmp",
2186 NonConstBB->getTerminator());
2187 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2188 InV = CmpInst::Create(CI->getOpcode(),
2190 PN->getIncomingValue(i), C, "phitmp",
2191 NonConstBB->getTerminator());
2193 llvm_unreachable("Unknown binop!");
2195 Worklist.Add(cast<Instruction>(InV));
2197 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2200 CastInst *CI = cast<CastInst>(&I);
2201 const Type *RetTy = CI->getType();
2202 for (unsigned i = 0; i != NumPHIValues; ++i) {
2204 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2205 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2207 assert(PN->getIncomingBlock(i) == NonConstBB);
2208 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2209 I.getType(), "phitmp",
2210 NonConstBB->getTerminator());
2211 Worklist.Add(cast<Instruction>(InV));
2213 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2216 return ReplaceInstUsesWith(I, NewPN);
2220 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2221 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2222 /// This basically requires proving that the add in the original type would not
2223 /// overflow to change the sign bit or have a carry out.
2224 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2225 // There are different heuristics we can use for this. Here are some simple
2228 // Add has the property that adding any two 2's complement numbers can only
2229 // have one carry bit which can change a sign. As such, if LHS and RHS each
2230 // have at least two sign bits, we know that the addition of the two values
2231 // will sign extend fine.
2232 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2236 // If one of the operands only has one non-zero bit, and if the other operand
2237 // has a known-zero bit in a more significant place than it (not including the
2238 // sign bit) the ripple may go up to and fill the zero, but won't change the
2239 // sign. For example, (X & ~4) + 1.
2247 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2248 bool Changed = SimplifyCommutative(I);
2249 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2251 if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
2252 I.hasNoUnsignedWrap(), TD))
2253 return ReplaceInstUsesWith(I, V);
2256 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2257 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2258 // X + (signbit) --> X ^ signbit
2259 const APInt& Val = CI->getValue();
2260 uint32_t BitWidth = Val.getBitWidth();
2261 if (Val == APInt::getSignBit(BitWidth))
2262 return BinaryOperator::CreateXor(LHS, RHS);
2264 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2265 // (X & 254)+1 -> (X&254)|1
2266 if (SimplifyDemandedInstructionBits(I))
2269 // zext(bool) + C -> bool ? C + 1 : C
2270 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2271 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2272 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2275 if (isa<PHINode>(LHS))
2276 if (Instruction *NV = FoldOpIntoPhi(I))
2279 ConstantInt *XorRHS = 0;
2281 if (isa<ConstantInt>(RHSC) &&
2282 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2283 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2284 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2286 uint32_t Size = TySizeBits / 2;
2287 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2288 APInt CFF80Val(-C0080Val);
2290 if (TySizeBits > Size) {
2291 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2292 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2293 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2294 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2295 // This is a sign extend if the top bits are known zero.
2296 if (!MaskedValueIsZero(XorLHS,
2297 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2298 Size = 0; // Not a sign ext, but can't be any others either.
2303 C0080Val = APIntOps::lshr(C0080Val, Size);
2304 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2305 } while (Size >= 1);
2307 // FIXME: This shouldn't be necessary. When the backends can handle types
2308 // with funny bit widths then this switch statement should be removed. It
2309 // is just here to get the size of the "middle" type back up to something
2310 // that the back ends can handle.
2311 const Type *MiddleType = 0;
2314 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2315 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2316 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2319 Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext");
2320 return new SExtInst(NewTrunc, I.getType(), I.getName());
2325 if (I.getType() == Type::getInt1Ty(*Context))
2326 return BinaryOperator::CreateXor(LHS, RHS);
2329 if (I.getType()->isInteger()) {
2330 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2333 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2334 if (RHSI->getOpcode() == Instruction::Sub)
2335 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2336 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2338 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2339 if (LHSI->getOpcode() == Instruction::Sub)
2340 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2341 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2346 // -A + -B --> -(A + B)
2347 if (Value *LHSV = dyn_castNegVal(LHS)) {
2348 if (LHS->getType()->isIntOrIntVector()) {
2349 if (Value *RHSV = dyn_castNegVal(RHS)) {
2350 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
2351 return BinaryOperator::CreateNeg(NewAdd);
2355 return BinaryOperator::CreateSub(RHS, LHSV);
2359 if (!isa<Constant>(RHS))
2360 if (Value *V = dyn_castNegVal(RHS))
2361 return BinaryOperator::CreateSub(LHS, V);
2365 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2366 if (X == RHS) // X*C + X --> X * (C+1)
2367 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2369 // X*C1 + X*C2 --> X * (C1+C2)
2371 if (X == dyn_castFoldableMul(RHS, C1))
2372 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2375 // X + X*C --> X * (C+1)
2376 if (dyn_castFoldableMul(RHS, C2) == LHS)
2377 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2379 // X + ~X --> -1 since ~X = -X-1
2380 if (dyn_castNotVal(LHS) == RHS ||
2381 dyn_castNotVal(RHS) == LHS)
2382 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2385 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2386 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2387 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2390 // A+B --> A|B iff A and B have no bits set in common.
2391 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2392 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2393 APInt LHSKnownOne(IT->getBitWidth(), 0);
2394 APInt LHSKnownZero(IT->getBitWidth(), 0);
2395 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2396 if (LHSKnownZero != 0) {
2397 APInt RHSKnownOne(IT->getBitWidth(), 0);
2398 APInt RHSKnownZero(IT->getBitWidth(), 0);
2399 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2401 // No bits in common -> bitwise or.
2402 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2403 return BinaryOperator::CreateOr(LHS, RHS);
2407 // W*X + Y*Z --> W * (X+Z) iff W == Y
2408 if (I.getType()->isIntOrIntVector()) {
2409 Value *W, *X, *Y, *Z;
2410 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2411 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2415 } else if (Y == X) {
2417 } else if (X == Z) {
2424 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
2425 return BinaryOperator::CreateMul(W, NewAdd);
2430 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2432 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2433 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2435 // (X & FF00) + xx00 -> (X+xx00) & FF00
2436 if (LHS->hasOneUse() &&
2437 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2438 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2439 if (Anded == CRHS) {
2440 // See if all bits from the first bit set in the Add RHS up are included
2441 // in the mask. First, get the rightmost bit.
2442 const APInt& AddRHSV = CRHS->getValue();
2444 // Form a mask of all bits from the lowest bit added through the top.
2445 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2447 // See if the and mask includes all of these bits.
2448 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2450 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2451 // Okay, the xform is safe. Insert the new add pronto.
2452 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
2453 return BinaryOperator::CreateAnd(NewAdd, C2);
2458 // Try to fold constant add into select arguments.
2459 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2460 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2464 // add (select X 0 (sub n A)) A --> select X A n
2466 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2469 SI = dyn_cast<SelectInst>(RHS);
2472 if (SI && SI->hasOneUse()) {
2473 Value *TV = SI->getTrueValue();
2474 Value *FV = SI->getFalseValue();
2477 // Can we fold the add into the argument of the select?
2478 // We check both true and false select arguments for a matching subtract.
2479 if (match(FV, m_Zero()) &&
2480 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2481 // Fold the add into the true select value.
2482 return SelectInst::Create(SI->getCondition(), N, A);
2483 if (match(TV, m_Zero()) &&
2484 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2485 // Fold the add into the false select value.
2486 return SelectInst::Create(SI->getCondition(), A, N);
2490 // Check for (add (sext x), y), see if we can merge this into an
2491 // integer add followed by a sext.
2492 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2493 // (add (sext x), cst) --> (sext (add x, cst'))
2494 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2496 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2497 if (LHSConv->hasOneUse() &&
2498 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2499 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2500 // Insert the new, smaller add.
2501 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2503 return new SExtInst(NewAdd, I.getType());
2507 // (add (sext x), (sext y)) --> (sext (add int x, y))
2508 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2509 // Only do this if x/y have the same type, if at last one of them has a
2510 // single use (so we don't increase the number of sexts), and if the
2511 // integer add will not overflow.
2512 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2513 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2514 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2515 RHSConv->getOperand(0))) {
2516 // Insert the new integer add.
2517 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2518 RHSConv->getOperand(0), "addconv");
2519 return new SExtInst(NewAdd, I.getType());
2524 return Changed ? &I : 0;
2527 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2528 bool Changed = SimplifyCommutative(I);
2529 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2531 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2533 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2534 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2535 (I.getType())->getValueAPF()))
2536 return ReplaceInstUsesWith(I, LHS);
2539 if (isa<PHINode>(LHS))
2540 if (Instruction *NV = FoldOpIntoPhi(I))
2545 // -A + -B --> -(A + B)
2546 if (Value *LHSV = dyn_castFNegVal(LHS))
2547 return BinaryOperator::CreateFSub(RHS, LHSV);
2550 if (!isa<Constant>(RHS))
2551 if (Value *V = dyn_castFNegVal(RHS))
2552 return BinaryOperator::CreateFSub(LHS, V);
2554 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2555 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2556 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2557 return ReplaceInstUsesWith(I, LHS);
2559 // Check for (add double (sitofp x), y), see if we can merge this into an
2560 // integer add followed by a promotion.
2561 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2562 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2563 // ... if the constant fits in the integer value. This is useful for things
2564 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2565 // requires a constant pool load, and generally allows the add to be better
2567 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2569 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2570 if (LHSConv->hasOneUse() &&
2571 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2572 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2573 // Insert the new integer add.
2574 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2576 return new SIToFPInst(NewAdd, I.getType());
2580 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2581 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2582 // Only do this if x/y have the same type, if at last one of them has a
2583 // single use (so we don't increase the number of int->fp conversions),
2584 // and if the integer add will not overflow.
2585 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2586 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2587 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2588 RHSConv->getOperand(0))) {
2589 // Insert the new integer add.
2590 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2591 RHSConv->getOperand(0),"addconv");
2592 return new SIToFPInst(NewAdd, I.getType());
2597 return Changed ? &I : 0;
2601 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
2602 /// code necessary to compute the offset from the base pointer (without adding
2603 /// in the base pointer). Return the result as a signed integer of intptr size.
2604 static Value *EmitGEPOffset(User *GEP, InstCombiner &IC) {
2605 TargetData &TD = *IC.getTargetData();
2606 gep_type_iterator GTI = gep_type_begin(GEP);
2607 const Type *IntPtrTy = TD.getIntPtrType(GEP->getContext());
2608 Value *Result = Constant::getNullValue(IntPtrTy);
2610 // Build a mask for high order bits.
2611 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2612 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
2614 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
2617 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
2618 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
2619 if (OpC->isZero()) continue;
2621 // Handle a struct index, which adds its field offset to the pointer.
2622 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
2623 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
2625 Result = IC.Builder->CreateAdd(Result,
2626 ConstantInt::get(IntPtrTy, Size),
2627 GEP->getName()+".offs");
2631 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
2633 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
2634 Scale = ConstantExpr::getMul(OC, Scale);
2635 // Emit an add instruction.
2636 Result = IC.Builder->CreateAdd(Result, Scale, GEP->getName()+".offs");
2639 // Convert to correct type.
2640 if (Op->getType() != IntPtrTy)
2641 Op = IC.Builder->CreateIntCast(Op, IntPtrTy, true, Op->getName()+".c");
2643 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
2644 // We'll let instcombine(mul) convert this to a shl if possible.
2645 Op = IC.Builder->CreateMul(Op, Scale, GEP->getName()+".idx");
2648 // Emit an add instruction.
2649 Result = IC.Builder->CreateAdd(Op, Result, GEP->getName()+".offs");
2655 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
2656 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
2657 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
2658 /// be complex, and scales are involved. The above expression would also be
2659 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
2660 /// This later form is less amenable to optimization though, and we are allowed
2661 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
2663 /// If we can't emit an optimized form for this expression, this returns null.
2665 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
2667 TargetData &TD = *IC.getTargetData();
2668 gep_type_iterator GTI = gep_type_begin(GEP);
2670 // Check to see if this gep only has a single variable index. If so, and if
2671 // any constant indices are a multiple of its scale, then we can compute this
2672 // in terms of the scale of the variable index. For example, if the GEP
2673 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
2674 // because the expression will cross zero at the same point.
2675 unsigned i, e = GEP->getNumOperands();
2677 for (i = 1; i != e; ++i, ++GTI) {
2678 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
2679 // Compute the aggregate offset of constant indices.
2680 if (CI->isZero()) continue;
2682 // Handle a struct index, which adds its field offset to the pointer.
2683 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
2684 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
2686 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
2687 Offset += Size*CI->getSExtValue();
2690 // Found our variable index.
2695 // If there are no variable indices, we must have a constant offset, just
2696 // evaluate it the general way.
2697 if (i == e) return 0;
2699 Value *VariableIdx = GEP->getOperand(i);
2700 // Determine the scale factor of the variable element. For example, this is
2701 // 4 if the variable index is into an array of i32.
2702 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
2704 // Verify that there are no other variable indices. If so, emit the hard way.
2705 for (++i, ++GTI; i != e; ++i, ++GTI) {
2706 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
2709 // Compute the aggregate offset of constant indices.
2710 if (CI->isZero()) continue;
2712 // Handle a struct index, which adds its field offset to the pointer.
2713 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
2714 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
2716 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
2717 Offset += Size*CI->getSExtValue();
2721 // Okay, we know we have a single variable index, which must be a
2722 // pointer/array/vector index. If there is no offset, life is simple, return
2724 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2726 // Cast to intptrty in case a truncation occurs. If an extension is needed,
2727 // we don't need to bother extending: the extension won't affect where the
2728 // computation crosses zero.
2729 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
2730 VariableIdx = new TruncInst(VariableIdx,
2731 TD.getIntPtrType(VariableIdx->getContext()),
2732 VariableIdx->getName(), &I);
2736 // Otherwise, there is an index. The computation we will do will be modulo
2737 // the pointer size, so get it.
2738 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
2740 Offset &= PtrSizeMask;
2741 VariableScale &= PtrSizeMask;
2743 // To do this transformation, any constant index must be a multiple of the
2744 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
2745 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
2746 // multiple of the variable scale.
2747 int64_t NewOffs = Offset / (int64_t)VariableScale;
2748 if (Offset != NewOffs*(int64_t)VariableScale)
2751 // Okay, we can do this evaluation. Start by converting the index to intptr.
2752 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
2753 if (VariableIdx->getType() != IntPtrTy)
2754 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
2756 VariableIdx->getName(), &I);
2757 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
2758 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
2762 /// Optimize pointer differences into the same array into a size. Consider:
2763 /// &A[10] - &A[0]: we should compile this to "10". LHS/RHS are the pointer
2764 /// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
2766 Value *InstCombiner::OptimizePointerDifference(Value *LHS, Value *RHS,
2768 assert(TD && "Must have target data info for this");
2770 // If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
2773 GetElementPtrInst *GEP;
2775 if ((GEP = dyn_cast<GetElementPtrInst>(LHS)) &&
2776 GEP->getOperand(0) == RHS)
2778 else if ((GEP = dyn_cast<GetElementPtrInst>(RHS)) &&
2779 GEP->getOperand(0) == LHS)
2784 // TODO: Could also optimize &A[i] - &A[j] -> "i-j".
2786 // Emit the offset of the GEP and an intptr_t.
2787 Value *Result = EmitGEPOffset(GEP, *this);
2789 // If we have p - gep(p, ...) then we have to negate the result.
2791 Result = Builder->CreateNeg(Result, "diff.neg");
2793 return Builder->CreateIntCast(Result, Ty, true);
2797 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2798 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2800 if (Op0 == Op1) // sub X, X -> 0
2801 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2803 // If this is a 'B = x-(-A)', change to B = x+A. This preserves NSW/NUW.
2804 if (Value *V = dyn_castNegVal(Op1)) {
2805 BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);
2806 Res->setHasNoSignedWrap(I.hasNoSignedWrap());
2807 Res->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
2811 if (isa<UndefValue>(Op0))
2812 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2813 if (isa<UndefValue>(Op1))
2814 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2815 if (I.getType() == Type::getInt1Ty(*Context))
2816 return BinaryOperator::CreateXor(Op0, Op1);
2818 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2819 // Replace (-1 - A) with (~A).
2820 if (C->isAllOnesValue())
2821 return BinaryOperator::CreateNot(Op1);
2823 // C - ~X == X + (1+C)
2825 if (match(Op1, m_Not(m_Value(X))))
2826 return BinaryOperator::CreateAdd(X, AddOne(C));
2828 // -(X >>u 31) -> (X >>s 31)
2829 // -(X >>s 31) -> (X >>u 31)
2831 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2832 if (SI->getOpcode() == Instruction::LShr) {
2833 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2834 // Check to see if we are shifting out everything but the sign bit.
2835 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2836 SI->getType()->getPrimitiveSizeInBits()-1) {
2837 // Ok, the transformation is safe. Insert AShr.
2838 return BinaryOperator::Create(Instruction::AShr,
2839 SI->getOperand(0), CU, SI->getName());
2842 } else if (SI->getOpcode() == Instruction::AShr) {
2843 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2844 // Check to see if we are shifting out everything but the sign bit.
2845 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2846 SI->getType()->getPrimitiveSizeInBits()-1) {
2847 // Ok, the transformation is safe. Insert LShr.
2848 return BinaryOperator::CreateLShr(
2849 SI->getOperand(0), CU, SI->getName());
2856 // Try to fold constant sub into select arguments.
2857 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2858 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2861 // C - zext(bool) -> bool ? C - 1 : C
2862 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2863 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2864 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2867 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2868 if (Op1I->getOpcode() == Instruction::Add) {
2869 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2870 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2872 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2873 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2875 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2876 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2877 // C1-(X+C2) --> (C1-C2)-X
2878 return BinaryOperator::CreateSub(
2879 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2883 if (Op1I->hasOneUse()) {
2884 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2885 // is not used by anyone else...
2887 if (Op1I->getOpcode() == Instruction::Sub) {
2888 // Swap the two operands of the subexpr...
2889 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2890 Op1I->setOperand(0, IIOp1);
2891 Op1I->setOperand(1, IIOp0);
2893 // Create the new top level add instruction...
2894 return BinaryOperator::CreateAdd(Op0, Op1);
2897 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2899 if (Op1I->getOpcode() == Instruction::And &&
2900 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2901 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2903 Value *NewNot = Builder->CreateNot(OtherOp, "B.not");
2904 return BinaryOperator::CreateAnd(Op0, NewNot);
2907 // 0 - (X sdiv C) -> (X sdiv -C)
2908 if (Op1I->getOpcode() == Instruction::SDiv)
2909 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2911 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2912 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2913 ConstantExpr::getNeg(DivRHS));
2915 // X - X*C --> X * (1-C)
2916 ConstantInt *C2 = 0;
2917 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2919 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2921 return BinaryOperator::CreateMul(Op0, CP1);
2926 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2927 if (Op0I->getOpcode() == Instruction::Add) {
2928 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2929 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2930 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2931 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2932 } else if (Op0I->getOpcode() == Instruction::Sub) {
2933 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2934 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2940 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2941 if (X == Op1) // X*C - X --> X * (C-1)
2942 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2944 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2945 if (X == dyn_castFoldableMul(Op1, C2))
2946 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2949 // Optimize pointer differences into the same array into a size. Consider:
2950 // &A[10] - &A[0]: we should compile this to "10".
2952 if (PtrToIntInst *LHS = dyn_cast<PtrToIntInst>(Op0))
2953 if (PtrToIntInst *RHS = dyn_cast<PtrToIntInst>(Op1))
2954 if (Value *Res = OptimizePointerDifference(LHS->getOperand(0),
2957 return ReplaceInstUsesWith(I, Res);
2959 // trunc(p)-trunc(q) -> trunc(p-q)
2960 if (TruncInst *LHST = dyn_cast<TruncInst>(Op0))
2961 if (TruncInst *RHST = dyn_cast<TruncInst>(Op1))
2962 if (PtrToIntInst *LHS = dyn_cast<PtrToIntInst>(LHST->getOperand(0)))
2963 if (PtrToIntInst *RHS = dyn_cast<PtrToIntInst>(RHST->getOperand(0)))
2964 if (Value *Res = OptimizePointerDifference(LHS->getOperand(0),
2967 return ReplaceInstUsesWith(I, Res);
2973 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2974 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2976 // If this is a 'B = x-(-A)', change to B = x+A...
2977 if (Value *V = dyn_castFNegVal(Op1))
2978 return BinaryOperator::CreateFAdd(Op0, V);
2980 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2981 if (Op1I->getOpcode() == Instruction::FAdd) {
2982 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2983 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2985 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2986 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2994 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2995 /// comparison only checks the sign bit. If it only checks the sign bit, set
2996 /// TrueIfSigned if the result of the comparison is true when the input value is
2998 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2999 bool &TrueIfSigned) {
3001 case ICmpInst::ICMP_SLT: // True if LHS s< 0
3002 TrueIfSigned = true;
3003 return RHS->isZero();
3004 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
3005 TrueIfSigned = true;
3006 return RHS->isAllOnesValue();
3007 case ICmpInst::ICMP_SGT: // True if LHS s> -1
3008 TrueIfSigned = false;
3009 return RHS->isAllOnesValue();
3010 case ICmpInst::ICMP_UGT:
3011 // True if LHS u> RHS and RHS == high-bit-mask - 1
3012 TrueIfSigned = true;
3013 return RHS->getValue() ==
3014 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
3015 case ICmpInst::ICMP_UGE:
3016 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
3017 TrueIfSigned = true;
3018 return RHS->getValue().isSignBit();
3024 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
3025 bool Changed = SimplifyCommutative(I);
3026 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3028 if (isa<UndefValue>(Op1)) // undef * X -> 0
3029 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3031 // Simplify mul instructions with a constant RHS.
3032 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3033 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1C)) {
3035 // ((X << C1)*C2) == (X * (C2 << C1))
3036 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
3037 if (SI->getOpcode() == Instruction::Shl)
3038 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
3039 return BinaryOperator::CreateMul(SI->getOperand(0),
3040 ConstantExpr::getShl(CI, ShOp));
3043 return ReplaceInstUsesWith(I, Op1C); // X * 0 == 0
3044 if (CI->equalsInt(1)) // X * 1 == X
3045 return ReplaceInstUsesWith(I, Op0);
3046 if (CI->isAllOnesValue()) // X * -1 == 0 - X
3047 return BinaryOperator::CreateNeg(Op0, I.getName());
3049 const APInt& Val = cast<ConstantInt>(CI)->getValue();
3050 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
3051 return BinaryOperator::CreateShl(Op0,
3052 ConstantInt::get(Op0->getType(), Val.logBase2()));
3054 } else if (isa<VectorType>(Op1C->getType())) {
3055 if (Op1C->isNullValue())
3056 return ReplaceInstUsesWith(I, Op1C);
3058 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
3059 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
3060 return BinaryOperator::CreateNeg(Op0, I.getName());
3062 // As above, vector X*splat(1.0) -> X in all defined cases.
3063 if (Constant *Splat = Op1V->getSplatValue()) {
3064 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
3065 if (CI->equalsInt(1))
3066 return ReplaceInstUsesWith(I, Op0);
3071 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3072 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
3073 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1C)) {
3074 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
3075 Value *Add = Builder->CreateMul(Op0I->getOperand(0), Op1C, "tmp");
3076 Value *C1C2 = Builder->CreateMul(Op1C, Op0I->getOperand(1));
3077 return BinaryOperator::CreateAdd(Add, C1C2);
3081 // Try to fold constant mul into select arguments.
3082 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3083 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3086 if (isa<PHINode>(Op0))
3087 if (Instruction *NV = FoldOpIntoPhi(I))
3091 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
3092 if (Value *Op1v = dyn_castNegVal(Op1))
3093 return BinaryOperator::CreateMul(Op0v, Op1v);
3095 // (X / Y) * Y = X - (X % Y)
3096 // (X / Y) * -Y = (X % Y) - X
3099 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
3101 (BO->getOpcode() != Instruction::UDiv &&
3102 BO->getOpcode() != Instruction::SDiv)) {
3104 BO = dyn_cast<BinaryOperator>(Op1);
3106 Value *Neg = dyn_castNegVal(Op1C);
3107 if (BO && BO->hasOneUse() &&
3108 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
3109 (BO->getOpcode() == Instruction::UDiv ||
3110 BO->getOpcode() == Instruction::SDiv)) {
3111 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
3113 // If the division is exact, X % Y is zero.
3114 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
3115 if (SDiv->isExact()) {
3117 return ReplaceInstUsesWith(I, Op0BO);
3118 return BinaryOperator::CreateNeg(Op0BO);
3122 if (BO->getOpcode() == Instruction::UDiv)
3123 Rem = Builder->CreateURem(Op0BO, Op1BO);
3125 Rem = Builder->CreateSRem(Op0BO, Op1BO);
3129 return BinaryOperator::CreateSub(Op0BO, Rem);
3130 return BinaryOperator::CreateSub(Rem, Op0BO);
3134 /// i1 mul -> i1 and.
3135 if (I.getType() == Type::getInt1Ty(*Context))
3136 return BinaryOperator::CreateAnd(Op0, Op1);
3138 // X*(1 << Y) --> X << Y
3139 // (1 << Y)*X --> X << Y
3142 if (match(Op0, m_Shl(m_One(), m_Value(Y))))
3143 return BinaryOperator::CreateShl(Op1, Y);
3144 if (match(Op1, m_Shl(m_One(), m_Value(Y))))
3145 return BinaryOperator::CreateShl(Op0, Y);
3148 // If one of the operands of the multiply is a cast from a boolean value, then
3149 // we know the bool is either zero or one, so this is a 'masking' multiply.
3150 // X * Y (where Y is 0 or 1) -> X & (0-Y)
3151 if (!isa<VectorType>(I.getType())) {
3152 // -2 is "-1 << 1" so it is all bits set except the low one.
3153 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
3155 Value *BoolCast = 0, *OtherOp = 0;
3156 if (MaskedValueIsZero(Op0, Negative2))
3157 BoolCast = Op0, OtherOp = Op1;
3158 else if (MaskedValueIsZero(Op1, Negative2))
3159 BoolCast = Op1, OtherOp = Op0;
3162 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
3164 return BinaryOperator::CreateAnd(V, OtherOp);
3168 return Changed ? &I : 0;
3171 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
3172 bool Changed = SimplifyCommutative(I);
3173 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3175 // Simplify mul instructions with a constant RHS...
3176 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
3177 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1C)) {
3178 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
3179 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
3180 if (Op1F->isExactlyValue(1.0))
3181 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
3182 } else if (isa<VectorType>(Op1C->getType())) {
3183 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
3184 // As above, vector X*splat(1.0) -> X in all defined cases.
3185 if (Constant *Splat = Op1V->getSplatValue()) {
3186 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
3187 if (F->isExactlyValue(1.0))
3188 return ReplaceInstUsesWith(I, Op0);
3193 // Try to fold constant mul into select arguments.
3194 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3195 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3198 if (isa<PHINode>(Op0))
3199 if (Instruction *NV = FoldOpIntoPhi(I))
3203 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
3204 if (Value *Op1v = dyn_castFNegVal(Op1))
3205 return BinaryOperator::CreateFMul(Op0v, Op1v);
3207 return Changed ? &I : 0;
3210 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
3212 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
3213 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
3215 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
3216 int NonNullOperand = -1;
3217 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3218 if (ST->isNullValue())
3220 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
3221 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3222 if (ST->isNullValue())
3225 if (NonNullOperand == -1)
3228 Value *SelectCond = SI->getOperand(0);
3230 // Change the div/rem to use 'Y' instead of the select.
3231 I.setOperand(1, SI->getOperand(NonNullOperand));
3233 // Okay, we know we replace the operand of the div/rem with 'Y' with no
3234 // problem. However, the select, or the condition of the select may have
3235 // multiple uses. Based on our knowledge that the operand must be non-zero,
3236 // propagate the known value for the select into other uses of it, and
3237 // propagate a known value of the condition into its other users.
3239 // If the select and condition only have a single use, don't bother with this,
3241 if (SI->use_empty() && SelectCond->hasOneUse())
3244 // Scan the current block backward, looking for other uses of SI.
3245 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
3247 while (BBI != BBFront) {
3249 // If we found a call to a function, we can't assume it will return, so
3250 // information from below it cannot be propagated above it.
3251 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
3254 // Replace uses of the select or its condition with the known values.
3255 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
3258 *I = SI->getOperand(NonNullOperand);
3260 } else if (*I == SelectCond) {
3261 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
3262 ConstantInt::getFalse(*Context);
3267 // If we past the instruction, quit looking for it.
3270 if (&*BBI == SelectCond)
3273 // If we ran out of things to eliminate, break out of the loop.
3274 if (SelectCond == 0 && SI == 0)
3282 /// This function implements the transforms on div instructions that work
3283 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
3284 /// used by the visitors to those instructions.
3285 /// @brief Transforms common to all three div instructions
3286 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
3287 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3289 // undef / X -> 0 for integer.
3290 // undef / X -> undef for FP (the undef could be a snan).
3291 if (isa<UndefValue>(Op0)) {
3292 if (Op0->getType()->isFPOrFPVector())
3293 return ReplaceInstUsesWith(I, Op0);
3294 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3297 // X / undef -> undef
3298 if (isa<UndefValue>(Op1))
3299 return ReplaceInstUsesWith(I, Op1);
3304 /// This function implements the transforms common to both integer division
3305 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3306 /// division instructions.
3307 /// @brief Common integer divide transforms
3308 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3309 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3311 // (sdiv X, X) --> 1 (udiv X, X) --> 1
3313 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
3314 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
3315 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
3316 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
3319 Constant *CI = ConstantInt::get(I.getType(), 1);
3320 return ReplaceInstUsesWith(I, CI);
3323 if (Instruction *Common = commonDivTransforms(I))
3326 // Handle cases involving: [su]div X, (select Cond, Y, Z)
3327 // This does not apply for fdiv.
3328 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3331 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3333 if (RHS->equalsInt(1))
3334 return ReplaceInstUsesWith(I, Op0);
3336 // (X / C1) / C2 -> X / (C1*C2)
3337 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3338 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3339 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3340 if (MultiplyOverflows(RHS, LHSRHS,
3341 I.getOpcode()==Instruction::SDiv))
3342 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3344 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3345 ConstantExpr::getMul(RHS, LHSRHS));
3348 if (!RHS->isZero()) { // avoid X udiv 0
3349 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3350 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3352 if (isa<PHINode>(Op0))
3353 if (Instruction *NV = FoldOpIntoPhi(I))
3358 // 0 / X == 0, we don't need to preserve faults!
3359 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3360 if (LHS->equalsInt(0))
3361 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3363 // It can't be division by zero, hence it must be division by one.
3364 if (I.getType() == Type::getInt1Ty(*Context))
3365 return ReplaceInstUsesWith(I, Op0);
3367 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3368 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3371 return ReplaceInstUsesWith(I, Op0);
3377 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3378 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3380 // Handle the integer div common cases
3381 if (Instruction *Common = commonIDivTransforms(I))
3384 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3385 // X udiv C^2 -> X >> C
3386 // Check to see if this is an unsigned division with an exact power of 2,
3387 // if so, convert to a right shift.
3388 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3389 return BinaryOperator::CreateLShr(Op0,
3390 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3392 // X udiv C, where C >= signbit
3393 if (C->getValue().isNegative()) {
3394 Value *IC = Builder->CreateICmpULT( Op0, C);
3395 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3396 ConstantInt::get(I.getType(), 1));
3400 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3401 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3402 if (RHSI->getOpcode() == Instruction::Shl &&
3403 isa<ConstantInt>(RHSI->getOperand(0))) {
3404 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3405 if (C1.isPowerOf2()) {
3406 Value *N = RHSI->getOperand(1);
3407 const Type *NTy = N->getType();
3408 if (uint32_t C2 = C1.logBase2())
3409 N = Builder->CreateAdd(N, ConstantInt::get(NTy, C2), "tmp");
3410 return BinaryOperator::CreateLShr(Op0, N);
3415 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3416 // where C1&C2 are powers of two.
3417 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3418 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3419 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3420 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3421 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3422 // Compute the shift amounts
3423 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3424 // Construct the "on true" case of the select
3425 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3426 Value *TSI = Builder->CreateLShr(Op0, TC, SI->getName()+".t");
3428 // Construct the "on false" case of the select
3429 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3430 Value *FSI = Builder->CreateLShr(Op0, FC, SI->getName()+".f");
3432 // construct the select instruction and return it.
3433 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3439 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3440 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3442 // Handle the integer div common cases
3443 if (Instruction *Common = commonIDivTransforms(I))
3446 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3448 if (RHS->isAllOnesValue())
3449 return BinaryOperator::CreateNeg(Op0);
3451 // sdiv X, C --> ashr X, log2(C)
3452 if (cast<SDivOperator>(&I)->isExact() &&
3453 RHS->getValue().isNonNegative() &&
3454 RHS->getValue().isPowerOf2()) {
3455 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3456 RHS->getValue().exactLogBase2());
3457 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3460 // -X/C --> X/-C provided the negation doesn't overflow.
3461 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3462 if (isa<Constant>(Sub->getOperand(0)) &&
3463 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3464 Sub->hasNoSignedWrap())
3465 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3466 ConstantExpr::getNeg(RHS));
3469 // If the sign bits of both operands are zero (i.e. we can prove they are
3470 // unsigned inputs), turn this into a udiv.
3471 if (I.getType()->isInteger()) {
3472 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3473 if (MaskedValueIsZero(Op0, Mask)) {
3474 if (MaskedValueIsZero(Op1, Mask)) {
3475 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3476 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3478 ConstantInt *ShiftedInt;
3479 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3480 ShiftedInt->getValue().isPowerOf2()) {
3481 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3482 // Safe because the only negative value (1 << Y) can take on is
3483 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3484 // the sign bit set.
3485 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3493 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3494 return commonDivTransforms(I);
3497 /// This function implements the transforms on rem instructions that work
3498 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3499 /// is used by the visitors to those instructions.
3500 /// @brief Transforms common to all three rem instructions
3501 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3502 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3504 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3505 if (I.getType()->isFPOrFPVector())
3506 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3507 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3509 if (isa<UndefValue>(Op1))
3510 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3512 // Handle cases involving: rem X, (select Cond, Y, Z)
3513 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3519 /// This function implements the transforms common to both integer remainder
3520 /// instructions (urem and srem). It is called by the visitors to those integer
3521 /// remainder instructions.
3522 /// @brief Common integer remainder transforms
3523 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3524 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3526 if (Instruction *common = commonRemTransforms(I))
3529 // 0 % X == 0 for integer, we don't need to preserve faults!
3530 if (Constant *LHS = dyn_cast<Constant>(Op0))
3531 if (LHS->isNullValue())
3532 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3534 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3535 // X % 0 == undef, we don't need to preserve faults!
3536 if (RHS->equalsInt(0))
3537 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3539 if (RHS->equalsInt(1)) // X % 1 == 0
3540 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3542 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3543 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3544 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3546 } else if (isa<PHINode>(Op0I)) {
3547 if (Instruction *NV = FoldOpIntoPhi(I))
3551 // See if we can fold away this rem instruction.
3552 if (SimplifyDemandedInstructionBits(I))
3560 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3561 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3563 if (Instruction *common = commonIRemTransforms(I))
3566 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3567 // X urem C^2 -> X and C
3568 // Check to see if this is an unsigned remainder with an exact power of 2,
3569 // if so, convert to a bitwise and.
3570 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3571 if (C->getValue().isPowerOf2())
3572 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3575 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3576 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3577 if (RHSI->getOpcode() == Instruction::Shl &&
3578 isa<ConstantInt>(RHSI->getOperand(0))) {
3579 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3580 Constant *N1 = Constant::getAllOnesValue(I.getType());
3581 Value *Add = Builder->CreateAdd(RHSI, N1, "tmp");
3582 return BinaryOperator::CreateAnd(Op0, Add);
3587 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3588 // where C1&C2 are powers of two.
3589 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3590 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3591 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3592 // STO == 0 and SFO == 0 handled above.
3593 if ((STO->getValue().isPowerOf2()) &&
3594 (SFO->getValue().isPowerOf2())) {
3595 Value *TrueAnd = Builder->CreateAnd(Op0, SubOne(STO),
3596 SI->getName()+".t");
3597 Value *FalseAnd = Builder->CreateAnd(Op0, SubOne(SFO),
3598 SI->getName()+".f");
3599 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3607 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3608 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3610 // Handle the integer rem common cases
3611 if (Instruction *Common = commonIRemTransforms(I))
3614 if (Value *RHSNeg = dyn_castNegVal(Op1))
3615 if (!isa<Constant>(RHSNeg) ||
3616 (isa<ConstantInt>(RHSNeg) &&
3617 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3619 Worklist.AddValue(I.getOperand(1));
3620 I.setOperand(1, RHSNeg);
3624 // If the sign bits of both operands are zero (i.e. we can prove they are
3625 // unsigned inputs), turn this into a urem.
3626 if (I.getType()->isInteger()) {
3627 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3628 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3629 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3630 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3634 // If it's a constant vector, flip any negative values positive.
3635 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3636 unsigned VWidth = RHSV->getNumOperands();
3638 bool hasNegative = false;
3639 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3640 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3641 if (RHS->getValue().isNegative())
3645 std::vector<Constant *> Elts(VWidth);
3646 for (unsigned i = 0; i != VWidth; ++i) {
3647 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3648 if (RHS->getValue().isNegative())
3649 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3655 Constant *NewRHSV = ConstantVector::get(Elts);
3656 if (NewRHSV != RHSV) {
3657 Worklist.AddValue(I.getOperand(1));
3658 I.setOperand(1, NewRHSV);
3667 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3668 return commonRemTransforms(I);
3671 // isOneBitSet - Return true if there is exactly one bit set in the specified
3673 static bool isOneBitSet(const ConstantInt *CI) {
3674 return CI->getValue().isPowerOf2();
3677 // isHighOnes - Return true if the constant is of the form 1+0+.
