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/MallocHelper.h"
46 #include "llvm/Analysis/ValueTracking.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Support/CallSite.h"
51 #include "llvm/Support/ConstantRange.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/ErrorHandling.h"
54 #include "llvm/Support/GetElementPtrTypeIterator.h"
55 #include "llvm/Support/InstVisitor.h"
56 #include "llvm/Support/IRBuilder.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/PatternMatch.h"
59 #include "llvm/Support/TargetFolder.h"
60 #include "llvm/Support/raw_ostream.h"
61 #include "llvm/ADT/DenseMap.h"
62 #include "llvm/ADT/SmallVector.h"
63 #include "llvm/ADT/SmallPtrSet.h"
64 #include "llvm/ADT/Statistic.h"
65 #include "llvm/ADT/STLExtras.h"
69 using namespace llvm::PatternMatch;
71 STATISTIC(NumCombined , "Number of insts combined");
72 STATISTIC(NumConstProp, "Number of constant folds");
73 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
74 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
75 STATISTIC(NumSunkInst , "Number of instructions sunk");
78 /// InstCombineWorklist - This is the worklist management logic for
80 class InstCombineWorklist {
81 SmallVector<Instruction*, 256> Worklist;
82 DenseMap<Instruction*, unsigned> WorklistMap;
84 void operator=(const InstCombineWorklist&RHS); // DO NOT IMPLEMENT
85 InstCombineWorklist(const InstCombineWorklist&); // DO NOT IMPLEMENT
87 InstCombineWorklist() {}
89 bool isEmpty() const { return Worklist.empty(); }
91 /// Add - Add the specified instruction to the worklist if it isn't already
93 void Add(Instruction *I) {
94 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second) {
95 DEBUG(errs() << "IC: ADD: " << *I << '\n');
96 Worklist.push_back(I);
100 void AddValue(Value *V) {
101 if (Instruction *I = dyn_cast<Instruction>(V))
105 /// AddInitialGroup - Add the specified batch of stuff in reverse order.
106 /// which should only be done when the worklist is empty and when the group
107 /// has no duplicates.
108 void AddInitialGroup(Instruction *const *List, unsigned NumEntries) {
109 assert(Worklist.empty() && "Worklist must be empty to add initial group");
110 Worklist.reserve(NumEntries+16);
111 DEBUG(errs() << "IC: ADDING: " << NumEntries << " instrs to worklist\n");
112 for (; NumEntries; --NumEntries) {
113 Instruction *I = List[NumEntries-1];
114 WorklistMap.insert(std::make_pair(I, Worklist.size()));
115 Worklist.push_back(I);
119 // Remove - remove I from the worklist if it exists.
120 void Remove(Instruction *I) {
121 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
122 if (It == WorklistMap.end()) return; // Not in worklist.
124 // Don't bother moving everything down, just null out the slot.
125 Worklist[It->second] = 0;
127 WorklistMap.erase(It);
130 Instruction *RemoveOne() {
131 Instruction *I = Worklist.back();
133 WorklistMap.erase(I);
137 /// AddUsersToWorkList - When an instruction is simplified, add all users of
138 /// the instruction to the work lists because they might get more simplified
141 void AddUsersToWorkList(Instruction &I) {
142 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
144 Add(cast<Instruction>(*UI));
148 /// Zap - check that the worklist is empty and nuke the backing store for
149 /// the map if it is large.
151 assert(WorklistMap.empty() && "Worklist empty, but map not?");
153 // Do an explicit clear, this shrinks the map if needed.
157 } // end anonymous namespace.
161 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works
162 /// just like the normal insertion helper, but also adds any new instructions
163 /// to the instcombine worklist.
164 class InstCombineIRInserter : public IRBuilderDefaultInserter<true> {
165 InstCombineWorklist &Worklist;
167 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
169 void InsertHelper(Instruction *I, const Twine &Name,
170 BasicBlock *BB, BasicBlock::iterator InsertPt) const {
171 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
175 } // end anonymous namespace
179 class InstCombiner : public FunctionPass,
180 public InstVisitor<InstCombiner, Instruction*> {
182 bool MustPreserveLCSSA;
185 /// Worklist - All of the instructions that need to be simplified.
186 InstCombineWorklist Worklist;
188 /// Builder - This is an IRBuilder that automatically inserts new
189 /// instructions into the worklist when they are created.
190 typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
193 static char ID; // Pass identification, replacement for typeid
194 InstCombiner() : FunctionPass(&ID), TD(0), Builder(0) {}
196 LLVMContext *Context;
197 LLVMContext *getContext() const { return Context; }
200 virtual bool runOnFunction(Function &F);
202 bool DoOneIteration(Function &F, unsigned ItNum);
204 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
205 AU.addPreservedID(LCSSAID);
206 AU.setPreservesCFG();
209 TargetData *getTargetData() const { return TD; }
211 // Visitation implementation - Implement instruction combining for different
212 // instruction types. The semantics are as follows:
214 // null - No change was made
215 // I - Change was made, I is still valid, I may be dead though
216 // otherwise - Change was made, replace I with returned instruction
218 Instruction *visitAdd(BinaryOperator &I);
219 Instruction *visitFAdd(BinaryOperator &I);
220 Instruction *visitSub(BinaryOperator &I);
221 Instruction *visitFSub(BinaryOperator &I);
222 Instruction *visitMul(BinaryOperator &I);
223 Instruction *visitFMul(BinaryOperator &I);
224 Instruction *visitURem(BinaryOperator &I);
225 Instruction *visitSRem(BinaryOperator &I);
226 Instruction *visitFRem(BinaryOperator &I);
227 bool SimplifyDivRemOfSelect(BinaryOperator &I);
228 Instruction *commonRemTransforms(BinaryOperator &I);
229 Instruction *commonIRemTransforms(BinaryOperator &I);
230 Instruction *commonDivTransforms(BinaryOperator &I);
231 Instruction *commonIDivTransforms(BinaryOperator &I);
232 Instruction *visitUDiv(BinaryOperator &I);
233 Instruction *visitSDiv(BinaryOperator &I);
234 Instruction *visitFDiv(BinaryOperator &I);
235 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
236 Instruction *FoldAndOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
237 Instruction *visitAnd(BinaryOperator &I);
238 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
239 Instruction *FoldOrOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
240 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
241 Value *A, Value *B, Value *C);
242 Instruction *visitOr (BinaryOperator &I);
243 Instruction *visitXor(BinaryOperator &I);
244 Instruction *visitShl(BinaryOperator &I);
245 Instruction *visitAShr(BinaryOperator &I);
246 Instruction *visitLShr(BinaryOperator &I);
247 Instruction *commonShiftTransforms(BinaryOperator &I);
248 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
250 Instruction *visitFCmpInst(FCmpInst &I);
251 Instruction *visitICmpInst(ICmpInst &I);
252 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
253 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
256 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
257 ConstantInt *DivRHS);
259 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
260 ICmpInst::Predicate Cond, Instruction &I);
261 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
263 Instruction *commonCastTransforms(CastInst &CI);
264 Instruction *commonIntCastTransforms(CastInst &CI);
265 Instruction *commonPointerCastTransforms(CastInst &CI);
266 Instruction *visitTrunc(TruncInst &CI);
267 Instruction *visitZExt(ZExtInst &CI);
268 Instruction *visitSExt(SExtInst &CI);
269 Instruction *visitFPTrunc(FPTruncInst &CI);
270 Instruction *visitFPExt(CastInst &CI);
271 Instruction *visitFPToUI(FPToUIInst &FI);
272 Instruction *visitFPToSI(FPToSIInst &FI);
273 Instruction *visitUIToFP(CastInst &CI);
274 Instruction *visitSIToFP(CastInst &CI);
275 Instruction *visitPtrToInt(PtrToIntInst &CI);
276 Instruction *visitIntToPtr(IntToPtrInst &CI);
277 Instruction *visitBitCast(BitCastInst &CI);
278 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
280 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
281 Instruction *visitSelectInst(SelectInst &SI);
282 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
283 Instruction *visitCallInst(CallInst &CI);
284 Instruction *visitInvokeInst(InvokeInst &II);
285 Instruction *visitPHINode(PHINode &PN);
286 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
287 Instruction *visitAllocaInst(AllocaInst &AI);
288 Instruction *visitFreeInst(FreeInst &FI);
289 Instruction *visitFree(Instruction &FI);
290 Instruction *visitLoadInst(LoadInst &LI);
291 Instruction *visitStoreInst(StoreInst &SI);
292 Instruction *visitBranchInst(BranchInst &BI);
293 Instruction *visitSwitchInst(SwitchInst &SI);
294 Instruction *visitInsertElementInst(InsertElementInst &IE);
295 Instruction *visitExtractElementInst(ExtractElementInst &EI);
296 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
297 Instruction *visitExtractValueInst(ExtractValueInst &EV);
299 // visitInstruction - Specify what to return for unhandled instructions...
300 Instruction *visitInstruction(Instruction &I) { return 0; }
303 Instruction *visitCallSite(CallSite CS);
304 bool transformConstExprCastCall(CallSite CS);
305 Instruction *transformCallThroughTrampoline(CallSite CS);
306 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
307 bool DoXform = true);
308 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
309 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
313 // InsertNewInstBefore - insert an instruction New before instruction Old
314 // in the program. Add the new instruction to the worklist.
316 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
317 assert(New && New->getParent() == 0 &&
318 "New instruction already inserted into a basic block!");
319 BasicBlock *BB = Old.getParent();
320 BB->getInstList().insert(&Old, New); // Insert inst
325 // ReplaceInstUsesWith - This method is to be used when an instruction is
326 // found to be dead, replacable with another preexisting expression. Here
327 // we add all uses of I to the worklist, replace all uses of I with the new
328 // value, then return I, so that the inst combiner will know that I was
331 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
332 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
334 // If we are replacing the instruction with itself, this must be in a
335 // segment of unreachable code, so just clobber the instruction.
337 V = UndefValue::get(I.getType());
339 I.replaceAllUsesWith(V);
343 // EraseInstFromFunction - When dealing with an instruction that has side
344 // effects or produces a void value, we can't rely on DCE to delete the
345 // instruction. Instead, visit methods should return the value returned by
347 Instruction *EraseInstFromFunction(Instruction &I) {
348 DEBUG(errs() << "IC: ERASE " << I << '\n');
350 assert(I.use_empty() && "Cannot erase instruction that is used!");
351 // Make sure that we reprocess all operands now that we reduced their
353 if (I.getNumOperands() < 8) {
354 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
355 if (Instruction *Op = dyn_cast<Instruction>(*i))
361 return 0; // Don't do anything with FI
364 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
365 APInt &KnownOne, unsigned Depth = 0) const {
366 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
369 bool MaskedValueIsZero(Value *V, const APInt &Mask,
370 unsigned Depth = 0) const {
371 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
373 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
374 return llvm::ComputeNumSignBits(Op, TD, Depth);
379 /// SimplifyCommutative - This performs a few simplifications for
380 /// commutative operators.
381 bool SimplifyCommutative(BinaryOperator &I);
383 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
384 /// most-complex to least-complex order.
385 bool SimplifyCompare(CmpInst &I);
387 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
388 /// based on the demanded bits.
389 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
390 APInt& KnownZero, APInt& KnownOne,
392 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
393 APInt& KnownZero, APInt& KnownOne,
396 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
397 /// SimplifyDemandedBits knows about. See if the instruction has any
398 /// properties that allow us to simplify its operands.
399 bool SimplifyDemandedInstructionBits(Instruction &Inst);
401 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
402 APInt& UndefElts, unsigned Depth = 0);
404 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
405 // which has a PHI node as operand #0, see if we can fold the instruction
406 // into the PHI (which is only possible if all operands to the PHI are
409 // If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
410 // that would normally be unprofitable because they strongly encourage jump
412 Instruction *FoldOpIntoPhi(Instruction &I, bool AllowAggressive = false);
414 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
415 // operator and they all are only used by the PHI, PHI together their
416 // inputs, and do the operation once, to the result of the PHI.
417 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
418 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
419 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
422 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
423 ConstantInt *AndRHS, BinaryOperator &TheAnd);
425 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
426 bool isSub, Instruction &I);
427 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
428 bool isSigned, bool Inside, Instruction &IB);
429 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
430 Instruction *MatchBSwap(BinaryOperator &I);
431 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
432 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
433 Instruction *SimplifyMemSet(MemSetInst *MI);
436 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
438 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
439 unsigned CastOpc, int &NumCastsRemoved);
440 unsigned GetOrEnforceKnownAlignment(Value *V,
441 unsigned PrefAlign = 0);
444 } // end anonymous namespace
446 char InstCombiner::ID = 0;
447 static RegisterPass<InstCombiner>
448 X("instcombine", "Combine redundant instructions");
450 // getComplexity: Assign a complexity or rank value to LLVM Values...
451 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
452 static unsigned getComplexity(Value *V) {
453 if (isa<Instruction>(V)) {
454 if (BinaryOperator::isNeg(V) ||
455 BinaryOperator::isFNeg(V) ||
456 BinaryOperator::isNot(V))
460 if (isa<Argument>(V)) return 3;
461 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
464 // isOnlyUse - Return true if this instruction will be deleted if we stop using
466 static bool isOnlyUse(Value *V) {
467 return V->hasOneUse() || isa<Constant>(V);
470 // getPromotedType - Return the specified type promoted as it would be to pass
471 // though a va_arg area...
472 static const Type *getPromotedType(const Type *Ty) {
473 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
474 if (ITy->getBitWidth() < 32)
475 return Type::getInt32Ty(Ty->getContext());
480 /// getBitCastOperand - If the specified operand is a CastInst, a constant
481 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
482 /// operand value, otherwise return null.
483 static Value *getBitCastOperand(Value *V) {
484 if (Operator *O = dyn_cast<Operator>(V)) {
485 if (O->getOpcode() == Instruction::BitCast)
486 return O->getOperand(0);
487 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
488 if (GEP->hasAllZeroIndices())
489 return GEP->getPointerOperand();
494 /// This function is a wrapper around CastInst::isEliminableCastPair. It
495 /// simply extracts arguments and returns what that function returns.
496 static Instruction::CastOps
497 isEliminableCastPair(
498 const CastInst *CI, ///< The first cast instruction
499 unsigned opcode, ///< The opcode of the second cast instruction
500 const Type *DstTy, ///< The target type for the second cast instruction
501 TargetData *TD ///< The target data for pointer size
504 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
505 const Type *MidTy = CI->getType(); // B from above
507 // Get the opcodes of the two Cast instructions
508 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
509 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
511 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
513 TD ? TD->getIntPtrType(CI->getContext()) : 0);
515 // We don't want to form an inttoptr or ptrtoint that converts to an integer
516 // type that differs from the pointer size.
517 if ((Res == Instruction::IntToPtr &&
518 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
519 (Res == Instruction::PtrToInt &&
520 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
523 return Instruction::CastOps(Res);
526 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
527 /// in any code being generated. It does not require codegen if V is simple
528 /// enough or if the cast can be folded into other casts.
529 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
530 const Type *Ty, TargetData *TD) {
531 if (V->getType() == Ty || isa<Constant>(V)) return false;
533 // If this is another cast that can be eliminated, it isn't codegen either.
534 if (const CastInst *CI = dyn_cast<CastInst>(V))
535 if (isEliminableCastPair(CI, opcode, Ty, TD))
540 // SimplifyCommutative - This performs a few simplifications for commutative
543 // 1. Order operands such that they are listed from right (least complex) to
544 // left (most complex). This puts constants before unary operators before
547 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
548 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
550 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
551 bool Changed = false;
552 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
553 Changed = !I.swapOperands();
555 if (!I.isAssociative()) return Changed;
556 Instruction::BinaryOps Opcode = I.getOpcode();
557 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
558 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
559 if (isa<Constant>(I.getOperand(1))) {
560 Constant *Folded = ConstantExpr::get(I.getOpcode(),
561 cast<Constant>(I.getOperand(1)),
562 cast<Constant>(Op->getOperand(1)));
563 I.setOperand(0, Op->getOperand(0));
564 I.setOperand(1, Folded);
566 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
567 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
568 isOnlyUse(Op) && isOnlyUse(Op1)) {
569 Constant *C1 = cast<Constant>(Op->getOperand(1));
570 Constant *C2 = cast<Constant>(Op1->getOperand(1));
572 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
573 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
574 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
578 I.setOperand(0, New);
579 I.setOperand(1, Folded);
586 /// SimplifyCompare - For a CmpInst this function just orders the operands
587 /// so that theyare listed from right (least complex) to left (most complex).
588 /// This puts constants before unary operators before binary operators.
589 bool InstCombiner::SimplifyCompare(CmpInst &I) {
590 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
593 // Compare instructions are not associative so there's nothing else we can do.
597 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
598 // if the LHS is a constant zero (which is the 'negate' form).
600 static inline Value *dyn_castNegVal(Value *V) {
601 if (BinaryOperator::isNeg(V))
602 return BinaryOperator::getNegArgument(V);
604 // Constants can be considered to be negated values if they can be folded.
605 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
606 return ConstantExpr::getNeg(C);
608 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
609 if (C->getType()->getElementType()->isInteger())
610 return ConstantExpr::getNeg(C);
615 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
616 // instruction if the LHS is a constant negative zero (which is the 'negate'
619 static inline Value *dyn_castFNegVal(Value *V) {
620 if (BinaryOperator::isFNeg(V))
621 return BinaryOperator::getFNegArgument(V);
623 // Constants can be considered to be negated values if they can be folded.