3678 // This is the same as lowones(~X).
3679 static bool isHighOnes(const ConstantInt *CI) {
3680 return (~CI->getValue() + 1).isPowerOf2();
3683 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3684 /// are carefully arranged to allow folding of expressions such as:
3686 /// (A < B) | (A > B) --> (A != B)
3688 /// Note that this is only valid if the first and second predicates have the
3689 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3691 /// Three bits are used to represent the condition, as follows:
3696 /// <=> Value Definition
3697 /// 000 0 Always false
3704 /// 111 7 Always true
3706 static unsigned getICmpCode(const ICmpInst *ICI) {
3707 switch (ICI->getPredicate()) {
3709 case ICmpInst::ICMP_UGT: return 1; // 001
3710 case ICmpInst::ICMP_SGT: return 1; // 001
3711 case ICmpInst::ICMP_EQ: return 2; // 010
3712 case ICmpInst::ICMP_UGE: return 3; // 011
3713 case ICmpInst::ICMP_SGE: return 3; // 011
3714 case ICmpInst::ICMP_ULT: return 4; // 100
3715 case ICmpInst::ICMP_SLT: return 4; // 100
3716 case ICmpInst::ICMP_NE: return 5; // 101
3717 case ICmpInst::ICMP_ULE: return 6; // 110
3718 case ICmpInst::ICMP_SLE: return 6; // 110
3721 llvm_unreachable("Invalid ICmp predicate!");
3726 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3727 /// predicate into a three bit mask. It also returns whether it is an ordered
3728 /// predicate by reference.
3729 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3732 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3733 case FCmpInst::FCMP_UNO: return 0; // 000
3734 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3735 case FCmpInst::FCMP_UGT: return 1; // 001
3736 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3737 case FCmpInst::FCMP_UEQ: return 2; // 010
3738 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3739 case FCmpInst::FCMP_UGE: return 3; // 011
3740 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3741 case FCmpInst::FCMP_ULT: return 4; // 100
3742 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3743 case FCmpInst::FCMP_UNE: return 5; // 101
3744 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3745 case FCmpInst::FCMP_ULE: return 6; // 110
3748 // Not expecting FCMP_FALSE and FCMP_TRUE;
3749 llvm_unreachable("Unexpected FCmp predicate!");
3754 /// getICmpValue - This is the complement of getICmpCode, which turns an
3755 /// opcode and two operands into either a constant true or false, or a brand
3756 /// new ICmp instruction. The sign is passed in to determine which kind
3757 /// of predicate to use in the new icmp instruction.
3758 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3759 LLVMContext *Context) {
3761 default: llvm_unreachable("Illegal ICmp code!");
3762 case 0: return ConstantInt::getFalse(*Context);
3765 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3767 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3768 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3771 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3773 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3776 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3778 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3779 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3782 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3784 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3785 case 7: return ConstantInt::getTrue(*Context);
3789 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3790 /// opcode and two operands into either a FCmp instruction. isordered is passed
3791 /// in to determine which kind of predicate to use in the new fcmp instruction.
3792 static Value *getFCmpValue(bool isordered, unsigned code,
3793 Value *LHS, Value *RHS, LLVMContext *Context) {
3795 default: llvm_unreachable("Illegal FCmp code!");
3798 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3800 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3803 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3805 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3808 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3810 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3813 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3815 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3818 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3820 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3823 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3825 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3828 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3830 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3831 case 7: return ConstantInt::getTrue(*Context);
3835 /// PredicatesFoldable - Return true if both predicates match sign or if at
3836 /// least one of them is an equality comparison (which is signless).
3837 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3838 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
3839 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
3840 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
3844 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3845 struct FoldICmpLogical {
3848 ICmpInst::Predicate pred;
3849 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3850 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3851 pred(ICI->getPredicate()) {}
3852 bool shouldApply(Value *V) const {
3853 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3854 if (PredicatesFoldable(pred, ICI->getPredicate()))
3855 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3856 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3859 Instruction *apply(Instruction &Log) const {
3860 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3861 if (ICI->getOperand(0) != LHS) {
3862 assert(ICI->getOperand(1) == LHS);
3863 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3866 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3867 unsigned LHSCode = getICmpCode(ICI);
3868 unsigned RHSCode = getICmpCode(RHSICI);
3870 switch (Log.getOpcode()) {
3871 case Instruction::And: Code = LHSCode & RHSCode; break;
3872 case Instruction::Or: Code = LHSCode | RHSCode; break;
3873 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3874 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3877 bool isSigned = RHSICI->isSigned() || ICI->isSigned();
3878 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3879 if (Instruction *I = dyn_cast<Instruction>(RV))
3881 // Otherwise, it's a constant boolean value...
3882 return IC.ReplaceInstUsesWith(Log, RV);
3885 } // end anonymous namespace
3887 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3888 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3889 // guaranteed to be a binary operator.
3890 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3892 ConstantInt *AndRHS,
3893 BinaryOperator &TheAnd) {
3894 Value *X = Op->getOperand(0);
3895 Constant *Together = 0;
3897 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3899 switch (Op->getOpcode()) {
3900 case Instruction::Xor:
3901 if (Op->hasOneUse()) {
3902 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3903 Value *And = Builder->CreateAnd(X, AndRHS);
3905 return BinaryOperator::CreateXor(And, Together);
3908 case Instruction::Or:
3909 if (Together == AndRHS) // (X | C) & C --> C
3910 return ReplaceInstUsesWith(TheAnd, AndRHS);
3912 if (Op->hasOneUse() && Together != OpRHS) {
3913 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3914 Value *Or = Builder->CreateOr(X, Together);
3916 return BinaryOperator::CreateAnd(Or, AndRHS);
3919 case Instruction::Add:
3920 if (Op->hasOneUse()) {
3921 // Adding a one to a single bit bit-field should be turned into an XOR
3922 // of the bit. First thing to check is to see if this AND is with a
3923 // single bit constant.
3924 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3926 // If there is only one bit set...
3927 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3928 // Ok, at this point, we know that we are masking the result of the
3929 // ADD down to exactly one bit. If the constant we are adding has
3930 // no bits set below this bit, then we can eliminate the ADD.
3931 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3933 // Check to see if any bits below the one bit set in AndRHSV are set.
3934 if ((AddRHS & (AndRHSV-1)) == 0) {
3935 // If not, the only thing that can effect the output of the AND is
3936 // the bit specified by AndRHSV. If that bit is set, the effect of
3937 // the XOR is to toggle the bit. If it is clear, then the ADD has
3939 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3940 TheAnd.setOperand(0, X);
3943 // Pull the XOR out of the AND.
3944 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
3945 NewAnd->takeName(Op);
3946 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3953 case Instruction::Shl: {
3954 // We know that the AND will not produce any of the bits shifted in, so if
3955 // the anded constant includes them, clear them now!
3957 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3958 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3959 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3960 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3962 if (CI->getValue() == ShlMask) {
3963 // Masking out bits that the shift already masks
3964 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3965 } else if (CI != AndRHS) { // Reducing bits set in and.
3966 TheAnd.setOperand(1, CI);
3971 case Instruction::LShr:
3973 // We know that the AND will not produce any of the bits shifted in, so if
3974 // the anded constant includes them, clear them now! This only applies to
3975 // unsigned shifts, because a signed shr may bring in set bits!
3977 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3978 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3979 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3980 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3982 if (CI->getValue() == ShrMask) {
3983 // Masking out bits that the shift already masks.
3984 return ReplaceInstUsesWith(TheAnd, Op);
3985 } else if (CI != AndRHS) {
3986 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3991 case Instruction::AShr:
3993 // See if this is shifting in some sign extension, then masking it out
3995 if (Op->hasOneUse()) {
3996 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3997 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3998 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3999 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
4000 if (C == AndRHS) { // Masking out bits shifted in.
4001 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
4002 // Make the argument unsigned.
4003 Value *ShVal = Op->getOperand(0);
4004 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
4005 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
4014 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
4015 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
4016 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
4017 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
4018 /// insert new instructions.
4019 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
4020 bool isSigned, bool Inside,
4022 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
4023 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
4024 "Lo is not <= Hi in range emission code!");
4027 if (Lo == Hi) // Trivially false.
4028 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
4030 // V >= Min && V < Hi --> V < Hi
4031 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
4032 ICmpInst::Predicate pred = (isSigned ?
4033 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
4034 return new ICmpInst(pred, V, Hi);
4037 // Emit V-Lo <u Hi-Lo
4038 Constant *NegLo = ConstantExpr::getNeg(Lo);
4039 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
4040 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
4041 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
4044 if (Lo == Hi) // Trivially true.
4045 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
4047 // V < Min || V >= Hi -> V > Hi-1
4048 Hi = SubOne(cast<ConstantInt>(Hi));
4049 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
4050 ICmpInst::Predicate pred = (isSigned ?
4051 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
4052 return new ICmpInst(pred, V, Hi);
4055 // Emit V-Lo >u Hi-1-Lo
4056 // Note that Hi has already had one subtracted from it, above.
4057 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
4058 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
4059 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
4060 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
4063 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
4064 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
4065 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
4066 // not, since all 1s are not contiguous.
4067 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
4068 const APInt& V = Val->getValue();
4069 uint32_t BitWidth = Val->getType()->getBitWidth();
4070 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
4072 // look for the first zero bit after the run of ones
4073 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
4074 // look for the first non-zero bit
4075 ME = V.getActiveBits();
4079 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
4080 /// where isSub determines whether the operator is a sub. If we can fold one of
4081 /// the following xforms:
4083 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
4084 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
4085 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
4087 /// return (A +/- B).
4089 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
4090 ConstantInt *Mask, bool isSub,
4092 Instruction *LHSI = dyn_cast<Instruction>(LHS);
4093 if (!LHSI || LHSI->getNumOperands() != 2 ||
4094 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
4096 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
4098 switch (LHSI->getOpcode()) {
4100 case Instruction::And:
4101 if (ConstantExpr::getAnd(N, Mask) == Mask) {
4102 // If the AndRHS is a power of two minus one (0+1+), this is simple.
4103 if ((Mask->getValue().countLeadingZeros() +
4104 Mask->getValue().countPopulation()) ==
4105 Mask->getValue().getBitWidth())
4108 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
4109 // part, we don't need any explicit masks to take them out of A. If that
4110 // is all N is, ignore it.
4111 uint32_t MB = 0, ME = 0;
4112 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
4113 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
4114 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
4115 if (MaskedValueIsZero(RHS, Mask))
4120 case Instruction::Or:
4121 case Instruction::Xor:
4122 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
4123 if ((Mask->getValue().countLeadingZeros() +
4124 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
4125 && ConstantExpr::getAnd(N, Mask)->isNullValue())
4131 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
4132 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
4135 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
4136 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
4137 ICmpInst *LHS, ICmpInst *RHS) {
4138 // (icmp eq A, null) & (icmp eq B, null) -->
4139 // (icmp eq (ptrtoint(A)|ptrtoint(B)), 0)
4141 LHS->getPredicate() == ICmpInst::ICMP_EQ &&
4142 RHS->getPredicate() == ICmpInst::ICMP_EQ &&
4143 isa<ConstantPointerNull>(LHS->getOperand(1)) &&
4144 isa<ConstantPointerNull>(RHS->getOperand(1))) {
4145 const Type *IntPtrTy = TD->getIntPtrType(I.getContext());
4146 Value *A = Builder->CreatePtrToInt(LHS->getOperand(0), IntPtrTy);
4147 Value *B = Builder->CreatePtrToInt(RHS->getOperand(0), IntPtrTy);
4148 Value *NewOr = Builder->CreateOr(A, B);
4149 return new ICmpInst(ICmpInst::ICMP_EQ, NewOr,
4150 Constant::getNullValue(IntPtrTy));
4154 ConstantInt *LHSCst, *RHSCst;
4155 ICmpInst::Predicate LHSCC, RHSCC;
4157 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
4158 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4159 m_ConstantInt(LHSCst))) ||
4160 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4161 m_ConstantInt(RHSCst))))
4164 if (LHSCst == RHSCst && LHSCC == RHSCC) {
4165 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
4166 // where C is a power of 2
4167 if (LHSCC == ICmpInst::ICMP_ULT &&
4168 LHSCst->getValue().isPowerOf2()) {
4169 Value *NewOr = Builder->CreateOr(Val, Val2);
4170 return new ICmpInst(LHSCC, NewOr, LHSCst);
4173 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
4174 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
4175 Value *NewOr = Builder->CreateOr(Val, Val2);
4176 return new ICmpInst(LHSCC, NewOr, LHSCst);
4180 // From here on, we only handle:
4181 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
4182 if (Val != Val2) return 0;
4184 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4185 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4186 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4187 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4188 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4191 // We can't fold (ugt x, C) & (sgt x, C2).
4192 if (!PredicatesFoldable(LHSCC, RHSCC))
4195 // Ensure that the larger constant is on the RHS.
4197 if (CmpInst::isSigned(LHSCC) ||
4198 (ICmpInst::isEquality(LHSCC) &&
4199 CmpInst::isSigned(RHSCC)))
4200 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4202 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4205 std::swap(LHS, RHS);
4206 std::swap(LHSCst, RHSCst);
4207 std::swap(LHSCC, RHSCC);
4210 // At this point, we know we have have two icmp instructions
4211 // comparing a value against two constants and and'ing the result
4212 // together. Because of the above check, we know that we only have
4213 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
4214 // (from the FoldICmpLogical check above), that the two constants
4215 // are not equal and that the larger constant is on the RHS
4216 assert(LHSCst != RHSCst && "Compares not folded above?");
4219 default: llvm_unreachable("Unknown integer condition code!");
4220 case ICmpInst::ICMP_EQ:
4222 default: llvm_unreachable("Unknown integer condition code!");
4223 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
4224 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
4225 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
4226 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4227 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
4228 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
4229 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
4230 return ReplaceInstUsesWith(I, LHS);
4232 case ICmpInst::ICMP_NE:
4234 default: llvm_unreachable("Unknown integer condition code!");
4235 case ICmpInst::ICMP_ULT:
4236 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
4237 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
4238 break; // (X != 13 & X u< 15) -> no change
4239 case ICmpInst::ICMP_SLT:
4240 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
4241 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
4242 break; // (X != 13 & X s< 15) -> no change
4243 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
4244 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
4245 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
4246 return ReplaceInstUsesWith(I, RHS);
4247 case ICmpInst::ICMP_NE:
4248 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
4249 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4250 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4251 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
4252 ConstantInt::get(Add->getType(), 1));
4254 break; // (X != 13 & X != 15) -> no change
4257 case ICmpInst::ICMP_ULT:
4259 default: llvm_unreachable("Unknown integer condition code!");
4260 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
4261 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
4262 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4263 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
4265 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
4266 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
4267 return ReplaceInstUsesWith(I, LHS);
4268 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
4272 case ICmpInst::ICMP_SLT:
4274 default: llvm_unreachable("Unknown integer condition code!");
4275 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
4276 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
4277 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4278 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
4280 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
4281 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
4282 return ReplaceInstUsesWith(I, LHS);
4283 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
4287 case ICmpInst::ICMP_UGT:
4289 default: llvm_unreachable("Unknown integer condition code!");
4290 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
4291 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
4292 return ReplaceInstUsesWith(I, RHS);
4293 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
4295 case ICmpInst::ICMP_NE:
4296 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
4297 return new ICmpInst(LHSCC, Val, RHSCst);
4298 break; // (X u> 13 & X != 15) -> no change
4299 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
4300 return InsertRangeTest(Val, AddOne(LHSCst),
4301 RHSCst, false, true, I);
4302 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4306 case ICmpInst::ICMP_SGT:
4308 default: llvm_unreachable("Unknown integer condition code!");
4309 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
4310 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4311 return ReplaceInstUsesWith(I, RHS);
4312 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4314 case ICmpInst::ICMP_NE:
4315 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4316 return new ICmpInst(LHSCC, Val, RHSCst);
4317 break; // (X s> 13 & X != 15) -> no change
4318 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
4319 return InsertRangeTest(Val, AddOne(LHSCst),
4320 RHSCst, true, true, I);
4321 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4330 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
4333 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4334 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4335 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4336 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4337 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4338 // If either of the constants are nans, then the whole thing returns
4340 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4341 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4342 return new FCmpInst(FCmpInst::FCMP_ORD,
4343 LHS->getOperand(0), RHS->getOperand(0));
4346 // Handle vector zeros. This occurs because the canonical form of
4347 // "fcmp ord x,x" is "fcmp ord x, 0".
4348 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4349 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4350 return new FCmpInst(FCmpInst::FCMP_ORD,
4351 LHS->getOperand(0), RHS->getOperand(0));
4355 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4356 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4357 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4360 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4361 // Swap RHS operands to match LHS.
4362 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4363 std::swap(Op1LHS, Op1RHS);
4366 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4367 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4369 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4371 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
4372 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4373 if (Op0CC == FCmpInst::FCMP_TRUE)
4374 return ReplaceInstUsesWith(I, RHS);
4375 if (Op1CC == FCmpInst::FCMP_TRUE)
4376 return ReplaceInstUsesWith(I, LHS);
4380 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4381 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4383 std::swap(LHS, RHS);
4384 std::swap(Op0Pred, Op1Pred);
4385 std::swap(Op0Ordered, Op1Ordered);
4388 // uno && ueq -> uno && (uno || eq) -> ueq
4389 // ord && olt -> ord && (ord && lt) -> olt
4390 if (Op0Ordered == Op1Ordered)
4391 return ReplaceInstUsesWith(I, RHS);
4393 // uno && oeq -> uno && (ord && eq) -> false
4394 // uno && ord -> false
4396 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4397 // ord && ueq -> ord && (uno || eq) -> oeq
4398 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4399 Op0LHS, Op0RHS, Context));
4407 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4408 bool Changed = SimplifyCommutative(I);
4409 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4411 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
4412 return ReplaceInstUsesWith(I, V);
4414 // See if we can simplify any instructions used by the instruction whose sole
4415 // purpose is to compute bits we don't care about.
4416 if (SimplifyDemandedInstructionBits(I))
4420 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4421 const APInt &AndRHSMask = AndRHS->getValue();
4422 APInt NotAndRHS(~AndRHSMask);
4424 // Optimize a variety of ((val OP C1) & C2) combinations...
4425 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4426 Value *Op0LHS = Op0I->getOperand(0);
4427 Value *Op0RHS = Op0I->getOperand(1);
4428 switch (Op0I->getOpcode()) {
4430 case Instruction::Xor:
4431 case Instruction::Or:
4432 // If the mask is only needed on one incoming arm, push it up.
4433 if (!Op0I->hasOneUse()) break;
4435 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4436 // Not masking anything out for the LHS, move to RHS.
4437 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
4438 Op0RHS->getName()+".masked");
4439 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
4441 if (!isa<Constant>(Op0RHS) &&
4442 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4443 // Not masking anything out for the RHS, move to LHS.
4444 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
4445 Op0LHS->getName()+".masked");
4446 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
4450 case Instruction::Add:
4451 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4452 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4453 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4454 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4455 return BinaryOperator::CreateAnd(V, AndRHS);
4456 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4457 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4460 case Instruction::Sub:
4461 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4462 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4463 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4464 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4465 return BinaryOperator::CreateAnd(V, AndRHS);
4467 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4468 // has 1's for all bits that the subtraction with A might affect.
4469 if (Op0I->hasOneUse()) {
4470 uint32_t BitWidth = AndRHSMask.getBitWidth();
4471 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4472 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4474 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4475 if (!(A && A->isZero()) && // avoid infinite recursion.
4476 MaskedValueIsZero(Op0LHS, Mask)) {
4477 Value *NewNeg = Builder->CreateNeg(Op0RHS);
4478 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4483 case Instruction::Shl:
4484 case Instruction::LShr:
4485 // (1 << x) & 1 --> zext(x == 0)
4486 // (1 >> x) & 1 --> zext(x == 0)
4487 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4489 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
4490 return new ZExtInst(NewICmp, I.getType());
4495 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4496 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4498 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4499 // If this is an integer truncation or change from signed-to-unsigned, and
4500 // if the source is an and/or with immediate, transform it. This
4501 // frequently occurs for bitfield accesses.
4502 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4503 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4504 CastOp->getNumOperands() == 2)
4505 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
4506 if (CastOp->getOpcode() == Instruction::And) {
4507 // Change: and (cast (and X, C1) to T), C2
4508 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4509 // This will fold the two constants together, which may allow
4510 // other simplifications.
4511 Value *NewCast = Builder->CreateTruncOrBitCast(
4512 CastOp->getOperand(0), I.getType(),
4513 CastOp->getName()+".shrunk");
4514 // trunc_or_bitcast(C1)&C2
4515 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4516 C3 = ConstantExpr::getAnd(C3, AndRHS);
4517 return BinaryOperator::CreateAnd(NewCast, C3);
4518 } else if (CastOp->getOpcode() == Instruction::Or) {
4519 // Change: and (cast (or X, C1) to T), C2
4520 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4521 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4522 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4524 return ReplaceInstUsesWith(I, AndRHS);
4530 // Try to fold constant and into select arguments.
4531 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4532 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4534 if (isa<PHINode>(Op0))
4535 if (Instruction *NV = FoldOpIntoPhi(I))
4540 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4541 if (Value *Op0NotVal = dyn_castNotVal(Op0))
4542 if (Value *Op1NotVal = dyn_castNotVal(Op1))
4543 if (Op0->hasOneUse() && Op1->hasOneUse()) {
4544 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
4545 I.getName()+".demorgan");
4546 return BinaryOperator::CreateNot(Or);
4550 Value *A = 0, *B = 0, *C = 0, *D = 0;
4551 // (A|B) & ~(A&B) -> A^B
4552 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
4553 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
4554 ((A == C && B == D) || (A == D && B == C)))
4555 return BinaryOperator::CreateXor(A, B);
4557 // ~(A&B) & (A|B) -> A^B
4558 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
4559 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
4560 ((A == C && B == D) || (A == D && B == C)))
4561 return BinaryOperator::CreateXor(A, B);
4563 if (Op0->hasOneUse() &&
4564 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4565 if (A == Op1) { // (A^B)&A -> A&(A^B)
4566 I.swapOperands(); // Simplify below
4567 std::swap(Op0, Op1);
4568 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4569 cast<BinaryOperator>(Op0)->swapOperands();
4570 I.swapOperands(); // Simplify below
4571 std::swap(Op0, Op1);
4575 if (Op1->hasOneUse() &&
4576 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4577 if (B == Op0) { // B&(A^B) -> B&(B^A)
4578 cast<BinaryOperator>(Op1)->swapOperands();
4581 if (A == Op0) // A&(A^B) -> A & ~B
4582 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
4585 // (A&((~A)|B)) -> A&B
4586 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4587 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4588 return BinaryOperator::CreateAnd(A, Op1);
4589 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4590 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4591 return BinaryOperator::CreateAnd(A, Op0);
4594 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4595 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4596 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4599 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4600 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4604 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4605 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4606 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4607 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4608 const Type *SrcTy = Op0C->getOperand(0)->getType();
4609 if (SrcTy == Op1C->getOperand(0)->getType() &&
4610 SrcTy->isIntOrIntVector() &&
4611 // Only do this if the casts both really cause code to be generated.
4612 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4614 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4616 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
4617 Op1C->getOperand(0), I.getName());
4618 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4622 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4623 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4624 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4625 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4626 SI0->getOperand(1) == SI1->getOperand(1) &&
4627 (SI0->hasOneUse() || SI1->hasOneUse())) {
4629 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
4631 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4632 SI1->getOperand(1));
4636 // If and'ing two fcmp, try combine them into one.
4637 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4638 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4639 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4643 return Changed ? &I : 0;
4646 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4647 /// capable of providing pieces of a bswap. The subexpression provides pieces
4648 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4649 /// the expression came from the corresponding "byte swapped" byte in some other
4650 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4651 /// we know that the expression deposits the low byte of %X into the high byte
4652 /// of the bswap result and that all other bytes are zero. This expression is
4653 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4656 /// This function returns true if the match was unsuccessful and false if so.
4657 /// On entry to the function the "OverallLeftShift" is a signed integer value
4658 /// indicating the number of bytes that the subexpression is later shifted. For
4659 /// example, if the expression is later right shifted by 16 bits, the
4660 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4661 /// byte of ByteValues is actually being set.
4663 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4664 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4665 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4666 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4667 /// always in the local (OverallLeftShift) coordinate space.
4669 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4670 SmallVector<Value*, 8> &ByteValues) {
4671 if (Instruction *I = dyn_cast<Instruction>(V)) {
4672 // If this is an or instruction, it may be an inner node of the bswap.
4673 if (I->getOpcode() == Instruction::Or) {
4674 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4676 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4680 // If this is a logical shift by a constant multiple of 8, recurse with
4681 // OverallLeftShift and ByteMask adjusted.
4682 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4684 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4685 // Ensure the shift amount is defined and of a byte value.
4686 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4689 unsigned ByteShift = ShAmt >> 3;
4690 if (I->getOpcode() == Instruction::Shl) {
4691 // X << 2 -> collect(X, +2)
4692 OverallLeftShift += ByteShift;
4693 ByteMask >>= ByteShift;
4695 // X >>u 2 -> collect(X, -2)
4696 OverallLeftShift -= ByteShift;
4697 ByteMask <<= ByteShift;
4698 ByteMask &= (~0U >> (32-ByteValues.size()));
4701 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4702 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4704 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4708 // If this is a logical 'and' with a mask that clears bytes, clear the
4709 // corresponding bytes in ByteMask.
4710 if (I->getOpcode() == Instruction::And &&
4711 isa<ConstantInt>(I->getOperand(1))) {
4712 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4713 unsigned NumBytes = ByteValues.size();
4714 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4715 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4717 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4718 // If this byte is masked out by a later operation, we don't care what
4720 if ((ByteMask & (1 << i)) == 0)
4723 // If the AndMask is all zeros for this byte, clear the bit.
4724 APInt MaskB = AndMask & Byte;
4726 ByteMask &= ~(1U << i);
4730 // If the AndMask is not all ones for this byte, it's not a bytezap.
4734 // Otherwise, this byte is kept.
4737 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4742 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4743 // the input value to the bswap. Some observations: 1) if more than one byte
4744 // is demanded from this input, then it could not be successfully assembled
4745 // into a byteswap. At least one of the two bytes would not be aligned with
4746 // their ultimate destination.
4747 if (!isPowerOf2_32(ByteMask)) return true;
4748 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4750 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4751 // is demanded, it needs to go into byte 0 of the result. This means that the
4752 // byte needs to be shifted until it lands in the right byte bucket. The
4753 // shift amount depends on the position: if the byte is coming from the high
4754 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4755 // low part, it must be shifted left.
4756 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4757 if (InputByteNo < ByteValues.size()/2) {
4758 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4761 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4765 // If the destination byte value is already defined, the values are or'd
4766 // together, which isn't a bswap (unless it's an or of the same bits).
4767 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4769 ByteValues[DestByteNo] = V;
4773 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4774 /// If so, insert the new bswap intrinsic and return it.
4775 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4776 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4777 if (!ITy || ITy->getBitWidth() % 16 ||
4778 // ByteMask only allows up to 32-byte values.
4779 ITy->getBitWidth() > 32*8)
4780 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4782 /// ByteValues - For each byte of the result, we keep track of which value
4783 /// defines each byte.
4784 SmallVector<Value*, 8> ByteValues;
4785 ByteValues.resize(ITy->getBitWidth()/8);
4787 // Try to find all the pieces corresponding to the bswap.
4788 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4789 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4792 // Check to see if all of the bytes come from the same value.
4793 Value *V = ByteValues[0];
4794 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4796 // Check to make sure that all of the bytes come from the same value.
4797 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4798 if (ByteValues[i] != V)
4800 const Type *Tys[] = { ITy };
4801 Module *M = I.getParent()->getParent()->getParent();
4802 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4803 return CallInst::Create(F, V);
4806 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4807 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4808 /// we can simplify this expression to "cond ? C : D or B".
4809 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4811 LLVMContext *Context) {
4812 // If A is not a select of -1/0, this cannot match.
4814 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4817 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4818 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4819 return SelectInst::Create(Cond, C, B);
4820 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4821 return SelectInst::Create(Cond, C, B);
4822 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4823 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4824 return SelectInst::Create(Cond, C, D);
4825 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4826 return SelectInst::Create(Cond, C, D);
4830 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4831 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4832 ICmpInst *LHS, ICmpInst *RHS) {
4833 // (icmp ne A, null) | (icmp ne B, null) -->
4834 // (icmp ne (ptrtoint(A)|ptrtoint(B)), 0)
4836 LHS->getPredicate() == ICmpInst::ICMP_NE &&
4837 RHS->getPredicate() == ICmpInst::ICMP_NE &&
4838 isa<ConstantPointerNull>(LHS->getOperand(1)) &&
4839 isa<ConstantPointerNull>(RHS->getOperand(1))) {
4840 const Type *IntPtrTy = TD->getIntPtrType(I.getContext());
4841 Value *A = Builder->CreatePtrToInt(LHS->getOperand(0), IntPtrTy);
4842 Value *B = Builder->CreatePtrToInt(RHS->getOperand(0), IntPtrTy);
4843 Value *NewOr = Builder->CreateOr(A, B);
4844 return new ICmpInst(ICmpInst::ICMP_NE, NewOr,
4845 Constant::getNullValue(IntPtrTy));
4849 ConstantInt *LHSCst, *RHSCst;
4850 ICmpInst::Predicate LHSCC, RHSCC;
4852 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4853 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4854 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4858 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
4859 if (LHSCst == RHSCst && LHSCC == RHSCC &&
4860 LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
4861 Value *NewOr = Builder->CreateOr(Val, Val2);
4862 return new ICmpInst(LHSCC, NewOr, LHSCst);
4865 // From here on, we only handle:
4866 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4867 if (Val != Val2) return 0;
4869 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4870 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4871 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4872 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4873 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4876 // We can't fold (ugt x, C) | (sgt x, C2).
4877 if (!PredicatesFoldable(LHSCC, RHSCC))
4880 // Ensure that the larger constant is on the RHS.
4882 if (CmpInst::isSigned(LHSCC) ||
4883 (ICmpInst::isEquality(LHSCC) &&
4884 CmpInst::isSigned(RHSCC)))
4885 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4887 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4890 std::swap(LHS, RHS);
4891 std::swap(LHSCst, RHSCst);
4892 std::swap(LHSCC, RHSCC);
4895 // At this point, we know we have have two icmp instructions
4896 // comparing a value against two constants and or'ing the result
4897 // together. Because of the above check, we know that we only have
4898 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4899 // FoldICmpLogical check above), that the two constants are not
4901 assert(LHSCst != RHSCst && "Compares not folded above?");
4904 default: llvm_unreachable("Unknown integer condition code!");
4905 case ICmpInst::ICMP_EQ:
4907 default: llvm_unreachable("Unknown integer condition code!");
4908 case ICmpInst::ICMP_EQ:
4909 if (LHSCst == SubOne(RHSCst)) {
4910 // (X == 13 | X == 14) -> X-13 <u 2
4911 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4912 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4913 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4914 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4916 break; // (X == 13 | X == 15) -> no change
4917 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4918 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4920 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4921 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4922 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4923 return ReplaceInstUsesWith(I, RHS);
4926 case ICmpInst::ICMP_NE:
4928 default: llvm_unreachable("Unknown integer condition code!");
4929 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4930 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4931 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4932 return ReplaceInstUsesWith(I, LHS);
4933 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4934 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4935 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4936 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4939 case ICmpInst::ICMP_ULT:
4941 default: llvm_unreachable("Unknown integer condition code!");
4942 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4944 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4945 // If RHSCst is [us]MAXINT, it is always false. Not handling
4946 // this can cause overflow.
4947 if (RHSCst->isMaxValue(false))
4948 return ReplaceInstUsesWith(I, LHS);
4949 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4951 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4953 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4954 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4955 return ReplaceInstUsesWith(I, RHS);
4956 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4960 case ICmpInst::ICMP_SLT:
4962 default: llvm_unreachable("Unknown integer condition code!");
4963 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4965 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4966 // If RHSCst is [us]MAXINT, it is always false. Not handling
4967 // this can cause overflow.
4968 if (RHSCst->isMaxValue(true))
4969 return ReplaceInstUsesWith(I, LHS);
4970 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4972 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4974 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4975 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4976 return ReplaceInstUsesWith(I, RHS);
4977 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4981 case ICmpInst::ICMP_UGT:
4983 default: llvm_unreachable("Unknown integer condition code!");
4984 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4985 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4986 return ReplaceInstUsesWith(I, LHS);
4987 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4989 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4990 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4991 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4992 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4996 case ICmpInst::ICMP_SGT:
4998 default: llvm_unreachable("Unknown integer condition code!");
4999 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
5000 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
5001 return ReplaceInstUsesWith(I, LHS);
5002 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
5004 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
5005 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
5006 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5007 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
5015 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
5017 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
5018 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
5019 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
5020 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
5021 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
5022 // If either of the constants are nans, then the whole thing returns
5024 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
5025 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5027 // Otherwise, no need to compare the two constants, compare the
5029 return new FCmpInst(FCmpInst::FCMP_UNO,
5030 LHS->getOperand(0), RHS->getOperand(0));
5033 // Handle vector zeros. This occurs because the canonical form of
5034 // "fcmp uno x,x" is "fcmp uno x, 0".
5035 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
5036 isa<ConstantAggregateZero>(RHS->getOperand(1)))
5037 return new FCmpInst(FCmpInst::FCMP_UNO,
5038 LHS->getOperand(0), RHS->getOperand(0));
5043 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
5044 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
5045 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
5047 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
5048 // Swap RHS operands to match LHS.
5049 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
5050 std::swap(Op1LHS, Op1RHS);
5052 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
5053 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
5055 return new FCmpInst((FCmpInst::Predicate)Op0CC,
5057 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
5058 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5059 if (Op0CC == FCmpInst::FCMP_FALSE)
5060 return ReplaceInstUsesWith(I, RHS);
5061 if (Op1CC == FCmpInst::FCMP_FALSE)
5062 return ReplaceInstUsesWith(I, LHS);
5065 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
5066 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
5067 if (Op0Ordered == Op1Ordered) {
5068 // If both are ordered or unordered, return a new fcmp with
5069 // or'ed predicates.
5070 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
5071 Op0LHS, Op0RHS, Context);
5072 if (Instruction *I = dyn_cast<Instruction>(RV))
5074 // Otherwise, it's a constant boolean value...
5075 return ReplaceInstUsesWith(I, RV);
5081 /// FoldOrWithConstants - This helper function folds:
5083 /// ((A | B) & C1) | (B & C2)
5089 /// when the XOR of the two constants is "all ones" (-1).
5090 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
5091 Value *A, Value *B, Value *C) {
5092 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
5096 ConstantInt *CI2 = 0;
5097 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
5099 APInt Xor = CI1->getValue() ^ CI2->getValue();
5100 if (!Xor.isAllOnesValue()) return 0;
5102 if (V1 == A || V1 == B) {
5103 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
5104 return BinaryOperator::CreateOr(NewOp, V1);
5110 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
5111 bool Changed = SimplifyCommutative(I);
5112 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5114 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
5115 return ReplaceInstUsesWith(I, V);
5118 // See if we can simplify any instructions used by the instruction whose sole
5119 // purpose is to compute bits we don't care about.
5120 if (SimplifyDemandedInstructionBits(I))
5123 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5124 ConstantInt *C1 = 0; Value *X = 0;
5125 // (X & C1) | C2 --> (X | C2) & (C1|C2)
5126 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
5128 Value *Or = Builder->CreateOr(X, RHS);
5130 return BinaryOperator::CreateAnd(Or,
5131 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
5134 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
5135 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
5137 Value *Or = Builder->CreateOr(X, RHS);
5139 return BinaryOperator::CreateXor(Or,
5140 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
5143 // Try to fold constant and into select arguments.
5144 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5145 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5147 if (isa<PHINode>(Op0))
5148 if (Instruction *NV = FoldOpIntoPhi(I))
5152 Value *A = 0, *B = 0;
5153 ConstantInt *C1 = 0, *C2 = 0;
5155 // (A | B) | C and A | (B | C) -> bswap if possible.
5156 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
5157 if (match(Op0, m_Or(m_Value(), m_Value())) ||
5158 match(Op1, m_Or(m_Value(), m_Value())) ||
5159 (match(Op0, m_Shift(m_Value(), m_Value())) &&
5160 match(Op1, m_Shift(m_Value(), m_Value())))) {
5161 if (Instruction *BSwap = MatchBSwap(I))
5165 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
5166 if (Op0->hasOneUse() &&
5167 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
5168 MaskedValueIsZero(Op1, C1->getValue())) {
5169 Value *NOr = Builder->CreateOr(A, Op1);
5171 return BinaryOperator::CreateXor(NOr, C1);
5174 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
5175 if (Op1->hasOneUse() &&
5176 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
5177 MaskedValueIsZero(Op0, C1->getValue())) {
5178 Value *NOr = Builder->CreateOr(A, Op0);
5180 return BinaryOperator::CreateXor(NOr, C1);
5184 Value *C = 0, *D = 0;
5185 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
5186 match(Op1, m_And(m_Value(B), m_Value(D)))) {
5187 Value *V1 = 0, *V2 = 0, *V3 = 0;
5188 C1 = dyn_cast<ConstantInt>(C);
5189 C2 = dyn_cast<ConstantInt>(D);
5190 if (C1 && C2) { // (A & C1)|(B & C2)
5191 // If we have: ((V + N) & C1) | (V & C2)
5192 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
5193 // replace with V+N.