624 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
625 return ConstantExpr::getFNeg(C);
627 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
628 if (C->getType()->getElementType()->isFloatingPoint())
629 return ConstantExpr::getFNeg(C);
634 static inline Value *dyn_castNotVal(Value *V) {
635 if (BinaryOperator::isNot(V))
636 return BinaryOperator::getNotArgument(V);
638 // Constants can be considered to be not'ed values...
639 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
640 return ConstantInt::get(C->getType(), ~C->getValue());
644 // dyn_castFoldableMul - If this value is a multiply that can be folded into
645 // other computations (because it has a constant operand), return the
646 // non-constant operand of the multiply, and set CST to point to the multiplier.
647 // Otherwise, return null.
649 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
650 if (V->hasOneUse() && V->getType()->isInteger())
651 if (Instruction *I = dyn_cast<Instruction>(V)) {
652 if (I->getOpcode() == Instruction::Mul)
653 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
654 return I->getOperand(0);
655 if (I->getOpcode() == Instruction::Shl)
656 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
657 // The multiplier is really 1 << CST.
658 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
659 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
660 CST = ConstantInt::get(V->getType()->getContext(),
661 APInt(BitWidth, 1).shl(CSTVal));
662 return I->getOperand(0);
668 /// AddOne - Add one to a ConstantInt
669 static Constant *AddOne(Constant *C) {
670 return ConstantExpr::getAdd(C,
671 ConstantInt::get(C->getType(), 1));
673 /// SubOne - Subtract one from a ConstantInt
674 static Constant *SubOne(ConstantInt *C) {
675 return ConstantExpr::getSub(C,
676 ConstantInt::get(C->getType(), 1));
678 /// MultiplyOverflows - True if the multiply can not be expressed in an int
680 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
681 uint32_t W = C1->getBitWidth();
682 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
691 APInt MulExt = LHSExt * RHSExt;
694 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
695 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
696 return MulExt.slt(Min) || MulExt.sgt(Max);
698 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
702 /// ShrinkDemandedConstant - Check to see if the specified operand of the
703 /// specified instruction is a constant integer. If so, check to see if there
704 /// are any bits set in the constant that are not demanded. If so, shrink the
705 /// constant and return true.
706 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
708 assert(I && "No instruction?");
709 assert(OpNo < I->getNumOperands() && "Operand index too large");
711 // If the operand is not a constant integer, nothing to do.
712 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
713 if (!OpC) return false;
715 // If there are no bits set that aren't demanded, nothing to do.
716 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
717 if ((~Demanded & OpC->getValue()) == 0)
720 // This instruction is producing bits that are not demanded. Shrink the RHS.
721 Demanded &= OpC->getValue();
722 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
726 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
727 // set of known zero and one bits, compute the maximum and minimum values that
728 // could have the specified known zero and known one bits, returning them in
730 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
731 const APInt& KnownOne,
732 APInt& Min, APInt& Max) {
733 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
734 KnownZero.getBitWidth() == Min.getBitWidth() &&
735 KnownZero.getBitWidth() == Max.getBitWidth() &&
736 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
737 APInt UnknownBits = ~(KnownZero|KnownOne);
739 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
740 // bit if it is unknown.
742 Max = KnownOne|UnknownBits;
744 if (UnknownBits.isNegative()) { // Sign bit is unknown
745 Min.set(Min.getBitWidth()-1);
746 Max.clear(Max.getBitWidth()-1);
750 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
751 // a set of known zero and one bits, compute the maximum and minimum values that
752 // could have the specified known zero and known one bits, returning them in
754 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
755 const APInt &KnownOne,
756 APInt &Min, APInt &Max) {
757 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
758 KnownZero.getBitWidth() == Min.getBitWidth() &&
759 KnownZero.getBitWidth() == Max.getBitWidth() &&
760 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
761 APInt UnknownBits = ~(KnownZero|KnownOne);
763 // The minimum value is when the unknown bits are all zeros.
765 // The maximum value is when the unknown bits are all ones.
766 Max = KnownOne|UnknownBits;
769 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
770 /// SimplifyDemandedBits knows about. See if the instruction has any
771 /// properties that allow us to simplify its operands.
772 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
773 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
774 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
775 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
777 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
778 KnownZero, KnownOne, 0);
779 if (V == 0) return false;
780 if (V == &Inst) return true;
781 ReplaceInstUsesWith(Inst, V);
785 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
786 /// specified instruction operand if possible, updating it in place. It returns
787 /// true if it made any change and false otherwise.
788 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
789 APInt &KnownZero, APInt &KnownOne,
791 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
792 KnownZero, KnownOne, Depth);
793 if (NewVal == 0) return false;
799 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
800 /// value based on the demanded bits. When this function is called, it is known
801 /// that only the bits set in DemandedMask of the result of V are ever used
802 /// downstream. Consequently, depending on the mask and V, it may be possible
803 /// to replace V with a constant or one of its operands. In such cases, this
804 /// function does the replacement and returns true. In all other cases, it
805 /// returns false after analyzing the expression and setting KnownOne and known
806 /// to be one in the expression. KnownZero contains all the bits that are known
807 /// to be zero in the expression. These are provided to potentially allow the
808 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
809 /// the expression. KnownOne and KnownZero always follow the invariant that
810 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
811 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
812 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
813 /// and KnownOne must all be the same.
815 /// This returns null if it did not change anything and it permits no
816 /// simplification. This returns V itself if it did some simplification of V's
817 /// operands based on the information about what bits are demanded. This returns
818 /// some other non-null value if it found out that V is equal to another value
819 /// in the context where the specified bits are demanded, but not for all users.
820 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
821 APInt &KnownZero, APInt &KnownOne,
823 assert(V != 0 && "Null pointer of Value???");
824 assert(Depth <= 6 && "Limit Search Depth");
825 uint32_t BitWidth = DemandedMask.getBitWidth();
826 const Type *VTy = V->getType();
827 assert((TD || !isa<PointerType>(VTy)) &&
828 "SimplifyDemandedBits needs to know bit widths!");
829 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
830 (!VTy->isIntOrIntVector() ||
831 VTy->getScalarSizeInBits() == BitWidth) &&
832 KnownZero.getBitWidth() == BitWidth &&
833 KnownOne.getBitWidth() == BitWidth &&
834 "Value *V, DemandedMask, KnownZero and KnownOne "
835 "must have same BitWidth");
836 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
837 // We know all of the bits for a constant!
838 KnownOne = CI->getValue() & DemandedMask;
839 KnownZero = ~KnownOne & DemandedMask;
842 if (isa<ConstantPointerNull>(V)) {
843 // We know all of the bits for a constant!
845 KnownZero = DemandedMask;
851 if (DemandedMask == 0) { // Not demanding any bits from V.
852 if (isa<UndefValue>(V))
854 return UndefValue::get(VTy);
857 if (Depth == 6) // Limit search depth.
860 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
861 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
863 Instruction *I = dyn_cast<Instruction>(V);
865 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
866 return 0; // Only analyze instructions.
869 // If there are multiple uses of this value and we aren't at the root, then
870 // we can't do any simplifications of the operands, because DemandedMask
871 // only reflects the bits demanded by *one* of the users.
872 if (Depth != 0 && !I->hasOneUse()) {
873 // Despite the fact that we can't simplify this instruction in all User's
874 // context, we can at least compute the knownzero/knownone bits, and we can
875 // do simplifications that apply to *just* the one user if we know that
876 // this instruction has a simpler value in that context.
877 if (I->getOpcode() == Instruction::And) {
878 // If either the LHS or the RHS are Zero, the result is zero.
879 ComputeMaskedBits(I->getOperand(1), DemandedMask,
880 RHSKnownZero, RHSKnownOne, Depth+1);
881 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
882 LHSKnownZero, LHSKnownOne, Depth+1);
884 // If all of the demanded bits are known 1 on one side, return the other.
885 // These bits cannot contribute to the result of the 'and' in this
887 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
888 (DemandedMask & ~LHSKnownZero))
889 return I->getOperand(0);
890 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
891 (DemandedMask & ~RHSKnownZero))
892 return I->getOperand(1);
894 // If all of the demanded bits in the inputs are known zeros, return zero.
895 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
896 return Constant::getNullValue(VTy);
898 } else if (I->getOpcode() == Instruction::Or) {
899 // We can simplify (X|Y) -> X or Y in the user's context if we know that
900 // only bits from X or Y are demanded.
902 // If either the LHS or the RHS are One, the result is One.
903 ComputeMaskedBits(I->getOperand(1), DemandedMask,
904 RHSKnownZero, RHSKnownOne, Depth+1);
905 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
906 LHSKnownZero, LHSKnownOne, Depth+1);
908 // If all of the demanded bits are known zero on one side, return the
909 // other. These bits cannot contribute to the result of the 'or' in this
911 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
912 (DemandedMask & ~LHSKnownOne))
913 return I->getOperand(0);
914 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
915 (DemandedMask & ~RHSKnownOne))
916 return I->getOperand(1);
918 // If all of the potentially set bits on one side are known to be set on
919 // the other side, just use the 'other' side.