5194 if (C1->getValue() == ~C2->getValue()) {
5195 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
5196 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
5197 // Add commutes, try both ways.
5198 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
5199 return ReplaceInstUsesWith(I, A);
5200 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
5201 return ReplaceInstUsesWith(I, A);
5203 // Or commutes, try both ways.
5204 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
5205 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
5206 // Add commutes, try both ways.
5207 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
5208 return ReplaceInstUsesWith(I, B);
5209 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
5210 return ReplaceInstUsesWith(I, B);
5213 V1 = 0; V2 = 0; V3 = 0;
5216 // Check to see if we have any common things being and'ed. If so, find the
5217 // terms for V1 & (V2|V3).
5218 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
5219 if (A == B) // (A & C)|(A & D) == A & (C|D)
5220 V1 = A, V2 = C, V3 = D;
5221 else if (A == D) // (A & C)|(B & A) == A & (B|C)
5222 V1 = A, V2 = B, V3 = C;
5223 else if (C == B) // (A & C)|(C & D) == C & (A|D)
5224 V1 = C, V2 = A, V3 = D;
5225 else if (C == D) // (A & C)|(B & C) == C & (A|B)
5226 V1 = C, V2 = A, V3 = B;
5229 Value *Or = Builder->CreateOr(V2, V3, "tmp");
5230 return BinaryOperator::CreateAnd(V1, Or);
5234 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
5235 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
5237 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
5239 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
5241 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
5244 // ((A&~B)|(~A&B)) -> A^B
5245 if ((match(C, m_Not(m_Specific(D))) &&
5246 match(B, m_Not(m_Specific(A)))))
5247 return BinaryOperator::CreateXor(A, D);
5248 // ((~B&A)|(~A&B)) -> A^B
5249 if ((match(A, m_Not(m_Specific(D))) &&
5250 match(B, m_Not(m_Specific(C)))))
5251 return BinaryOperator::CreateXor(C, D);
5252 // ((A&~B)|(B&~A)) -> A^B
5253 if ((match(C, m_Not(m_Specific(B))) &&
5254 match(D, m_Not(m_Specific(A)))))
5255 return BinaryOperator::CreateXor(A, B);
5256 // ((~B&A)|(B&~A)) -> A^B
5257 if ((match(A, m_Not(m_Specific(B))) &&
5258 match(D, m_Not(m_Specific(C)))))
5259 return BinaryOperator::CreateXor(C, B);
5262 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
5263 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
5264 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
5265 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
5266 SI0->getOperand(1) == SI1->getOperand(1) &&
5267 (SI0->hasOneUse() || SI1->hasOneUse())) {
5268 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
5270 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
5271 SI1->getOperand(1));
5275 // ((A|B)&1)|(B&-2) -> (A&1) | B
5276 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
5277 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
5278 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
5279 if (Ret) return Ret;
5281 // (B&-2)|((A|B)&1) -> (A&1) | B
5282 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
5283 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
5284 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
5285 if (Ret) return Ret;
5288 // (~A | ~B) == (~(A & B)) - De Morgan's Law
5289 if (Value *Op0NotVal = dyn_castNotVal(Op0))
5290 if (Value *Op1NotVal = dyn_castNotVal(Op1))
5291 if (Op0->hasOneUse() && Op1->hasOneUse()) {
5292 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
5293 I.getName()+".demorgan");
5294 return BinaryOperator::CreateNot(And);
5297 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
5298 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
5299 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5302 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
5303 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
5307 // fold (or (cast A), (cast B)) -> (cast (or A, B))
5308 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5309 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5310 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
5311 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
5312 !isa<ICmpInst>(Op1C->getOperand(0))) {
5313 const Type *SrcTy = Op0C->getOperand(0)->getType();
5314 if (SrcTy == Op1C->getOperand(0)->getType() &&
5315 SrcTy->isIntOrIntVector() &&
5316 // Only do this if the casts both really cause code to be
5318 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5320 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5322 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
5323 Op1C->getOperand(0), I.getName());
5324 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5331 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
5332 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
5333 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
5334 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5338 return Changed ? &I : 0;
5343 // XorSelf - Implements: X ^ X --> 0
5346 XorSelf(Value *rhs) : RHS(rhs) {}
5347 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5348 Instruction *apply(BinaryOperator &Xor) const {
5355 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5356 bool Changed = SimplifyCommutative(I);
5357 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5359 if (isa<UndefValue>(Op1)) {
5360 if (isa<UndefValue>(Op0))
5361 // Handle undef ^ undef -> 0 special case. This is a common
5363 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5364 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5367 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5368 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5369 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5370 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5373 // See if we can simplify any instructions used by the instruction whose sole
5374 // purpose is to compute bits we don't care about.
5375 if (SimplifyDemandedInstructionBits(I))
5377 if (isa<VectorType>(I.getType()))
5378 if (isa<ConstantAggregateZero>(Op1))
5379 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5381 // Is this a ~ operation?
5382 if (Value *NotOp = dyn_castNotVal(&I)) {
5383 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5384 if (Op0I->getOpcode() == Instruction::And ||
5385 Op0I->getOpcode() == Instruction::Or) {
5386 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5387 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5388 if (dyn_castNotVal(Op0I->getOperand(1)))
5389 Op0I->swapOperands();
5390 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5392 Builder->CreateNot(Op0I->getOperand(1),
5393 Op0I->getOperand(1)->getName()+".not");
5394 if (Op0I->getOpcode() == Instruction::And)
5395 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5396 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5399 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
5400 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
5401 if (isFreeToInvert(Op0I->getOperand(0)) &&
5402 isFreeToInvert(Op0I->getOperand(1))) {
5404 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
5406 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
5407 if (Op0I->getOpcode() == Instruction::And)
5408 return BinaryOperator::CreateOr(NotX, NotY);
5409 return BinaryOperator::CreateAnd(NotX, NotY);
5416 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5417 if (RHS->isOne() && Op0->hasOneUse()) {
5418 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5419 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5420 return new ICmpInst(ICI->getInversePredicate(),
5421 ICI->getOperand(0), ICI->getOperand(1));
5423 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5424 return new FCmpInst(FCI->getInversePredicate(),
5425 FCI->getOperand(0), FCI->getOperand(1));
5428 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5429 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5430 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5431 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5432 Instruction::CastOps Opcode = Op0C->getOpcode();
5433 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
5434 (RHS == ConstantExpr::getCast(Opcode,
5435 ConstantInt::getTrue(*Context),
5436 Op0C->getDestTy()))) {
5437 CI->setPredicate(CI->getInversePredicate());
5438 return CastInst::Create(Opcode, CI, Op0C->getType());
5444 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5445 // ~(c-X) == X-c-1 == X+(-c-1)
5446 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5447 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5448 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5449 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5450 ConstantInt::get(I.getType(), 1));
5451 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5454 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5455 if (Op0I->getOpcode() == Instruction::Add) {
5456 // ~(X-c) --> (-c-1)-X
5457 if (RHS->isAllOnesValue()) {
5458 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5459 return BinaryOperator::CreateSub(
5460 ConstantExpr::getSub(NegOp0CI,
5461 ConstantInt::get(I.getType(), 1)),
5462 Op0I->getOperand(0));
5463 } else if (RHS->getValue().isSignBit()) {
5464 // (X + C) ^ signbit -> (X + C + signbit)
5465 Constant *C = ConstantInt::get(*Context,
5466 RHS->getValue() + Op0CI->getValue());
5467 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5470 } else if (Op0I->getOpcode() == Instruction::Or) {
5471 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5472 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5473 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5474 // Anything in both C1 and C2 is known to be zero, remove it from
5476 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5477 NewRHS = ConstantExpr::getAnd(NewRHS,
5478 ConstantExpr::getNot(CommonBits));
5480 I.setOperand(0, Op0I->getOperand(0));
5481 I.setOperand(1, NewRHS);
5488 // Try to fold constant and into select arguments.
5489 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5490 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5492 if (isa<PHINode>(Op0))
5493 if (Instruction *NV = FoldOpIntoPhi(I))
5497 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5499 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5501 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5503 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5506 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5509 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5510 if (A == Op0) { // B^(B|A) == (A|B)^B
5511 Op1I->swapOperands();
5513 std::swap(Op0, Op1);
5514 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5515 I.swapOperands(); // Simplified below.
5516 std::swap(Op0, Op1);
5518 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5519 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5520 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5521 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5522 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5524 if (A == Op0) { // A^(A&B) -> A^(B&A)
5525 Op1I->swapOperands();
5528 if (B == Op0) { // A^(B&A) -> (B&A)^A
5529 I.swapOperands(); // Simplified below.
5530 std::swap(Op0, Op1);
5535 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5538 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5539 Op0I->hasOneUse()) {
5540 if (A == Op1) // (B|A)^B == (A|B)^B
5542 if (B == Op1) // (A|B)^B == A & ~B
5543 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
5544 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5545 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5546 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5547 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5548 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5550 if (A == Op1) // (A&B)^A -> (B&A)^A
5552 if (B == Op1 && // (B&A)^A == ~B & A
5553 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5554 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
5559 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5560 if (Op0I && Op1I && Op0I->isShift() &&
5561 Op0I->getOpcode() == Op1I->getOpcode() &&
5562 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5563 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5565 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
5567 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5568 Op1I->getOperand(1));
5572 Value *A, *B, *C, *D;
5573 // (A & B)^(A | B) -> A ^ B
5574 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5575 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5576 if ((A == C && B == D) || (A == D && B == C))
5577 return BinaryOperator::CreateXor(A, B);
5579 // (A | B)^(A & B) -> A ^ B
5580 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5581 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5582 if ((A == C && B == D) || (A == D && B == C))
5583 return BinaryOperator::CreateXor(A, B);
5587 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5588 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5589 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5590 // (X & Y)^(X & Y) -> (Y^Z) & X
5591 Value *X = 0, *Y = 0, *Z = 0;
5593 X = A, Y = B, Z = D;
5595 X = A, Y = B, Z = C;
5597 X = B, Y = A, Z = D;
5599 X = B, Y = A, Z = C;
5602 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
5603 return BinaryOperator::CreateAnd(NewOp, X);
5608 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5609 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5610 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5613 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5614 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5615 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5616 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5617 const Type *SrcTy = Op0C->getOperand(0)->getType();
5618 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5619 // Only do this if the casts both really cause code to be generated.
5620 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5622 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5624 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
5625 Op1C->getOperand(0), I.getName());
5626 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5631 return Changed ? &I : 0;
5634 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5635 LLVMContext *Context) {
5636 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5639 static bool HasAddOverflow(ConstantInt *Result,
5640 ConstantInt *In1, ConstantInt *In2,
5643 if (In2->getValue().isNegative())
5644 return Result->getValue().sgt(In1->getValue());
5646 return Result->getValue().slt(In1->getValue());
5648 return Result->getValue().ult(In1->getValue());
5651 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5652 /// overflowed for this type.
5653 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5654 Constant *In2, LLVMContext *Context,
5655 bool IsSigned = false) {
5656 Result = ConstantExpr::getAdd(In1, In2);
5658 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5659 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5660 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5661 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5662 ExtractElement(In1, Idx, Context),
5663 ExtractElement(In2, Idx, Context),
5670 return HasAddOverflow(cast<ConstantInt>(Result),
5671 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5675 static bool HasSubOverflow(ConstantInt *Result,
5676 ConstantInt *In1, ConstantInt *In2,
5679 if (In2->getValue().isNegative())
5680 return Result->getValue().slt(In1->getValue());
5682 return Result->getValue().sgt(In1->getValue());
5684 return Result->getValue().ugt(In1->getValue());
5687 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5688 /// overflowed for this type.
5689 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5690 Constant *In2, LLVMContext *Context,
5691 bool IsSigned = false) {
5692 Result = ConstantExpr::getSub(In1, In2);
5694 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5695 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5696 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5697 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5698 ExtractElement(In1, Idx, Context),
5699 ExtractElement(In2, Idx, Context),
5706 return HasSubOverflow(cast<ConstantInt>(Result),
5707 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5712 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5713 /// else. At this point we know that the GEP is on the LHS of the comparison.
5714 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5715 ICmpInst::Predicate Cond,
5717 // Look through bitcasts.
5718 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5719 RHS = BCI->getOperand(0);
5721 Value *PtrBase = GEPLHS->getOperand(0);
5722 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5723 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5724 // This transformation (ignoring the base and scales) is valid because we
5725 // know pointers can't overflow since the gep is inbounds. See if we can
5726 // output an optimized form.
5727 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5729 // If not, synthesize the offset the hard way.
5731 Offset = EmitGEPOffset(GEPLHS, *this);
5732 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5733 Constant::getNullValue(Offset->getType()));
5734 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5735 // If the base pointers are different, but the indices are the same, just
5736 // compare the base pointer.
5737 if (PtrBase != GEPRHS->getOperand(0)) {
5738 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5739 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5740 GEPRHS->getOperand(0)->getType();
5742 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5743 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5744 IndicesTheSame = false;
5748 // If all indices are the same, just compare the base pointers.
5750 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5751 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5753 // Otherwise, the base pointers are different and the indices are
5754 // different, bail out.
5758 // If one of the GEPs has all zero indices, recurse.
5759 bool AllZeros = true;
5760 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5761 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5762 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5767 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5768 ICmpInst::getSwappedPredicate(Cond), I);
5770 // If the other GEP has all zero indices, recurse.
5772 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5773 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5774 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5779 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5781 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5782 // If the GEPs only differ by one index, compare it.
5783 unsigned NumDifferences = 0; // Keep track of # differences.
5784 unsigned DiffOperand = 0; // The operand that differs.
5785 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5786 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5787 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5788 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5789 // Irreconcilable differences.
5793 if (NumDifferences++) break;
5798 if (NumDifferences == 0) // SAME GEP?
5799 return ReplaceInstUsesWith(I, // No comparison is needed here.
5800 ConstantInt::get(Type::getInt1Ty(*Context),
5801 ICmpInst::isTrueWhenEqual(Cond)));
5803 else if (NumDifferences == 1) {
5804 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5805 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5806 // Make sure we do a signed comparison here.
5807 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5811 // Only lower this if the icmp is the only user of the GEP or if we expect
5812 // the result to fold to a constant!
5814 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5815 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5816 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5817 Value *L = EmitGEPOffset(GEPLHS, *this);
5818 Value *R = EmitGEPOffset(GEPRHS, *this);
5819 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5825 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5827 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5830 if (!isa<ConstantFP>(RHSC)) return 0;
5831 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5833 // Get the width of the mantissa. We don't want to hack on conversions that
5834 // might lose information from the integer, e.g. "i64 -> float"
5835 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5836 if (MantissaWidth == -1) return 0; // Unknown.
5838 // Check to see that the input is converted from an integer type that is small
5839 // enough that preserves all bits. TODO: check here for "known" sign bits.
5840 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5841 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5843 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5844 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5848 // If the conversion would lose info, don't hack on this.
5849 if ((int)InputSize > MantissaWidth)
5852 // Otherwise, we can potentially simplify the comparison. We know that it
5853 // will always come through as an integer value and we know the constant is
5854 // not a NAN (it would have been previously simplified).
5855 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5857 ICmpInst::Predicate Pred;
5858 switch (I.getPredicate()) {
5859 default: llvm_unreachable("Unexpected predicate!");
5860 case FCmpInst::FCMP_UEQ:
5861 case FCmpInst::FCMP_OEQ:
5862 Pred = ICmpInst::ICMP_EQ;
5864 case FCmpInst::FCMP_UGT:
5865 case FCmpInst::FCMP_OGT:
5866 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5868 case FCmpInst::FCMP_UGE:
5869 case FCmpInst::FCMP_OGE:
5870 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5872 case FCmpInst::FCMP_ULT:
5873 case FCmpInst::FCMP_OLT:
5874 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5876 case FCmpInst::FCMP_ULE:
5877 case FCmpInst::FCMP_OLE:
5878 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5880 case FCmpInst::FCMP_UNE:
5881 case FCmpInst::FCMP_ONE:
5882 Pred = ICmpInst::ICMP_NE;
5884 case FCmpInst::FCMP_ORD:
5885 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5886 case FCmpInst::FCMP_UNO:
5887 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5890 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5892 // Now we know that the APFloat is a normal number, zero or inf.
5894 // See if the FP constant is too large for the integer. For example,
5895 // comparing an i8 to 300.0.
5896 unsigned IntWidth = IntTy->getScalarSizeInBits();
5899 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5900 // and large values.
5901 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5902 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5903 APFloat::rmNearestTiesToEven);
5904 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5905 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5906 Pred == ICmpInst::ICMP_SLE)
5907 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5908 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5911 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5912 // +INF and large values.
5913 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5914 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5915 APFloat::rmNearestTiesToEven);
5916 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5917 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5918 Pred == ICmpInst::ICMP_ULE)
5919 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5920 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5925 // See if the RHS value is < SignedMin.
5926 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5927 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5928 APFloat::rmNearestTiesToEven);
5929 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5930 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5931 Pred == ICmpInst::ICMP_SGE)
5932 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5933 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5937 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5938 // [0, UMAX], but it may still be fractional. See if it is fractional by
5939 // casting the FP value to the integer value and back, checking for equality.
5940 // Don't do this for zero, because -0.0 is not fractional.
5941 Constant *RHSInt = LHSUnsigned
5942 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5943 : ConstantExpr::getFPToSI(RHSC, IntTy);
5944 if (!RHS.isZero()) {
5945 bool Equal = LHSUnsigned
5946 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5947 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5949 // If we had a comparison against a fractional value, we have to adjust
5950 // the compare predicate and sometimes the value. RHSC is rounded towards
5951 // zero at this point.
5953 default: llvm_unreachable("Unexpected integer comparison!");
5954 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5955 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5956 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5957 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5958 case ICmpInst::ICMP_ULE:
5959 // (float)int <= 4.4 --> int <= 4
5960 // (float)int <= -4.4 --> false
5961 if (RHS.isNegative())
5962 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5964 case ICmpInst::ICMP_SLE:
5965 // (float)int <= 4.4 --> int <= 4
5966 // (float)int <= -4.4 --> int < -4
5967 if (RHS.isNegative())
5968 Pred = ICmpInst::ICMP_SLT;
5970 case ICmpInst::ICMP_ULT:
5971 // (float)int < -4.4 --> false
5972 // (float)int < 4.4 --> int <= 4
5973 if (RHS.isNegative())
5974 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5975 Pred = ICmpInst::ICMP_ULE;
5977 case ICmpInst::ICMP_SLT:
5978 // (float)int < -4.4 --> int < -4
5979 // (float)int < 4.4 --> int <= 4
5980 if (!RHS.isNegative())
5981 Pred = ICmpInst::ICMP_SLE;
5983 case ICmpInst::ICMP_UGT:
5984 // (float)int > 4.4 --> int > 4
5985 // (float)int > -4.4 --> true
5986 if (RHS.isNegative())
5987 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5989 case ICmpInst::ICMP_SGT:
5990 // (float)int > 4.4 --> int > 4
5991 // (float)int > -4.4 --> int >= -4
5992 if (RHS.isNegative())
5993 Pred = ICmpInst::ICMP_SGE;
5995 case ICmpInst::ICMP_UGE:
5996 // (float)int >= -4.4 --> true
5997 // (float)int >= 4.4 --> int > 4
5998 if (!RHS.isNegative())
5999 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6000 Pred = ICmpInst::ICMP_UGT;
6002 case ICmpInst::ICMP_SGE:
6003 // (float)int >= -4.4 --> int >= -4
6004 // (float)int >= 4.4 --> int > 4
6005 if (!RHS.isNegative())
6006 Pred = ICmpInst::ICMP_SGT;
6012 // Lower this FP comparison into an appropriate integer version of the
6014 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
6017 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
6018 bool Changed = false;
6020 /// Orders the operands of the compare so that they are listed from most
6021 /// complex to least complex. This puts constants before unary operators,
6022 /// before binary operators.
6023 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
6028 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6030 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
6031 return ReplaceInstUsesWith(I, V);
6033 // Simplify 'fcmp pred X, X'
6035 switch (I.getPredicate()) {
6036 default: llvm_unreachable("Unknown predicate!");
6037 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
6038 case FCmpInst::FCMP_ULT: // True if unordered or less than
6039 case FCmpInst::FCMP_UGT: // True if unordered or greater than
6040 case FCmpInst::FCMP_UNE: // True if unordered or not equal
6041 // Canonicalize these to be 'fcmp uno %X, 0.0'.
6042 I.setPredicate(FCmpInst::FCMP_UNO);
6043 I.setOperand(1, Constant::getNullValue(Op0->getType()));
6046 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
6047 case FCmpInst::FCMP_OEQ: // True if ordered and equal
6048 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
6049 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
6050 // Canonicalize these to be 'fcmp ord %X, 0.0'.
6051 I.setPredicate(FCmpInst::FCMP_ORD);
6052 I.setOperand(1, Constant::getNullValue(Op0->getType()));
6057 // Handle fcmp with constant RHS
6058 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6059 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6060 switch (LHSI->getOpcode()) {
6061 case Instruction::PHI:
6062 // Only fold fcmp into the PHI if the phi and fcmp are in the same
6063 // block. If in the same block, we're encouraging jump threading. If
6064 // not, we are just pessimizing the code by making an i1 phi.
6065 if (LHSI->getParent() == I.getParent())
6066 if (Instruction *NV = FoldOpIntoPhi(I, true))
6069 case Instruction::SIToFP:
6070 case Instruction::UIToFP:
6071 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
6074 case Instruction::Select:
6075 // If either operand of the select is a constant, we can fold the
6076 // comparison into the select arms, which will cause one to be
6077 // constant folded and the select turned into a bitwise or.
6078 Value *Op1 = 0, *Op2 = 0;
6079 if (LHSI->hasOneUse()) {
6080 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6081 // Fold the known value into the constant operand.
6082 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
6083 // Insert a new FCmp of the other select operand.
6084 Op2 = Builder->CreateFCmp(I.getPredicate(),
6085 LHSI->getOperand(2), RHSC, I.getName());
6086 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6087 // Fold the known value into the constant operand.
6088 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
6089 // Insert a new FCmp of the other select operand.
6090 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
6096 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6101 return Changed ? &I : 0;
6104 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
6105 bool Changed = false;
6107 /// Orders the operands of the compare so that they are listed from most
6108 /// complex to least complex. This puts constants before unary operators,
6109 /// before binary operators.
6110 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
6115 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6117 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
6118 return ReplaceInstUsesWith(I, V);
6120 const Type *Ty = Op0->getType();
6122 // icmp's with boolean values can always be turned into bitwise operations
6123 if (Ty == Type::getInt1Ty(*Context)) {
6124 switch (I.getPredicate()) {
6125 default: llvm_unreachable("Invalid icmp instruction!");
6126 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6127 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
6128 return BinaryOperator::CreateNot(Xor);
6130 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6131 return BinaryOperator::CreateXor(Op0, Op1);
6133 case ICmpInst::ICMP_UGT:
6134 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6136 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6137 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6138 return BinaryOperator::CreateAnd(Not, Op1);
6140 case ICmpInst::ICMP_SGT:
6141 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6143 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6144 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6145 return BinaryOperator::CreateAnd(Not, Op0);
6147 case ICmpInst::ICMP_UGE:
6148 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6150 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6151 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6152 return BinaryOperator::CreateOr(Not, Op1);
6154 case ICmpInst::ICMP_SGE:
6155 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6157 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6158 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6159 return BinaryOperator::CreateOr(Not, Op0);
6164 unsigned BitWidth = 0;
6166 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6167 else if (Ty->isIntOrIntVector())
6168 BitWidth = Ty->getScalarSizeInBits();
6170 bool isSignBit = false;
6172 // See if we are doing a comparison with a constant.
6173 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6174 Value *A = 0, *B = 0;
6176 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6177 if (I.isEquality() && CI->isNullValue() &&
6178 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6179 // (icmp cond A B) if cond is equality
6180 return new ICmpInst(I.getPredicate(), A, B);
6183 // If we have an icmp le or icmp ge instruction, turn it into the
6184 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6185 // them being folded in the code below. The SimplifyICmpInst code has
6186 // already handled the edge cases for us, so we just assert on them.
6187 switch (I.getPredicate()) {
6189 case ICmpInst::ICMP_ULE:
6190 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
6191 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
6193 case ICmpInst::ICMP_SLE:
6194 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
6195 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6197 case ICmpInst::ICMP_UGE:
6198 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
6199 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
6201 case ICmpInst::ICMP_SGE:
6202 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
6203 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6207 // If this comparison is a normal comparison, it demands all
6208 // bits, if it is a sign bit comparison, it only demands the sign bit.
6210 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6213 // See if we can fold the comparison based on range information we can get
6214 // by checking whether bits are known to be zero or one in the input.
6215 if (BitWidth != 0) {
6216 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6217 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6219 if (SimplifyDemandedBits(I.getOperandUse(0),
6220 isSignBit ? APInt::getSignBit(BitWidth)
6221 : APInt::getAllOnesValue(BitWidth),
6222 Op0KnownZero, Op0KnownOne, 0))
6224 if (SimplifyDemandedBits(I.getOperandUse(1),
6225 APInt::getAllOnesValue(BitWidth),
6226 Op1KnownZero, Op1KnownOne, 0))
6229 // Given the known and unknown bits, compute a range that the LHS could be
6230 // in. Compute the Min, Max and RHS values based on the known bits. For the
6231 // EQ and NE we use unsigned values.
6232 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6233 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6235 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6237 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6240 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6242 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6246 // If Min and Max are known to be the same, then SimplifyDemandedBits
6247 // figured out that the LHS is a constant. Just constant fold this now so
6248 // that code below can assume that Min != Max.
6249 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6250 return new ICmpInst(I.getPredicate(),
6251 ConstantInt::get(*Context, Op0Min), Op1);
6252 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6253 return new ICmpInst(I.getPredicate(), Op0,
6254 ConstantInt::get(*Context, Op1Min));
6256 // Based on the range information we know about the LHS, see if we can
6257 // simplify this comparison. For example, (x&4) < 8 is always true.
6258 switch (I.getPredicate()) {
6259 default: llvm_unreachable("Unknown icmp opcode!");
6260 case ICmpInst::ICMP_EQ:
6261 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6262 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6264 case ICmpInst::ICMP_NE:
6265 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6266 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6268 case ICmpInst::ICMP_ULT:
6269 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6270 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6271 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6272 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6273 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6274 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6275 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6276 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6277 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6280 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6281 if (CI->isMinValue(true))
6282 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6283 Constant::getAllOnesValue(Op0->getType()));
6286 case ICmpInst::ICMP_UGT:
6287 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6288 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6289 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6290 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6292 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6293 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6294 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6295 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6296 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6299 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6300 if (CI->isMaxValue(true))
6301 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6302 Constant::getNullValue(Op0->getType()));
6305 case ICmpInst::ICMP_SLT:
6306 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6307 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6308 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6309 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6310 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6311 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6312 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6313 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6314 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6318 case ICmpInst::ICMP_SGT:
6319 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6320 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6321 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6322 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6324 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6325 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6326 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6327 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6328 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6332 case ICmpInst::ICMP_SGE:
6333 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6334 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6335 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6336 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6337 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6339 case ICmpInst::ICMP_SLE:
6340 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6341 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6342 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6343 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6344 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6346 case ICmpInst::ICMP_UGE:
6347 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6348 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6349 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6350 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6351 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6353 case ICmpInst::ICMP_ULE:
6354 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6355 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6356 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6357 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6358 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6362 // Turn a signed comparison into an unsigned one if both operands
6363 // are known to have the same sign.
6365 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6366 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6367 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6370 // Test if the ICmpInst instruction is used exclusively by a select as
6371 // part of a minimum or maximum operation. If so, refrain from doing
6372 // any other folding. This helps out other analyses which understand
6373 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6374 // and CodeGen. And in this case, at least one of the comparison
6375 // operands has at least one user besides the compare (the select),
6376 // which would often largely negate the benefit of folding anyway.
6378 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6379 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6380 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6383 // See if we are doing a comparison between a constant and an instruction that
6384 // can be folded into the comparison.
6385 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6386 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6387 // instruction, see if that instruction also has constants so that the
6388 // instruction can be folded into the icmp
6389 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6390 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6394 // Handle icmp with constant (but not simple integer constant) RHS
6395 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6396 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6397 switch (LHSI->getOpcode()) {
6398 case Instruction::GetElementPtr:
6399 if (RHSC->isNullValue()) {
6400 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6401 bool isAllZeros = true;
6402 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6403 if (!isa<Constant>(LHSI->getOperand(i)) ||
6404 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6409 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6410 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6414 case Instruction::PHI:
6415 // Only fold icmp into the PHI if the phi and icmp are in the same
6416 // block. If in the same block, we're encouraging jump threading. If
6417 // not, we are just pessimizing the code by making an i1 phi.
6418 if (LHSI->getParent() == I.getParent())
6419 if (Instruction *NV = FoldOpIntoPhi(I, true))
6422 case Instruction::Select: {
6423 // If either operand of the select is a constant, we can fold the
6424 // comparison into the select arms, which will cause one to be
6425 // constant folded and the select turned into a bitwise or.
6426 Value *Op1 = 0, *Op2 = 0;
6427 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
6428 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6429 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
6430 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6432 // We only want to perform this transformation if it will not lead to
6433 // additional code. This is true if either both sides of the select
6434 // fold to a constant (in which case the icmp is replaced with a select
6435 // which will usually simplify) or this is the only user of the
6436 // select (in which case we are trading a select+icmp for a simpler
6438 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
6440 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
6443 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
6445 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6449 case Instruction::Call:
6450 // If we have (malloc != null), and if the malloc has a single use, we
6451 // can assume it is successful and remove the malloc.
6452 if (isMalloc(LHSI) && LHSI->hasOneUse() &&
6453 isa<ConstantPointerNull>(RHSC)) {
6454 // Need to explicitly erase malloc call here, instead of adding it to
6455 // Worklist, because it won't get DCE'd from the Worklist since
6456 // isInstructionTriviallyDead() returns false for function calls.
6457 // It is OK to replace LHSI/MallocCall with Undef because the
6458 // instruction that uses it will be erased via Worklist.
6459 if (extractMallocCall(LHSI)) {
6460 LHSI->replaceAllUsesWith(UndefValue::get(LHSI->getType()));
6461 EraseInstFromFunction(*LHSI);
6462 return ReplaceInstUsesWith(I,
6463 ConstantInt::get(Type::getInt1Ty(*Context),
6464 !I.isTrueWhenEqual()));
6466 if (CallInst* MallocCall = extractMallocCallFromBitCast(LHSI))
6467 if (MallocCall->hasOneUse()) {
6468 MallocCall->replaceAllUsesWith(
6469 UndefValue::get(MallocCall->getType()));
6470 EraseInstFromFunction(*MallocCall);
6471 Worklist.Add(LHSI); // The malloc's bitcast use.
6472 return ReplaceInstUsesWith(I,
6473 ConstantInt::get(Type::getInt1Ty(*Context),
6474 !I.isTrueWhenEqual()));
6481 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6482 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6483 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6485 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6486 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6487 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6490 // Test to see if the operands of the icmp are casted versions of other
6491 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6493 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6494 if (isa<PointerType>(Op0->getType()) &&
6495 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6496 // We keep moving the cast from the left operand over to the right
6497 // operand, where it can often be eliminated completely.
6498 Op0 = CI->getOperand(0);
6500 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6501 // so eliminate it as well.
6502 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6503 Op1 = CI2->getOperand(0);
6505 // If Op1 is a constant, we can fold the cast into the constant.
6506 if (Op0->getType() != Op1->getType()) {
6507 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6508 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6510 // Otherwise, cast the RHS right before the icmp
6511 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
6514 return new ICmpInst(I.getPredicate(), Op0, Op1);
6518 if (isa<CastInst>(Op0)) {
6519 // Handle the special case of: icmp (cast bool to X), <cst>
6520 // This comes up when you have code like
6523 // For generality, we handle any zero-extension of any operand comparison
6524 // with a constant or another cast from the same type.
6525 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
6526 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6530 // See if it's the same type of instruction on the left and right.
6531 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6532 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6533 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6534 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6535 switch (Op0I->getOpcode()) {
6537 case Instruction::Add:
6538 case Instruction::Sub:
6539 case Instruction::Xor:
6540 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6541 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6542 Op1I->getOperand(0));
6543 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6544 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6545 if (CI->getValue().isSignBit()) {
6546 ICmpInst::Predicate Pred = I.isSigned()
6547 ? I.getUnsignedPredicate()
6548 : I.getSignedPredicate();
6549 return new ICmpInst(Pred, Op0I->getOperand(0),
6550 Op1I->getOperand(0));
6553 if (CI->getValue().isMaxSignedValue()) {
6554 ICmpInst::Predicate Pred = I.isSigned()
6555 ? I.getUnsignedPredicate()
6556 : I.getSignedPredicate();
6557 Pred = I.getSwappedPredicate(Pred);
6558 return new ICmpInst(Pred, Op0I->getOperand(0),
6559 Op1I->getOperand(0));
6563 case Instruction::Mul:
6564 if (!I.isEquality())
6567 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6568 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6569 // Mask = -1 >> count-trailing-zeros(Cst).
6570 if (!CI->isZero() && !CI->isOne()) {
6571 const APInt &AP = CI->getValue();
6572 ConstantInt *Mask = ConstantInt::get(*Context,
6573 APInt::getLowBitsSet(AP.getBitWidth(),
6575 AP.countTrailingZeros()));
6576 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
6577 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
6578 return new ICmpInst(I.getPredicate(), And1, And2);
6587 // ~x < ~y --> y < x
6589 if (match(Op0, m_Not(m_Value(A))) &&
6590 match(Op1, m_Not(m_Value(B))))
6591 return new ICmpInst(I.getPredicate(), B, A);
6594 if (I.isEquality()) {
6595 Value *A, *B, *C, *D;
6597 // -x == -y --> x == y
6598 if (match(Op0, m_Neg(m_Value(A))) &&
6599 match(Op1, m_Neg(m_Value(B))))
6600 return new ICmpInst(I.getPredicate(), A, B);
6602 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6603 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6604 Value *OtherVal = A == Op1 ? B : A;
6605 return new ICmpInst(I.getPredicate(), OtherVal,
6606 Constant::getNullValue(A->getType()));
6609 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6610 // A^c1 == C^c2 --> A == C^(c1^c2)
6611 ConstantInt *C1, *C2;
6612 if (match(B, m_ConstantInt(C1)) &&
6613 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6615 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6616 Value *Xor = Builder->CreateXor(C, NC, "tmp");
6617 return new ICmpInst(I.getPredicate(), A, Xor);
6620 // A^B == A^D -> B == D
6621 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6622 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6623 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6624 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6628 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6629 (A == Op0 || B == Op0)) {
6630 // A == (A^B) -> B == 0
6631 Value *OtherVal = A == Op0 ? B : A;
6632 return new ICmpInst(I.getPredicate(), OtherVal,
6633 Constant::getNullValue(A->getType()));
6636 // (A-B) == A -> B == 0
6637 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6638 return new ICmpInst(I.getPredicate(), B,
6639 Constant::getNullValue(B->getType()));
6641 // A == (A-B) -> B == 0
6642 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6643 return new ICmpInst(I.getPredicate(), B,
6644 Constant::getNullValue(B->getType()));
6646 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6647 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6648 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6649 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6650 Value *X = 0, *Y = 0, *Z = 0;
6653 X = B; Y = D; Z = A;
6654 } else if (A == D) {
6655 X = B; Y = C; Z = A;
6656 } else if (B == C) {
6657 X = A; Y = D; Z = B;
6658 } else if (B == D) {
6659 X = A; Y = C; Z = B;
6662 if (X) { // Build (X^Y) & Z
6663 Op1 = Builder->CreateXor(X, Y, "tmp");
6664 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
6665 I.setOperand(0, Op1);
6666 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6673 Value *X; ConstantInt *Cst;
6675 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
6676 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
6679 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
6680 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
6682 return Changed ? &I : 0;
6685 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
6686 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
6687 Value *X, ConstantInt *CI,
6688 ICmpInst::Predicate Pred,
6690 // If we have X+0, exit early (simplifying logic below) and let it get folded
6691 // elsewhere. icmp X+0, X -> icmp X, X
6693 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
6694 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
6697 // (X+4) == X -> false.
6698 if (Pred == ICmpInst::ICMP_EQ)
6699 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
6701 // (X+4) != X -> true.
6702 if (Pred == ICmpInst::ICMP_NE)
6703 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
6705 // If this is an instruction (as opposed to constantexpr) get NUW/NSW info.
6706 bool isNUW = false, isNSW = false;
6707 if (BinaryOperator *Add = dyn_cast<BinaryOperator>(TheAdd)) {
6708 isNUW = Add->hasNoUnsignedWrap();
6709 isNSW = Add->hasNoSignedWrap();
6712 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
6713 // so the values can never be equal. Similiarly for all other "or equals"
6716 // (X+1) <u X --> X >u (MAXUINT-1) --> X != 255
6717 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
6718 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
6719 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
6720 // If this is an NUW add, then this is always false.
6722 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
6724 Value *R = ConstantExpr::getSub(ConstantInt::get(CI->getType(), -1ULL), CI);
6725 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
6728 // (X+1) >u X --> X <u (0-1) --> X != 255
6729 // (X+2) >u X --> X <u (0-2) --> X <u 254
6730 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
6731 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
6732 // If this is an NUW add, then this is always true.
6734 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
6735 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
6738 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
6739 ConstantInt *SMax = ConstantInt::get(X->getContext(),
6740 APInt::getSignedMaxValue(BitWidth));
6742 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
6743 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
6744 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
6745 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
6746 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
6747 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
6748 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
6749 // If this is an NSW add, then we have two cases: if the constant is
6750 // positive, then this is always false, if negative, this is always true.
6752 bool isTrue = CI->getValue().isNegative();
6753 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
6756 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
6759 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
6760 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
6761 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
6762 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
6763 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
6764 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
6766 // If this is an NSW add, then we have two cases: if the constant is
6767 // positive, then this is always true, if negative, this is always false.
6769 bool isTrue = !CI->getValue().isNegative();
6770 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
6773 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
6774 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
6775 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
6778 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6779 /// and CmpRHS are both known to be integer constants.
6780 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6781 ConstantInt *DivRHS) {
6782 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6783 const APInt &CmpRHSV = CmpRHS->getValue();
6785 // FIXME: If the operand types don't match the type of the divide
6786 // then don't attempt this transform. The code below doesn't have the
6787 // logic to deal with a signed divide and an unsigned compare (and
6788 // vice versa). This is because (x /s C1) <s C2 produces different
6789 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6790 // (x /u C1) <u C2. Simply casting the operands and result won't
6791 // work. :( The if statement below tests that condition and bails
6793 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6794 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
6796 if (DivRHS->isZero())
6797 return 0; // The ProdOV computation fails on divide by zero.
6798 if (DivIsSigned && DivRHS->isAllOnesValue())
6799 return 0; // The overflow computation also screws up here
6800 if (DivRHS->isOne())
6801 return 0; // Not worth bothering, and eliminates some funny cases
6804 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6805 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6806 // C2 (CI). By solving for X we can turn this into a range check
6807 // instead of computing a divide.
6808 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6810 // Determine if the product overflows by seeing if the product is
6811 // not equal to the divide. Make sure we do the same kind of divide
6812 // as in the LHS instruction that we're folding.
6813 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6814 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6816 // Get the ICmp opcode
6817 ICmpInst::Predicate Pred = ICI.getPredicate();
6819 // Figure out the interval that is being checked. For example, a comparison
6820 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6821 // Compute this interval based on the constants involved and the signedness of
6822 // the compare/divide. This computes a half-open interval, keeping track of
6823 // whether either value in the interval overflows. After analysis each
6824 // overflow variable is set to 0 if it's corresponding bound variable is valid
6825 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6826 int LoOverflow = 0, HiOverflow = 0;
6827 Constant *LoBound = 0, *HiBound = 0;
6829 if (!DivIsSigned) { // udiv
6830 // e.g. X/5 op 3 --> [15, 20)
6832 HiOverflow = LoOverflow = ProdOV;
6834 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6835 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6836 if (CmpRHSV == 0) { // (X / pos) op 0
6837 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6838 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6840 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6841 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6842 HiOverflow = LoOverflow = ProdOV;
6844 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6845 } else { // (X / pos) op neg
6846 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6847 HiBound = AddOne(Prod);
6848 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6850 ConstantInt* DivNeg =
6851 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6852 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6856 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6857 if (CmpRHSV == 0) { // (X / neg) op 0
6858 // e.g. X/-5 op 0 --> [-4, 5)
6859 LoBound = AddOne(DivRHS);
6860 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6861 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6862 HiOverflow = 1; // [INTMIN+1, overflow)
6863 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6865 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6866 // e.g. X/-5 op 3 --> [-19, -14)
6867 HiBound = AddOne(Prod);
6868 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6870 LoOverflow = AddWithOverflow(LoBound, HiBound,
6871 DivRHS, Context, true) ? -1 : 0;
6872 } else { // (X / neg) op neg
6873 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6874 LoOverflow = HiOverflow = ProdOV;
6876 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6879 // Dividing by a negative swaps the condition. LT <-> GT
6880 Pred = ICmpInst::getSwappedPredicate(Pred);
6883 Value *X = DivI->getOperand(0);
6885 default: llvm_unreachable("Unhandled icmp opcode!");
6886 case ICmpInst::ICMP_EQ:
6887 if (LoOverflow && HiOverflow)
6888 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6889 else if (HiOverflow)
6890 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6891 ICmpInst::ICMP_UGE, X, LoBound);
6892 else if (LoOverflow)
6893 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6894 ICmpInst::ICMP_ULT, X, HiBound);
6896 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6897 case ICmpInst::ICMP_NE:
6898 if (LoOverflow && HiOverflow)
6899 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6900 else if (HiOverflow)
6901 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6902 ICmpInst::ICMP_ULT, X, LoBound);
6903 else if (LoOverflow)
6904 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6905 ICmpInst::ICMP_UGE, X, HiBound);
6907 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6908 case ICmpInst::ICMP_ULT:
6909 case ICmpInst::ICMP_SLT:
6910 if (LoOverflow == +1) // Low bound is greater than input range.
6911 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6912 if (LoOverflow == -1) // Low bound is less than input range.
6913 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6914 return new ICmpInst(Pred, X, LoBound);
6915 case ICmpInst::ICMP_UGT:
6916 case ICmpInst::ICMP_SGT:
6917 if (HiOverflow == +1) // High bound greater than input range.
6918 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6919 else if (HiOverflow == -1) // High bound less than input range.
6920 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6921 if (Pred == ICmpInst::ICMP_UGT)
6922 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6924 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6929 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6931 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6934 const APInt &RHSV = RHS->getValue();
6936 switch (LHSI->getOpcode()) {
6937 case Instruction::Trunc:
6938 if (ICI.isEquality() && LHSI->hasOneUse()) {
6939 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6940 // of the high bits truncated out of x are known.
6941 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6942 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6943 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6944 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6945 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6947 // If all the high bits are known, we can do this xform.
6948 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6949 // Pull in the high bits from known-ones set.
6950 APInt NewRHS(RHS->getValue());
6951 NewRHS.zext(SrcBits);
6953 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6954 ConstantInt::get(*Context, NewRHS));
6959 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6960 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6961 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6963 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6964 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6965 Value *CompareVal = LHSI->getOperand(0);
6967 // If the sign bit of the XorCST is not set, there is no change to
6968 // the operation, just stop using the Xor.
6969 if (!XorCST->getValue().isNegative()) {
6970 ICI.setOperand(0, CompareVal);
6975 // Was the old condition true if the operand is positive?
6976 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6978 // If so, the new one isn't.
6979 isTrueIfPositive ^= true;
6981 if (isTrueIfPositive)
6982 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
6985 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
6989 if (LHSI->hasOneUse()) {
6990 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6991 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6992 const APInt &SignBit = XorCST->getValue();
6993 ICmpInst::Predicate Pred = ICI.isSigned()
6994 ? ICI.getUnsignedPredicate()
6995 : ICI.getSignedPredicate();
6996 return new ICmpInst(Pred, LHSI->getOperand(0),
6997 ConstantInt::get(*Context, RHSV ^ SignBit));
7000 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
7001 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
7002 const APInt &NotSignBit = XorCST->getValue();
7003 ICmpInst::Predicate Pred = ICI.isSigned()
7004 ? ICI.getUnsignedPredicate()
7005 : ICI.getSignedPredicate();
7006 Pred = ICI.getSwappedPredicate(Pred);
7007 return new ICmpInst(Pred, LHSI->getOperand(0),
7008 ConstantInt::get(*Context, RHSV ^ NotSignBit));
7013 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
7014 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
7015 LHSI->getOperand(0)->hasOneUse()) {
7016 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
7018 // If the LHS is an AND of a truncating cast, we can widen the
7019 // and/compare to be the input width without changing the value
7020 // produced, eliminating a cast.
7021 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
7022 // We can do this transformation if either the AND constant does not
7023 // have its sign bit set or if it is an equality comparison.
7024 // Extending a relational comparison when we're checking the sign
7025 // bit would not work.
7026 if (Cast->hasOneUse() &&
7027 (ICI.isEquality() ||
7028 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
7030 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
7031 APInt NewCST = AndCST->getValue();
7032 NewCST.zext(BitWidth);
7034 NewCI.zext(BitWidth);
7036 Builder->CreateAnd(Cast->getOperand(0),
7037 ConstantInt::get(*Context, NewCST), LHSI->getName());
7038 return new ICmpInst(ICI.getPredicate(), NewAnd,
7039 ConstantInt::get(*Context, NewCI));
7043 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
7044 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
7045 // happens a LOT in code produced by the C front-end, for bitfield
7047 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
7048 if (Shift && !Shift->isShift())
7052 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
7053 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
7054 const Type *AndTy = AndCST->getType(); // Type of the and.
7056 // We can fold this as long as we can't shift unknown bits
7057 // into the mask. This can only happen with signed shift
7058 // rights, as they sign-extend.
7060 bool CanFold = Shift->isLogicalShift();
7062 // To test for the bad case of the signed shr, see if any
7063 // of the bits shifted in could be tested after the mask.
7064 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
7065 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
7067 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
7068 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
7069 AndCST->getValue()) == 0)
7075 if (Shift->getOpcode() == Instruction::Shl)
7076 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
7078 NewCst = ConstantExpr::getShl(RHS, ShAmt);
7080 // Check to see if we are shifting out any of the bits being
7082 if (ConstantExpr::get(Shift->getOpcode(),
7083 NewCst, ShAmt) != RHS) {
7084 // If we shifted bits out, the fold is not going to work out.
7085 // As a special case, check to see if this means that the
7086 // result is always true or false now.
7087 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7088 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7089 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7090 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7092 ICI.setOperand(1, NewCst);
7093 Constant *NewAndCST;
7094 if (Shift->getOpcode() == Instruction::Shl)
7095 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
7097 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
7098 LHSI->setOperand(1, NewAndCST);
7099 LHSI->setOperand(0, Shift->getOperand(0));
7100 Worklist.Add(Shift); // Shift is dead.
7106 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
7107 // preferable because it allows the C<<Y expression to be hoisted out
7108 // of a loop if Y is invariant and X is not.
7109 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
7110 ICI.isEquality() && !Shift->isArithmeticShift() &&
7111 !isa<Constant>(Shift->getOperand(0))) {
7114 if (Shift->getOpcode() == Instruction::LShr) {
7115 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
7117 // Insert a logical shift.
7118 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
7121 // Compute X & (C << Y).
7123 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
7125 ICI.setOperand(0, NewAnd);
7131 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
7132 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7135 uint32_t TypeBits = RHSV.getBitWidth();
7137 // Check that the shift amount is in range. If not, don't perform
7138 // undefined shifts. When the shift is visited it will be
7140 if (ShAmt->uge(TypeBits))
7143 if (ICI.isEquality()) {
7144 // If we are comparing against bits always shifted out, the
7145 // comparison cannot succeed.
7147 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
7149 if (Comp != RHS) {// Comparing against a bit that we know is zero.
7150 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7151 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
7152 return ReplaceInstUsesWith(ICI, Cst);
7155 if (LHSI->hasOneUse()) {
7156 // Otherwise strength reduce the shift into an and.
7157 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
7159 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
7160 TypeBits-ShAmtVal));
7163 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
7164 return new ICmpInst(ICI.getPredicate(), And,
7165 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
7169 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
7170 bool TrueIfSigned = false;
7171 if (LHSI->hasOneUse() &&
7172 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
7173 // (X << 31) <s 0 --> (X&1) != 0
7174 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
7175 (TypeBits-ShAmt->getZExtValue()-1));
7177 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
7178 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
7179 And, Constant::getNullValue(And->getType()));
7184 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
7185 case Instruction::AShr: {
7186 // Only handle equality comparisons of shift-by-constant.
7187 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7188 if (!ShAmt || !ICI.isEquality()) break;
7190 // Check that the shift amount is in range. If not, don't perform
7191 // undefined shifts. When the shift is visited it will be
7193 uint32_t TypeBits = RHSV.getBitWidth();
7194 if (ShAmt->uge(TypeBits))
7197 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
7199 // If we are comparing against bits always shifted out, the
7200 // comparison cannot succeed.
7201 APInt Comp = RHSV << ShAmtVal;
7202 if (LHSI->getOpcode() == Instruction::LShr)
7203 Comp = Comp.lshr(ShAmtVal);
7205 Comp = Comp.ashr(ShAmtVal);
7207 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
7208 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7209 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
7210 return ReplaceInstUsesWith(ICI, Cst);
7213 // Otherwise, check to see if the bits shifted out are known to be zero.
7214 // If so, we can compare against the unshifted value:
7215 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
7216 if (LHSI->hasOneUse() &&
7217 MaskedValueIsZero(LHSI->getOperand(0),
7218 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
7219 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
7220 ConstantExpr::getShl(RHS, ShAmt));
7223 if (LHSI->hasOneUse()) {
7224 // Otherwise strength reduce the shift into an and.
7225 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7226 Constant *Mask = ConstantInt::get(*Context, Val);
7228 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
7229 Mask, LHSI->getName()+".mask");
7230 return new ICmpInst(ICI.getPredicate(), And,
7231 ConstantExpr::getShl(RHS, ShAmt));
7236 case Instruction::SDiv:
7237 case Instruction::UDiv:
7238 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7239 // Fold this div into the comparison, producing a range check.
7240 // Determine, based on the divide type, what the range is being
7241 // checked. If there is an overflow on the low or high side, remember
7242 // it, otherwise compute the range [low, hi) bounding the new value.
7243 // See: InsertRangeTest above for the kinds of replacements possible.
7244 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7245 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7250 case Instruction::Add:
7251 // Fold: icmp pred (add X, C1), C2
7252 if (!ICI.isEquality()) {
7253 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7255 const APInt &LHSV = LHSC->getValue();
7257 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7260 if (ICI.isSigned()) {
7261 if (CR.getLower().isSignBit()) {
7262 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7263 ConstantInt::get(*Context, CR.getUpper()));
7264 } else if (CR.getUpper().isSignBit()) {
7265 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7266 ConstantInt::get(*Context, CR.getLower()));
7269 if (CR.getLower().isMinValue()) {
7270 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7271 ConstantInt::get(*Context, CR.getUpper()));
7272 } else if (CR.getUpper().isMinValue()) {
7273 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7274 ConstantInt::get(*Context, CR.getLower()));
7281 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7282 if (ICI.isEquality()) {
7283 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7285 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7286 // the second operand is a constant, simplify a bit.
7287 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7288 switch (BO->getOpcode()) {
7289 case Instruction::SRem:
7290 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7291 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7292 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7293 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7295 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
7297 return new ICmpInst(ICI.getPredicate(), NewRem,
7298 Constant::getNullValue(BO->getType()));
7302 case Instruction::Add:
7303 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7304 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7305 if (BO->hasOneUse())
7306 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7307 ConstantExpr::getSub(RHS, BOp1C));
7308 } else if (RHSV == 0) {
7309 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7310 // efficiently invertible, or if the add has just this one use.
7311 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7313 if (Value *NegVal = dyn_castNegVal(BOp1))
7314 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
7315 else if (Value *NegVal = dyn_castNegVal(BOp0))
7316 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
7317 else if (BO->hasOneUse()) {
7318 Value *Neg = Builder->CreateNeg(BOp1);
7320 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
7324 case Instruction::Xor:
7325 // For the xor case, we can xor two constants together, eliminating
7326 // the explicit xor.
7327 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7328 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7329 ConstantExpr::getXor(RHS, BOC));
7332 case Instruction::Sub:
7333 // Replace (([sub|xor] A, B) != 0) with (A != B)
7335 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7339 case Instruction::Or:
7340 // If bits are being or'd in that are not present in the constant we
7341 // are comparing against, then the comparison could never succeed!
7342 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7343 Constant *NotCI = ConstantExpr::getNot(RHS);
7344 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7345 return ReplaceInstUsesWith(ICI,
7346 ConstantInt::get(Type::getInt1Ty(*Context),
7351 case Instruction::And:
7352 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7353 // If bits are being compared against that are and'd out, then the
7354 // comparison can never succeed!
7355 if ((RHSV & ~BOC->getValue()) != 0)
7356 return ReplaceInstUsesWith(ICI,
7357 ConstantInt::get(Type::getInt1Ty(*Context),
7360 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7361 if (RHS == BOC && RHSV.isPowerOf2())
7362 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7363 ICmpInst::ICMP_NE, LHSI,
7364 Constant::getNullValue(RHS->getType()));
7366 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7367 if (BOC->getValue().isSignBit()) {
7368 Value *X = BO->getOperand(0);
7369 Constant *Zero = Constant::getNullValue(X->getType());
7370 ICmpInst::Predicate pred = isICMP_NE ?
7371 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7372 return new ICmpInst(pred, X, Zero);
7375 // ((X & ~7) == 0) --> X < 8
7376 if (RHSV == 0 && isHighOnes(BOC)) {
7377 Value *X = BO->getOperand(0);
7378 Constant *NegX = ConstantExpr::getNeg(BOC);
7379 ICmpInst::Predicate pred = isICMP_NE ?
7380 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7381 return new ICmpInst(pred, X, NegX);
7386 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7387 // Handle icmp {eq|ne} <intrinsic>, intcst.
7388 if (II->getIntrinsicID() == Intrinsic::bswap) {
7390 ICI.setOperand(0, II->getOperand(1));
7391 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7399 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7400 /// We only handle extending casts so far.
7402 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7403 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7404 Value *LHSCIOp = LHSCI->getOperand(0);
7405 const Type *SrcTy = LHSCIOp->getType();
7406 const Type *DestTy = LHSCI->getType();
7409 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7410 // integer type is the same size as the pointer type.
7411 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7412 TD->getPointerSizeInBits() ==
7413 cast<IntegerType>(DestTy)->getBitWidth()) {
7415 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7416 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7417 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7418 RHSOp = RHSC->getOperand(0);
7419 // If the pointer types don't match, insert a bitcast.
7420 if (LHSCIOp->getType() != RHSOp->getType())
7421 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
7425 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7428 // The code below only handles extension cast instructions, so far.
7430 if (LHSCI->getOpcode() != Instruction::ZExt &&
7431 LHSCI->getOpcode() != Instruction::SExt)
7434 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7435 bool isSignedCmp = ICI.isSigned();
7437 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7438 // Not an extension from the same type?
7439 RHSCIOp = CI->getOperand(0);
7440 if (RHSCIOp->getType() != LHSCIOp->getType())
7443 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7444 // and the other is a zext), then we can't handle this.
7445 if (CI->getOpcode() != LHSCI->getOpcode())
7448 // Deal with equality cases early.
7449 if (ICI.isEquality())
7450 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7452 // A signed comparison of sign extended values simplifies into a
7453 // signed comparison.
7454 if (isSignedCmp && isSignedExt)
7455 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7457 // The other three cases all fold into an unsigned comparison.
7458 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7461 // If we aren't dealing with a constant on the RHS, exit early
7462 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7466 // Compute the constant that would happen if we truncated to SrcTy then
7467 // reextended to DestTy.
7468 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7469 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7472 // If the re-extended constant didn't change...
7474 // Deal with equality cases early.
7475 if (ICI.isEquality())
7476 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7478 // A signed comparison of sign extended values simplifies into a
7479 // signed comparison.
7480 if (isSignedExt && isSignedCmp)
7481 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7483 // The other three cases all fold into an unsigned comparison.
7484 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
7487 // The re-extended constant changed so the constant cannot be represented
7488 // in the shorter type. Consequently, we cannot emit a simple comparison.
7490 // First, handle some easy cases. We know the result cannot be equal at this
7491 // point so handle the ICI.isEquality() cases
7492 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7493 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7494 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7495 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7497 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7498 // should have been folded away previously and not enter in here.
7501 // We're performing a signed comparison.
7502 if (cast<ConstantInt>(CI)->getValue().isNegative())
7503 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7505 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7507 // We're performing an unsigned comparison.
7509 // We're performing an unsigned comp with a sign extended value.
7510 // This is true if the input is >= 0. [aka >s -1]
7511 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7512 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
7514 // Unsigned extend & unsigned compare -> always true.
7515 Result = ConstantInt::getTrue(*Context);
7519 // Finally, return the value computed.
7520 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7521 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7522 return ReplaceInstUsesWith(ICI, Result);
7524 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7525 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7526 "ICmp should be folded!");
7527 if (Constant *CI = dyn_cast<Constant>(Result))
7528 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7529 return BinaryOperator::CreateNot(Result);
7532 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7533 return commonShiftTransforms(I);
7536 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7537 return commonShiftTransforms(I);
7540 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7541 if (Instruction *R = commonShiftTransforms(I))
7544 Value *Op0 = I.getOperand(0);
7546 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7547 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7548 if (CSI->isAllOnesValue())
7549 return ReplaceInstUsesWith(I, CSI);
7551 // See if we can turn a signed shr into an unsigned shr.
7552 if (MaskedValueIsZero(Op0,
7553 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7554 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7556 // Arithmetic shifting an all-sign-bit value is a no-op.
7557 unsigned NumSignBits = ComputeNumSignBits(Op0);
7558 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7559 return ReplaceInstUsesWith(I, Op0);
7564 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7565 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7566 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7568 // shl X, 0 == X and shr X, 0 == X
7569 // shl 0, X == 0 and shr 0, X == 0
7570 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7571 Op0 == Constant::getNullValue(Op0->getType()))
7572 return ReplaceInstUsesWith(I, Op0);
7574 if (isa<UndefValue>(Op0)) {
7575 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7576 return ReplaceInstUsesWith(I, Op0);
7577 else // undef << X -> 0, undef >>u X -> 0
7578 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7580 if (isa<UndefValue>(Op1)) {
7581 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7582 return ReplaceInstUsesWith(I, Op0);
7583 else // X << undef, X >>u undef -> 0
7584 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7587 // See if we can fold away this shift.
7588 if (SimplifyDemandedInstructionBits(I))
7591 // Try to fold constant and into select arguments.
7592 if (isa<Constant>(Op0))
7593 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7594 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7597 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7598 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7603 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7604 BinaryOperator &I) {
7605 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7607 // See if we can simplify any instructions used by the instruction whose sole
7608 // purpose is to compute bits we don't care about.
7609 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7611 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7614 if (Op1->uge(TypeBits)) {
7615 if (I.getOpcode() != Instruction::AShr)
7616 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7618 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7623 // ((X*C1) << C2) == (X * (C1 << C2))
7624 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7625 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7626 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7627 return BinaryOperator::CreateMul(BO->getOperand(0),
7628 ConstantExpr::getShl(BOOp, Op1));
7630 // Try to fold constant and into select arguments.
7631 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7632 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7634 if (isa<PHINode>(Op0))
7635 if (Instruction *NV = FoldOpIntoPhi(I))
7638 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7639 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7640 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7641 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7642 // place. Don't try to do this transformation in this case. Also, we
7643 // require that the input operand is a shift-by-constant so that we have
7644 // confidence that the shifts will get folded together. We could do this
7645 // xform in more cases, but it is unlikely to be profitable.
7646 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7647 isa<ConstantInt>(TrOp->getOperand(1))) {
7648 // Okay, we'll do this xform. Make the shift of shift.
7649 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7650 // (shift2 (shift1 & 0x00FF), c2)
7651 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
7653 // For logical shifts, the truncation has the effect of making the high
7654 // part of the register be zeros. Emulate this by inserting an AND to
7655 // clear the top bits as needed. This 'and' will usually be zapped by
7656 // other xforms later if dead.
7657 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7658 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7659 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7661 // The mask we constructed says what the trunc would do if occurring
7662 // between the shifts. We want to know the effect *after* the second
7663 // shift. We know that it is a logical shift by a constant, so adjust the
7664 // mask as appropriate.
7665 if (I.getOpcode() == Instruction::Shl)
7666 MaskV <<= Op1->getZExtValue();
7668 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7669 MaskV = MaskV.lshr(Op1->getZExtValue());
7673 Value *And = Builder->CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7676 // Return the value truncated to the interesting size.
7677 return new TruncInst(And, I.getType());
7681 if (Op0->hasOneUse()) {
7682 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7683 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7686 switch (Op0BO->getOpcode()) {
7688 case Instruction::Add:
7689 case Instruction::And:
7690 case Instruction::Or:
7691 case Instruction::Xor: {
7692 // These operators commute.
7693 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7694 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7695 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7696 m_Specific(Op1)))) {
7697 Value *YS = // (Y << C)
7698 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
7700 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
7701 Op0BO->getOperand(1)->getName());
7702 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7703 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7704 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7707 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7708 Value *Op0BOOp1 = Op0BO->getOperand(1);
7709 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7711 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7712 m_ConstantInt(CC))) &&
7713 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7714 Value *YS = // (Y << C)
7715 Builder->CreateShl(Op0BO->getOperand(0), Op1,
7718 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7719 V1->getName()+".mask");
7720 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7725 case Instruction::Sub: {
7726 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7727 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7728 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7729 m_Specific(Op1)))) {
7730 Value *YS = // (Y << C)
7731 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7733 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
7734 Op0BO->getOperand(0)->getName());
7735 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7736 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7737 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7740 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7741 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7742 match(Op0BO->getOperand(0),
7743 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7744 m_ConstantInt(CC))) && V2 == Op1 &&
7745 cast<BinaryOperator>(Op0BO->getOperand(0))
7746 ->getOperand(0)->hasOneUse()) {
7747 Value *YS = // (Y << C)
7748 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7750 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7751 V1->getName()+".mask");
7753 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7761 // If the operand is an bitwise operator with a constant RHS, and the
7762 // shift is the only use, we can pull it out of the shift.
7763 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7764 bool isValid = true; // Valid only for And, Or, Xor
7765 bool highBitSet = false; // Transform if high bit of constant set?
7767 switch (Op0BO->getOpcode()) {
7768 default: isValid = false; break; // Do not perform transform!
7769 case Instruction::Add:
7770 isValid = isLeftShift;
7772 case Instruction::Or:
7773 case Instruction::Xor:
7776 case Instruction::And:
7781 // If this is a signed shift right, and the high bit is modified
7782 // by the logical operation, do not perform the transformation.
7783 // The highBitSet boolean indicates the value of the high bit of
7784 // the constant which would cause it to be modified for this
7787 if (isValid && I.getOpcode() == Instruction::AShr)
7788 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7791 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7794 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
7795 NewShift->takeName(Op0BO);
7797 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7804 // Find out if this is a shift of a shift by a constant.
7805 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7806 if (ShiftOp && !ShiftOp->isShift())
7809 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7810 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7811 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7812 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7813 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7814 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7815 Value *X = ShiftOp->getOperand(0);
7817 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7819 const IntegerType *Ty = cast<IntegerType>(I.getType());
7821 // Check for (X << c1) << c2 and (X >> c1) >> c2
7822 if (I.getOpcode() == ShiftOp->getOpcode()) {
7823 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7825 if (AmtSum >= TypeBits) {
7826 if (I.getOpcode() != Instruction::AShr)
7827 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7828 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7831 return BinaryOperator::Create(I.getOpcode(), X,
7832 ConstantInt::get(Ty, AmtSum));
7835 if (ShiftOp->getOpcode() == Instruction::LShr &&
7836 I.getOpcode() == Instruction::AShr) {
7837 if (AmtSum >= TypeBits)
7838 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7840 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7841 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7844 if (ShiftOp->getOpcode() == Instruction::AShr &&
7845 I.getOpcode() == Instruction::LShr) {
7846 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7847 if (AmtSum >= TypeBits)
7848 AmtSum = TypeBits-1;
7850 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7852 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7853 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7856 // Okay, if we get here, one shift must be left, and the other shift must be
7857 // right. See if the amounts are equal.
7858 if (ShiftAmt1 == ShiftAmt2) {
7859 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7860 if (I.getOpcode() == Instruction::Shl) {
7861 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7862 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7864 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7865 if (I.getOpcode() == Instruction::LShr) {
7866 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7867 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7869 // We can simplify ((X << C) >>s C) into a trunc + sext.
7870 // NOTE: we could do this for any C, but that would make 'unusual' integer
7871 // types. For now, just stick to ones well-supported by the code
7873 const Type *SExtType = 0;
7874 switch (Ty->getBitWidth() - ShiftAmt1) {
7881 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
7886 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
7887 // Otherwise, we can't handle it yet.
7888 } else if (ShiftAmt1 < ShiftAmt2) {
7889 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7891 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7892 if (I.getOpcode() == Instruction::Shl) {
7893 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7894 ShiftOp->getOpcode() == Instruction::AShr);
7895 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7897 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7898 return BinaryOperator::CreateAnd(Shift,
7899 ConstantInt::get(*Context, Mask));
7902 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7903 if (I.getOpcode() == Instruction::LShr) {
7904 assert(ShiftOp->getOpcode() == Instruction::Shl);
7905 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7907 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7908 return BinaryOperator::CreateAnd(Shift,
7909 ConstantInt::get(*Context, Mask));
7912 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7914 assert(ShiftAmt2 < ShiftAmt1);
7915 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7917 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7918 if (I.getOpcode() == Instruction::Shl) {
7919 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7920 ShiftOp->getOpcode() == Instruction::AShr);
7921 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
7922 ConstantInt::get(Ty, ShiftDiff));
7924 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7925 return BinaryOperator::CreateAnd(Shift,
7926 ConstantInt::get(*Context, Mask));
7929 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7930 if (I.getOpcode() == Instruction::LShr) {
7931 assert(ShiftOp->getOpcode() == Instruction::Shl);
7932 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7934 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7935 return BinaryOperator::CreateAnd(Shift,
7936 ConstantInt::get(*Context, Mask));
7939 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7946 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7947 /// expression. If so, decompose it, returning some value X, such that Val is
7950 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7951 int &Offset, LLVMContext *Context) {
7952 assert(Val->getType() == Type::getInt32Ty(*Context) &&
7953 "Unexpected allocation size type!");
7954 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7955 Offset = CI->getZExtValue();
7957 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
7958 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7959 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7960 if (I->getOpcode() == Instruction::Shl) {
7961 // This is a value scaled by '1 << the shift amt'.
7962 Scale = 1U << RHS->getZExtValue();
7964 return I->getOperand(0);
7965 } else if (I->getOpcode() == Instruction::Mul) {
7966 // This value is scaled by 'RHS'.
7967 Scale = RHS->getZExtValue();
7969 return I->getOperand(0);
7970 } else if (I->getOpcode() == Instruction::Add) {
7971 // We have X+C. Check to see if we really have (X*C2)+C1,
7972 // where C1 is divisible by C2.
7975 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7977 Offset += RHS->getZExtValue();
7984 // Otherwise, we can't look past this.
7991 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7992 /// try to eliminate the cast by moving the type information into the alloc.
7993 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7995 const PointerType *PTy = cast<PointerType>(CI.getType());
7997 BuilderTy AllocaBuilder(*Builder);
7998 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
8000 // Remove any uses of AI that are dead.
8001 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
8003 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
8004 Instruction *User = cast<Instruction>(*UI++);
8005 if (isInstructionTriviallyDead(User)) {
8006 while (UI != E && *UI == User)
8007 ++UI; // If this instruction uses AI more than once, don't break UI.
8010 DEBUG(errs() << "IC: DCE: " << *User << '\n');
8011 EraseInstFromFunction(*User);
8015 // This requires TargetData to get the alloca alignment and size information.
8018 // Get the type really allocated and the type casted to.
8019 const Type *AllocElTy = AI.getAllocatedType();
8020 const Type *CastElTy = PTy->getElementType();
8021 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
8023 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
8024 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
8025 if (CastElTyAlign < AllocElTyAlign) return 0;
8027 // If the allocation has multiple uses, only promote it if we are strictly
8028 // increasing the alignment of the resultant allocation. If we keep it the
8029 // same, we open the door to infinite loops of various kinds. (A reference
8030 // from a dbg.declare doesn't count as a use for this purpose.)
8031 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
8032 CastElTyAlign == AllocElTyAlign) return 0;
8034 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
8035 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
8036 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
8038 // See if we can satisfy the modulus by pulling a scale out of the array
8040 unsigned ArraySizeScale;
8042 Value *NumElements = // See if the array size is a decomposable linear expr.
8043 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
8044 ArrayOffset, Context);
8046 // If we can now satisfy the modulus, by using a non-1 scale, we really can
8048 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
8049 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
8051 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
8056 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
8057 // Insert before the alloca, not before the cast.
8058 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
8061 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
8062 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
8063 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
8066 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
8067 New->setAlignment(AI.getAlignment());
8070 // If the allocation has one real use plus a dbg.declare, just remove the
8072 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
8073 EraseInstFromFunction(*DI);
8075 // If the allocation has multiple real uses, insert a cast and change all
8076 // things that used it to use the new cast. This will also hack on CI, but it
8078 else if (!AI.hasOneUse()) {
8079 // New is the allocation instruction, pointer typed. AI is the original
8080 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
8081 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
8082 AI.replaceAllUsesWith(NewCast);
8084 return ReplaceInstUsesWith(CI, New);
8087 /// CanEvaluateInDifferentType - Return true if we can take the specified value
8088 /// and return it as type Ty without inserting any new casts and without
8089 /// changing the computed value. This is used by code that tries to decide
8090 /// whether promoting or shrinking integer operations to wider or smaller types
8091 /// will allow us to eliminate a truncate or extend.
8093 /// This is a truncation operation if Ty is smaller than V->getType(), or an
8094 /// extension operation if Ty is larger.
8096 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
8097 /// should return true if trunc(V) can be computed by computing V in the smaller
8098 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
8099 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
8100 /// efficiently truncated.
8102 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
8103 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
8104 /// the final result.
8105 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
8107 int &NumCastsRemoved){
8108 // We can always evaluate constants in another type.
8109 if (isa<Constant>(V))
8112 Instruction *I = dyn_cast<Instruction>(V);
8113 if (!I) return false;
8115 const Type *OrigTy = V->getType();
8117 // If this is an extension or truncate, we can often eliminate it.
8118 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
8119 // If this is a cast from the destination type, we can trivially eliminate
8120 // it, and this will remove a cast overall.
8121 if (I->getOperand(0)->getType() == Ty) {
8122 // If the first operand is itself a cast, and is eliminable, do not count
8123 // this as an eliminable cast. We would prefer to eliminate those two
8125 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
8131 // We can't extend or shrink something that has multiple uses: doing so would
8132 // require duplicating the instruction in general, which isn't profitable.
8133 if (!I->hasOneUse()) return false;
8135 unsigned Opc = I->getOpcode();
8137 case Instruction::Add:
8138 case Instruction::Sub:
8139 case Instruction::Mul:
8140 case Instruction::And:
8141 case Instruction::Or:
8142 case Instruction::Xor:
8143 // These operators can all arbitrarily be extended or truncated.
8144 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8146 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
8149 case Instruction::UDiv:
8150 case Instruction::URem: {
8151 // UDiv and URem can be truncated if all the truncated bits are zero.
8152 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8153 uint32_t BitWidth = Ty->getScalarSizeInBits();
8154 if (BitWidth < OrigBitWidth) {
8155 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
8156 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
8157 MaskedValueIsZero(I->getOperand(1), Mask)) {
8158 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8160 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
8166 case Instruction::Shl:
8167 // If we are truncating the result of this SHL, and if it's a shift of a
8168 // constant amount, we can always perform a SHL in a smaller type.
8169 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8170 uint32_t BitWidth = Ty->getScalarSizeInBits();
8171 if (BitWidth < OrigTy->getScalarSizeInBits() &&
8172 CI->getLimitedValue(BitWidth) < BitWidth)
8173 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8177 case Instruction::LShr:
8178 // If this is a truncate of a logical shr, we can truncate it to a smaller
8179 // lshr iff we know that the bits we would otherwise be shifting in are
8181 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8182 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8183 uint32_t BitWidth = Ty->getScalarSizeInBits();
8184 if (BitWidth < OrigBitWidth &&
8185 MaskedValueIsZero(I->getOperand(0),
8186 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8187 CI->getLimitedValue(BitWidth) < BitWidth) {
8188 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8193 case Instruction::ZExt:
8194 case Instruction::SExt:
8195 case Instruction::Trunc:
8196 // If this is the same kind of case as our original (e.g. zext+zext), we
8197 // can safely replace it. Note that replacing it does not reduce the number
8198 // of casts in the input.
8202 // sext (zext ty1), ty2 -> zext ty2
8203 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8206 case Instruction::Select: {
8207 SelectInst *SI = cast<SelectInst>(I);
8208 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8210 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8213 case Instruction::PHI: {
8214 // We can change a phi if we can change all operands.
8215 PHINode *PN = cast<PHINode>(I);
8216 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8217 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8223 // TODO: Can handle more cases here.
8230 /// EvaluateInDifferentType - Given an expression that
8231 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8232 /// evaluate the expression.
8233 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8235 if (Constant *C = dyn_cast<Constant>(V))
8236 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
8238 // Otherwise, it must be an instruction.
8239 Instruction *I = cast<Instruction>(V);
8240 Instruction *Res = 0;
8241 unsigned Opc = I->getOpcode();
8243 case Instruction::Add:
8244 case Instruction::Sub:
8245 case Instruction::Mul:
8246 case Instruction::And:
8247 case Instruction::Or:
8248 case Instruction::Xor:
8249 case Instruction::AShr:
8250 case Instruction::LShr:
8251 case Instruction::Shl:
8252 case Instruction::UDiv:
8253 case Instruction::URem: {
8254 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8255 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8256 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8259 case Instruction::Trunc:
8260 case Instruction::ZExt:
8261 case Instruction::SExt:
8262 // If the source type of the cast is the type we're trying for then we can
8263 // just return the source. There's no need to insert it because it is not
8265 if (I->getOperand(0)->getType() == Ty)
8266 return I->getOperand(0);
8268 // Otherwise, must be the same type of cast, so just reinsert a new one.