920 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
921 (DemandedMask & (~RHSKnownZero)))
922 return I->getOperand(0);
923 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
924 (DemandedMask & (~LHSKnownZero)))
925 return I->getOperand(1);
928 // Compute the KnownZero/KnownOne bits to simplify things downstream.
929 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
933 // If this is the root being simplified, allow it to have multiple uses,
934 // just set the DemandedMask to all bits so that we can try to simplify the
935 // operands. This allows visitTruncInst (for example) to simplify the
936 // operand of a trunc without duplicating all the logic below.
937 if (Depth == 0 && !V->hasOneUse())
938 DemandedMask = APInt::getAllOnesValue(BitWidth);
940 switch (I->getOpcode()) {
942 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
944 case Instruction::And:
945 // If either the LHS or the RHS are Zero, the result is zero.
946 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
947 RHSKnownZero, RHSKnownOne, Depth+1) ||
948 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
949 LHSKnownZero, LHSKnownOne, Depth+1))
951 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
952 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
954 // If all of the demanded bits are known 1 on one side, return the other.
955 // These bits cannot contribute to the result of the 'and'.
956 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
957 (DemandedMask & ~LHSKnownZero))
958 return I->getOperand(0);
959 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
960 (DemandedMask & ~RHSKnownZero))
961 return I->getOperand(1);
963 // If all of the demanded bits in the inputs are known zeros, return zero.
964 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
965 return Constant::getNullValue(VTy);
967 // If the RHS is a constant, see if we can simplify it.
968 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
971 // Output known-1 bits are only known if set in both the LHS & RHS.
972 RHSKnownOne &= LHSKnownOne;
973 // Output known-0 are known to be clear if zero in either the LHS | RHS.
974 RHSKnownZero |= LHSKnownZero;
976 case Instruction::Or:
977 // If either the LHS or the RHS are One, the result is One.
978 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
979 RHSKnownZero, RHSKnownOne, Depth+1) ||
980 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
981 LHSKnownZero, LHSKnownOne, Depth+1))
983 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
984 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
986 // If all of the demanded bits are known zero on one side, return the other.
987 // These bits cannot contribute to the result of the 'or'.
988 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
989 (DemandedMask & ~LHSKnownOne))
990 return I->getOperand(0);
991 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
992 (DemandedMask & ~RHSKnownOne))
993 return I->getOperand(1);
995 // If all of the potentially set bits on one side are known to be set on
996 // the other side, just use the 'other' side.
997 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
998 (DemandedMask & (~RHSKnownZero)))
999 return I->getOperand(0);
1000 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1001 (DemandedMask & (~LHSKnownZero)))
1002 return I->getOperand(1);
1004 // If the RHS is a constant, see if we can simplify it.
1005 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1008 // Output known-0 bits are only known if clear in both the LHS & RHS.
1009 RHSKnownZero &= LHSKnownZero;
1010 // Output known-1 are known to be set if set in either the LHS | RHS.
1011 RHSKnownOne |= LHSKnownOne;
1013 case Instruction::Xor: {
1014 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1015 RHSKnownZero, RHSKnownOne, Depth+1) ||
1016 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1017 LHSKnownZero, LHSKnownOne, Depth+1))
1019 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1020 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1022 // If all of the demanded bits are known zero on one side, return the other.
1023 // These bits cannot contribute to the result of the 'xor'.
1024 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1025 return I->getOperand(0);
1026 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1027 return I->getOperand(1);
1029 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1030 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1031 (RHSKnownOne & LHSKnownOne);
1032 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1033 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1034 (RHSKnownOne & LHSKnownZero);
1036 // If all of the demanded bits are known to be zero on one side or the
1037 // other, turn this into an *inclusive* or.
1038 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1039 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1041 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1043 return InsertNewInstBefore(Or, *I);
1046 // If all of the demanded bits on one side are known, and all of the set
1047 // bits on that side are also known to be set on the other side, turn this
1048 // into an AND, as we know the bits will be cleared.
1049 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1050 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1052 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1053 Constant *AndC = Constant::getIntegerValue(VTy,
1054 ~RHSKnownOne & DemandedMask);
1056 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1057 return InsertNewInstBefore(And, *I);
1061 // If the RHS is a constant, see if we can simplify it.
1062 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1063 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1066 // If our LHS is an 'and' and if it has one use, and if any of the bits we
1067 // are flipping are known to be set, then the xor is just resetting those
1068 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
1069 // simplifying both of them.
1070 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0)))
1071 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
1072 isa<ConstantInt>(I->getOperand(1)) &&
1073 isa<ConstantInt>(LHSInst->getOperand(1)) &&
1074 (LHSKnownOne & RHSKnownOne & DemandedMask) != 0) {
1075 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
1076 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
1077 APInt NewMask = ~(LHSKnownOne & RHSKnownOne & DemandedMask);
1080 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
1081 Instruction *NewAnd =
1082 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1083 InsertNewInstBefore(NewAnd, *I);
1086 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
1087 Instruction *NewXor =
1088 BinaryOperator::CreateXor(NewAnd, XorC, "tmp");
1089 return InsertNewInstBefore(NewXor, *I);
1093 RHSKnownZero = KnownZeroOut;
1094 RHSKnownOne = KnownOneOut;
1097 case Instruction::Select:
1098 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1099 RHSKnownZero, RHSKnownOne, Depth+1) ||
1100 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1101 LHSKnownZero, LHSKnownOne, Depth+1))
1103 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1104 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1106 // If the operands are constants, see if we can simplify them.
1107 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1108 ShrinkDemandedConstant(I, 2, DemandedMask))
1111 // Only known if known in both the LHS and RHS.
1112 RHSKnownOne &= LHSKnownOne;
1113 RHSKnownZero &= LHSKnownZero;
1115 case Instruction::Trunc: {
1116 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1117 DemandedMask.zext(truncBf);
1118 RHSKnownZero.zext(truncBf);
1119 RHSKnownOne.zext(truncBf);
1120 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1121 RHSKnownZero, RHSKnownOne, Depth+1))
1123 DemandedMask.trunc(BitWidth);
1124 RHSKnownZero.trunc(BitWidth);
1125 RHSKnownOne.trunc(BitWidth);
1126 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1129 case Instruction::BitCast:
1130 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1131 return false; // vector->int or fp->int?
1133 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1134 if (const VectorType *SrcVTy =
1135 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1136 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1137 // Don't touch a bitcast between vectors of different element counts.
1140 // Don't touch a scalar-to-vector bitcast.
1142 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1143 // Don't touch a vector-to-scalar bitcast.
1146 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1147 RHSKnownZero, RHSKnownOne, Depth+1))
1149 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1151 case Instruction::ZExt: {
1152 // Compute the bits in the result that are not present in the input.
1153 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1155 DemandedMask.trunc(SrcBitWidth);
1156 RHSKnownZero.trunc(SrcBitWidth);
1157 RHSKnownOne.trunc(SrcBitWidth);
1158 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1159 RHSKnownZero, RHSKnownOne, Depth+1))
1161 DemandedMask.zext(BitWidth);
1162 RHSKnownZero.zext(BitWidth);
1163 RHSKnownOne.zext(BitWidth);
1164 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1165 // The top bits are known to be zero.
1166 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1169 case Instruction::SExt: {
1170 // Compute the bits in the result that are not present in the input.
1171 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1173 APInt InputDemandedBits = DemandedMask &
1174 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1176 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1177 // If any of the sign extended bits are demanded, we know that the sign
1179 if ((NewBits & DemandedMask) != 0)
1180 InputDemandedBits.set(SrcBitWidth-1);
1182 InputDemandedBits.trunc(SrcBitWidth);
1183 RHSKnownZero.trunc(SrcBitWidth);
1184 RHSKnownOne.trunc(SrcBitWidth);
1185 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1186 RHSKnownZero, RHSKnownOne, Depth+1))
1188 InputDemandedBits.zext(BitWidth);
1189 RHSKnownZero.zext(BitWidth);
1190 RHSKnownOne.zext(BitWidth);
1191 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1193 // If the sign bit of the input is known set or clear, then we know the
1194 // top bits of the result.
1196 // If the input sign bit is known zero, or if the NewBits are not demanded
1197 // convert this into a zero extension.
1198 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1199 // Convert to ZExt cast
1200 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1201 return InsertNewInstBefore(NewCast, *I);
1202 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1203 RHSKnownOne |= NewBits;
1207 case Instruction::Add: {
1208 // Figure out what the input bits are. If the top bits of the and result
1209 // are not demanded, then the add doesn't demand them from its input
1211 unsigned NLZ = DemandedMask.countLeadingZeros();
1213 // If there is a constant on the RHS, there are a variety of xformations
1215 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1216 // If null, this should be simplified elsewhere. Some of the xforms here
1217 // won't work if the RHS is zero.
1221 // If the top bit of the output is demanded, demand everything from the
1222 // input. Otherwise, we demand all the input bits except NLZ top bits.
1223 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1225 // Find information about known zero/one bits in the input.
1226 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1227 LHSKnownZero, LHSKnownOne, Depth+1))
1230 // If the RHS of the add has bits set that can't affect the input, reduce
1232 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1235 // Avoid excess work.
1236 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1239 // Turn it into OR if input bits are zero.
1240 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1242 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1244 return InsertNewInstBefore(Or, *I);
1247 // We can say something about the output known-zero and known-one bits,
1248 // depending on potential carries from the input constant and the
1249 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1250 // bits set and the RHS constant is 0x01001, then we know we have a known
1251 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1253 // To compute this, we first compute the potential carry bits. These are
1254 // the bits which may be modified. I'm not aware of a better way to do
1256 const APInt &RHSVal = RHS->getValue();
1257 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1259 // Now that we know which bits have carries, compute the known-1/0 sets.
1261 // Bits are known one if they are known zero in one operand and one in the
1262 // other, and there is no input carry.
1263 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1264 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1266 // Bits are known zero if they are known zero in both operands and there
1267 // is no input carry.
1268 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1270 // If the high-bits of this ADD are not demanded, then it does not demand
1271 // the high bits of its LHS or RHS.
1272 if (DemandedMask[BitWidth-1] == 0) {
1273 // Right fill the mask of bits for this ADD to demand the most
1274 // significant bit and all those below it.
1275 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1276 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1277 LHSKnownZero, LHSKnownOne, Depth+1) ||
1278 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1279 LHSKnownZero, LHSKnownOne, Depth+1))
1285 case Instruction::Sub:
1286 // If the high-bits of this SUB are not demanded, then it does not demand
1287 // the high bits of its LHS or RHS.
1288 if (DemandedMask[BitWidth-1] == 0) {
1289 // Right fill the mask of bits for this SUB to demand the most
1290 // significant bit and all those below it.
1291 uint32_t NLZ = DemandedMask.countLeadingZeros();
1292 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1293 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1294 LHSKnownZero, LHSKnownOne, Depth+1) ||
1295 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1296 LHSKnownZero, LHSKnownOne, Depth+1))
1299 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1300 // the known zeros and ones.
1301 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1303 case Instruction::Shl:
1304 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1305 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1306 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1307 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1308 RHSKnownZero, RHSKnownOne, Depth+1))
1310 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1311 RHSKnownZero <<= ShiftAmt;
1312 RHSKnownOne <<= ShiftAmt;
1313 // low bits known zero.
1315 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1318 case Instruction::LShr:
1319 // For a logical shift right
1320 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1321 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1323 // Unsigned shift right.
1324 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1325 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1326 RHSKnownZero, RHSKnownOne, Depth+1))
1328 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1329 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1330 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1332 // Compute the new bits that are at the top now.
1333 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1334 RHSKnownZero |= HighBits; // high bits known zero.
1338 case Instruction::AShr:
1339 // If this is an arithmetic shift right and only the low-bit is set, we can
1340 // always convert this into a logical shr, even if the shift amount is
1341 // variable. The low bit of the shift cannot be an input sign bit unless
1342 // the shift amount is >= the size of the datatype, which is undefined.
1343 if (DemandedMask == 1) {
1344 // Perform the logical shift right.
1345 Instruction *NewVal = BinaryOperator::CreateLShr(
1346 I->getOperand(0), I->getOperand(1), I->getName());
1347 return InsertNewInstBefore(NewVal, *I);
1350 // If the sign bit is the only bit demanded by this ashr, then there is no
1351 // need to do it, the shift doesn't change the high bit.
1352 if (DemandedMask.isSignBit())
1353 return I->getOperand(0);
1355 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1356 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1358 // Signed shift right.
1359 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1360 // If any of the "high bits" are demanded, we should set the sign bit as
1362 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1363 DemandedMaskIn.set(BitWidth-1);
1364 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1365 RHSKnownZero, RHSKnownOne, Depth+1))
1367 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1368 // Compute the new bits that are at the top now.
1369 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1370 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1371 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1373 // Handle the sign bits.
1374 APInt SignBit(APInt::getSignBit(BitWidth));
1375 // Adjust to where it is now in the mask.
1376 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1378 // If the input sign bit is known to be zero, or if none of the top bits
1379 // are demanded, turn this into an unsigned shift right.
1380 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1381 (HighBits & ~DemandedMask) == HighBits) {
1382 // Perform the logical shift right.
1383 Instruction *NewVal = BinaryOperator::CreateLShr(
1384 I->getOperand(0), SA, I->getName());
1385 return InsertNewInstBefore(NewVal, *I);
1386 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1387 RHSKnownOne |= HighBits;
1391 case Instruction::SRem:
1392 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1393 APInt RA = Rem->getValue().abs();
1394 if (RA.isPowerOf2()) {
1395 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1396 return I->getOperand(0);
1398 APInt LowBits = RA - 1;
1399 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1400 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1401 LHSKnownZero, LHSKnownOne, Depth+1))
1404 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1405 LHSKnownZero |= ~LowBits;
1407 KnownZero |= LHSKnownZero & DemandedMask;
1409 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1413 case Instruction::URem: {
1414 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1415 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1416 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1417 KnownZero2, KnownOne2, Depth+1) ||
1418 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1419 KnownZero2, KnownOne2, Depth+1))
1422 unsigned Leaders = KnownZero2.countLeadingOnes();
1423 Leaders = std::max(Leaders,
1424 KnownZero2.countLeadingOnes());
1425 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1428 case Instruction::Call:
1429 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1430 switch (II->getIntrinsicID()) {
1432 case Intrinsic::bswap: {
1433 // If the only bits demanded come from one byte of the bswap result,
1434 // just shift the input byte into position to eliminate the bswap.
1435 unsigned NLZ = DemandedMask.countLeadingZeros();
1436 unsigned NTZ = DemandedMask.countTrailingZeros();
1438 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1439 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1440 // have 14 leading zeros, round to 8.
1443 // If we need exactly one byte, we can do this transformation.
1444 if (BitWidth-NLZ-NTZ == 8) {
1445 unsigned ResultBit = NTZ;
1446 unsigned InputBit = BitWidth-NTZ-8;
1448 // Replace this with either a left or right shift to get the byte into
1450 Instruction *NewVal;
1451 if (InputBit > ResultBit)
1452 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1453 ConstantInt::get(I->getType(), InputBit-ResultBit));
1455 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1456 ConstantInt::get(I->getType(), ResultBit-InputBit));
1457 NewVal->takeName(I);
1458 return InsertNewInstBefore(NewVal, *I);
1461 // TODO: Could compute known zero/one bits based on the input.
1466 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1470 // If the client is only demanding bits that we know, return the known
1472 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1473 return Constant::getIntegerValue(VTy, RHSKnownOne);
1478 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1479 /// any number of elements. DemandedElts contains the set of elements that are
1480 /// actually used by the caller. This method analyzes which elements of the
1481 /// operand are undef and returns that information in UndefElts.
1483 /// If the information about demanded elements can be used to simplify the
1484 /// operation, the operation is simplified, then the resultant value is
1485 /// returned. This returns null if no change was made.
1486 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1489 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1490 APInt EltMask(APInt::getAllOnesValue(VWidth));
1491 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1493 if (isa<UndefValue>(V)) {
1494 // If the entire vector is undefined, just return this info.
1495 UndefElts = EltMask;
1497 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1498 UndefElts = EltMask;
1499 return UndefValue::get(V->getType());
1503 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1504 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1505 Constant *Undef = UndefValue::get(EltTy);
1507 std::vector<Constant*> Elts;
1508 for (unsigned i = 0; i != VWidth; ++i)
1509 if (!DemandedElts[i]) { // If not demanded, set to undef.
1510 Elts.push_back(Undef);
1512 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1513 Elts.push_back(Undef);
1515 } else { // Otherwise, defined.
1516 Elts.push_back(CP->getOperand(i));
1519 // If we changed the constant, return it.
1520 Constant *NewCP = ConstantVector::get(Elts);
1521 return NewCP != CP ? NewCP : 0;
1522 } else if (isa<ConstantAggregateZero>(V)) {
1523 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1526 // Check if this is identity. If so, return 0 since we are not simplifying
1528 if (DemandedElts == ((1ULL << VWidth) -1))
1531 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1532 Constant *Zero = Constant::getNullValue(EltTy);
1533 Constant *Undef = UndefValue::get(EltTy);
1534 std::vector<Constant*> Elts;
1535 for (unsigned i = 0; i != VWidth; ++i) {
1536 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1537 Elts.push_back(Elt);
1539 UndefElts = DemandedElts ^ EltMask;
1540 return ConstantVector::get(Elts);
1543 // Limit search depth.