8269 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty);
8271 case Instruction::Select: {
8272 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8273 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8274 Res = SelectInst::Create(I->getOperand(0), True, False);
8277 case Instruction::PHI: {
8278 PHINode *OPN = cast<PHINode>(I);
8279 PHINode *NPN = PHINode::Create(Ty);
8280 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8281 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8282 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8288 // TODO: Can handle more cases here.
8289 llvm_unreachable("Unreachable!");
8294 return InsertNewInstBefore(Res, *I);
8297 /// @brief Implement the transforms common to all CastInst visitors.
8298 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8299 Value *Src = CI.getOperand(0);
8301 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8302 // eliminate it now.
8303 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8304 if (Instruction::CastOps opc =
8305 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8306 // The first cast (CSrc) is eliminable so we need to fix up or replace
8307 // the second cast (CI). CSrc will then have a good chance of being dead.
8308 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8312 // If we are casting a select then fold the cast into the select
8313 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8314 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8317 // If we are casting a PHI then fold the cast into the PHI
8318 if (isa<PHINode>(Src)) {
8319 // We don't do this if this would create a PHI node with an illegal type if
8320 // it is currently legal.
8321 if (!isa<IntegerType>(Src->getType()) ||
8322 !isa<IntegerType>(CI.getType()) ||
8323 ShouldChangeType(CI.getType(), Src->getType(), TD))
8324 if (Instruction *NV = FoldOpIntoPhi(CI))
8331 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8332 /// or not there is a sequence of GEP indices into the type that will land us at
8333 /// the specified offset. If so, fill them into NewIndices and return the
8334 /// resultant element type, otherwise return null.
8335 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8336 SmallVectorImpl<Value*> &NewIndices,
8337 const TargetData *TD,
8338 LLVMContext *Context) {
8340 if (!Ty->isSized()) return 0;
8342 // Start with the index over the outer type. Note that the type size
8343 // might be zero (even if the offset isn't zero) if the indexed type
8344 // is something like [0 x {int, int}]
8345 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8346 int64_t FirstIdx = 0;
8347 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8348 FirstIdx = Offset/TySize;
8349 Offset -= FirstIdx*TySize;
8351 // Handle hosts where % returns negative instead of values [0..TySize).
8355 assert(Offset >= 0);
8357 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8360 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8362 // Index into the types. If we fail, set OrigBase to null.
8364 // Indexing into tail padding between struct/array elements.
8365 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8368 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8369 const StructLayout *SL = TD->getStructLayout(STy);
8370 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8371 "Offset must stay within the indexed type");
8373 unsigned Elt = SL->getElementContainingOffset(Offset);
8374 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8376 Offset -= SL->getElementOffset(Elt);
8377 Ty = STy->getElementType(Elt);
8378 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8379 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8380 assert(EltSize && "Cannot index into a zero-sized array");
8381 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8383 Ty = AT->getElementType();
8385 // Otherwise, we can't index into the middle of this atomic type, bail.
8393 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8394 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8395 Value *Src = CI.getOperand(0);
8397 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8398 // If casting the result of a getelementptr instruction with no offset, turn
8399 // this into a cast of the original pointer!
8400 if (GEP->hasAllZeroIndices()) {
8401 // Changing the cast operand is usually not a good idea but it is safe
8402 // here because the pointer operand is being replaced with another
8403 // pointer operand so the opcode doesn't need to change.
8405 CI.setOperand(0, GEP->getOperand(0));
8409 // If the GEP has a single use, and the base pointer is a bitcast, and the
8410 // GEP computes a constant offset, see if we can convert these three
8411 // instructions into fewer. This typically happens with unions and other
8412 // non-type-safe code.
8413 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8414 if (GEP->hasAllConstantIndices()) {
8415 // We are guaranteed to get a constant from EmitGEPOffset.
8416 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, *this));
8417 int64_t Offset = OffsetV->getSExtValue();
8419 // Get the base pointer input of the bitcast, and the type it points to.
8420 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8421 const Type *GEPIdxTy =
8422 cast<PointerType>(OrigBase->getType())->getElementType();
8423 SmallVector<Value*, 8> NewIndices;
8424 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8425 // If we were able to index down into an element, create the GEP
8426 // and bitcast the result. This eliminates one bitcast, potentially
8428 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
8429 Builder->CreateInBoundsGEP(OrigBase,
8430 NewIndices.begin(), NewIndices.end()) :
8431 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
8432 NGEP->takeName(GEP);
8434 if (isa<BitCastInst>(CI))
8435 return new BitCastInst(NGEP, CI.getType());
8436 assert(isa<PtrToIntInst>(CI));
8437 return new PtrToIntInst(NGEP, CI.getType());
8443 return commonCastTransforms(CI);
8446 /// commonIntCastTransforms - This function implements the common transforms
8447 /// for trunc, zext, and sext.
8448 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8449 if (Instruction *Result = commonCastTransforms(CI))
8452 Value *Src = CI.getOperand(0);
8453 const Type *SrcTy = Src->getType();
8454 const Type *DestTy = CI.getType();
8455 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8456 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8458 // See if we can simplify any instructions used by the LHS whose sole
8459 // purpose is to compute bits we don't care about.
8460 if (SimplifyDemandedInstructionBits(CI))
8463 // If the source isn't an instruction or has more than one use then we
8464 // can't do anything more.
8465 Instruction *SrcI = dyn_cast<Instruction>(Src);
8466 if (!SrcI || !Src->hasOneUse())
8469 // Attempt to propagate the cast into the instruction for int->int casts.
8470 int NumCastsRemoved = 0;
8471 // Only do this if the dest type is a simple type, don't convert the
8472 // expression tree to something weird like i93 unless the source is also
8474 if ((isa<VectorType>(DestTy) ||
8475 ShouldChangeType(SrcI->getType(), DestTy, TD)) &&
8476 CanEvaluateInDifferentType(SrcI, DestTy,
8477 CI.getOpcode(), NumCastsRemoved)) {
8478 // If this cast is a truncate, evaluting in a different type always
8479 // eliminates the cast, so it is always a win. If this is a zero-extension,
8480 // we need to do an AND to maintain the clear top-part of the computation,
8481 // so we require that the input have eliminated at least one cast. If this
8482 // is a sign extension, we insert two new casts (to do the extension) so we
8483 // require that two casts have been eliminated.
8484 bool DoXForm = false;
8485 bool JustReplace = false;
8486 switch (CI.getOpcode()) {
8488 // All the others use floating point so we shouldn't actually
8489 // get here because of the check above.
8490 llvm_unreachable("Unknown cast type");
8491 case Instruction::Trunc:
8494 case Instruction::ZExt: {
8495 DoXForm = NumCastsRemoved >= 1;
8497 if (!DoXForm && 0) {
8498 // If it's unnecessary to issue an AND to clear the high bits, it's
8499 // always profitable to do this xform.
8500 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8501 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8502 if (MaskedValueIsZero(TryRes, Mask))
8503 return ReplaceInstUsesWith(CI, TryRes);
8505 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8506 if (TryI->use_empty())
8507 EraseInstFromFunction(*TryI);
8511 case Instruction::SExt: {
8512 DoXForm = NumCastsRemoved >= 2;
8513 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8514 // If we do not have to emit the truncate + sext pair, then it's always
8515 // profitable to do this xform.
8517 // It's not safe to eliminate the trunc + sext pair if one of the
8518 // eliminated cast is a truncate. e.g.
8519 // t2 = trunc i32 t1 to i16
8520 // t3 = sext i16 t2 to i32
8523 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8524 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8525 if (NumSignBits > (DestBitSize - SrcBitSize))
8526 return ReplaceInstUsesWith(CI, TryRes);
8528 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8529 if (TryI->use_empty())
8530 EraseInstFromFunction(*TryI);
8537 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8538 " to avoid cast: " << CI);
8539 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8540 CI.getOpcode() == Instruction::SExt);
8542 // Just replace this cast with the result.
8543 return ReplaceInstUsesWith(CI, Res);
8545 assert(Res->getType() == DestTy);
8546 switch (CI.getOpcode()) {
8547 default: llvm_unreachable("Unknown cast type!");
8548 case Instruction::Trunc:
8549 // Just replace this cast with the result.
8550 return ReplaceInstUsesWith(CI, Res);
8551 case Instruction::ZExt: {
8552 assert(SrcBitSize < DestBitSize && "Not a zext?");
8554 // If the high bits are already zero, just replace this cast with the
8556 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8557 if (MaskedValueIsZero(Res, Mask))
8558 return ReplaceInstUsesWith(CI, Res);
8560 // We need to emit an AND to clear the high bits.
8561 Constant *C = ConstantInt::get(*Context,
8562 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8563 return BinaryOperator::CreateAnd(Res, C);
8565 case Instruction::SExt: {
8566 // If the high bits are already filled with sign bit, just replace this
8567 // cast with the result.
8568 unsigned NumSignBits = ComputeNumSignBits(Res);
8569 if (NumSignBits > (DestBitSize - SrcBitSize))
8570 return ReplaceInstUsesWith(CI, Res);
8572 // We need to emit a cast to truncate, then a cast to sext.
8573 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
8579 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8580 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8582 switch (SrcI->getOpcode()) {
8583 case Instruction::Add:
8584 case Instruction::Mul:
8585 case Instruction::And:
8586 case Instruction::Or:
8587 case Instruction::Xor:
8588 // If we are discarding information, rewrite.
8589 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8590 // Don't insert two casts unless at least one can be eliminated.
8591 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8592 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8593 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8594 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8595 return BinaryOperator::Create(
8596 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8600 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8601 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8602 SrcI->getOpcode() == Instruction::Xor &&
8603 Op1 == ConstantInt::getTrue(*Context) &&
8604 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8605 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName());
8606 return BinaryOperator::CreateXor(New,
8607 ConstantInt::get(CI.getType(), 1));
8611 case Instruction::Shl: {
8612 // Canonicalize trunc inside shl, if we can.
8613 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8614 if (CI && DestBitSize < SrcBitSize &&
8615 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8616 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8617 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8618 return BinaryOperator::CreateShl(Op0c, Op1c);
8626 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8627 if (Instruction *Result = commonIntCastTransforms(CI))
8630 Value *Src = CI.getOperand(0);
8631 const Type *Ty = CI.getType();
8632 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8633 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8635 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8636 if (DestBitWidth == 1) {
8637 Constant *One = ConstantInt::get(Src->getType(), 1);
8638 Src = Builder->CreateAnd(Src, One, "tmp");
8639 Value *Zero = Constant::getNullValue(Src->getType());
8640 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8643 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8644 ConstantInt *ShAmtV = 0;
8646 if (Src->hasOneUse() &&
8647 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8648 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8650 // Get a mask for the bits shifting in.
8651 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8652 if (MaskedValueIsZero(ShiftOp, Mask)) {
8653 if (ShAmt >= DestBitWidth) // All zeros.
8654 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8656 // Okay, we can shrink this. Truncate the input, then return a new
8658 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName());
8659 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8660 return BinaryOperator::CreateLShr(V1, V2);
8667 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8668 /// in order to eliminate the icmp.
8669 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8671 // If we are just checking for a icmp eq of a single bit and zext'ing it
8672 // to an integer, then shift the bit to the appropriate place and then
8673 // cast to integer to avoid the comparison.
8674 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8675 const APInt &Op1CV = Op1C->getValue();
8677 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8678 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8679 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8680 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8681 if (!DoXform) return ICI;
8683 Value *In = ICI->getOperand(0);
8684 Value *Sh = ConstantInt::get(In->getType(),
8685 In->getType()->getScalarSizeInBits()-1);
8686 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
8687 if (In->getType() != CI.getType())
8688 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
8690 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8691 Constant *One = ConstantInt::get(In->getType(), 1);
8692 In = Builder->CreateXor(In, One, In->getName()+".not");
8695 return ReplaceInstUsesWith(CI, In);
8700 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8701 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8702 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8703 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8704 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8705 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8706 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8707 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8708 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8709 // This only works for EQ and NE
8710 ICI->isEquality()) {
8711 // If Op1C some other power of two, convert:
8712 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8713 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8714 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8715 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8717 APInt KnownZeroMask(~KnownZero);
8718 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8719 if (!DoXform) return ICI;
8721 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8722 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8723 // (X&4) == 2 --> false
8724 // (X&4) != 2 --> true
8725 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8726 Res = ConstantExpr::getZExt(Res, CI.getType());
8727 return ReplaceInstUsesWith(CI, Res);
8730 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8731 Value *In = ICI->getOperand(0);
8733 // Perform a logical shr by shiftamt.
8734 // Insert the shift to put the result in the low bit.
8735 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
8736 In->getName()+".lobit");
8739 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8740 Constant *One = ConstantInt::get(In->getType(), 1);
8741 In = Builder->CreateXor(In, One, "tmp");
8744 if (CI.getType() == In->getType())
8745 return ReplaceInstUsesWith(CI, In);
8747 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8752 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
8753 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
8754 // may lead to additional simplifications.
8755 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
8756 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
8757 uint32_t BitWidth = ITy->getBitWidth();
8758 Value *LHS = ICI->getOperand(0);
8759 Value *RHS = ICI->getOperand(1);
8761 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
8762 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
8763 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8764 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
8765 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
8767 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
8768 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
8769 APInt UnknownBit = ~KnownBits;
8770 if (UnknownBit.countPopulation() == 1) {
8771 if (!DoXform) return ICI;
8773 Value *Result = Builder->CreateXor(LHS, RHS);
8775 // Mask off any bits that are set and won't be shifted away.
8776 if (KnownOneLHS.uge(UnknownBit))
8777 Result = Builder->CreateAnd(Result,
8778 ConstantInt::get(ITy, UnknownBit));
8780 // Shift the bit we're testing down to the lsb.
8781 Result = Builder->CreateLShr(
8782 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
8784 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8785 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
8786 Result->takeName(ICI);
8787 return ReplaceInstUsesWith(CI, Result);
8796 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8797 // If one of the common conversion will work ..
8798 if (Instruction *Result = commonIntCastTransforms(CI))
8801 Value *Src = CI.getOperand(0);
8803 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8804 // types and if the sizes are just right we can convert this into a logical
8805 // 'and' which will be much cheaper than the pair of casts.
8806 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8807 // Get the sizes of the types involved. We know that the intermediate type
8808 // will be smaller than A or C, but don't know the relation between A and C.
8809 Value *A = CSrc->getOperand(0);
8810 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8811 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8812 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8813 // If we're actually extending zero bits, then if
8814 // SrcSize < DstSize: zext(a & mask)
8815 // SrcSize == DstSize: a & mask
8816 // SrcSize > DstSize: trunc(a) & mask
8817 if (SrcSize < DstSize) {
8818 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8819 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8820 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
8821 return new ZExtInst(And, CI.getType());
8824 if (SrcSize == DstSize) {
8825 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8826 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8829 if (SrcSize > DstSize) {
8830 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
8831 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8832 return BinaryOperator::CreateAnd(Trunc,
8833 ConstantInt::get(Trunc->getType(),
8838 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8839 return transformZExtICmp(ICI, CI);
8841 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8842 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8843 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8844 // of the (zext icmp) will be transformed.
8845 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8846 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8847 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8848 (transformZExtICmp(LHS, CI, false) ||
8849 transformZExtICmp(RHS, CI, false))) {
8850 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
8851 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
8852 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8856 // zext(trunc(t) & C) -> (t & zext(C)).
8857 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8858 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8859 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8860 Value *TI0 = TI->getOperand(0);
8861 if (TI0->getType() == CI.getType())
8863 BinaryOperator::CreateAnd(TI0,
8864 ConstantExpr::getZExt(C, CI.getType()));
8867 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8868 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8869 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8870 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8871 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8872 And->getOperand(1) == C)
8873 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8874 Value *TI0 = TI->getOperand(0);
8875 if (TI0->getType() == CI.getType()) {
8876 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8877 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
8878 return BinaryOperator::CreateXor(NewAnd, ZC);
8885 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8886 if (Instruction *I = commonIntCastTransforms(CI))
8889 Value *Src = CI.getOperand(0);
8891 // Canonicalize sign-extend from i1 to a select.
8892 if (Src->getType() == Type::getInt1Ty(*Context))
8893 return SelectInst::Create(Src,
8894 Constant::getAllOnesValue(CI.getType()),
8895 Constant::getNullValue(CI.getType()));
8897 // See if the value being truncated is already sign extended. If so, just
8898 // eliminate the trunc/sext pair.
8899 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8900 Value *Op = cast<User>(Src)->getOperand(0);
8901 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8902 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8903 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8904 unsigned NumSignBits = ComputeNumSignBits(Op);
8906 if (OpBits == DestBits) {
8907 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8908 // bits, it is already ready.
8909 if (NumSignBits > DestBits-MidBits)
8910 return ReplaceInstUsesWith(CI, Op);
8911 } else if (OpBits < DestBits) {
8912 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8913 // bits, just sext from i32.
8914 if (NumSignBits > OpBits-MidBits)
8915 return new SExtInst(Op, CI.getType(), "tmp");
8917 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8918 // bits, just truncate to i32.
8919 if (NumSignBits > OpBits-MidBits)
8920 return new TruncInst(Op, CI.getType(), "tmp");
8924 // If the input is a shl/ashr pair of a same constant, then this is a sign
8925 // extension from a smaller value. If we could trust arbitrary bitwidth
8926 // integers, we could turn this into a truncate to the smaller bit and then
8927 // use a sext for the whole extension. Since we don't, look deeper and check
8928 // for a truncate. If the source and dest are the same type, eliminate the
8929 // trunc and extend and just do shifts. For example, turn:
8930 // %a = trunc i32 %i to i8
8931 // %b = shl i8 %a, 6
8932 // %c = ashr i8 %b, 6
8933 // %d = sext i8 %c to i32
8935 // %a = shl i32 %i, 30
8936 // %d = ashr i32 %a, 30
8938 ConstantInt *BA = 0, *CA = 0;
8939 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8940 m_ConstantInt(CA))) &&
8941 BA == CA && isa<TruncInst>(A)) {
8942 Value *I = cast<TruncInst>(A)->getOperand(0);
8943 if (I->getType() == CI.getType()) {
8944 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8945 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8946 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8947 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8948 I = Builder->CreateShl(I, ShAmtV, CI.getName());
8949 return BinaryOperator::CreateAShr(I, ShAmtV);
8956 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8957 /// in the specified FP type without changing its value.
8958 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8959 LLVMContext *Context) {
8961 APFloat F = CFP->getValueAPF();
8962 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8964 return ConstantFP::get(*Context, F);
8968 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8969 /// through it until we get the source value.
8970 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8971 if (Instruction *I = dyn_cast<Instruction>(V))
8972 if (I->getOpcode() == Instruction::FPExt)
8973 return LookThroughFPExtensions(I->getOperand(0), Context);
8975 // If this value is a constant, return the constant in the smallest FP type
8976 // that can accurately represent it. This allows us to turn
8977 // (float)((double)X+2.0) into x+2.0f.
8978 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8979 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
8980 return V; // No constant folding of this.
8981 // See if the value can be truncated to float and then reextended.
8982 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8984 if (CFP->getType() == Type::getDoubleTy(*Context))
8985 return V; // Won't shrink.
8986 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8988 // Don't try to shrink to various long double types.
8994 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8995 if (Instruction *I = commonCastTransforms(CI))
8998 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8999 // smaller than the destination type, we can eliminate the truncate by doing
9000 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
9001 // many builtins (sqrt, etc).
9002 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
9003 if (OpI && OpI->hasOneUse()) {
9004 switch (OpI->getOpcode()) {
9006 case Instruction::FAdd:
9007 case Instruction::FSub:
9008 case Instruction::FMul:
9009 case Instruction::FDiv:
9010 case Instruction::FRem:
9011 const Type *SrcTy = OpI->getType();
9012 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
9013 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
9014 if (LHSTrunc->getType() != SrcTy &&
9015 RHSTrunc->getType() != SrcTy) {
9016 unsigned DstSize = CI.getType()->getScalarSizeInBits();
9017 // If the source types were both smaller than the destination type of
9018 // the cast, do this xform.
9019 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
9020 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
9021 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
9022 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
9023 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
9032 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
9033 return commonCastTransforms(CI);
9036 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
9037 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
9039 return commonCastTransforms(FI);
9041 // fptoui(uitofp(X)) --> X
9042 // fptoui(sitofp(X)) --> X
9043 // This is safe if the intermediate type has enough bits in its mantissa to
9044 // accurately represent all values of X. For example, do not do this with
9045 // i64->float->i64. This is also safe for sitofp case, because any negative
9046 // 'X' value would cause an undefined result for the fptoui.
9047 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
9048 OpI->getOperand(0)->getType() == FI.getType() &&
9049 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
9050 OpI->getType()->getFPMantissaWidth())
9051 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
9053 return commonCastTransforms(FI);
9056 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
9057 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
9059 return commonCastTransforms(FI);
9061 // fptosi(sitofp(X)) --> X
9062 // fptosi(uitofp(X)) --> X
9063 // This is safe if the intermediate type has enough bits in its mantissa to
9064 // accurately represent all values of X. For example, do not do this with
9065 // i64->float->i64. This is also safe for sitofp case, because any negative
9066 // 'X' value would cause an undefined result for the fptoui.
9067 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
9068 OpI->getOperand(0)->getType() == FI.getType() &&
9069 (int)FI.getType()->getScalarSizeInBits() <=
9070 OpI->getType()->getFPMantissaWidth())
9071 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
9073 return commonCastTransforms(FI);
9076 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
9077 return commonCastTransforms(CI);
9080 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
9081 return commonCastTransforms(CI);
9084 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
9085 // If the destination integer type is smaller than the intptr_t type for
9086 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
9087 // trunc to be exposed to other transforms. Don't do this for extending
9088 // ptrtoint's, because we don't know if the target sign or zero extends its
9091 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
9092 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
9093 TD->getIntPtrType(CI.getContext()),
9095 return new TruncInst(P, CI.getType());
9098 return commonPointerCastTransforms(CI);
9101 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
9102 // If the source integer type is larger than the intptr_t type for
9103 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
9104 // allows the trunc to be exposed to other transforms. Don't do this for
9105 // extending inttoptr's, because we don't know if the target sign or zero
9106 // extends to pointers.
9107 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
9108 TD->getPointerSizeInBits()) {
9109 Value *P = Builder->CreateTrunc(CI.getOperand(0),
9110 TD->getIntPtrType(CI.getContext()), "tmp");
9111 return new IntToPtrInst(P, CI.getType());
9114 if (Instruction *I = commonCastTransforms(CI))
9120 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
9121 // If the operands are integer typed then apply the integer transforms,
9122 // otherwise just apply the common ones.
9123 Value *Src = CI.getOperand(0);
9124 const Type *SrcTy = Src->getType();
9125 const Type *DestTy = CI.getType();
9127 if (isa<PointerType>(SrcTy)) {
9128 if (Instruction *I = commonPointerCastTransforms(CI))
9131 if (Instruction *Result = commonCastTransforms(CI))
9136 // Get rid of casts from one type to the same type. These are useless and can
9137 // be replaced by the operand.
9138 if (DestTy == Src->getType())
9139 return ReplaceInstUsesWith(CI, Src);
9141 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
9142 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
9143 const Type *DstElTy = DstPTy->getElementType();
9144 const Type *SrcElTy = SrcPTy->getElementType();
9146 // If the address spaces don't match, don't eliminate the bitcast, which is
9147 // required for changing types.
9148 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
9151 // If we are casting a alloca to a pointer to a type of the same
9152 // size, rewrite the allocation instruction to allocate the "right" type.
9153 // There is no need to modify malloc calls because it is their bitcast that
9154 // needs to be cleaned up.
9155 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
9156 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
9159 // If the source and destination are pointers, and this cast is equivalent
9160 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
9161 // This can enhance SROA and other transforms that want type-safe pointers.
9162 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
9163 unsigned NumZeros = 0;
9164 while (SrcElTy != DstElTy &&
9165 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
9166 SrcElTy->getNumContainedTypes() /* not "{}" */) {
9167 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
9171 // If we found a path from the src to dest, create the getelementptr now.
9172 if (SrcElTy == DstElTy) {
9173 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
9174 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(), "",
9175 ((Instruction*) NULL));
9179 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
9180 if (DestVTy->getNumElements() == 1) {
9181 if (!isa<VectorType>(SrcTy)) {
9182 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
9183 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
9184 Constant::getNullValue(Type::getInt32Ty(*Context)));
9186 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
9190 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
9191 if (SrcVTy->getNumElements() == 1) {
9192 if (!isa<VectorType>(DestTy)) {
9194 Builder->CreateExtractElement(Src,
9195 Constant::getNullValue(Type::getInt32Ty(*Context)));
9196 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
9201 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9202 if (SVI->hasOneUse()) {
9203 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9204 // a bitconvert to a vector with the same # elts.
9205 if (isa<VectorType>(DestTy) &&
9206 cast<VectorType>(DestTy)->getNumElements() ==
9207 SVI->getType()->getNumElements() &&
9208 SVI->getType()->getNumElements() ==
9209 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9211 // If either of the operands is a cast from CI.getType(), then
9212 // evaluating the shuffle in the casted destination's type will allow
9213 // us to eliminate at least one cast.
9214 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9215 Tmp->getOperand(0)->getType() == DestTy) ||
9216 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9217 Tmp->getOperand(0)->getType() == DestTy)) {
9218 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
9219 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
9220 // Return a new shuffle vector. Use the same element ID's, as we
9221 // know the vector types match #elts.
9222 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9230 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9232 /// %D = select %cond, %C, %A
9234 /// %C = select %cond, %B, 0
9237 /// Assuming that the specified instruction is an operand to the select, return
9238 /// a bitmask indicating which operands of this instruction are foldable if they
9239 /// equal the other incoming value of the select.
9241 static unsigned GetSelectFoldableOperands(Instruction *I) {
9242 switch (I->getOpcode()) {
9243 case Instruction::Add:
9244 case Instruction::Mul:
9245 case Instruction::And:
9246 case Instruction::Or:
9247 case Instruction::Xor:
9248 return 3; // Can fold through either operand.
9249 case Instruction::Sub: // Can only fold on the amount subtracted.
9250 case Instruction::Shl: // Can only fold on the shift amount.
9251 case Instruction::LShr:
9252 case Instruction::AShr:
9255 return 0; // Cannot fold
9259 /// GetSelectFoldableConstant - For the same transformation as the previous
9260 /// function, return the identity constant that goes into the select.
9261 static Constant *GetSelectFoldableConstant(Instruction *I,
9262 LLVMContext *Context) {
9263 switch (I->getOpcode()) {
9264 default: llvm_unreachable("This cannot happen!");
9265 case Instruction::Add:
9266 case Instruction::Sub:
9267 case Instruction::Or:
9268 case Instruction::Xor:
9269 case Instruction::Shl:
9270 case Instruction::LShr:
9271 case Instruction::AShr:
9272 return Constant::getNullValue(I->getType());
9273 case Instruction::And:
9274 return Constant::getAllOnesValue(I->getType());
9275 case Instruction::Mul:
9276 return ConstantInt::get(I->getType(), 1);
9280 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9281 /// have the same opcode and only one use each. Try to simplify this.
9282 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9284 if (TI->getNumOperands() == 1) {
9285 // If this is a non-volatile load or a cast from the same type,
9288 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9291 return 0; // unknown unary op.
9294 // Fold this by inserting a select from the input values.
9295 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9296 FI->getOperand(0), SI.getName()+".v");
9297 InsertNewInstBefore(NewSI, SI);
9298 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9302 // Only handle binary operators here.
9303 if (!isa<BinaryOperator>(TI))
9306 // Figure out if the operations have any operands in common.
9307 Value *MatchOp, *OtherOpT, *OtherOpF;
9309 if (TI->getOperand(0) == FI->getOperand(0)) {
9310 MatchOp = TI->getOperand(0);
9311 OtherOpT = TI->getOperand(1);
9312 OtherOpF = FI->getOperand(1);
9313 MatchIsOpZero = true;
9314 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9315 MatchOp = TI->getOperand(1);
9316 OtherOpT = TI->getOperand(0);
9317 OtherOpF = FI->getOperand(0);
9318 MatchIsOpZero = false;
9319 } else if (!TI->isCommutative()) {
9321 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9322 MatchOp = TI->getOperand(0);
9323 OtherOpT = TI->getOperand(1);
9324 OtherOpF = FI->getOperand(0);
9325 MatchIsOpZero = true;
9326 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9327 MatchOp = TI->getOperand(1);
9328 OtherOpT = TI->getOperand(0);
9329 OtherOpF = FI->getOperand(1);
9330 MatchIsOpZero = true;
9335 // If we reach here, they do have operations in common.
9336 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9337 OtherOpF, SI.getName()+".v");
9338 InsertNewInstBefore(NewSI, SI);
9340 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9342 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9344 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9346 llvm_unreachable("Shouldn't get here");
9350 static bool isSelect01(Constant *C1, Constant *C2) {
9351 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9354 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9357 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9360 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9361 /// facilitate further optimization.
9362 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9364 // See the comment above GetSelectFoldableOperands for a description of the
9365 // transformation we are doing here.
9366 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9367 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9368 !isa<Constant>(FalseVal)) {
9369 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9370 unsigned OpToFold = 0;
9371 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9373 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9378 Constant *C = GetSelectFoldableConstant(TVI, Context);
9379 Value *OOp = TVI->getOperand(2-OpToFold);
9380 // Avoid creating select between 2 constants unless it's selecting
9382 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9383 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9384 InsertNewInstBefore(NewSel, SI);
9385 NewSel->takeName(TVI);
9386 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9387 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9388 llvm_unreachable("Unknown instruction!!");
9395 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9396 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9397 !isa<Constant>(TrueVal)) {
9398 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9399 unsigned OpToFold = 0;
9400 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9402 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9407 Constant *C = GetSelectFoldableConstant(FVI, Context);
9408 Value *OOp = FVI->getOperand(2-OpToFold);
9409 // Avoid creating select between 2 constants unless it's selecting
9411 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9412 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9413 InsertNewInstBefore(NewSel, SI);
9414 NewSel->takeName(FVI);
9415 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9416 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9417 llvm_unreachable("Unknown instruction!!");
9427 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9428 /// ICmpInst as its first operand.
9430 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9432 bool Changed = false;
9433 ICmpInst::Predicate Pred = ICI->getPredicate();
9434 Value *CmpLHS = ICI->getOperand(0);
9435 Value *CmpRHS = ICI->getOperand(1);
9436 Value *TrueVal = SI.getTrueValue();
9437 Value *FalseVal = SI.getFalseValue();
9439 // Check cases where the comparison is with a constant that
9440 // can be adjusted to fit the min/max idiom. We may edit ICI in
9441 // place here, so make sure the select is the only user.
9442 if (ICI->hasOneUse())
9443 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9446 case ICmpInst::ICMP_ULT:
9447 case ICmpInst::ICMP_SLT: {
9448 // X < MIN ? T : F --> F
9449 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9450 return ReplaceInstUsesWith(SI, FalseVal);
9451 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9452 Constant *AdjustedRHS = SubOne(CI);
9453 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9454 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9455 Pred = ICmpInst::getSwappedPredicate(Pred);
9456 CmpRHS = AdjustedRHS;
9457 std::swap(FalseVal, TrueVal);
9458 ICI->setPredicate(Pred);
9459 ICI->setOperand(1, CmpRHS);
9460 SI.setOperand(1, TrueVal);
9461 SI.setOperand(2, FalseVal);
9466 case ICmpInst::ICMP_UGT:
9467 case ICmpInst::ICMP_SGT: {
9468 // X > MAX ? T : F --> F
9469 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9470 return ReplaceInstUsesWith(SI, FalseVal);
9471 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9472 Constant *AdjustedRHS = AddOne(CI);
9473 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9474 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9475 Pred = ICmpInst::getSwappedPredicate(Pred);
9476 CmpRHS = AdjustedRHS;
9477 std::swap(FalseVal, TrueVal);
9478 ICI->setPredicate(Pred);
9479 ICI->setOperand(1, CmpRHS);
9480 SI.setOperand(1, TrueVal);
9481 SI.setOperand(2, FalseVal);
9488 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9489 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9490 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9491 if (match(TrueVal, m_ConstantInt<-1>()) &&
9492 match(FalseVal, m_ConstantInt<0>()))
9493 Pred = ICI->getPredicate();
9494 else if (match(TrueVal, m_ConstantInt<0>()) &&
9495 match(FalseVal, m_ConstantInt<-1>()))
9496 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9498 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9499 // If we are just checking for a icmp eq of a single bit and zext'ing it
9500 // to an integer, then shift the bit to the appropriate place and then
9501 // cast to integer to avoid the comparison.
9502 const APInt &Op1CV = CI->getValue();
9504 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9505 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9506 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9507 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9508 Value *In = ICI->getOperand(0);
9509 Value *Sh = ConstantInt::get(In->getType(),
9510 In->getType()->getScalarSizeInBits()-1);
9511 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9512 In->getName()+".lobit"),
9514 if (In->getType() != SI.getType())
9515 In = CastInst::CreateIntegerCast(In, SI.getType(),
9516 true/*SExt*/, "tmp", ICI);
9518 if (Pred == ICmpInst::ICMP_SGT)
9519 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9520 In->getName()+".not"), *ICI);
9522 return ReplaceInstUsesWith(SI, In);
9527 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9528 // Transform (X == Y) ? X : Y -> Y
9529 if (Pred == ICmpInst::ICMP_EQ)
9530 return ReplaceInstUsesWith(SI, FalseVal);
9531 // Transform (X != Y) ? X : Y -> X
9532 if (Pred == ICmpInst::ICMP_NE)
9533 return ReplaceInstUsesWith(SI, TrueVal);
9534 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9536 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9537 // Transform (X == Y) ? Y : X -> X
9538 if (Pred == ICmpInst::ICMP_EQ)
9539 return ReplaceInstUsesWith(SI, FalseVal);
9540 // Transform (X != Y) ? Y : X -> Y
9541 if (Pred == ICmpInst::ICMP_NE)
9542 return ReplaceInstUsesWith(SI, TrueVal);
9543 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9545 return Changed ? &SI : 0;
9549 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a
9550 /// PHI node (but the two may be in different blocks). See if the true/false
9551 /// values (V) are live in all of the predecessor blocks of the PHI. For
9552 /// example, cases like this cannot be mapped:
9554 /// X = phi [ C1, BB1], [C2, BB2]
9556 /// Z = select X, Y, 0
9558 /// because Y is not live in BB1/BB2.
9560 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V,
9561 const SelectInst &SI) {
9562 // If the value is a non-instruction value like a constant or argument, it
9563 // can always be mapped.
9564 const Instruction *I = dyn_cast<Instruction>(V);
9565 if (I == 0) return true;
9567 // If V is a PHI node defined in the same block as the condition PHI, we can
9568 // map the arguments.
9569 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
9571 if (const PHINode *VP = dyn_cast<PHINode>(I))
9572 if (VP->getParent() == CondPHI->getParent())
9575 // Otherwise, if the PHI and select are defined in the same block and if V is
9576 // defined in a different block, then we can transform it.
9577 if (SI.getParent() == CondPHI->getParent() &&
9578 I->getParent() != CondPHI->getParent())
9581 // Otherwise we have a 'hard' case and we can't tell without doing more
9582 // detailed dominator based analysis, punt.
9586 /// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form:
9587 /// SPF2(SPF1(A, B), C)
9588 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner,
9589 SelectPatternFlavor SPF1,
9592 SelectPatternFlavor SPF2, Value *C) {
9593 if (C == A || C == B) {
9594 // MAX(MAX(A, B), B) -> MAX(A, B)
9595 // MIN(MIN(a, b), a) -> MIN(a, b)
9597 return ReplaceInstUsesWith(Outer, Inner);
9599 // MAX(MIN(a, b), a) -> a
9600 // MIN(MAX(a, b), a) -> a
9601 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
9602 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
9603 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
9604 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
9605 return ReplaceInstUsesWith(Outer, C);
9608 // TODO: MIN(MIN(A, 23), 97)
9615 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9616 Value *CondVal = SI.getCondition();
9617 Value *TrueVal = SI.getTrueValue();
9618 Value *FalseVal = SI.getFalseValue();
9620 // select true, X, Y -> X
9621 // select false, X, Y -> Y
9622 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9623 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9625 // select C, X, X -> X
9626 if (TrueVal == FalseVal)
9627 return ReplaceInstUsesWith(SI, TrueVal);
9629 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9630 return ReplaceInstUsesWith(SI, FalseVal);
9631 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9632 return ReplaceInstUsesWith(SI, TrueVal);
9633 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9634 if (isa<Constant>(TrueVal))
9635 return ReplaceInstUsesWith(SI, TrueVal);
9637 return ReplaceInstUsesWith(SI, FalseVal);
9640 if (SI.getType() == Type::getInt1Ty(*Context)) {
9641 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9642 if (C->getZExtValue()) {
9643 // Change: A = select B, true, C --> A = or B, C
9644 return BinaryOperator::CreateOr(CondVal, FalseVal);
9646 // Change: A = select B, false, C --> A = and !B, C
9648 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9649 "not."+CondVal->getName()), SI);
9650 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9652 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9653 if (C->getZExtValue() == false) {
9654 // Change: A = select B, C, false --> A = and B, C
9655 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9657 // Change: A = select B, C, true --> A = or !B, C
9659 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9660 "not."+CondVal->getName()), SI);
9661 return BinaryOperator::CreateOr(NotCond, TrueVal);
9665 // select a, b, a -> a&b
9666 // select a, a, b -> a|b
9667 if (CondVal == TrueVal)
9668 return BinaryOperator::CreateOr(CondVal, FalseVal);
9669 else if (CondVal == FalseVal)
9670 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9673 // Selecting between two integer constants?
9674 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9675 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9676 // select C, 1, 0 -> zext C to int
9677 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9678 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9679 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9680 // select C, 0, 1 -> zext !C to int
9682 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9683 "not."+CondVal->getName()), SI);
9684 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9687 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9688 // If one of the constants is zero (we know they can't both be) and we
9689 // have an icmp instruction with zero, and we have an 'and' with the
9690 // non-constant value, eliminate this whole mess. This corresponds to
9691 // cases like this: ((X & 27) ? 27 : 0)
9692 if (TrueValC->isZero() || FalseValC->isZero())
9693 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9694 cast<Constant>(IC->getOperand(1))->isNullValue())
9695 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9696 if (ICA->getOpcode() == Instruction::And &&
9697 isa<ConstantInt>(ICA->getOperand(1)) &&
9698 (ICA->getOperand(1) == TrueValC ||
9699 ICA->getOperand(1) == FalseValC) &&
9700 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9701 // Okay, now we know that everything is set up, we just don't
9702 // know whether we have a icmp_ne or icmp_eq and whether the
9703 // true or false val is the zero.
9704 bool ShouldNotVal = !TrueValC->isZero();
9705 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9708 V = InsertNewInstBefore(BinaryOperator::Create(
9709 Instruction::Xor, V, ICA->getOperand(1)), SI);
9710 return ReplaceInstUsesWith(SI, V);
9715 // See if we are selecting two values based on a comparison of the two values.
9716 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9717 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9718 // Transform (X == Y) ? X : Y -> Y
9719 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9720 // This is not safe in general for floating point:
9721 // consider X== -0, Y== +0.
9722 // It becomes safe if either operand is a nonzero constant.
9723 ConstantFP *CFPt, *CFPf;
9724 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9725 !CFPt->getValueAPF().isZero()) ||
9726 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9727 !CFPf->getValueAPF().isZero()))
9728 return ReplaceInstUsesWith(SI, FalseVal);
9730 // Transform (X != Y) ? X : Y -> X
9731 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9732 return ReplaceInstUsesWith(SI, TrueVal);
9733 // NOTE: if we wanted to, this is where to detect MIN/MAX
9735 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9736 // Transform (X == Y) ? Y : X -> X
9737 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9738 // This is not safe in general for floating point:
9739 // consider X== -0, Y== +0.
9740 // It becomes safe if either operand is a nonzero constant.
9741 ConstantFP *CFPt, *CFPf;
9742 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9743 !CFPt->getValueAPF().isZero()) ||
9744 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9745 !CFPf->getValueAPF().isZero()))
9746 return ReplaceInstUsesWith(SI, FalseVal);
9748 // Transform (X != Y) ? Y : X -> Y
9749 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9750 return ReplaceInstUsesWith(SI, TrueVal);
9751 // NOTE: if we wanted to, this is where to detect MIN/MAX
9753 // NOTE: if we wanted to, this is where to detect ABS
9756 // See if we are selecting two values based on a comparison of the two values.
9757 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9758 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9761 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9762 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9763 if (TI->hasOneUse() && FI->hasOneUse()) {
9764 Instruction *AddOp = 0, *SubOp = 0;
9766 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9767 if (TI->getOpcode() == FI->getOpcode())
9768 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9771 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9772 // even legal for FP.
9773 if ((TI->getOpcode() == Instruction::Sub &&
9774 FI->getOpcode() == Instruction::Add) ||
9775 (TI->getOpcode() == Instruction::FSub &&
9776 FI->getOpcode() == Instruction::FAdd)) {
9777 AddOp = FI; SubOp = TI;
9778 } else if ((FI->getOpcode() == Instruction::Sub &&
9779 TI->getOpcode() == Instruction::Add) ||
9780 (FI->getOpcode() == Instruction::FSub &&
9781 TI->getOpcode() == Instruction::FAdd)) {
9782 AddOp = TI; SubOp = FI;
9786 Value *OtherAddOp = 0;
9787 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9788 OtherAddOp = AddOp->getOperand(1);
9789 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9790 OtherAddOp = AddOp->getOperand(0);
9794 // So at this point we know we have (Y -> OtherAddOp):
9795 // select C, (add X, Y), (sub X, Z)
9796 Value *NegVal; // Compute -Z
9797 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9798 NegVal = ConstantExpr::getNeg(C);
9800 NegVal = InsertNewInstBefore(
9801 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9805 Value *NewTrueOp = OtherAddOp;
9806 Value *NewFalseOp = NegVal;
9808 std::swap(NewTrueOp, NewFalseOp);
9809 Instruction *NewSel =
9810 SelectInst::Create(CondVal, NewTrueOp,
9811 NewFalseOp, SI.getName() + ".p");
9813 NewSel = InsertNewInstBefore(NewSel, SI);
9814 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9819 // See if we can fold the select into one of our operands.
9820 if (SI.getType()->isInteger()) {
9821 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal))
9824 // MAX(MAX(a, b), a) -> MAX(a, b)
9825 // MIN(MIN(a, b), a) -> MIN(a, b)
9826 // MAX(MIN(a, b), a) -> a
9827 // MIN(MAX(a, b), a) -> a
9828 Value *LHS, *RHS, *LHS2, *RHS2;
9829 if (SelectPatternFlavor SPF = MatchSelectPattern(&SI, LHS, RHS)) {
9830 if (SelectPatternFlavor SPF2 = MatchSelectPattern(LHS, LHS2, RHS2))
9831 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2,
9834 if (SelectPatternFlavor SPF2 = MatchSelectPattern(RHS, LHS2, RHS2))
9835 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2,
9841 // ABS(-X) -> ABS(X)
9842 // ABS(ABS(X)) -> ABS(X)
9845 // See if we can fold the select into a phi node if the condition is a select.
9846 if (isa<PHINode>(SI.getCondition()))
9847 // The true/false values have to be live in the PHI predecessor's blocks.
9848 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
9849 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
9850 if (Instruction *NV = FoldOpIntoPhi(SI))
9853 if (BinaryOperator::isNot(CondVal)) {
9854 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9855 SI.setOperand(1, FalseVal);
9856 SI.setOperand(2, TrueVal);
9863 /// EnforceKnownAlignment - If the specified pointer points to an object that
9864 /// we control, modify the object's alignment to PrefAlign. This isn't
9865 /// often possible though. If alignment is important, a more reliable approach
9866 /// is to simply align all global variables and allocation instructions to
9867 /// their preferred alignment from the beginning.
9869 static unsigned EnforceKnownAlignment(Value *V,
9870 unsigned Align, unsigned PrefAlign) {
9872 User *U = dyn_cast<User>(V);
9873 if (!U) return Align;
9875 switch (Operator::getOpcode(U)) {
9877 case Instruction::BitCast:
9878 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9879 case Instruction::GetElementPtr: {
9880 // If all indexes are zero, it is just the alignment of the base pointer.
9881 bool AllZeroOperands = true;
9882 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9883 if (!isa<Constant>(*i) ||
9884 !cast<Constant>(*i)->isNullValue()) {
9885 AllZeroOperands = false;
9889 if (AllZeroOperands) {
9890 // Treat this like a bitcast.
9891 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9897 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9898 // If there is a large requested alignment and we can, bump up the alignment
9900 if (!GV->isDeclaration()) {
9901 if (GV->getAlignment() >= PrefAlign)
9902 Align = GV->getAlignment();
9904 GV->setAlignment(PrefAlign);
9908 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
9909 // If there is a requested alignment and if this is an alloca, round up.
9910 if (AI->getAlignment() >= PrefAlign)
9911 Align = AI->getAlignment();
9913 AI->setAlignment(PrefAlign);
9921 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9922 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9923 /// and it is more than the alignment of the ultimate object, see if we can
9924 /// increase the alignment of the ultimate object, making this check succeed.
9925 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9926 unsigned PrefAlign) {
9927 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9928 sizeof(PrefAlign) * CHAR_BIT;
9929 APInt Mask = APInt::getAllOnesValue(BitWidth);
9930 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9931 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9932 unsigned TrailZ = KnownZero.countTrailingOnes();
9933 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9935 if (PrefAlign > Align)
9936 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9938 // We don't need to make any adjustment.
9942 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9943 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9944 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9945 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9946 unsigned CopyAlign = MI->getAlignment();
9948 if (CopyAlign < MinAlign) {
9949 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9954 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9956 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9957 if (MemOpLength == 0) return 0;
9959 // Source and destination pointer types are always "i8*" for intrinsic. See
9960 // if the size is something we can handle with a single primitive load/store.
9961 // A single load+store correctly handles overlapping memory in the memmove
9963 unsigned Size = MemOpLength->getZExtValue();
9964 if (Size == 0) return MI; // Delete this mem transfer.
9966 if (Size > 8 || (Size&(Size-1)))
9967 return 0; // If not 1/2/4/8 bytes, exit.
9969 // Use an integer load+store unless we can find something better.
9971 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
9973 // Memcpy forces the use of i8* for the source and destination. That means
9974 // that if you're using memcpy to move one double around, you'll get a cast
9975 // from double* to i8*. We'd much rather use a double load+store rather than
9976 // an i64 load+store, here because this improves the odds that the source or
9977 // dest address will be promotable. See if we can find a better type than the
9978 // integer datatype.
9979 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9980 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9981 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9982 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9983 // down through these levels if so.
9984 while (!SrcETy->isSingleValueType()) {
9985 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9986 if (STy->getNumElements() == 1)
9987 SrcETy = STy->getElementType(0);
9990 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9991 if (ATy->getNumElements() == 1)
9992 SrcETy = ATy->getElementType();
9999 if (SrcETy->isSingleValueType())
10000 NewPtrTy = PointerType::getUnqual(SrcETy);
10005 // If the memcpy/memmove provides better alignment info than we can
10007 SrcAlign = std::max(SrcAlign, CopyAlign);
10008 DstAlign = std::max(DstAlign, CopyAlign);
10010 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy);
10011 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy);
10012 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
10013 InsertNewInstBefore(L, *MI);
10014 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
10016 // Set the size of the copy to 0, it will be deleted on the next iteration.
10017 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
10021 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
10022 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
10023 if (MI->getAlignment() < Alignment) {
10024 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
10025 Alignment, false));
10029 // Extract the length and alignment and fill if they are constant.
10030 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
10031 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
10032 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
10034 uint64_t Len = LenC->getZExtValue();
10035 Alignment = MI->getAlignment();
10037 // If the length is zero, this is a no-op
10038 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
10040 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
10041 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
10042 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
10044 Value *Dest = MI->getDest();
10045 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
10047 // Alignment 0 is identity for alignment 1 for memset, but not store.
10048 if (Alignment == 0) Alignment = 1;
10050 // Extract the fill value and store.
10051 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
10052 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
10053 Dest, false, Alignment), *MI);
10055 // Set the size of the copy to 0, it will be deleted on the next iteration.
10056 MI->setLength(Constant::getNullValue(LenC->getType()));
10064 /// visitCallInst - CallInst simplification. This mostly only handles folding
10065 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
10066 /// the heavy lifting.
10068 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
10069 if (isFreeCall(&CI))
10070 return visitFree(CI);
10072 // If the caller function is nounwind, mark the call as nounwind, even if the
10074 if (CI.getParent()->getParent()->doesNotThrow() &&
10075 !CI.doesNotThrow()) {
10076 CI.setDoesNotThrow();
10080 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
10081 if (!II) return visitCallSite(&CI);
10083 // Intrinsics cannot occur in an invoke, so handle them here instead of in
10085 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
10086 bool Changed = false;
10088 // memmove/cpy/set of zero bytes is a noop.
10089 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
10090 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
10092 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
10093 if (CI->getZExtValue() == 1) {
10094 // Replace the instruction with just byte operations. We would
10095 // transform other cases to loads/stores, but we don't know if
10096 // alignment is sufficient.
10100 // If we have a memmove and the source operation is a constant global,
10101 // then the source and dest pointers can't alias, so we can change this
10102 // into a call to memcpy.
10103 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
10104 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
10105 if (GVSrc->isConstant()) {
10106 Module *M = CI.getParent()->getParent()->getParent();
10107 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
10108 const Type *Tys[1];
10109 Tys[0] = CI.getOperand(3)->getType();
10111 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
10116 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
10117 // memmove(x,x,size) -> noop.
10118 if (MTI->getSource() == MTI->getDest())
10119 return EraseInstFromFunction(CI);
10122 // If we can determine a pointer alignment that is bigger than currently
10123 // set, update the alignment.
10124 if (isa<MemTransferInst>(MI)) {
10125 if (Instruction *I = SimplifyMemTransfer(MI))
10127 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
10128 if (Instruction *I = SimplifyMemSet(MSI))
10132 if (Changed) return II;
10135 switch (II->getIntrinsicID()) {
10137 case Intrinsic::bswap:
10138 // bswap(bswap(x)) -> x
10139 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
10140 if (Operand->getIntrinsicID() == Intrinsic::bswap)
10141 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
10143 case Intrinsic::uadd_with_overflow: {
10144 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
10145 const IntegerType *IT = cast<IntegerType>(II->getOperand(1)->getType());
10146 uint32_t BitWidth = IT->getBitWidth();
10147 APInt Mask = APInt::getSignBit(BitWidth);
10148 APInt LHSKnownZero(BitWidth, 0);
10149 APInt LHSKnownOne(BitWidth, 0);
10150 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
10151 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
10152 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
10154 if (LHSKnownNegative || LHSKnownPositive) {
10155 APInt RHSKnownZero(BitWidth, 0);
10156 APInt RHSKnownOne(BitWidth, 0);
10157 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
10158 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
10159 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
10160 if (LHSKnownNegative && RHSKnownNegative) {
10161 // The sign bit is set in both cases: this MUST overflow.
10162 // Create a simple add instruction, and insert it into the struct.
10163 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
10166 UndefValue::get(LHS->getType()), ConstantInt::getTrue(*Context)
10168 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10169 return InsertValueInst::Create(Struct, Add, 0);
10172 if (LHSKnownPositive && RHSKnownPositive) {
10173 // The sign bit is clear in both cases: this CANNOT overflow.
10174 // Create a simple add instruction, and insert it into the struct.
10175 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
10178 UndefValue::get(LHS->getType()), ConstantInt::getFalse(*Context)
10180 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10181 return InsertValueInst::Create(Struct, Add, 0);
10185 // FALL THROUGH uadd into sadd
10186 case Intrinsic::sadd_with_overflow:
10187 // Canonicalize constants into the RHS.
10188 if (isa<Constant>(II->getOperand(1)) &&
10189 !isa<Constant>(II->getOperand(2))) {
10190 Value *LHS = II->getOperand(1);
10191 II->setOperand(1, II->getOperand(2));
10192 II->setOperand(2, LHS);
10196 // X + undef -> undef
10197 if (isa<UndefValue>(II->getOperand(2)))
10198 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
10200 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) {
10201 // X + 0 -> {X, false}
10202 if (RHS->isZero()) {
10204 UndefValue::get(II->getOperand(0)->getType()),
10205 ConstantInt::getFalse(*Context)
10207 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10208 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
10212 case Intrinsic::usub_with_overflow:
10213 case Intrinsic::ssub_with_overflow:
10214 // undef - X -> undef
10215 // X - undef -> undef
10216 if (isa<UndefValue>(II->getOperand(1)) ||
10217 isa<UndefValue>(II->getOperand(2)))
10218 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
10220 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) {
10221 // X - 0 -> {X, false}
10222 if (RHS->isZero()) {
10224 UndefValue::get(II->getOperand(1)->getType()),
10225 ConstantInt::getFalse(*Context)
10227 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10228 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
10232 case Intrinsic::umul_with_overflow:
10233 case Intrinsic::smul_with_overflow:
10234 // Canonicalize constants into the RHS.
10235 if (isa<Constant>(II->getOperand(1)) &&
10236 !isa<Constant>(II->getOperand(2))) {
10237 Value *LHS = II->getOperand(1);
10238 II->setOperand(1, II->getOperand(2));
10239 II->setOperand(2, LHS);
10243 // X * undef -> undef
10244 if (isa<UndefValue>(II->getOperand(2)))
10245 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
10247 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getOperand(2))) {
10248 // X*0 -> {0, false}
10249 if (RHSI->isZero())
10250 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
10252 // X * 1 -> {X, false}
10253 if (RHSI->equalsInt(1)) {
10255 UndefValue::get(II->getOperand(1)->getType()),
10256 ConstantInt::getFalse(*Context)
10258 Constant *Struct = ConstantStruct::get(*Context, V, 2, false);
10259 return InsertValueInst::Create(Struct, II->getOperand(1), 0);
10263 case Intrinsic::ppc_altivec_lvx:
10264 case Intrinsic::ppc_altivec_lvxl:
10265 case Intrinsic::x86_sse_loadu_ps:
10266 case Intrinsic::x86_sse2_loadu_pd:
10267 case Intrinsic::x86_sse2_loadu_dq:
10268 // Turn PPC lvx -> load if the pointer is known aligned.
10269 // Turn X86 loadups -> load if the pointer is known aligned.
10270 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
10271 Value *Ptr = Builder->CreateBitCast(II->getOperand(1),
10272 PointerType::getUnqual(II->getType()));
10273 return new LoadInst(Ptr);
10276 case Intrinsic::ppc_altivec_stvx:
10277 case Intrinsic::ppc_altivec_stvxl:
10278 // Turn stvx -> store if the pointer is known aligned.
10279 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
10280 const Type *OpPtrTy =
10281 PointerType::getUnqual(II->getOperand(1)->getType());
10282 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy);
10283 return new StoreInst(II->getOperand(1), Ptr);
10286 case Intrinsic::x86_sse_storeu_ps:
10287 case Intrinsic::x86_sse2_storeu_pd:
10288 case Intrinsic::x86_sse2_storeu_dq:
10289 // Turn X86 storeu -> store if the pointer is known aligned.
10290 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
10291 const Type *OpPtrTy =
10292 PointerType::getUnqual(II->getOperand(2)->getType());
10293 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
10294 return new StoreInst(II->getOperand(2), Ptr);
10298 case Intrinsic::x86_sse_cvttss2si: {
10299 // These intrinsics only demands the 0th element of its input vector. If
10300 // we can simplify the input based on that, do so now.
10302 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
10303 APInt DemandedElts(VWidth, 1);
10304 APInt UndefElts(VWidth, 0);
10305 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
10307 II->setOperand(1, V);
10313 case Intrinsic::ppc_altivec_vperm:
10314 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
10315 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
10316 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
10318 // Check that all of the elements are integer constants or undefs.
10319 bool AllEltsOk = true;
10320 for (unsigned i = 0; i != 16; ++i) {
10321 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
10322 !isa<UndefValue>(Mask->getOperand(i))) {
10329 // Cast the input vectors to byte vectors.
10330 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
10331 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType());
10332 Value *Result = UndefValue::get(Op0->getType());
10334 // Only extract each element once.
10335 Value *ExtractedElts[32];
10336 memset(ExtractedElts, 0, sizeof(ExtractedElts));
10338 for (unsigned i = 0; i != 16; ++i) {
10339 if (isa<UndefValue>(Mask->getOperand(i)))
10341 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
10342 Idx &= 31; // Match the hardware behavior.
10344 if (ExtractedElts[Idx] == 0) {
10345 ExtractedElts[Idx] =
10346 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
10347 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false),
10351 // Insert this value into the result vector.
10352 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
10353 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
10356 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
10361 case Intrinsic::stackrestore: {
10362 // If the save is right next to the restore, remove the restore. This can
10363 // happen when variable allocas are DCE'd.
10364 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
10365 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
10366 BasicBlock::iterator BI = SS;
10368 return EraseInstFromFunction(CI);
10372 // Scan down this block to see if there is another stack restore in the
10373 // same block without an intervening call/alloca.
10374 BasicBlock::iterator BI = II;
10375 TerminatorInst *TI = II->getParent()->getTerminator();
10376 bool CannotRemove = false;
10377 for (++BI; &*BI != TI; ++BI) {
10378 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
10379 CannotRemove = true;
10382 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
10383 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
10384 // If there is a stackrestore below this one, remove this one.
10385 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10386 return EraseInstFromFunction(CI);
10387 // Otherwise, ignore the intrinsic.
10389 // If we found a non-intrinsic call, we can't remove the stack
10391 CannotRemove = true;
10397 // If the stack restore is in a return/unwind block and if there are no
10398 // allocas or calls between the restore and the return, nuke the restore.
10399 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10400 return EraseInstFromFunction(CI);
10405 return visitCallSite(II);
10408 // InvokeInst simplification
10410 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10411 return visitCallSite(&II);
10414 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10415 /// passed through the varargs area, we can eliminate the use of the cast.
10416 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10417 const CastInst * const CI,
10418 const TargetData * const TD,
10420 if (!CI->isLosslessCast())
10423 // The size of ByVal arguments is derived from the type, so we
10424 // can't change to a type with a different size. If the size were
10425 // passed explicitly we could avoid this check.
10426 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10429 const Type* SrcTy =
10430 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10431 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10432 if (!SrcTy->isSized() || !DstTy->isSized())
10434 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10439 // visitCallSite - Improvements for call and invoke instructions.
10441 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10442 bool Changed = false;
10444 // If the callee is a constexpr cast of a function, attempt to move the cast
10445 // to the arguments of the call/invoke.
10446 if (transformConstExprCastCall(CS)) return 0;
10448 Value *Callee = CS.getCalledValue();
10450 if (Function *CalleeF = dyn_cast<Function>(Callee))
10451 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10452 Instruction *OldCall = CS.getInstruction();
10453 // If the call and callee calling conventions don't match, this call must
10454 // be unreachable, as the call is undefined.
10455 new StoreInst(ConstantInt::getTrue(*Context),
10456 UndefValue::get(Type::getInt1PtrTy(*Context)),
10458 // If OldCall dues not return void then replaceAllUsesWith undef.
10459 // This allows ValueHandlers and custom metadata to adjust itself.
10460 if (!OldCall->getType()->isVoidTy())
10461 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10462 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10463 return EraseInstFromFunction(*OldCall);
10467 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10468 // This instruction is not reachable, just remove it. We insert a store to
10469 // undef so that we know that this code is not reachable, despite the fact
10470 // that we can't modify the CFG here.
10471 new StoreInst(ConstantInt::getTrue(*Context),
10472 UndefValue::get(Type::getInt1PtrTy(*Context)),
10473 CS.getInstruction());
10475 // If CS dues not return void then replaceAllUsesWith undef.
10476 // This allows ValueHandlers and custom metadata to adjust itself.
10477 if (!CS.getInstruction()->getType()->isVoidTy())
10478 CS.getInstruction()->
10479 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10481 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10482 // Don't break the CFG, insert a dummy cond branch.
10483 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10484 ConstantInt::getTrue(*Context), II);
10486 return EraseInstFromFunction(*CS.getInstruction());
10489 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10490 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10491 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10492 return transformCallThroughTrampoline(CS);
10494 const PointerType *PTy = cast<PointerType>(Callee->getType());
10495 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10496 if (FTy->isVarArg()) {
10497 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10498 // See if we can optimize any arguments passed through the varargs area of
10500 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10501 E = CS.arg_end(); I != E; ++I, ++ix) {
10502 CastInst *CI = dyn_cast<CastInst>(*I);
10503 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10504 *I = CI->getOperand(0);
10510 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10511 // Inline asm calls cannot throw - mark them 'nounwind'.
10512 CS.setDoesNotThrow();
10516 return Changed ? CS.getInstruction() : 0;
10519 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10520 // attempt to move the cast to the arguments of the call/invoke.
10522 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10523 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10524 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10525 if (CE->getOpcode() != Instruction::BitCast ||
10526 !isa<Function>(CE->getOperand(0)))
10528 Function *Callee = cast<Function>(CE->getOperand(0));
10529 Instruction *Caller = CS.getInstruction();
10530 const AttrListPtr &CallerPAL = CS.getAttributes();
10532 // Okay, this is a cast from a function to a different type. Unless doing so
10533 // would cause a type conversion of one of our arguments, change this call to
10534 // be a direct call with arguments casted to the appropriate types.
10536 const FunctionType *FT = Callee->getFunctionType();
10537 const Type *OldRetTy = Caller->getType();
10538 const Type *NewRetTy = FT->getReturnType();
10540 if (isa<StructType>(NewRetTy))
10541 return false; // TODO: Handle multiple return values.
10543 // Check to see if we are changing the return type...
10544 if (OldRetTy != NewRetTy) {
10545 if (Callee->isDeclaration() &&
10546 // Conversion is ok if changing from one pointer type to another or from
10547 // a pointer to an integer of the same size.
10548 !((isa<PointerType>(OldRetTy) || !TD ||
10549 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
10550 (isa<PointerType>(NewRetTy) || !TD ||
10551 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
10552 return false; // Cannot transform this return value.
10554 if (!Caller->use_empty() &&
10555 // void -> non-void is handled specially
10556 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
10557 return false; // Cannot transform this return value.
10559 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10560 Attributes RAttrs = CallerPAL.getRetAttributes();
10561 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10562 return false; // Attribute not compatible with transformed value.
10565 // If the callsite is an invoke instruction, and the return value is used by
10566 // a PHI node in a successor, we cannot change the return type of the call
10567 // because there is no place to put the cast instruction (without breaking
10568 // the critical edge). Bail out in this case.
10569 if (!Caller->use_empty())
10570 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10571 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10573 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10574 if (PN->getParent() == II->getNormalDest() ||
10575 PN->getParent() == II->getUnwindDest())
10579 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10580 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10582 CallSite::arg_iterator AI = CS.arg_begin();
10583 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10584 const Type *ParamTy = FT->getParamType(i);
10585 const Type *ActTy = (*AI)->getType();
10587 if (!CastInst::isCastable(ActTy, ParamTy))
10588 return false; // Cannot transform this parameter value.
10590 if (CallerPAL.getParamAttributes(i + 1)
10591 & Attribute::typeIncompatible(ParamTy))
10592 return false; // Attribute not compatible with transformed value.
10594 // Converting from one pointer type to another or between a pointer and an
10595 // integer of the same size is safe even if we do not have a body.
10596 bool isConvertible = ActTy == ParamTy ||
10597 (TD && ((isa<PointerType>(ParamTy) ||
10598 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10599 (isa<PointerType>(ActTy) ||
10600 ActTy == TD->getIntPtrType(Caller->getContext()))));
10601 if (Callee->isDeclaration() && !isConvertible) return false;
10604 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10605 Callee->isDeclaration())
10606 return false; // Do not delete arguments unless we have a function body.
10608 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10609 !CallerPAL.isEmpty())
10610 // In this case we have more arguments than the new function type, but we
10611 // won't be dropping them. Check that these extra arguments have attributes
10612 // that are compatible with being a vararg call argument.
10613 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10614 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10616 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10617 if (PAttrs & Attribute::VarArgsIncompatible)
10621 // Okay, we decided that this is a safe thing to do: go ahead and start
10622 // inserting cast instructions as necessary...
10623 std::vector<Value*> Args;
10624 Args.reserve(NumActualArgs);
10625 SmallVector<AttributeWithIndex, 8> attrVec;
10626 attrVec.reserve(NumCommonArgs);
10628 // Get any return attributes.
10629 Attributes RAttrs = CallerPAL.getRetAttributes();
10631 // If the return value is not being used, the type may not be compatible
10632 // with the existing attributes. Wipe out any problematic attributes.
10633 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10635 // Add the new return attributes.
10637 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10639 AI = CS.arg_begin();
10640 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10641 const Type *ParamTy = FT->getParamType(i);
10642 if ((*AI)->getType() == ParamTy) {
10643 Args.push_back(*AI);
10645 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10646 false, ParamTy, false);
10647 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
10650 // Add any parameter attributes.
10651 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10652 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10655 // If the function takes more arguments than the call was taking, add them
10657 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10658 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10660 // If we are removing arguments to the function, emit an obnoxious warning.
10661 if (FT->getNumParams() < NumActualArgs) {
10662 if (!FT->isVarArg()) {
10663 errs() << "WARNING: While resolving call to function '"
10664 << Callee->getName() << "' arguments were dropped!\n";
10666 // Add all of the arguments in their promoted form to the arg list.
10667 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10668 const Type *PTy = getPromotedType((*AI)->getType());
10669 if (PTy != (*AI)->getType()) {
10670 // Must promote to pass through va_arg area!
10671 Instruction::CastOps opcode =
10672 CastInst::getCastOpcode(*AI, false, PTy, false);
10673 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
10675 Args.push_back(*AI);
10678 // Add any parameter attributes.
10679 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10680 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10685 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10686 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10688 if (NewRetTy->isVoidTy())
10689 Caller->setName(""); // Void type should not have a name.
10691 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10695 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10696 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10697 Args.begin(), Args.end(),
10698 Caller->getName(), Caller);
10699 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10700 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10702 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10703 Caller->getName(), Caller);
10704 CallInst *CI = cast<CallInst>(Caller);
10705 if (CI->isTailCall())
10706 cast<CallInst>(NC)->setTailCall();
10707 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10708 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10711 // Insert a cast of the return type as necessary.
10713 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10714 if (!NV->getType()->isVoidTy()) {
10715 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10717 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10719 // If this is an invoke instruction, we should insert it after the first
10720 // non-phi, instruction in the normal successor block.
10721 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10722 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10723 InsertNewInstBefore(NC, *I);
10725 // Otherwise, it's a call, just insert cast right after the call instr
10726 InsertNewInstBefore(NC, *Caller);
10728 Worklist.AddUsersToWorkList(*Caller);
10730 NV = UndefValue::get(Caller->getType());
10735 if (!Caller->use_empty())
10736 Caller->replaceAllUsesWith(NV);
10738 EraseInstFromFunction(*Caller);
10742 // transformCallThroughTrampoline - Turn a call to a function created by the
10743 // init_trampoline intrinsic into a direct call to the underlying function.
10745 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10746 Value *Callee = CS.getCalledValue();
10747 const PointerType *PTy = cast<PointerType>(Callee->getType());
10748 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10749 const AttrListPtr &Attrs = CS.getAttributes();
10751 // If the call already has the 'nest' attribute somewhere then give up -
10752 // otherwise 'nest' would occur twice after splicing in the chain.
10753 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10756 IntrinsicInst *Tramp =
10757 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10759 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10760 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10761 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10763 const AttrListPtr &NestAttrs = NestF->getAttributes();
10764 if (!NestAttrs.isEmpty()) {
10765 unsigned NestIdx = 1;
10766 const Type *NestTy = 0;
10767 Attributes NestAttr = Attribute::None;
10769 // Look for a parameter marked with the 'nest' attribute.
10770 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10771 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10772 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10773 // Record the parameter type and any other attributes.
10775 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10780 Instruction *Caller = CS.getInstruction();
10781 std::vector<Value*> NewArgs;
10782 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10784 SmallVector<AttributeWithIndex, 8> NewAttrs;
10785 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10787 // Insert the nest argument into the call argument list, which may
10788 // mean appending it. Likewise for attributes.
10790 // Add any result attributes.
10791 if (Attributes Attr = Attrs.getRetAttributes())
10792 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10796 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10798 if (Idx == NestIdx) {
10799 // Add the chain argument and attributes.
10800 Value *NestVal = Tramp->getOperand(3);
10801 if (NestVal->getType() != NestTy)
10802 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10803 NewArgs.push_back(NestVal);
10804 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10810 // Add the original argument and attributes.
10811 NewArgs.push_back(*I);
10812 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10814 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10820 // Add any function attributes.
10821 if (Attributes Attr = Attrs.getFnAttributes())
10822 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10824 // The trampoline may have been bitcast to a bogus type (FTy).
10825 // Handle this by synthesizing a new function type, equal to FTy
10826 // with the chain parameter inserted.
10828 std::vector<const Type*> NewTypes;
10829 NewTypes.reserve(FTy->getNumParams()+1);
10831 // Insert the chain's type into the list of parameter types, which may
10832 // mean appending it.
10835 FunctionType::param_iterator I = FTy->param_begin(),
10836 E = FTy->param_end();
10839 if (Idx == NestIdx)
10840 // Add the chain's type.
10841 NewTypes.push_back(NestTy);
10846 // Add the original type.
10847 NewTypes.push_back(*I);
10853 // Replace the trampoline call with a direct call. Let the generic
10854 // code sort out any function type mismatches.
10855 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10857 Constant *NewCallee =
10858 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10859 NestF : ConstantExpr::getBitCast(NestF,
10860 PointerType::getUnqual(NewFTy));
10861 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10864 Instruction *NewCaller;
10865 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10866 NewCaller = InvokeInst::Create(NewCallee,
10867 II->getNormalDest(), II->getUnwindDest(),
10868 NewArgs.begin(), NewArgs.end(),
10869 Caller->getName(), Caller);
10870 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10871 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10873 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10874 Caller->getName(), Caller);
10875 if (cast<CallInst>(Caller)->isTailCall())
10876 cast<CallInst>(NewCaller)->setTailCall();
10877 cast<CallInst>(NewCaller)->
10878 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10879 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10881 if (!Caller->getType()->isVoidTy())
10882 Caller->replaceAllUsesWith(NewCaller);
10883 Caller->eraseFromParent();
10884 Worklist.Remove(Caller);
10889 // Replace the trampoline call with a direct call. Since there is no 'nest'
10890 // parameter, there is no need to adjust the argument list. Let the generic
10891 // code sort out any function type mismatches.
10892 Constant *NewCallee =
10893 NestF->getType() == PTy ? NestF :
10894 ConstantExpr::getBitCast(NestF, PTy);
10895 CS.setCalledFunction(NewCallee);
10896 return CS.getInstruction();
10899 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
10900 /// and if a/b/c and the add's all have a single use, turn this into a phi
10901 /// and a single binop.
10902 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10903 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10904 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10905 unsigned Opc = FirstInst->getOpcode();
10906 Value *LHSVal = FirstInst->getOperand(0);
10907 Value *RHSVal = FirstInst->getOperand(1);
10909 const Type *LHSType = LHSVal->getType();
10910 const Type *RHSType = RHSVal->getType();
10912 // Scan to see if all operands are the same opcode, and all have one use.
10913 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10914 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10915 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10916 // Verify type of the LHS matches so we don't fold cmp's of different
10917 // types or GEP's with different index types.
10918 I->getOperand(0)->getType() != LHSType ||
10919 I->getOperand(1)->getType() != RHSType)
10922 // If they are CmpInst instructions, check their predicates
10923 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10924 if (cast<CmpInst>(I)->getPredicate() !=
10925 cast<CmpInst>(FirstInst)->getPredicate())
10928 // Keep track of which operand needs a phi node.
10929 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10930 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10933 // If both LHS and RHS would need a PHI, don't do this transformation,
10934 // because it would increase the number of PHIs entering the block,
10935 // which leads to higher register pressure. This is especially
10936 // bad when the PHIs are in the header of a loop.
10937 if (!LHSVal && !RHSVal)
10940 // Otherwise, this is safe to transform!
10942 Value *InLHS = FirstInst->getOperand(0);
10943 Value *InRHS = FirstInst->getOperand(1);
10944 PHINode *NewLHS = 0, *NewRHS = 0;
10946 NewLHS = PHINode::Create(LHSType,
10947 FirstInst->getOperand(0)->getName() + ".pn");
10948 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10949 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10950 InsertNewInstBefore(NewLHS, PN);
10955 NewRHS = PHINode::Create(RHSType,
10956 FirstInst->getOperand(1)->getName() + ".pn");
10957 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10958 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10959 InsertNewInstBefore(NewRHS, PN);
10963 // Add all operands to the new PHIs.
10964 if (NewLHS || NewRHS) {
10965 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10966 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10968 Value *NewInLHS = InInst->getOperand(0);
10969 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10972 Value *NewInRHS = InInst->getOperand(1);
10973 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10978 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10979 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10980 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10981 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10985 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10986 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10988 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10989 FirstInst->op_end());
10990 // This is true if all GEP bases are allocas and if all indices into them are
10992 bool AllBasePointersAreAllocas = true;
10994 // We don't want to replace this phi if the replacement would require
10995 // more than one phi, which leads to higher register pressure. This is
10996 // especially bad when the PHIs are in the header of a loop.
10997 bool NeededPhi = false;
10999 // Scan to see if all operands are the same opcode, and all have one use.
11000 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
11001 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
11002 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
11003 GEP->getNumOperands() != FirstInst->getNumOperands())
11006 // Keep track of whether or not all GEPs are of alloca pointers.
11007 if (AllBasePointersAreAllocas &&
11008 (!isa<AllocaInst>(GEP->getOperand(0)) ||
11009 !GEP->hasAllConstantIndices()))
11010 AllBasePointersAreAllocas = false;
11012 // Compare the operand lists.
11013 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
11014 if (FirstInst->getOperand(op) == GEP->getOperand(op))
11017 // Don't merge two GEPs when two operands differ (introducing phi nodes)
11018 // if one of the PHIs has a constant for the index. The index may be
11019 // substantially cheaper to compute for the constants, so making it a
11020 // variable index could pessimize the path. This also handles the case
11021 // for struct indices, which must always be constant.