1547 // If multiple users are using the root value, procede with
1548 // simplification conservatively assuming that all elements
1550 if (!V->hasOneUse()) {
1551 // Quit if we find multiple users of a non-root value though.
1552 // They'll be handled when it's their turn to be visited by
1553 // the main instcombine process.
1555 // TODO: Just compute the UndefElts information recursively.
1558 // Conservatively assume that all elements are needed.
1559 DemandedElts = EltMask;
1562 Instruction *I = dyn_cast<Instruction>(V);
1563 if (!I) return 0; // Only analyze instructions.
1565 bool MadeChange = false;
1566 APInt UndefElts2(VWidth, 0);
1568 switch (I->getOpcode()) {
1571 case Instruction::InsertElement: {
1572 // If this is a variable index, we don't know which element it overwrites.
1573 // demand exactly the same input as we produce.
1574 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1576 // Note that we can't propagate undef elt info, because we don't know
1577 // which elt is getting updated.
1578 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1579 UndefElts2, Depth+1);
1580 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1584 // If this is inserting an element that isn't demanded, remove this
1586 unsigned IdxNo = Idx->getZExtValue();
1587 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1589 return I->getOperand(0);
1592 // Otherwise, the element inserted overwrites whatever was there, so the
1593 // input demanded set is simpler than the output set.
1594 APInt DemandedElts2 = DemandedElts;
1595 DemandedElts2.clear(IdxNo);
1596 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1597 UndefElts, Depth+1);
1598 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1600 // The inserted element is defined.
1601 UndefElts.clear(IdxNo);
1604 case Instruction::ShuffleVector: {
1605 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1606 uint64_t LHSVWidth =
1607 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1608 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1609 for (unsigned i = 0; i < VWidth; i++) {
1610 if (DemandedElts[i]) {
1611 unsigned MaskVal = Shuffle->getMaskValue(i);
1612 if (MaskVal != -1u) {
1613 assert(MaskVal < LHSVWidth * 2 &&
1614 "shufflevector mask index out of range!");
1615 if (MaskVal < LHSVWidth)
1616 LeftDemanded.set(MaskVal);
1618 RightDemanded.set(MaskVal - LHSVWidth);
1623 APInt UndefElts4(LHSVWidth, 0);
1624 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1625 UndefElts4, Depth+1);
1626 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1628 APInt UndefElts3(LHSVWidth, 0);
1629 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1630 UndefElts3, Depth+1);
1631 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1633 bool NewUndefElts = false;
1634 for (unsigned i = 0; i < VWidth; i++) {
1635 unsigned MaskVal = Shuffle->getMaskValue(i);
1636 if (MaskVal == -1u) {
1638 } else if (MaskVal < LHSVWidth) {
1639 if (UndefElts4[MaskVal]) {
1640 NewUndefElts = true;
1644 if (UndefElts3[MaskVal - LHSVWidth]) {
1645 NewUndefElts = true;
1652 // Add additional discovered undefs.
1653 std::vector<Constant*> Elts;
1654 for (unsigned i = 0; i < VWidth; ++i) {
1656 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1658 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1659 Shuffle->getMaskValue(i)));
1661 I->setOperand(2, ConstantVector::get(Elts));
1666 case Instruction::BitCast: {
1667 // Vector->vector casts only.
1668 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1670 unsigned InVWidth = VTy->getNumElements();
1671 APInt InputDemandedElts(InVWidth, 0);
1674 if (VWidth == InVWidth) {
1675 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1676 // elements as are demanded of us.
1678 InputDemandedElts = DemandedElts;
1679 } else if (VWidth > InVWidth) {
1683 // If there are more elements in the result than there are in the source,
1684 // then an input element is live if any of the corresponding output
1685 // elements are live.
1686 Ratio = VWidth/InVWidth;
1687 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1688 if (DemandedElts[OutIdx])
1689 InputDemandedElts.set(OutIdx/Ratio);
1695 // If there are more elements in the source than there are in the result,
1696 // then an input element is live if the corresponding output element is
1698 Ratio = InVWidth/VWidth;
1699 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1700 if (DemandedElts[InIdx/Ratio])
1701 InputDemandedElts.set(InIdx);
1704 // div/rem demand all inputs, because they don't want divide by zero.
1705 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1706 UndefElts2, Depth+1);
1708 I->setOperand(0, TmpV);
1712 UndefElts = UndefElts2;
1713 if (VWidth > InVWidth) {
1714 llvm_unreachable("Unimp");
1715 // If there are more elements in the result than there are in the source,
1716 // then an output element is undef if the corresponding input element is
1718 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1719 if (UndefElts2[OutIdx/Ratio])
1720 UndefElts.set(OutIdx);
1721 } else if (VWidth < InVWidth) {
1722 llvm_unreachable("Unimp");
1723 // If there are more elements in the source than there are in the result,
1724 // then a result element is undef if all of the corresponding input
1725 // elements are undef.
1726 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1727 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1728 if (!UndefElts2[InIdx]) // Not undef?
1729 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1733 case Instruction::And:
1734 case Instruction::Or:
1735 case Instruction::Xor:
1736 case Instruction::Add:
1737 case Instruction::Sub:
1738 case Instruction::Mul:
1739 // div/rem demand all inputs, because they don't want divide by zero.
1740 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1741 UndefElts, Depth+1);
1742 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1743 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1744 UndefElts2, Depth+1);
1745 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1747 // Output elements are undefined if both are undefined. Consider things
1748 // like undef&0. The result is known zero, not undef.
1749 UndefElts &= UndefElts2;
1752 case Instruction::Call: {
1753 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1755 switch (II->getIntrinsicID()) {
1758 // Binary vector operations that work column-wise. A dest element is a
1759 // function of the corresponding input elements from the two inputs.
1760 case Intrinsic::x86_sse_sub_ss:
1761 case Intrinsic::x86_sse_mul_ss:
1762 case Intrinsic::x86_sse_min_ss:
1763 case Intrinsic::x86_sse_max_ss:
1764 case Intrinsic::x86_sse2_sub_sd:
1765 case Intrinsic::x86_sse2_mul_sd:
1766 case Intrinsic::x86_sse2_min_sd:
1767 case Intrinsic::x86_sse2_max_sd:
1768 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1769 UndefElts, Depth+1);
1770 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1771 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1772 UndefElts2, Depth+1);
1773 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1775 // If only the low elt is demanded and this is a scalarizable intrinsic,
1776 // scalarize it now.
1777 if (DemandedElts == 1) {
1778 switch (II->getIntrinsicID()) {
1780 case Intrinsic::x86_sse_sub_ss:
1781 case Intrinsic::x86_sse_mul_ss:
1782 case Intrinsic::x86_sse2_sub_sd:
1783 case Intrinsic::x86_sse2_mul_sd:
1784 // TODO: Lower MIN/MAX/ABS/etc
1785 Value *LHS = II->getOperand(1);
1786 Value *RHS = II->getOperand(2);
1787 // Extract the element as scalars.
1788 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1789 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1790 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1791 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1793 switch (II->getIntrinsicID()) {
1794 default: llvm_unreachable("Case stmts out of sync!");
1795 case Intrinsic::x86_sse_sub_ss:
1796 case Intrinsic::x86_sse2_sub_sd:
1797 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1798 II->getName()), *II);
1800 case Intrinsic::x86_sse_mul_ss:
1801 case Intrinsic::x86_sse2_mul_sd:
1802 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1803 II->getName()), *II);
1808 InsertElementInst::Create(
1809 UndefValue::get(II->getType()), TmpV,
1810 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1811 InsertNewInstBefore(New, *II);
1816 // Output elements are undefined if both are undefined. Consider things
1817 // like undef&0. The result is known zero, not undef.
1818 UndefElts &= UndefElts2;
1824 return MadeChange ? I : 0;
1828 /// AssociativeOpt - Perform an optimization on an associative operator. This
1829 /// function is designed to check a chain of associative operators for a
1830 /// potential to apply a certain optimization. Since the optimization may be
1831 /// applicable if the expression was reassociated, this checks the chain, then
1832 /// reassociates the expression as necessary to expose the optimization
1833 /// opportunity. This makes use of a special Functor, which must define
1834 /// 'shouldApply' and 'apply' methods.
1836 template<typename Functor>
1837 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1838 unsigned Opcode = Root.getOpcode();
1839 Value *LHS = Root.getOperand(0);
1841 // Quick check, see if the immediate LHS matches...
1842 if (F.shouldApply(LHS))
1843 return F.apply(Root);
1845 // Otherwise, if the LHS is not of the same opcode as the root, return.
1846 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1847 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1848 // Should we apply this transform to the RHS?
1849 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1851 // If not to the RHS, check to see if we should apply to the LHS...
1852 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1853 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1857 // If the functor wants to apply the optimization to the RHS of LHSI,
1858 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1860 // Now all of the instructions are in the current basic block, go ahead
1861 // and perform the reassociation.