11022 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
11023 isa<ConstantInt>(GEP->getOperand(op)))
11026 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
11029 // If we already needed a PHI for an earlier operand, and another operand
11030 // also requires a PHI, we'd be introducing more PHIs than we're
11031 // eliminating, which increases register pressure on entry to the PHI's
11036 FixedOperands[op] = 0; // Needs a PHI.
11041 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
11042 // bother doing this transformation. At best, this will just save a bit of
11043 // offset calculation, but all the predecessors will have to materialize the
11044 // stack address into a register anyway. We'd actually rather *clone* the
11045 // load up into the predecessors so that we have a load of a gep of an alloca,
11046 // which can usually all be folded into the load.
11047 if (AllBasePointersAreAllocas)
11050 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
11051 // that is variable.
11052 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
11054 bool HasAnyPHIs = false;
11055 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
11056 if (FixedOperands[i]) continue; // operand doesn't need a phi.
11057 Value *FirstOp = FirstInst->getOperand(i);
11058 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
11059 FirstOp->getName()+".pn");
11060 InsertNewInstBefore(NewPN, PN);
11062 NewPN->reserveOperandSpace(e);
11063 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
11064 OperandPhis[i] = NewPN;
11065 FixedOperands[i] = NewPN;
11070 // Add all operands to the new PHIs.
11072 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11073 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
11074 BasicBlock *InBB = PN.getIncomingBlock(i);
11076 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
11077 if (PHINode *OpPhi = OperandPhis[op])
11078 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
11082 Value *Base = FixedOperands[0];
11083 return cast<GEPOperator>(FirstInst)->isInBounds() ?
11084 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
11085 FixedOperands.end()) :
11086 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
11087 FixedOperands.end());
11091 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
11092 /// sink the load out of the block that defines it. This means that it must be
11093 /// obvious the value of the load is not changed from the point of the load to
11094 /// the end of the block it is in.
11096 /// Finally, it is safe, but not profitable, to sink a load targetting a
11097 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
11099 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
11100 BasicBlock::iterator BBI = L, E = L->getParent()->end();
11102 for (++BBI; BBI != E; ++BBI)
11103 if (BBI->mayWriteToMemory())
11106 // Check for non-address taken alloca. If not address-taken already, it isn't
11107 // profitable to do this xform.
11108 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
11109 bool isAddressTaken = false;
11110 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
11112 if (isa<LoadInst>(UI)) continue;
11113 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
11114 // If storing TO the alloca, then the address isn't taken.
11115 if (SI->getOperand(1) == AI) continue;
11117 isAddressTaken = true;
11121 if (!isAddressTaken && AI->isStaticAlloca())
11125 // If this load is a load from a GEP with a constant offset from an alloca,
11126 // then we don't want to sink it. In its present form, it will be
11127 // load [constant stack offset]. Sinking it will cause us to have to
11128 // materialize the stack addresses in each predecessor in a register only to
11129 // do a shared load from register in the successor.
11130 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
11131 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
11132 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
11138 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
11139 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
11141 // When processing loads, we need to propagate two bits of information to the
11142 // sunk load: whether it is volatile, and what its alignment is. We currently
11143 // don't sink loads when some have their alignment specified and some don't.
11144 // visitLoadInst will propagate an alignment onto the load when TD is around,
11145 // and if TD isn't around, we can't handle the mixed case.
11146 bool isVolatile = FirstLI->isVolatile();
11147 unsigned LoadAlignment = FirstLI->getAlignment();
11149 // We can't sink the load if the loaded value could be modified between the
11150 // load and the PHI.
11151 if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
11152 !isSafeAndProfitableToSinkLoad(FirstLI))
11155 // If the PHI is of volatile loads and the load block has multiple
11156 // successors, sinking it would remove a load of the volatile value from
11157 // the path through the other successor.
11159 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
11162 // Check to see if all arguments are the same operation.
11163 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11164 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
11165 if (!LI || !LI->hasOneUse())
11168 // We can't sink the load if the loaded value could be modified between
11169 // the load and the PHI.
11170 if (LI->isVolatile() != isVolatile ||
11171 LI->getParent() != PN.getIncomingBlock(i) ||
11172 !isSafeAndProfitableToSinkLoad(LI))
11175 // If some of the loads have an alignment specified but not all of them,
11176 // we can't do the transformation.
11177 if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
11180 LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
11182 // If the PHI is of volatile loads and the load block has multiple
11183 // successors, sinking it would remove a load of the volatile value from
11184 // the path through the other successor.
11186 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
11190 // Okay, they are all the same operation. Create a new PHI node of the
11191 // correct type, and PHI together all of the LHS's of the instructions.
11192 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
11193 PN.getName()+".in");
11194 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
11196 Value *InVal = FirstLI->getOperand(0);
11197 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
11199 // Add all operands to the new PHI.
11200 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11201 Value *NewInVal = cast<LoadInst>(PN.getIncomingValue(i))->getOperand(0);
11202 if (NewInVal != InVal)
11204 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
11209 // The new PHI unions all of the same values together. This is really
11210 // common, so we handle it intelligently here for compile-time speed.
11214 InsertNewInstBefore(NewPN, PN);
11218 // If this was a volatile load that we are merging, make sure to loop through
11219 // and mark all the input loads as non-volatile. If we don't do this, we will
11220 // insert a new volatile load and the old ones will not be deletable.
11222 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
11223 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
11225 return new LoadInst(PhiVal, "", isVolatile, LoadAlignment);
11230 /// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
11231 /// operator and they all are only used by the PHI, PHI together their
11232 /// inputs, and do the operation once, to the result of the PHI.
11233 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
11234 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
11236 if (isa<GetElementPtrInst>(FirstInst))
11237 return FoldPHIArgGEPIntoPHI(PN);
11238 if (isa<LoadInst>(FirstInst))
11239 return FoldPHIArgLoadIntoPHI(PN);
11241 // Scan the instruction, looking for input operations that can be folded away.
11242 // If all input operands to the phi are the same instruction (e.g. a cast from
11243 // the same type or "+42") we can pull the operation through the PHI, reducing
11244 // code size and simplifying code.
11245 Constant *ConstantOp = 0;
11246 const Type *CastSrcTy = 0;
11248 if (isa<CastInst>(FirstInst)) {
11249 CastSrcTy = FirstInst->getOperand(0)->getType();
11251 // Be careful about transforming integer PHIs. We don't want to pessimize
11252 // the code by turning an i32 into an i1293.
11253 if (isa<IntegerType>(PN.getType()) && isa<IntegerType>(CastSrcTy)) {
11254 if (!ShouldChangeType(PN.getType(), CastSrcTy, TD))
11257 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
11258 // Can fold binop, compare or shift here if the RHS is a constant,
11259 // otherwise call FoldPHIArgBinOpIntoPHI.
11260 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
11261 if (ConstantOp == 0)
11262 return FoldPHIArgBinOpIntoPHI(PN);
11264 return 0; // Cannot fold this operation.
11267 // Check to see if all arguments are the same operation.
11268 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11269 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
11270 if (I == 0 || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
11273 if (I->getOperand(0)->getType() != CastSrcTy)
11274 return 0; // Cast operation must match.
11275 } else if (I->getOperand(1) != ConstantOp) {
11280 // Okay, they are all the same operation. Create a new PHI node of the
11281 // correct type, and PHI together all of the LHS's of the instructions.
11282 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
11283 PN.getName()+".in");
11284 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
11286 Value *InVal = FirstInst->getOperand(0);
11287 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
11289 // Add all operands to the new PHI.
11290 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
11291 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
11292 if (NewInVal != InVal)
11294 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
11299 // The new PHI unions all of the same values together. This is really
11300 // common, so we handle it intelligently here for compile-time speed.
11304 InsertNewInstBefore(NewPN, PN);
11308 // Insert and return the new operation.
11309 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst))
11310 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
11312 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
11313 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
11315 CmpInst *CIOp = cast<CmpInst>(FirstInst);
11316 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
11317 PhiVal, ConstantOp);
11320 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
11322 static bool DeadPHICycle(PHINode *PN,
11323 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
11324 if (PN->use_empty()) return true;
11325 if (!PN->hasOneUse()) return false;
11327 // Remember this node, and if we find the cycle, return.
11328 if (!PotentiallyDeadPHIs.insert(PN))
11331 // Don't scan crazily complex things.
11332 if (PotentiallyDeadPHIs.size() == 16)
11335 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
11336 return DeadPHICycle(PU, PotentiallyDeadPHIs);
11341 /// PHIsEqualValue - Return true if this phi node is always equal to
11342 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
11343 /// z = some value; x = phi (y, z); y = phi (x, z)
11344 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
11345 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
11346 // See if we already saw this PHI node.
11347 if (!ValueEqualPHIs.insert(PN))
11350 // Don't scan crazily complex things.
11351 if (ValueEqualPHIs.size() == 16)
11354 // Scan the operands to see if they are either phi nodes or are equal to
11356 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
11357 Value *Op = PN->getIncomingValue(i);
11358 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
11359 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
11361 } else if (Op != NonPhiInVal)
11370 struct PHIUsageRecord {
11371 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)
11372 unsigned Shift; // The amount shifted.
11373 Instruction *Inst; // The trunc instruction.
11375 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
11376 : PHIId(pn), Shift(Sh), Inst(User) {}
11378 bool operator<(const PHIUsageRecord &RHS) const {
11379 if (PHIId < RHS.PHIId) return true;
11380 if (PHIId > RHS.PHIId) return false;
11381 if (Shift < RHS.Shift) return true;
11382 if (Shift > RHS.Shift) return false;
11383 return Inst->getType()->getPrimitiveSizeInBits() <
11384 RHS.Inst->getType()->getPrimitiveSizeInBits();
11388 struct LoweredPHIRecord {
11389 PHINode *PN; // The PHI that was lowered.
11390 unsigned Shift; // The amount shifted.
11391 unsigned Width; // The width extracted.
11393 LoweredPHIRecord(PHINode *pn, unsigned Sh, const Type *Ty)
11394 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
11396 // Ctor form used by DenseMap.
11397 LoweredPHIRecord(PHINode *pn, unsigned Sh)
11398 : PN(pn), Shift(Sh), Width(0) {}
11404 struct DenseMapInfo<LoweredPHIRecord> {
11405 static inline LoweredPHIRecord getEmptyKey() {
11406 return LoweredPHIRecord(0, 0);
11408 static inline LoweredPHIRecord getTombstoneKey() {
11409 return LoweredPHIRecord(0, 1);
11411 static unsigned getHashValue(const LoweredPHIRecord &Val) {
11412 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
11415 static bool isEqual(const LoweredPHIRecord &LHS,
11416 const LoweredPHIRecord &RHS) {
11417 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
11418 LHS.Width == RHS.Width;
11422 struct isPodLike<LoweredPHIRecord> { static const bool value = true; };
11426 /// SliceUpIllegalIntegerPHI - This is an integer PHI and we know that it has an
11427 /// illegal type: see if it is only used by trunc or trunc(lshr) operations. If
11428 /// so, we split the PHI into the various pieces being extracted. This sort of
11429 /// thing is introduced when SROA promotes an aggregate to large integer values.
11431 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
11432 /// inttoptr. We should produce new PHIs in the right type.
11434 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
11435 // PHIUsers - Keep track of all of the truncated values extracted from a set
11436 // of PHIs, along with their offset. These are the things we want to rewrite.
11437 SmallVector<PHIUsageRecord, 16> PHIUsers;
11439 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
11440 // nodes which are extracted from. PHIsToSlice is a set we use to avoid
11441 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
11442 // check the uses of (to ensure they are all extracts).
11443 SmallVector<PHINode*, 8> PHIsToSlice;
11444 SmallPtrSet<PHINode*, 8> PHIsInspected;
11446 PHIsToSlice.push_back(&FirstPhi);
11447 PHIsInspected.insert(&FirstPhi);
11449 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
11450 PHINode *PN = PHIsToSlice[PHIId];
11452 // Scan the input list of the PHI. If any input is an invoke, and if the
11453 // input is defined in the predecessor, then we won't be split the critical
11454 // edge which is required to insert a truncate. Because of this, we have to
11456 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
11457 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
11458 if (II == 0) continue;
11459 if (II->getParent() != PN->getIncomingBlock(i))
11462 // If we have a phi, and if it's directly in the predecessor, then we have
11463 // a critical edge where we need to put the truncate. Since we can't
11464 // split the edge in instcombine, we have to bail out.
11469 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
11471 Instruction *User = cast<Instruction>(*UI);
11473 // If the user is a PHI, inspect its uses recursively.
11474 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
11475 if (PHIsInspected.insert(UserPN))
11476 PHIsToSlice.push_back(UserPN);
11480 // Truncates are always ok.
11481 if (isa<TruncInst>(User)) {
11482 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, User));
11486 // Otherwise it must be a lshr which can only be used by one trunc.
11487 if (User->getOpcode() != Instruction::LShr ||
11488 !User->hasOneUse() || !isa<TruncInst>(User->use_back()) ||
11489 !isa<ConstantInt>(User->getOperand(1)))
11492 unsigned Shift = cast<ConstantInt>(User->getOperand(1))->getZExtValue();
11493 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, User->use_back()));
11497 // If we have no users, they must be all self uses, just nuke the PHI.
11498 if (PHIUsers.empty())
11499 return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
11501 // If this phi node is transformable, create new PHIs for all the pieces
11502 // extracted out of it. First, sort the users by their offset and size.
11503 array_pod_sort(PHIUsers.begin(), PHIUsers.end());
11505 DEBUG(errs() << "SLICING UP PHI: " << FirstPhi << '\n';
11506 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
11507 errs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] <<'\n';
11510 // PredValues - This is a temporary used when rewriting PHI nodes. It is
11511 // hoisted out here to avoid construction/destruction thrashing.
11512 DenseMap<BasicBlock*, Value*> PredValues;
11514 // ExtractedVals - Each new PHI we introduce is saved here so we don't
11515 // introduce redundant PHIs.
11516 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
11518 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
11519 unsigned PHIId = PHIUsers[UserI].PHIId;
11520 PHINode *PN = PHIsToSlice[PHIId];
11521 unsigned Offset = PHIUsers[UserI].Shift;
11522 const Type *Ty = PHIUsers[UserI].Inst->getType();
11526 // If we've already lowered a user like this, reuse the previously lowered
11528 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == 0) {
11530 // Otherwise, Create the new PHI node for this user.
11531 EltPHI = PHINode::Create(Ty, PN->getName()+".off"+Twine(Offset), PN);
11532 assert(EltPHI->getType() != PN->getType() &&
11533 "Truncate didn't shrink phi?");
11535 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
11536 BasicBlock *Pred = PN->getIncomingBlock(i);
11537 Value *&PredVal = PredValues[Pred];
11539 // If we already have a value for this predecessor, reuse it.
11541 EltPHI->addIncoming(PredVal, Pred);
11545 // Handle the PHI self-reuse case.
11546 Value *InVal = PN->getIncomingValue(i);
11549 EltPHI->addIncoming(PredVal, Pred);
11553 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
11554 // If the incoming value was a PHI, and if it was one of the PHIs we
11555 // already rewrote it, just use the lowered value.
11556 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
11558 EltPHI->addIncoming(PredVal, Pred);
11563 // Otherwise, do an extract in the predecessor.
11564 Builder->SetInsertPoint(Pred, Pred->getTerminator());
11565 Value *Res = InVal;
11567 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
11568 Offset), "extract");
11569 Res = Builder->CreateTrunc(Res, Ty, "extract.t");
11571 EltPHI->addIncoming(Res, Pred);
11573 // If the incoming value was a PHI, and if it was one of the PHIs we are
11574 // rewriting, we will ultimately delete the code we inserted. This
11575 // means we need to revisit that PHI to make sure we extract out the
11577 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
11578 if (PHIsInspected.count(OldInVal)) {
11579 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
11580 OldInVal)-PHIsToSlice.begin();
11581 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
11582 cast<Instruction>(Res)));
11586 PredValues.clear();
11588 DEBUG(errs() << " Made element PHI for offset " << Offset << ": "
11589 << *EltPHI << '\n');
11590 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
11593 // Replace the use of this piece with the PHI node.
11594 ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
11597 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
11599 Value *Undef = UndefValue::get(FirstPhi.getType());
11600 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
11601 ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
11602 return ReplaceInstUsesWith(FirstPhi, Undef);
11605 // PHINode simplification
11607 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
11608 // If LCSSA is around, don't mess with Phi nodes
11609 if (MustPreserveLCSSA) return 0;
11611 if (Value *V = PN.hasConstantValue())
11612 return ReplaceInstUsesWith(PN, V);
11614 // If all PHI operands are the same operation, pull them through the PHI,
11615 // reducing code size.
11616 if (isa<Instruction>(PN.getIncomingValue(0)) &&
11617 isa<Instruction>(PN.getIncomingValue(1)) &&
11618 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
11619 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
11620 // FIXME: The hasOneUse check will fail for PHIs that use the value more
11621 // than themselves more than once.
11622 PN.getIncomingValue(0)->hasOneUse())
11623 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
11626 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
11627 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
11628 // PHI)... break the cycle.
11629 if (PN.hasOneUse()) {
11630 Instruction *PHIUser = cast<Instruction>(PN.use_back());
11631 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
11632 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
11633 PotentiallyDeadPHIs.insert(&PN);
11634 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
11635 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
11638 // If this phi has a single use, and if that use just computes a value for
11639 // the next iteration of a loop, delete the phi. This occurs with unused
11640 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
11641 // common case here is good because the only other things that catch this
11642 // are induction variable analysis (sometimes) and ADCE, which is only run
11644 if (PHIUser->hasOneUse() &&
11645 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
11646 PHIUser->use_back() == &PN) {
11647 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
11651 // We sometimes end up with phi cycles that non-obviously end up being the
11652 // same value, for example:
11653 // z = some value; x = phi (y, z); y = phi (x, z)
11654 // where the phi nodes don't necessarily need to be in the same block. Do a
11655 // quick check to see if the PHI node only contains a single non-phi value, if
11656 // so, scan to see if the phi cycle is actually equal to that value.
11658 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
11659 // Scan for the first non-phi operand.
11660 while (InValNo != NumOperandVals &&
11661 isa<PHINode>(PN.getIncomingValue(InValNo)))
11664 if (InValNo != NumOperandVals) {
11665 Value *NonPhiInVal = PN.getOperand(InValNo);
11667 // Scan the rest of the operands to see if there are any conflicts, if so
11668 // there is no need to recursively scan other phis.
11669 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
11670 Value *OpVal = PN.getIncomingValue(InValNo);
11671 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
11675 // If we scanned over all operands, then we have one unique value plus
11676 // phi values. Scan PHI nodes to see if they all merge in each other or
11678 if (InValNo == NumOperandVals) {
11679 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
11680 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
11681 return ReplaceInstUsesWith(PN, NonPhiInVal);
11686 // If there are multiple PHIs, sort their operands so that they all list
11687 // the blocks in the same order. This will help identical PHIs be eliminated
11688 // by other passes. Other passes shouldn't depend on this for correctness
11690 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
11691 if (&PN != FirstPN)
11692 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
11693 BasicBlock *BBA = PN.getIncomingBlock(i);
11694 BasicBlock *BBB = FirstPN->getIncomingBlock(i);
11696 Value *VA = PN.getIncomingValue(i);
11697 unsigned j = PN.getBasicBlockIndex(BBB);
11698 Value *VB = PN.getIncomingValue(j);
11699 PN.setIncomingBlock(i, BBB);
11700 PN.setIncomingValue(i, VB);
11701 PN.setIncomingBlock(j, BBA);
11702 PN.setIncomingValue(j, VA);
11703 // NOTE: Instcombine normally would want us to "return &PN" if we
11704 // modified any of the operands of an instruction. However, since we
11705 // aren't adding or removing uses (just rearranging them) we don't do
11706 // this in this case.
11710 // If this is an integer PHI and we know that it has an illegal type, see if
11711 // it is only used by trunc or trunc(lshr) operations. If so, we split the
11712 // PHI into the various pieces being extracted. This sort of thing is
11713 // introduced when SROA promotes an aggregate to a single large integer type.
11714 if (isa<IntegerType>(PN.getType()) && TD &&
11715 !TD->isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
11716 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
11722 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11723 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
11725 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
11726 return ReplaceInstUsesWith(GEP, V);
11728 Value *PtrOp = GEP.getOperand(0);
11730 if (isa<UndefValue>(GEP.getOperand(0)))
11731 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
11733 // Eliminate unneeded casts for indices.
11735 bool MadeChange = false;
11736 unsigned PtrSize = TD->getPointerSizeInBits();
11738 gep_type_iterator GTI = gep_type_begin(GEP);
11739 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
11740 I != E; ++I, ++GTI) {
11741 if (!isa<SequentialType>(*GTI)) continue;
11743 // If we are using a wider index than needed for this platform, shrink it
11744 // to what we need. If narrower, sign-extend it to what we need. This
11745 // explicit cast can make subsequent optimizations more obvious.
11746 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
11747 if (OpBits == PtrSize)
11750 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
11753 if (MadeChange) return &GEP;
11756 // Combine Indices - If the source pointer to this getelementptr instruction
11757 // is a getelementptr instruction, combine the indices of the two
11758 // getelementptr instructions into a single instruction.
11760 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
11761 // Note that if our source is a gep chain itself that we wait for that
11762 // chain to be resolved before we perform this transformation. This
11763 // avoids us creating a TON of code in some cases.
11765 if (GetElementPtrInst *SrcGEP =
11766 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
11767 if (SrcGEP->getNumOperands() == 2)
11768 return 0; // Wait until our source is folded to completion.
11770 SmallVector<Value*, 8> Indices;
11772 // Find out whether the last index in the source GEP is a sequential idx.
11773 bool EndsWithSequential = false;
11774 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
11776 EndsWithSequential = !isa<StructType>(*I);
11778 // Can we combine the two pointer arithmetics offsets?
11779 if (EndsWithSequential) {
11780 // Replace: gep (gep %P, long B), long A, ...
11781 // With: T = long A+B; gep %P, T, ...
11784 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
11785 Value *GO1 = GEP.getOperand(1);
11786 if (SO1 == Constant::getNullValue(SO1->getType())) {
11788 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
11791 // If they aren't the same type, then the input hasn't been processed
11792 // by the loop above yet (which canonicalizes sequential index types to
11793 // intptr_t). Just avoid transforming this until the input has been
11795 if (SO1->getType() != GO1->getType())
11797 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11800 // Update the GEP in place if possible.
11801 if (Src->getNumOperands() == 2) {
11802 GEP.setOperand(0, Src->getOperand(0));
11803 GEP.setOperand(1, Sum);
11806 Indices.append(Src->op_begin()+1, Src->op_end()-1);
11807 Indices.push_back(Sum);
11808 Indices.append(GEP.op_begin()+2, GEP.op_end());
11809 } else if (isa<Constant>(*GEP.idx_begin()) &&
11810 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11811 Src->getNumOperands() != 1) {
11812 // Otherwise we can do the fold if the first index of the GEP is a zero
11813 Indices.append(Src->op_begin()+1, Src->op_end());
11814 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
11817 if (!Indices.empty())
11818 return (cast<GEPOperator>(&GEP)->isInBounds() &&
11819 Src->isInBounds()) ?
11820 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
11821 Indices.end(), GEP.getName()) :
11822 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
11823 Indices.end(), GEP.getName());
11826 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
11827 if (Value *X = getBitCastOperand(PtrOp)) {
11828 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
11830 // If the input bitcast is actually "bitcast(bitcast(x))", then we don't
11831 // want to change the gep until the bitcasts are eliminated.
11832 if (getBitCastOperand(X)) {
11833 Worklist.AddValue(PtrOp);
11837 bool HasZeroPointerIndex = false;
11838 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
11839 HasZeroPointerIndex = C->isZero();
11841 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11842 // into : GEP [10 x i8]* X, i32 0, ...
11844 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11845 // into : GEP i8* X, ...
11847 // This occurs when the program declares an array extern like "int X[];"
11848 if (HasZeroPointerIndex) {
11849 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11850 const PointerType *XTy = cast<PointerType>(X->getType());
11851 if (const ArrayType *CATy =
11852 dyn_cast<ArrayType>(CPTy->getElementType())) {
11853 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11854 if (CATy->getElementType() == XTy->getElementType()) {
11855 // -> GEP i8* X, ...
11856 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11857 return cast<GEPOperator>(&GEP)->isInBounds() ?
11858 GetElementPtrInst::CreateInBounds(X, Indices.begin(), Indices.end(),
11860 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11864 if (const ArrayType *XATy = dyn_cast<ArrayType>(XTy->getElementType())){
11865 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11866 if (CATy->getElementType() == XATy->getElementType()) {
11867 // -> GEP [10 x i8]* X, i32 0, ...
11868 // At this point, we know that the cast source type is a pointer
11869 // to an array of the same type as the destination pointer
11870 // array. Because the array type is never stepped over (there
11871 // is a leading zero) we can fold the cast into this GEP.
11872 GEP.setOperand(0, X);
11877 } else if (GEP.getNumOperands() == 2) {
11878 // Transform things like:
11879 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11880 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11881 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11882 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11883 if (TD && isa<ArrayType>(SrcElTy) &&
11884 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11885 TD->getTypeAllocSize(ResElTy)) {
11887 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11888 Idx[1] = GEP.getOperand(1);
11889 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11890 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11891 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11892 // V and GEP are both pointer types --> BitCast
11893 return new BitCastInst(NewGEP, GEP.getType());
11896 // Transform things like:
11897 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11898 // (where tmp = 8*tmp2) into:
11899 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11901 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
11902 uint64_t ArrayEltSize =
11903 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11905 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11906 // allow either a mul, shift, or constant here.
11908 ConstantInt *Scale = 0;
11909 if (ArrayEltSize == 1) {
11910 NewIdx = GEP.getOperand(1);
11911 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11912 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11913 NewIdx = ConstantInt::get(CI->getType(), 1);
11915 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11916 if (Inst->getOpcode() == Instruction::Shl &&
11917 isa<ConstantInt>(Inst->getOperand(1))) {
11918 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11919 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11920 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11922 NewIdx = Inst->getOperand(0);
11923 } else if (Inst->getOpcode() == Instruction::Mul &&
11924 isa<ConstantInt>(Inst->getOperand(1))) {
11925 Scale = cast<ConstantInt>(Inst->getOperand(1));
11926 NewIdx = Inst->getOperand(0);
11930 // If the index will be to exactly the right offset with the scale taken
11931 // out, perform the transformation. Note, we don't know whether Scale is
11932 // signed or not. We'll use unsigned version of division/modulo
11933 // operation after making sure Scale doesn't have the sign bit set.
11934 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11935 Scale->getZExtValue() % ArrayEltSize == 0) {
11936 Scale = ConstantInt::get(Scale->getType(),
11937 Scale->getZExtValue() / ArrayEltSize);
11938 if (Scale->getZExtValue() != 1) {
11939 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11941 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
11944 // Insert the new GEP instruction.
11946 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11948 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11949 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11950 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11951 // The NewGEP must be pointer typed, so must the old one -> BitCast
11952 return new BitCastInst(NewGEP, GEP.getType());
11958 /// See if we can simplify:
11959 /// X = bitcast A* to B*
11960 /// Y = gep X, <...constant indices...>
11961 /// into a gep of the original struct. This is important for SROA and alias
11962 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11963 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11965 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11966 // Determine how much the GEP moves the pointer. We are guaranteed to get
11967 // a constant back from EmitGEPOffset.
11968 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP, *this));
11969 int64_t Offset = OffsetV->getSExtValue();
11971 // If this GEP instruction doesn't move the pointer, just replace the GEP
11972 // with a bitcast of the real input to the dest type.
11974 // If the bitcast is of an allocation, and the allocation will be
11975 // converted to match the type of the cast, don't touch this.
11976 if (isa<AllocaInst>(BCI->getOperand(0)) ||
11977 isMalloc(BCI->getOperand(0))) {
11978 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11979 if (Instruction *I = visitBitCast(*BCI)) {
11982 BCI->getParent()->getInstList().insert(BCI, I);
11983 ReplaceInstUsesWith(*BCI, I);
11988 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11991 // Otherwise, if the offset is non-zero, we need to find out if there is a
11992 // field at Offset in 'A's type. If so, we can pull the cast through the
11994 SmallVector<Value*, 8> NewIndices;
11996 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11997 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11998 Value *NGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11999 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
12000 NewIndices.end()) :
12001 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
12004 if (NGEP->getType() == GEP.getType())
12005 return ReplaceInstUsesWith(GEP, NGEP);
12006 NGEP->takeName(&GEP);
12007 return new BitCastInst(NGEP, GEP.getType());
12015 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
12016 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
12017 if (AI.isArrayAllocation()) { // Check C != 1
12018 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
12019 const Type *NewTy =
12020 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
12021 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
12022 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
12023 New->setAlignment(AI.getAlignment());
12025 // Scan to the end of the allocation instructions, to skip over a block of
12026 // allocas if possible...also skip interleaved debug info
12028 BasicBlock::iterator It = New;
12029 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
12031 // Now that I is pointing to the first non-allocation-inst in the block,
12032 // insert our getelementptr instruction...
12034 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
12038 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
12039 New->getName()+".sub", It);
12041 // Now make everything use the getelementptr instead of the original
12043 return ReplaceInstUsesWith(AI, V);
12044 } else if (isa<UndefValue>(AI.getArraySize())) {
12045 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
12049 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
12050 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
12051 // Note that we only do this for alloca's, because malloc should allocate
12052 // and return a unique pointer, even for a zero byte allocation.
12053 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
12054 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
12056 // If the alignment is 0 (unspecified), assign it the preferred alignment.
12057 if (AI.getAlignment() == 0)
12058 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
12064 Instruction *InstCombiner::visitFree(Instruction &FI) {
12065 Value *Op = FI.getOperand(1);
12067 // free undef -> unreachable.
12068 if (isa<UndefValue>(Op)) {
12069 // Insert a new store to null because we cannot modify the CFG here.
12070 new StoreInst(ConstantInt::getTrue(*Context),
12071 UndefValue::get(Type::getInt1PtrTy(*Context)), &FI);
12072 return EraseInstFromFunction(FI);
12075 // If we have 'free null' delete the instruction. This can happen in stl code
12076 // when lots of inlining happens.
12077 if (isa<ConstantPointerNull>(Op))
12078 return EraseInstFromFunction(FI);
12080 // If we have a malloc call whose only use is a free call, delete both.
12081 if (isMalloc(Op)) {
12082 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
12083 if (Op->hasOneUse() && CI->hasOneUse()) {
12084 EraseInstFromFunction(FI);
12085 EraseInstFromFunction(*CI);
12086 return EraseInstFromFunction(*cast<Instruction>(Op));
12089 // Op is a call to malloc
12090 if (Op->hasOneUse()) {
12091 EraseInstFromFunction(FI);
12092 return EraseInstFromFunction(*cast<Instruction>(Op));
12100 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
12101 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
12102 const TargetData *TD) {
12103 User *CI = cast<User>(LI.getOperand(0));
12104 Value *CastOp = CI->getOperand(0);
12105 LLVMContext *Context = IC.getContext();
12107 const PointerType *DestTy = cast<PointerType>(CI->getType());
12108 const Type *DestPTy = DestTy->getElementType();
12109 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
12111 // If the address spaces don't match, don't eliminate the cast.
12112 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
12115 const Type *SrcPTy = SrcTy->getElementType();
12117 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
12118 isa<VectorType>(DestPTy)) {
12119 // If the source is an array, the code below will not succeed. Check to
12120 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
12122 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
12123 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
12124 if (ASrcTy->getNumElements() != 0) {
12126 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
12128 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
12129 SrcTy = cast<PointerType>(CastOp->getType());
12130 SrcPTy = SrcTy->getElementType();
12133 if (IC.getTargetData() &&
12134 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
12135 isa<VectorType>(SrcPTy)) &&
12136 // Do not allow turning this into a load of an integer, which is then
12137 // casted to a pointer, this pessimizes pointer analysis a lot.
12138 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
12139 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
12140 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
12142 // Okay, we are casting from one integer or pointer type to another of
12143 // the same size. Instead of casting the pointer before the load, cast
12144 // the result of the loaded value.
12146 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
12147 // Now cast the result of the load.
12148 return new BitCastInst(NewLoad, LI.getType());
12155 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
12156 Value *Op = LI.getOperand(0);
12158 // Attempt to improve the alignment.
12160 unsigned KnownAlign =
12161 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
12163 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
12164 LI.getAlignment()))
12165 LI.setAlignment(KnownAlign);
12168 // load (cast X) --> cast (load X) iff safe.
12169 if (isa<CastInst>(Op))
12170 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
12173 // None of the following transforms are legal for volatile loads.
12174 if (LI.isVolatile()) return 0;
12176 // Do really simple store-to-load forwarding and load CSE, to catch cases
12177 // where there are several consequtive memory accesses to the same location,
12178 // separated by a few arithmetic operations.
12179 BasicBlock::iterator BBI = &LI;
12180 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
12181 return ReplaceInstUsesWith(LI, AvailableVal);
12183 // load(gep null, ...) -> unreachable
12184 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
12185 const Value *GEPI0 = GEPI->getOperand(0);
12186 // TODO: Consider a target hook for valid address spaces for this xform.
12187 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
12188 // Insert a new store to null instruction before the load to indicate
12189 // that this code is not reachable. We do this instead of inserting
12190 // an unreachable instruction directly because we cannot modify the
12192 new StoreInst(UndefValue::get(LI.getType()),
12193 Constant::getNullValue(Op->getType()), &LI);
12194 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
12198 // load null/undef -> unreachable
12199 // TODO: Consider a target hook for valid address spaces for this xform.
12200 if (isa<UndefValue>(Op) ||
12201 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
12202 // Insert a new store to null instruction before the load to indicate that
12203 // this code is not reachable. We do this instead of inserting an
12204 // unreachable instruction directly because we cannot modify the CFG.
12205 new StoreInst(UndefValue::get(LI.getType()),
12206 Constant::getNullValue(Op->getType()), &LI);
12207 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
12210 // Instcombine load (constantexpr_cast global) -> cast (load global)
12211 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
12213 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
12216 if (Op->hasOneUse()) {
12217 // Change select and PHI nodes to select values instead of addresses: this
12218 // helps alias analysis out a lot, allows many others simplifications, and
12219 // exposes redundancy in the code.
12221 // Note that we cannot do the transformation unless we know that the
12222 // introduced loads cannot trap! Something like this is valid as long as
12223 // the condition is always false: load (select bool %C, int* null, int* %G),
12224 // but it would not be valid if we transformed it to load from null
12225 // unconditionally.
12227 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
12228 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
12229 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
12230 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
12231 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
12232 SI->getOperand(1)->getName()+".val");
12233 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
12234 SI->getOperand(2)->getName()+".val");
12235 return SelectInst::Create(SI->getCondition(), V1, V2);
12238 // load (select (cond, null, P)) -> load P
12239 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
12240 if (C->isNullValue()) {
12241 LI.setOperand(0, SI->getOperand(2));
12245 // load (select (cond, P, null)) -> load P
12246 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
12247 if (C->isNullValue()) {
12248 LI.setOperand(0, SI->getOperand(1));
12256 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
12257 /// when possible. This makes it generally easy to do alias analysis and/or
12258 /// SROA/mem2reg of the memory object.
12259 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
12260 User *CI = cast<User>(SI.getOperand(1));
12261 Value *CastOp = CI->getOperand(0);
12263 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
12264 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
12265 if (SrcTy == 0) return 0;
12267 const Type *SrcPTy = SrcTy->getElementType();
12269 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
12272 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
12273 /// to its first element. This allows us to handle things like:
12274 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
12275 /// on 32-bit hosts.
12276 SmallVector<Value*, 4> NewGEPIndices;
12278 // If the source is an array, the code below will not succeed. Check to
12279 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
12281 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
12282 // Index through pointer.
12283 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
12284 NewGEPIndices.push_back(Zero);
12287 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
12288 if (!STy->getNumElements()) /* Struct can be empty {} */
12290 NewGEPIndices.push_back(Zero);
12291 SrcPTy = STy->getElementType(0);
12292 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
12293 NewGEPIndices.push_back(Zero);
12294 SrcPTy = ATy->getElementType();
12300 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
12303 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
12306 // If the pointers point into different address spaces or if they point to
12307 // values with different sizes, we can't do the transformation.
12308 if (!IC.getTargetData() ||
12309 SrcTy->getAddressSpace() !=
12310 cast<PointerType>(CI->getType())->getAddressSpace() ||
12311 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
12312 IC.getTargetData()->getTypeSizeInBits(DestPTy))
12315 // Okay, we are casting from one integer or pointer type to another of
12316 // the same size. Instead of casting the pointer before
12317 // the store, cast the value to be stored.
12319 Value *SIOp0 = SI.getOperand(0);
12320 Instruction::CastOps opcode = Instruction::BitCast;
12321 const Type* CastSrcTy = SIOp0->getType();
12322 const Type* CastDstTy = SrcPTy;
12323 if (isa<PointerType>(CastDstTy)) {
12324 if (CastSrcTy->isInteger())
12325 opcode = Instruction::IntToPtr;
12326 } else if (isa<IntegerType>(CastDstTy)) {
12327 if (isa<PointerType>(SIOp0->getType()))
12328 opcode = Instruction::PtrToInt;
12331 // SIOp0 is a pointer to aggregate and this is a store to the first field,
12332 // emit a GEP to index into its first field.