1862 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1864 // First move the selected RHS to the LHS of the root...
1865 Root.setOperand(0, LHSI->getOperand(1));
1867 // Make what used to be the LHS of the root be the user of the root...
1868 Value *ExtraOperand = TmpLHSI->getOperand(1);
1869 if (&Root == TmpLHSI) {
1870 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1873 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1874 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1875 BasicBlock::iterator ARI = &Root; ++ARI;
1876 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1879 // Now propagate the ExtraOperand down the chain of instructions until we
1881 while (TmpLHSI != LHSI) {
1882 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1883 // Move the instruction to immediately before the chain we are
1884 // constructing to avoid breaking dominance properties.
1885 NextLHSI->moveBefore(ARI);
1888 Value *NextOp = NextLHSI->getOperand(1);
1889 NextLHSI->setOperand(1, ExtraOperand);
1891 ExtraOperand = NextOp;
1894 // Now that the instructions are reassociated, have the functor perform
1895 // the transformation...
1896 return F.apply(Root);
1899 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1906 // AddRHS - Implements: X + X --> X << 1
1909 explicit AddRHS(Value *rhs) : RHS(rhs) {}
1910 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1911 Instruction *apply(BinaryOperator &Add) const {
1912 return BinaryOperator::CreateShl(Add.getOperand(0),
1913 ConstantInt::get(Add.getType(), 1));
1917 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1919 struct AddMaskingAnd {
1921 explicit AddMaskingAnd(Constant *c) : C2(c) {}
1922 bool shouldApply(Value *LHS) const {
1924 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1925 ConstantExpr::getAnd(C1, C2)->isNullValue();
1927 Instruction *apply(BinaryOperator &Add) const {
1928 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1934 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1936 if (CastInst *CI = dyn_cast<CastInst>(&I))
1937 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
1939 // Figure out if the constant is the left or the right argument.
1940 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1941 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1943 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1945 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1946 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1949 Value *Op0 = SO, *Op1 = ConstOperand;
1951 std::swap(Op0, Op1);
1953 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1954 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
1955 SO->getName()+".op");
1956 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
1957 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1958 SO->getName()+".cmp");
1959 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
1960 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1961 SO->getName()+".cmp");
1962 llvm_unreachable("Unknown binary instruction type!");
1965 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1966 // constant as the other operand, try to fold the binary operator into the
1967 // select arguments. This also works for Cast instructions, which obviously do
1968 // not have a second operand.
1969 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1971 // Don't modify shared select instructions
1972 if (!SI->hasOneUse()) return 0;
1973 Value *TV = SI->getOperand(1);
1974 Value *FV = SI->getOperand(2);
1976 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1977 // Bool selects with constant operands can be folded to logical ops.
1978 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
1980 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1981 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1983 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1990 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
1991 /// has a PHI node as operand #0, see if we can fold the instruction into the
1992 /// PHI (which is only possible if all operands to the PHI are constants).
1994 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
1995 /// that would normally be unprofitable because they strongly encourage jump
1997 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
1998 bool AllowAggressive) {
1999 AllowAggressive = false;
2000 PHINode *PN = cast<PHINode>(I.getOperand(0));
2001 unsigned NumPHIValues = PN->getNumIncomingValues();
2002 if (NumPHIValues == 0 ||
2003 // We normally only transform phis with a single use, unless we're trying
2004 // hard to make jump threading happen.
2005 (!PN->hasOneUse() && !AllowAggressive))
2009 // Check to see if all of the operands of the PHI are simple constants
2010 // (constantint/constantfp/undef). If there is one non-constant value,
2011 // remember the BB it is in. If there is more than one or if *it* is a PHI,
2012 // bail out. We don't do arbitrary constant expressions here because moving
2013 // their computation can be expensive without a cost model.
2014 BasicBlock *NonConstBB = 0;
2015 for (unsigned i = 0; i != NumPHIValues; ++i)
2016 if (!isa<Constant>(PN->getIncomingValue(i)) ||
2017 isa<ConstantExpr>(PN->getIncomingValue(i))) {
2018 if (NonConstBB) return 0; // More than one non-const value.
2019 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2020 NonConstBB = PN->getIncomingBlock(i);
2022 // If the incoming non-constant value is in I's block, we have an infinite
2024 if (NonConstBB == I.getParent())
2028 // If there is exactly one non-constant value, we can insert a copy of the
2029 // operation in that block. However, if this is a critical edge, we would be
2030 // inserting the computation one some other paths (e.g. inside a loop). Only
2031 // do this if the pred block is unconditionally branching into the phi block.
2032 if (NonConstBB != 0 && !AllowAggressive) {
2033 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2034 if (!BI || !BI->isUnconditional()) return 0;
2037 // Okay, we can do the transformation: create the new PHI node.
2038 PHINode *NewPN = PHINode::Create(I.getType(), "");
2039 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2040 InsertNewInstBefore(NewPN, *PN);
2041 NewPN->takeName(PN);
2043 // Next, add all of the operands to the PHI.
2044 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
2045 // We only currently try to fold the condition of a select when it is a phi,
2046 // not the true/false values.
2047 Value *TrueV = SI->getTrueValue();
2048 Value *FalseV = SI->getFalseValue();
2049 BasicBlock *PhiTransBB = PN->getParent();
2050 for (unsigned i = 0; i != NumPHIValues; ++i) {
2051 BasicBlock *ThisBB = PN->getIncomingBlock(i);
2052 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
2053 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
2055 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2056 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
2058 assert(PN->getIncomingBlock(i) == NonConstBB);
2059 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
2061 "phitmp", NonConstBB->getTerminator());
2062 Worklist.Add(cast<Instruction>(InV));
2064 NewPN->addIncoming(InV, ThisBB);
2066 } else if (I.getNumOperands() == 2) {
2067 Constant *C = cast<Constant>(I.getOperand(1));
2068 for (unsigned i = 0; i != NumPHIValues; ++i) {
2070 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2071 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2072 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2074 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2076 assert(PN->getIncomingBlock(i) == NonConstBB);
2077 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2078 InV = BinaryOperator::Create(BO->getOpcode(),
2079 PN->getIncomingValue(i), C, "phitmp",
2080 NonConstBB->getTerminator());
2081 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2082 InV = CmpInst::Create(CI->getOpcode(),
2084 PN->getIncomingValue(i), C, "phitmp",
2085 NonConstBB->getTerminator());
2087 llvm_unreachable("Unknown binop!");
2089 Worklist.Add(cast<Instruction>(InV));
2091 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2094 CastInst *CI = cast<CastInst>(&I);
2095 const Type *RetTy = CI->getType();
2096 for (unsigned i = 0; i != NumPHIValues; ++i) {
2098 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2099 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2101 assert(PN->getIncomingBlock(i) == NonConstBB);
2102 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2103 I.getType(), "phitmp",
2104 NonConstBB->getTerminator());
2105 Worklist.Add(cast<Instruction>(InV));
2107 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2110 return ReplaceInstUsesWith(I, NewPN);
2114 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2115 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2116 /// This basically requires proving that the add in the original type would not
2117 /// overflow to change the sign bit or have a carry out.
2118 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2119 // There are different heuristics we can use for this. Here are some simple
2122 // Add has the property that adding any two 2's complement numbers can only
2123 // have one carry bit which can change a sign. As such, if LHS and RHS each
2124 // have at least two sign bits, we know that the addition of the two values will
2125 // sign extend fine.
2126 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2130 // If one of the operands only has one non-zero bit, and if the other operand
2131 // has a known-zero bit in a more significant place than it (not including the
2132 // sign bit) the ripple may go up to and fill the zero, but won't change the
2133 // sign. For example, (X & ~4) + 1.
2141 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2142 bool Changed = SimplifyCommutative(I);
2143 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2145 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2146 // X + undef -> undef
2147 if (isa<UndefValue>(RHS))
2148 return ReplaceInstUsesWith(I, RHS);
2151 if (RHSC->isNullValue())
2152 return ReplaceInstUsesWith(I, LHS);
2154 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2155 // X + (signbit) --> X ^ signbit
2156 const APInt& Val = CI->getValue();
2157 uint32_t BitWidth = Val.getBitWidth();
2158 if (Val == APInt::getSignBit(BitWidth))
2159 return BinaryOperator::CreateXor(LHS, RHS);
2161 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2162 // (X & 254)+1 -> (X&254)|1
2163 if (SimplifyDemandedInstructionBits(I))
2166 // zext(bool) + C -> bool ? C + 1 : C
2167 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2168 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2169 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2172 if (isa<PHINode>(LHS))
2173 if (Instruction *NV = FoldOpIntoPhi(I))
2176 ConstantInt *XorRHS = 0;
2178 if (isa<ConstantInt>(RHSC) &&
2179 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2180 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2181 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2183 uint32_t Size = TySizeBits / 2;
2184 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2185 APInt CFF80Val(-C0080Val);
2187 if (TySizeBits > Size) {
2188 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2189 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2190 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2191 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2192 // This is a sign extend if the top bits are known zero.