12333 if (!NewGEPIndices.empty())
12334 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
12335 NewGEPIndices.end());
12337 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
12338 SIOp0->getName()+".c");
12339 return new StoreInst(NewCast, CastOp);
12342 /// equivalentAddressValues - Test if A and B will obviously have the same
12343 /// value. This includes recognizing that %t0 and %t1 will have the same
12344 /// value in code like this:
12345 /// %t0 = getelementptr \@a, 0, 3
12346 /// store i32 0, i32* %t0
12347 /// %t1 = getelementptr \@a, 0, 3
12348 /// %t2 = load i32* %t1
12350 static bool equivalentAddressValues(Value *A, Value *B) {
12351 // Test if the values are trivially equivalent.
12352 if (A == B) return true;
12354 // Test if the values come form identical arithmetic instructions.
12355 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
12356 // its only used to compare two uses within the same basic block, which
12357 // means that they'll always either have the same value or one of them
12358 // will have an undefined value.
12359 if (isa<BinaryOperator>(A) ||
12360 isa<CastInst>(A) ||
12362 isa<GetElementPtrInst>(A))
12363 if (Instruction *BI = dyn_cast<Instruction>(B))
12364 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
12367 // Otherwise they may not be equivalent.
12371 // If this instruction has two uses, one of which is a llvm.dbg.declare,
12372 // return the llvm.dbg.declare.
12373 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
12374 if (!V->hasNUses(2))
12376 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
12378 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
12380 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
12381 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
12388 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
12389 Value *Val = SI.getOperand(0);
12390 Value *Ptr = SI.getOperand(1);
12392 // If the RHS is an alloca with a single use, zapify the store, making the
12394 // If the RHS is an alloca with a two uses, the other one being a
12395 // llvm.dbg.declare, zapify the store and the declare, making the
12396 // alloca dead. We must do this to prevent declare's from affecting
12398 if (!SI.isVolatile()) {
12399 if (Ptr->hasOneUse()) {
12400 if (isa<AllocaInst>(Ptr)) {
12401 EraseInstFromFunction(SI);
12405 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
12406 if (isa<AllocaInst>(GEP->getOperand(0))) {
12407 if (GEP->getOperand(0)->hasOneUse()) {
12408 EraseInstFromFunction(SI);
12412 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
12413 EraseInstFromFunction(*DI);
12414 EraseInstFromFunction(SI);
12421 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
12422 EraseInstFromFunction(*DI);
12423 EraseInstFromFunction(SI);
12429 // Attempt to improve the alignment.
12431 unsigned KnownAlign =
12432 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
12434 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
12435 SI.getAlignment()))
12436 SI.setAlignment(KnownAlign);
12439 // Do really simple DSE, to catch cases where there are several consecutive
12440 // stores to the same location, separated by a few arithmetic operations. This
12441 // situation often occurs with bitfield accesses.
12442 BasicBlock::iterator BBI = &SI;
12443 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
12446 // Don't count debug info directives, lest they affect codegen,
12447 // and we skip pointer-to-pointer bitcasts, which are NOPs.
12448 // It is necessary for correctness to skip those that feed into a
12449 // llvm.dbg.declare, as these are not present when debugging is off.
12450 if (isa<DbgInfoIntrinsic>(BBI) ||
12451 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12456 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
12457 // Prev store isn't volatile, and stores to the same location?
12458 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
12459 SI.getOperand(1))) {
12462 EraseInstFromFunction(*PrevSI);
12468 // If this is a load, we have to stop. However, if the loaded value is from
12469 // the pointer we're loading and is producing the pointer we're storing,
12470 // then *this* store is dead (X = load P; store X -> P).
12471 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
12472 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
12473 !SI.isVolatile()) {
12474 EraseInstFromFunction(SI);
12478 // Otherwise, this is a load from some other location. Stores before it
12479 // may not be dead.
12483 // Don't skip over loads or things that can modify memory.
12484 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
12489 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
12491 // store X, null -> turns into 'unreachable' in SimplifyCFG
12492 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
12493 if (!isa<UndefValue>(Val)) {
12494 SI.setOperand(0, UndefValue::get(Val->getType()));
12495 if (Instruction *U = dyn_cast<Instruction>(Val))
12496 Worklist.Add(U); // Dropped a use.
12499 return 0; // Do not modify these!
12502 // store undef, Ptr -> noop
12503 if (isa<UndefValue>(Val)) {
12504 EraseInstFromFunction(SI);
12509 // If the pointer destination is a cast, see if we can fold the cast into the
12511 if (isa<CastInst>(Ptr))
12512 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
12514 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
12516 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
12520 // If this store is the last instruction in the basic block (possibly
12521 // excepting debug info instructions and the pointer bitcasts that feed
12522 // into them), and if the block ends with an unconditional branch, try
12523 // to move it to the successor block.
12527 } while (isa<DbgInfoIntrinsic>(BBI) ||
12528 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
12529 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
12530 if (BI->isUnconditional())
12531 if (SimplifyStoreAtEndOfBlock(SI))
12532 return 0; // xform done!
12537 /// SimplifyStoreAtEndOfBlock - Turn things like:
12538 /// if () { *P = v1; } else { *P = v2 }
12539 /// into a phi node with a store in the successor.
12541 /// Simplify things like:
12542 /// *P = v1; if () { *P = v2; }
12543 /// into a phi node with a store in the successor.
12545 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
12546 BasicBlock *StoreBB = SI.getParent();
12548 // Check to see if the successor block has exactly two incoming edges. If
12549 // so, see if the other predecessor contains a store to the same location.
12550 // if so, insert a PHI node (if needed) and move the stores down.
12551 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
12553 // Determine whether Dest has exactly two predecessors and, if so, compute
12554 // the other predecessor.
12555 pred_iterator PI = pred_begin(DestBB);
12556 BasicBlock *OtherBB = 0;
12557 if (*PI != StoreBB)
12560 if (PI == pred_end(DestBB))
12563 if (*PI != StoreBB) {
12568 if (++PI != pred_end(DestBB))
12571 // Bail out if all the relevant blocks aren't distinct (this can happen,
12572 // for example, if SI is in an infinite loop)
12573 if (StoreBB == DestBB || OtherBB == DestBB)
12576 // Verify that the other block ends in a branch and is not otherwise empty.
12577 BasicBlock::iterator BBI = OtherBB->getTerminator();
12578 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12579 if (!OtherBr || BBI == OtherBB->begin())
12582 // If the other block ends in an unconditional branch, check for the 'if then
12583 // else' case. there is an instruction before the branch.
12584 StoreInst *OtherStore = 0;
12585 if (OtherBr->isUnconditional()) {
12587 // Skip over debugging info.
12588 while (isa<DbgInfoIntrinsic>(BBI) ||
12589 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12590 if (BBI==OtherBB->begin())
12594 // If this isn't a store, isn't a store to the same location, or if the
12595 // alignments differ, bail out.
12596 OtherStore = dyn_cast<StoreInst>(BBI);
12597 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
12598 OtherStore->getAlignment() != SI.getAlignment())
12601 // Otherwise, the other block ended with a conditional branch. If one of the
12602 // destinations is StoreBB, then we have the if/then case.
12603 if (OtherBr->getSuccessor(0) != StoreBB &&
12604 OtherBr->getSuccessor(1) != StoreBB)
12607 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12608 // if/then triangle. See if there is a store to the same ptr as SI that
12609 // lives in OtherBB.
12611 // Check to see if we find the matching store.
12612 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12613 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
12614 OtherStore->getAlignment() != SI.getAlignment())
12618 // If we find something that may be using or overwriting the stored
12619 // value, or if we run out of instructions, we can't do the xform.
12620 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12621 BBI == OtherBB->begin())
12625 // In order to eliminate the store in OtherBr, we have to
12626 // make sure nothing reads or overwrites the stored value in
12628 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12629 // FIXME: This should really be AA driven.
12630 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12635 // Insert a PHI node now if we need it.
12636 Value *MergedVal = OtherStore->getOperand(0);
12637 if (MergedVal != SI.getOperand(0)) {
12638 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12639 PN->reserveOperandSpace(2);
12640 PN->addIncoming(SI.getOperand(0), SI.getParent());
12641 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12642 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12645 // Advance to a place where it is safe to insert the new store and
12647 BBI = DestBB->getFirstNonPHI();
12648 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12649 OtherStore->isVolatile(),
12650 SI.getAlignment()), *BBI);
12652 // Nuke the old stores.
12653 EraseInstFromFunction(SI);
12654 EraseInstFromFunction(*OtherStore);
12660 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12661 // Change br (not X), label True, label False to: br X, label False, True
12663 BasicBlock *TrueDest;
12664 BasicBlock *FalseDest;
12665 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
12666 !isa<Constant>(X)) {
12667 // Swap Destinations and condition...
12668 BI.setCondition(X);
12669 BI.setSuccessor(0, FalseDest);
12670 BI.setSuccessor(1, TrueDest);
12674 // Cannonicalize fcmp_one -> fcmp_oeq
12675 FCmpInst::Predicate FPred; Value *Y;
12676 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12677 TrueDest, FalseDest)) &&
12678 BI.getCondition()->hasOneUse())
12679 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12680 FPred == FCmpInst::FCMP_OGE) {
12681 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
12682 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
12684 // Swap Destinations and condition.
12685 BI.setSuccessor(0, FalseDest);
12686 BI.setSuccessor(1, TrueDest);
12687 Worklist.Add(Cond);
12691 // Cannonicalize icmp_ne -> icmp_eq
12692 ICmpInst::Predicate IPred;
12693 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12694 TrueDest, FalseDest)) &&
12695 BI.getCondition()->hasOneUse())
12696 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12697 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12698 IPred == ICmpInst::ICMP_SGE) {
12699 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
12700 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
12701 // Swap Destinations and condition.
12702 BI.setSuccessor(0, FalseDest);
12703 BI.setSuccessor(1, TrueDest);
12704 Worklist.Add(Cond);
12711 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12712 Value *Cond = SI.getCondition();
12713 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12714 if (I->getOpcode() == Instruction::Add)
12715 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12716 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12717 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12719 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12721 SI.setOperand(0, I->getOperand(0));
12729 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12730 Value *Agg = EV.getAggregateOperand();
12732 if (!EV.hasIndices())
12733 return ReplaceInstUsesWith(EV, Agg);
12735 if (Constant *C = dyn_cast<Constant>(Agg)) {
12736 if (isa<UndefValue>(C))
12737 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12739 if (isa<ConstantAggregateZero>(C))
12740 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12742 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12743 // Extract the element indexed by the first index out of the constant
12744 Value *V = C->getOperand(*EV.idx_begin());
12745 if (EV.getNumIndices() > 1)
12746 // Extract the remaining indices out of the constant indexed by the
12748 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12750 return ReplaceInstUsesWith(EV, V);
12752 return 0; // Can't handle other constants
12754 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12755 // We're extracting from an insertvalue instruction, compare the indices
12756 const unsigned *exti, *exte, *insi, *inse;
12757 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12758 exte = EV.idx_end(), inse = IV->idx_end();
12759 exti != exte && insi != inse;
12761 if (*insi != *exti)
12762 // The insert and extract both reference distinctly different elements.
12763 // This means the extract is not influenced by the insert, and we can
12764 // replace the aggregate operand of the extract with the aggregate
12765 // operand of the insert. i.e., replace
12766 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12767 // %E = extractvalue { i32, { i32 } } %I, 0
12769 // %E = extractvalue { i32, { i32 } } %A, 0
12770 return ExtractValueInst::Create(IV->getAggregateOperand(),
12771 EV.idx_begin(), EV.idx_end());
12773 if (exti == exte && insi == inse)
12774 // Both iterators are at the end: Index lists are identical. Replace
12775 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12776 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12778 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12779 if (exti == exte) {
12780 // The extract list is a prefix of the insert list. i.e. replace
12781 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12782 // %E = extractvalue { i32, { i32 } } %I, 1
12784 // %X = extractvalue { i32, { i32 } } %A, 1
12785 // %E = insertvalue { i32 } %X, i32 42, 0
12786 // by switching the order of the insert and extract (though the
12787 // insertvalue should be left in, since it may have other uses).
12788 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
12789 EV.idx_begin(), EV.idx_end());
12790 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12794 // The insert list is a prefix of the extract list
12795 // We can simply remove the common indices from the extract and make it
12796 // operate on the inserted value instead of the insertvalue result.
12798 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12799 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12801 // %E extractvalue { i32 } { i32 42 }, 0
12802 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12805 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
12806 // We're extracting from an intrinsic, see if we're the only user, which
12807 // allows us to simplify multiple result intrinsics to simpler things that
12808 // just get one value..
12809 if (II->hasOneUse()) {
12810 // Check if we're grabbing the overflow bit or the result of a 'with
12811 // overflow' intrinsic. If it's the latter we can remove the intrinsic
12812 // and replace it with a traditional binary instruction.
12813 switch (II->getIntrinsicID()) {
12814 case Intrinsic::uadd_with_overflow:
12815 case Intrinsic::sadd_with_overflow:
12816 if (*EV.idx_begin() == 0) { // Normal result.
12817 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
12818 II->replaceAllUsesWith(UndefValue::get(II->getType()));
12819 EraseInstFromFunction(*II);
12820 return BinaryOperator::CreateAdd(LHS, RHS);
12823 case Intrinsic::usub_with_overflow:
12824 case Intrinsic::ssub_with_overflow:
12825 if (*EV.idx_begin() == 0) { // Normal result.
12826 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
12827 II->replaceAllUsesWith(UndefValue::get(II->getType()));
12828 EraseInstFromFunction(*II);
12829 return BinaryOperator::CreateSub(LHS, RHS);
12832 case Intrinsic::umul_with_overflow:
12833 case Intrinsic::smul_with_overflow:
12834 if (*EV.idx_begin() == 0) { // Normal result.
12835 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
12836 II->replaceAllUsesWith(UndefValue::get(II->getType()));
12837 EraseInstFromFunction(*II);
12838 return BinaryOperator::CreateMul(LHS, RHS);
12846 // Can't simplify extracts from other values. Note that nested extracts are
12847 // already simplified implicitely by the above (extract ( extract (insert) )
12848 // will be translated into extract ( insert ( extract ) ) first and then just
12849 // the value inserted, if appropriate).
12853 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12854 /// is to leave as a vector operation.
12855 static bool CheapToScalarize(Value *V, bool isConstant) {
12856 if (isa<ConstantAggregateZero>(V))
12858 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12859 if (isConstant) return true;
12860 // If all elts are the same, we can extract.
12861 Constant *Op0 = C->getOperand(0);
12862 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12863 if (C->getOperand(i) != Op0)
12867 Instruction *I = dyn_cast<Instruction>(V);
12868 if (!I) return false;
12870 // Insert element gets simplified to the inserted element or is deleted if
12871 // this is constant idx extract element and its a constant idx insertelt.
12872 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12873 isa<ConstantInt>(I->getOperand(2)))
12875 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12877 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12878 if (BO->hasOneUse() &&
12879 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12880 CheapToScalarize(BO->getOperand(1), isConstant)))
12882 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12883 if (CI->hasOneUse() &&
12884 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12885 CheapToScalarize(CI->getOperand(1), isConstant)))
12891 /// Read and decode a shufflevector mask.
12893 /// It turns undef elements into values that are larger than the number of
12894 /// elements in the input.
12895 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12896 unsigned NElts = SVI->getType()->getNumElements();
12897 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12898 return std::vector<unsigned>(NElts, 0);
12899 if (isa<UndefValue>(SVI->getOperand(2)))
12900 return std::vector<unsigned>(NElts, 2*NElts);
12902 std::vector<unsigned> Result;
12903 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12904 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12905 if (isa<UndefValue>(*i))
12906 Result.push_back(NElts*2); // undef -> 8
12908 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12912 /// FindScalarElement - Given a vector and an element number, see if the scalar
12913 /// value is already around as a register, for example if it were inserted then
12914 /// extracted from the vector.
12915 static Value *FindScalarElement(Value *V, unsigned EltNo,
12916 LLVMContext *Context) {
12917 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12918 const VectorType *PTy = cast<VectorType>(V->getType());
12919 unsigned Width = PTy->getNumElements();
12920 if (EltNo >= Width) // Out of range access.
12921 return UndefValue::get(PTy->getElementType());
12923 if (isa<UndefValue>(V))
12924 return UndefValue::get(PTy->getElementType());
12925 else if (isa<ConstantAggregateZero>(V))
12926 return Constant::getNullValue(PTy->getElementType());
12927 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12928 return CP->getOperand(EltNo);
12929 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12930 // If this is an insert to a variable element, we don't know what it is.
12931 if (!isa<ConstantInt>(III->getOperand(2)))
12933 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12935 // If this is an insert to the element we are looking for, return the
12937 if (EltNo == IIElt)
12938 return III->getOperand(1);
12940 // Otherwise, the insertelement doesn't modify the value, recurse on its
12942 return FindScalarElement(III->getOperand(0), EltNo, Context);
12943 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12944 unsigned LHSWidth =
12945 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12946 unsigned InEl = getShuffleMask(SVI)[EltNo];
12947 if (InEl < LHSWidth)
12948 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12949 else if (InEl < LHSWidth*2)
12950 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12952 return UndefValue::get(PTy->getElementType());
12955 // Otherwise, we don't know.
12959 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12960 // If vector val is undef, replace extract with scalar undef.
12961 if (isa<UndefValue>(EI.getOperand(0)))
12962 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12964 // If vector val is constant 0, replace extract with scalar 0.
12965 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12966 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12968 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12969 // If vector val is constant with all elements the same, replace EI with
12970 // that element. When the elements are not identical, we cannot replace yet
12971 // (we do that below, but only when the index is constant).
12972 Constant *op0 = C->getOperand(0);
12973 for (unsigned i = 1; i != C->getNumOperands(); ++i)
12974 if (C->getOperand(i) != op0) {
12979 return ReplaceInstUsesWith(EI, op0);
12982 // If extracting a specified index from the vector, see if we can recursively
12983 // find a previously computed scalar that was inserted into the vector.
12984 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12985 unsigned IndexVal = IdxC->getZExtValue();
12986 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
12988 // If this is extracting an invalid index, turn this into undef, to avoid
12989 // crashing the code below.
12990 if (IndexVal >= VectorWidth)
12991 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12993 // This instruction only demands the single element from the input vector.
12994 // If the input vector has a single use, simplify it based on this use
12996 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12997 APInt UndefElts(VectorWidth, 0);
12998 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12999 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
13000 DemandedMask, UndefElts)) {
13001 EI.setOperand(0, V);
13006 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
13007 return ReplaceInstUsesWith(EI, Elt);
13009 // If the this extractelement is directly using a bitcast from a vector of
13010 // the same number of elements, see if we can find the source element from
13011 // it. In this case, we will end up needing to bitcast the scalars.
13012 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
13013 if (const VectorType *VT =
13014 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
13015 if (VT->getNumElements() == VectorWidth)
13016 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
13017 IndexVal, Context))
13018 return new BitCastInst(Elt, EI.getType());
13022 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
13023 // Push extractelement into predecessor operation if legal and
13024 // profitable to do so
13025 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
13026 if (I->hasOneUse() &&
13027 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
13029 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
13030 EI.getName()+".lhs");
13032 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
13033 EI.getName()+".rhs");
13034 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
13036 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
13037 // Extracting the inserted element?
13038 if (IE->getOperand(2) == EI.getOperand(1))
13039 return ReplaceInstUsesWith(EI, IE->getOperand(1));
13040 // If the inserted and extracted elements are constants, they must not
13041 // be the same value, extract from the pre-inserted value instead.
13042 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
13043 Worklist.AddValue(EI.getOperand(0));
13044 EI.setOperand(0, IE->getOperand(0));
13047 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
13048 // If this is extracting an element from a shufflevector, figure out where
13049 // it came from and extract from the appropriate input element instead.
13050 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
13051 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
13053 unsigned LHSWidth =
13054 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
13056 if (SrcIdx < LHSWidth)
13057 Src = SVI->getOperand(0);
13058 else if (SrcIdx < LHSWidth*2) {
13059 SrcIdx -= LHSWidth;
13060 Src = SVI->getOperand(1);
13062 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
13064 return ExtractElementInst::Create(Src,
13065 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx,
13069 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
13074 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
13075 /// elements from either LHS or RHS, return the shuffle mask and true.
13076 /// Otherwise, return false.
13077 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
13078 std::vector<Constant*> &Mask,
13079 LLVMContext *Context) {
13080 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
13081 "Invalid CollectSingleShuffleElements");
13082 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
13084 if (isa<UndefValue>(V)) {
13085 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
13087 } else if (V == LHS) {
13088 for (unsigned i = 0; i != NumElts; ++i)
13089 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
13091 } else if (V == RHS) {
13092 for (unsigned i = 0; i != NumElts; ++i)
13093 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
13095 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
13096 // If this is an insert of an extract from some other vector, include it.
13097 Value *VecOp = IEI->getOperand(0);
13098 Value *ScalarOp = IEI->getOperand(1);
13099 Value *IdxOp = IEI->getOperand(2);
13101 if (!isa<ConstantInt>(IdxOp))
13103 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
13105 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
13106 // Okay, we can handle this if the vector we are insertinting into is
13107 // transitively ok.
13108 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
13109 // If so, update the mask to reflect the inserted undef.
13110 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
13113 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
13114 if (isa<ConstantInt>(EI->getOperand(1)) &&
13115 EI->getOperand(0)->getType() == V->getType()) {
13116 unsigned ExtractedIdx =
13117 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
13119 // This must be extracting from either LHS or RHS.
13120 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
13121 // Okay, we can handle this if the vector we are insertinting into is
13122 // transitively ok.
13123 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
13124 // If so, update the mask to reflect the inserted value.
13125 if (EI->getOperand(0) == LHS) {
13126 Mask[InsertedIdx % NumElts] =
13127 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
13129 assert(EI->getOperand(0) == RHS);
13130 Mask[InsertedIdx % NumElts] =
13131 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
13140 // TODO: Handle shufflevector here!
13145 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
13146 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
13147 /// that computes V and the LHS value of the shuffle.
13148 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
13149 Value *&RHS, LLVMContext *Context) {
13150 assert(isa<VectorType>(V->getType()) &&
13151 (RHS == 0 || V->getType() == RHS->getType()) &&
13152 "Invalid shuffle!");
13153 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
13155 if (isa<UndefValue>(V)) {
13156 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
13158 } else if (isa<ConstantAggregateZero>(V)) {
13159 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
13161 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
13162 // If this is an insert of an extract from some other vector, include it.
13163 Value *VecOp = IEI->getOperand(0);
13164 Value *ScalarOp = IEI->getOperand(1);
13165 Value *IdxOp = IEI->getOperand(2);
13167 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
13168 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
13169 EI->getOperand(0)->getType() == V->getType()) {
13170 unsigned ExtractedIdx =
13171 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
13172 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
13174 // Either the extracted from or inserted into vector must be RHSVec,
13175 // otherwise we'd end up with a shuffle of three inputs.
13176 if (EI->getOperand(0) == RHS || RHS == 0) {
13177 RHS = EI->getOperand(0);
13178 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
13179 Mask[InsertedIdx % NumElts] =
13180 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
13184 if (VecOp == RHS) {
13185 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
13187 // Everything but the extracted element is replaced with the RHS.
13188 for (unsigned i = 0; i != NumElts; ++i) {
13189 if (i != InsertedIdx)
13190 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
13195 // If this insertelement is a chain that comes from exactly these two
13196 // vectors, return the vector and the effective shuffle.
13197 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
13199 return EI->getOperand(0);
13204 // TODO: Handle shufflevector here!
13206 // Otherwise, can't do anything fancy. Return an identity vector.
13207 for (unsigned i = 0; i != NumElts; ++i)
13208 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
13212 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
13213 Value *VecOp = IE.getOperand(0);
13214 Value *ScalarOp = IE.getOperand(1);
13215 Value *IdxOp = IE.getOperand(2);
13217 // Inserting an undef or into an undefined place, remove this.
13218 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
13219 ReplaceInstUsesWith(IE, VecOp);
13221 // If the inserted element was extracted from some other vector, and if the
13222 // indexes are constant, try to turn this into a shufflevector operation.
13223 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
13224 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
13225 EI->getOperand(0)->getType() == IE.getType()) {
13226 unsigned NumVectorElts = IE.getType()->getNumElements();
13227 unsigned ExtractedIdx =
13228 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
13229 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
13231 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
13232 return ReplaceInstUsesWith(IE, VecOp);
13234 if (InsertedIdx >= NumVectorElts) // Out of range insert.
13235 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
13237 // If we are extracting a value from a vector, then inserting it right
13238 // back into the same place, just use the input vector.
13239 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
13240 return ReplaceInstUsesWith(IE, VecOp);
13242 // If this insertelement isn't used by some other insertelement, turn it
13243 // (and any insertelements it points to), into one big shuffle.
13244 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
13245 std::vector<Constant*> Mask;
13247 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
13248 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
13249 // We now have a shuffle of LHS, RHS, Mask.
13250 return new ShuffleVectorInst(LHS, RHS,
13251 ConstantVector::get(Mask));
13256 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
13257 APInt UndefElts(VWidth, 0);
13258 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
13259 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
13266 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
13267 Value *LHS = SVI.getOperand(0);
13268 Value *RHS = SVI.getOperand(1);
13269 std::vector<unsigned> Mask = getShuffleMask(&SVI);
13271 bool MadeChange = false;
13273 // Undefined shuffle mask -> undefined value.
13274 if (isa<UndefValue>(SVI.getOperand(2)))
13275 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
13277 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
13279 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
13282 APInt UndefElts(VWidth, 0);
13283 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
13284 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
13285 LHS = SVI.getOperand(0);
13286 RHS = SVI.getOperand(1);
13290 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
13291 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
13292 if (LHS == RHS || isa<UndefValue>(LHS)) {
13293 if (isa<UndefValue>(LHS) && LHS == RHS) {
13294 // shuffle(undef,undef,mask) -> undef.
13295 return ReplaceInstUsesWith(SVI, LHS);
13298 // Remap any references to RHS to use LHS.
13299 std::vector<Constant*> Elts;
13300 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
13301 if (Mask[i] >= 2*e)
13302 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
13304 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
13305 (Mask[i] < e && isa<UndefValue>(LHS))) {
13306 Mask[i] = 2*e; // Turn into undef.
13307 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
13309 Mask[i] = Mask[i] % e; // Force to LHS.
13310 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
13314 SVI.setOperand(0, SVI.getOperand(1));
13315 SVI.setOperand(1, UndefValue::get(RHS->getType()));
13316 SVI.setOperand(2, ConstantVector::get(Elts));
13317 LHS = SVI.getOperand(0);
13318 RHS = SVI.getOperand(1);
13322 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
13323 bool isLHSID = true, isRHSID = true;
13325 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
13326 if (Mask[i] >= e*2) continue; // Ignore undef values.
13327 // Is this an identity shuffle of the LHS value?
13328 isLHSID &= (Mask[i] == i);
13330 // Is this an identity shuffle of the RHS value?
13331 isRHSID &= (Mask[i]-e == i);
13334 // Eliminate identity shuffles.
13335 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
13336 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
13338 // If the LHS is a shufflevector itself, see if we can combine it with this
13339 // one without producing an unusual shuffle. Here we are really conservative:
13340 // we are absolutely afraid of producing a shuffle mask not in the input
13341 // program, because the code gen may not be smart enough to turn a merged
13342 // shuffle into two specific shuffles: it may produce worse code. As such,
13343 // we only merge two shuffles if the result is one of the two input shuffle
13344 // masks. In this case, merging the shuffles just removes one instruction,
13345 // which we know is safe. This is good for things like turning:
13346 // (splat(splat)) -> splat.
13347 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
13348 if (isa<UndefValue>(RHS)) {
13349 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
13351 if (LHSMask.size() == Mask.size()) {
13352 std::vector<unsigned> NewMask;
13353 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
13355 NewMask.push_back(2*e);
13357 NewMask.push_back(LHSMask[Mask[i]]);
13359 // If the result mask is equal to the src shuffle or this
13360 // shuffle mask, do the replacement.
13361 if (NewMask == LHSMask || NewMask == Mask) {
13362 unsigned LHSInNElts =
13363 cast<VectorType>(LHSSVI->getOperand(0)->getType())->
13365 std::vector<Constant*> Elts;
13366 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
13367 if (NewMask[i] >= LHSInNElts*2) {
13368 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
13370 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
13374 return new ShuffleVectorInst(LHSSVI->getOperand(0),
13375 LHSSVI->getOperand(1),
13376 ConstantVector::get(Elts));
13382 return MadeChange ? &SVI : 0;
13388 /// TryToSinkInstruction - Try to move the specified instruction from its
13389 /// current block into the beginning of DestBlock, which can only happen if it's
13390 /// safe to move the instruction past all of the instructions between it and the
13391 /// end of its block.
13392 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
13393 assert(I->hasOneUse() && "Invariants didn't hold!");
13395 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
13396 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
13399 // Do not sink alloca instructions out of the entry block.
13400 if (isa<AllocaInst>(I) && I->getParent() ==
13401 &DestBlock->getParent()->getEntryBlock())
13404 // We can only sink load instructions if there is nothing between the load and
13405 // the end of block that could change the value.
13406 if (I->mayReadFromMemory()) {
13407 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
13409 if (Scan->mayWriteToMemory())
13413 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
13415 CopyPrecedingStopPoint(I, InsertPos);
13416 I->moveBefore(InsertPos);
13422 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
13423 /// all reachable code to the worklist.
13425 /// This has a couple of tricks to make the code faster and more powerful. In
13426 /// particular, we constant fold and DCE instructions as we go, to avoid adding
13427 /// them to the worklist (this significantly speeds up instcombine on code where
13428 /// many instructions are dead or constant). Additionally, if we find a branch
13429 /// whose condition is a known constant, we only visit the reachable successors.
13431 static bool AddReachableCodeToWorklist(BasicBlock *BB,
13432 SmallPtrSet<BasicBlock*, 64> &Visited,
13434 const TargetData *TD) {
13435 bool MadeIRChange = false;
13436 SmallVector<BasicBlock*, 256> Worklist;
13437 Worklist.push_back(BB);
13439 std::vector<Instruction*> InstrsForInstCombineWorklist;
13440 InstrsForInstCombineWorklist.reserve(128);
13442 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
13444 while (!Worklist.empty()) {
13445 BB = Worklist.back();
13446 Worklist.pop_back();
13448 // We have now visited this block! If we've already been here, ignore it.
13449 if (!Visited.insert(BB)) continue;
13451 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
13452 Instruction *Inst = BBI++;
13454 // DCE instruction if trivially dead.
13455 if (isInstructionTriviallyDead(Inst)) {
13457 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
13458 Inst->eraseFromParent();
13462 // ConstantProp instruction if trivially constant.
13463 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
13464 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
13465 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
13467 Inst->replaceAllUsesWith(C);
13469 Inst->eraseFromParent();
13476 // See if we can constant fold its operands.
13477 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
13479 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
13480 if (CE == 0) continue;
13482 // If we already folded this constant, don't try again.
13483 if (!FoldedConstants.insert(CE))
13486 Constant *NewC = ConstantFoldConstantExpression(CE, TD);
13487 if (NewC && NewC != CE) {
13489 MadeIRChange = true;
13495 InstrsForInstCombineWorklist.push_back(Inst);
13498 // Recursively visit successors. If this is a branch or switch on a
13499 // constant, only visit the reachable successor.
13500 TerminatorInst *TI = BB->getTerminator();
13501 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
13502 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
13503 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
13504 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
13505 Worklist.push_back(ReachableBB);
13508 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
13509 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
13510 // See if this is an explicit destination.
13511 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
13512 if (SI->getCaseValue(i) == Cond) {
13513 BasicBlock *ReachableBB = SI->getSuccessor(i);
13514 Worklist.push_back(ReachableBB);
13518 // Otherwise it is the default destination.
13519 Worklist.push_back(SI->getSuccessor(0));
13524 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
13525 Worklist.push_back(TI->getSuccessor(i));
13528 // Once we've found all of the instructions to add to instcombine's worklist,
13529 // add them in reverse order. This way instcombine will visit from the top
13530 // of the function down. This jives well with the way that it adds all uses
13531 // of instructions to the worklist after doing a transformation, thus avoiding
13532 // some N^2 behavior in pathological cases.
13533 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
13534 InstrsForInstCombineWorklist.size());
13536 return MadeIRChange;
13539 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
13540 MadeIRChange = false;
13542 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
13543 << F.getNameStr() << "\n");
13546 // Do a depth-first traversal of the function, populate the worklist with
13547 // the reachable instructions. Ignore blocks that are not reachable. Keep
13548 // track of which blocks we visit.
13549 SmallPtrSet<BasicBlock*, 64> Visited;
13550 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
13552 // Do a quick scan over the function. If we find any blocks that are
13553 // unreachable, remove any instructions inside of them. This prevents
13554 // the instcombine code from having to deal with some bad special cases.
13555 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
13556 if (!Visited.count(BB)) {
13557 Instruction *Term = BB->getTerminator();
13558 while (Term != BB->begin()) { // Remove instrs bottom-up
13559 BasicBlock::iterator I = Term; --I;
13561 DEBUG(errs() << "IC: DCE: " << *I << '\n');
13562 // A debug intrinsic shouldn't force another iteration if we weren't
13563 // going to do one without it.
13564 if (!isa<DbgInfoIntrinsic>(I)) {
13566 MadeIRChange = true;
13569 // If I is not void type then replaceAllUsesWith undef.
13570 // This allows ValueHandlers and custom metadata to adjust itself.
13571 if (!I->getType()->isVoidTy())
13572 I->replaceAllUsesWith(UndefValue::get(I->getType()));
13573 I->eraseFromParent();
13578 while (!Worklist.isEmpty()) {
13579 Instruction *I = Worklist.RemoveOne();
13580 if (I == 0) continue; // skip null values.
13582 // Check to see if we can DCE the instruction.
13583 if (isInstructionTriviallyDead(I)) {
13584 DEBUG(errs() << "IC: DCE: " << *I << '\n');
13585 EraseInstFromFunction(*I);
13587 MadeIRChange = true;
13591 // Instruction isn't dead, see if we can constant propagate it.
13592 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
13593 if (Constant *C = ConstantFoldInstruction(I, TD)) {
13594 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
13596 // Add operands to the worklist.
13597 ReplaceInstUsesWith(*I, C);
13599 EraseInstFromFunction(*I);
13600 MadeIRChange = true;
13604 // See if we can trivially sink this instruction to a successor basic block.
13605 if (I->hasOneUse()) {
13606 BasicBlock *BB = I->getParent();
13607 Instruction *UserInst = cast<Instruction>(I->use_back());
13608 BasicBlock *UserParent;
13610 // Get the block the use occurs in.
13611 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
13612 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
13614 UserParent = UserInst->getParent();
13616 if (UserParent != BB) {
13617 bool UserIsSuccessor = false;
13618 // See if the user is one of our successors.
13619 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13620 if (*SI == UserParent) {
13621 UserIsSuccessor = true;
13625 // If the user is one of our immediate successors, and if that successor
13626 // only has us as a predecessors (we'd have to split the critical edge
13627 // otherwise), we can keep going.
13628 if (UserIsSuccessor && UserParent->getSinglePredecessor())
13629 // Okay, the CFG is simple enough, try to sink this instruction.
13630 MadeIRChange |= TryToSinkInstruction(I, UserParent);
13634 // Now that we have an instruction, try combining it to simplify it.
13635 Builder->SetInsertPoint(I->getParent(), I);
13640 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
13641 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
13643 if (Instruction *Result = visit(*I)) {
13645 // Should we replace the old instruction with a new one?
13647 DEBUG(errs() << "IC: Old = " << *I << '\n'
13648 << " New = " << *Result << '\n');
13650 // Everything uses the new instruction now.
13651 I->replaceAllUsesWith(Result);
13653 // Push the new instruction and any users onto the worklist.
13654 Worklist.Add(Result);
13655 Worklist.AddUsersToWorkList(*Result);
13657 // Move the name to the new instruction first.
13658 Result->takeName(I);
13660 // Insert the new instruction into the basic block...
13661 BasicBlock *InstParent = I->getParent();
13662 BasicBlock::iterator InsertPos = I;
13664 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13665 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13668 InstParent->getInstList().insert(InsertPos, Result);
13670 EraseInstFromFunction(*I);
13673 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
13674 << " New = " << *I << '\n');
13677 // If the instruction was modified, it's possible that it is now dead.
13678 // if so, remove it.
13679 if (isInstructionTriviallyDead(I)) {
13680 EraseInstFromFunction(*I);
13683 Worklist.AddUsersToWorkList(*I);
13686 MadeIRChange = true;
13691 return MadeIRChange;
13695 bool InstCombiner::runOnFunction(Function &F) {
13696 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13697 Context = &F.getContext();
13698 TD = getAnalysisIfAvailable<TargetData>();
13701 /// Builder - This is an IRBuilder that automatically inserts new
13702 /// instructions into the worklist when they are created.
13703 IRBuilder<true, TargetFolder, InstCombineIRInserter>
13704 TheBuilder(F.getContext(), TargetFolder(TD),
13705 InstCombineIRInserter(Worklist));
13706 Builder = &TheBuilder;
13708 bool EverMadeChange = false;
13710 // Iterate while there is work to do.
13711 unsigned Iteration = 0;
13712 while (DoOneIteration(F, Iteration++))
13713 EverMadeChange = true;
13716 return EverMadeChange;
13719 FunctionPass *llvm::createInstructionCombiningPass() {
13720 return new InstCombiner();