2193 if (!MaskedValueIsZero(XorLHS,
2194 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2195 Size = 0; // Not a sign ext, but can't be any others either.
2200 C0080Val = APIntOps::lshr(C0080Val, Size);
2201 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2202 } while (Size >= 1);
2204 // FIXME: This shouldn't be necessary. When the backends can handle types
2205 // with funny bit widths then this switch statement should be removed. It
2206 // is just here to get the size of the "middle" type back up to something
2207 // that the back ends can handle.
2208 const Type *MiddleType = 0;
2211 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2212 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2213 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2216 Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext");
2217 return new SExtInst(NewTrunc, I.getType(), I.getName());
2222 if (I.getType() == Type::getInt1Ty(*Context))
2223 return BinaryOperator::CreateXor(LHS, RHS);
2226 if (I.getType()->isInteger()) {
2227 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2230 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2231 if (RHSI->getOpcode() == Instruction::Sub)
2232 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2233 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2235 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2236 if (LHSI->getOpcode() == Instruction::Sub)
2237 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2238 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2243 // -A + -B --> -(A + B)
2244 if (Value *LHSV = dyn_castNegVal(LHS)) {
2245 if (LHS->getType()->isIntOrIntVector()) {
2246 if (Value *RHSV = dyn_castNegVal(RHS)) {
2247 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
2248 return BinaryOperator::CreateNeg(NewAdd);
2252 return BinaryOperator::CreateSub(RHS, LHSV);
2256 if (!isa<Constant>(RHS))
2257 if (Value *V = dyn_castNegVal(RHS))
2258 return BinaryOperator::CreateSub(LHS, V);
2262 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2263 if (X == RHS) // X*C + X --> X * (C+1)
2264 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2266 // X*C1 + X*C2 --> X * (C1+C2)
2268 if (X == dyn_castFoldableMul(RHS, C1))
2269 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2272 // X + X*C --> X * (C+1)
2273 if (dyn_castFoldableMul(RHS, C2) == LHS)
2274 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2276 // X + ~X --> -1 since ~X = -X-1
2277 if (dyn_castNotVal(LHS) == RHS ||
2278 dyn_castNotVal(RHS) == LHS)
2279 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2282 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2283 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2284 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2287 // A+B --> A|B iff A and B have no bits set in common.
2288 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2289 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2290 APInt LHSKnownOne(IT->getBitWidth(), 0);
2291 APInt LHSKnownZero(IT->getBitWidth(), 0);
2292 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2293 if (LHSKnownZero != 0) {
2294 APInt RHSKnownOne(IT->getBitWidth(), 0);
2295 APInt RHSKnownZero(IT->getBitWidth(), 0);
2296 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2298 // No bits in common -> bitwise or.
2299 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2300 return BinaryOperator::CreateOr(LHS, RHS);
2304 // W*X + Y*Z --> W * (X+Z) iff W == Y
2305 if (I.getType()->isIntOrIntVector()) {
2306 Value *W, *X, *Y, *Z;
2307 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2308 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2312 } else if (Y == X) {
2314 } else if (X == Z) {
2321 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
2322 return BinaryOperator::CreateMul(W, NewAdd);
2327 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2329 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2330 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2332 // (X & FF00) + xx00 -> (X+xx00) & FF00
2333 if (LHS->hasOneUse() &&
2334 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2335 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2336 if (Anded == CRHS) {
2337 // See if all bits from the first bit set in the Add RHS up are included
2338 // in the mask. First, get the rightmost bit.
2339 const APInt& AddRHSV = CRHS->getValue();
2341 // Form a mask of all bits from the lowest bit added through the top.
2342 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2344 // See if the and mask includes all of these bits.
2345 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2347 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2348 // Okay, the xform is safe. Insert the new add pronto.
2349 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
2350 return BinaryOperator::CreateAnd(NewAdd, C2);
2355 // Try to fold constant add into select arguments.
2356 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2357 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2361 // add (select X 0 (sub n A)) A --> select X A n
2363 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2366 SI = dyn_cast<SelectInst>(RHS);
2369 if (SI && SI->hasOneUse()) {
2370 Value *TV = SI->getTrueValue();
2371 Value *FV = SI->getFalseValue();
2374 // Can we fold the add into the argument of the select?
2375 // We check both true and false select arguments for a matching subtract.
2376 if (match(FV, m_Zero()) &&
2377 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2378 // Fold the add into the true select value.
2379 return SelectInst::Create(SI->getCondition(), N, A);
2380 if (match(TV, m_Zero()) &&
2381 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2382 // Fold the add into the false select value.
2383 return SelectInst::Create(SI->getCondition(), A, N);
2387 // Check for (add (sext x), y), see if we can merge this into an
2388 // integer add followed by a sext.
2389 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2390 // (add (sext x), cst) --> (sext (add x, cst'))
2391 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2393 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2394 if (LHSConv->hasOneUse() &&
2395 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2396 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2397 // Insert the new, smaller add.
2398 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2400 return new SExtInst(NewAdd, I.getType());
2404 // (add (sext x), (sext y)) --> (sext (add int x, y))
2405 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2406 // Only do this if x/y have the same type, if at last one of them has a
2407 // single use (so we don't increase the number of sexts), and if the
2408 // integer add will not overflow.
2409 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2410 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2411 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2412 RHSConv->getOperand(0))) {
2413 // Insert the new integer add.
2414 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2415 RHSConv->getOperand(0), "addconv");
2416 return new SExtInst(NewAdd, I.getType());
2421 return Changed ? &I : 0;
2424 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2425 bool Changed = SimplifyCommutative(I);
2426 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2428 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2430 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2431 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2432 (I.getType())->getValueAPF()))
2433 return ReplaceInstUsesWith(I, LHS);
2436 if (isa<PHINode>(LHS))
2437 if (Instruction *NV = FoldOpIntoPhi(I))
2442 // -A + -B --> -(A + B)
2443 if (Value *LHSV = dyn_castFNegVal(LHS))
2444 return BinaryOperator::CreateFSub(RHS, LHSV);
2447 if (!isa<Constant>(RHS))
2448 if (Value *V = dyn_castFNegVal(RHS))
2449 return BinaryOperator::CreateFSub(LHS, V);
2451 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2452 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2453 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2454 return ReplaceInstUsesWith(I, LHS);
2456 // Check for (add double (sitofp x), y), see if we can merge this into an
2457 // integer add followed by a promotion.
2458 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2459 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2460 // ... if the constant fits in the integer value. This is useful for things
2461 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2462 // requires a constant pool load, and generally allows the add to be better
2464 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2466 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2467 if (LHSConv->hasOneUse() &&
2468 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2469 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2470 // Insert the new integer add.
2471 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2473 return new SIToFPInst(NewAdd, I.getType());
2477 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2478 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2479 // Only do this if x/y have the same type, if at last one of them has a
2480 // single use (so we don't increase the number of int->fp conversions),
2481 // and if the integer add will not overflow.
2482 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2483 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2484 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2485 RHSConv->getOperand(0))) {
2486 // Insert the new integer add.
2487 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2488 RHSConv->getOperand(0), "addconv");
2489 return new SIToFPInst(NewAdd, I.getType());
2494 return Changed ? &I : 0;
2497 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2498 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2500 if (Op0 == Op1) // sub X, X -> 0
2501 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2503 // If this is a 'B = x-(-A)', change to B = x+A...
2504 if (Value *V = dyn_castNegVal(Op1))
2505 return BinaryOperator::CreateAdd(Op0, V);
2507 if (isa<UndefValue>(Op0))
2508 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2509 if (isa<UndefValue>(Op1))
2510 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2512 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2513 // Replace (-1 - A) with (~A)...
2514 if (C->isAllOnesValue())
2515 return BinaryOperator::CreateNot(Op1);
2517 // C - ~X == X + (1+C)
2519 if (match(Op1, m_Not(m_Value(X))))
2520 return BinaryOperator::CreateAdd(X, AddOne(C));
2522 // -(X >>u 31) -> (X >>s 31)
2523 // -(X >>s 31) -> (X >>u 31)
2525 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2526 if (SI->getOpcode() == Instruction::LShr) {
2527 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2528 // Check to see if we are shifting out everything but the sign bit.
2529 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2530 SI->getType()->getPrimitiveSizeInBits()-1) {
2531 // Ok, the transformation is safe. Insert AShr.
2532 return BinaryOperator::Create(Instruction::AShr,
2533 SI->getOperand(0), CU, SI->getName());
2537 else if (SI->getOpcode() == Instruction::AShr) {
2538 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2539 // Check to see if we are shifting out everything but the sign bit.
2540 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2541 SI->getType()->getPrimitiveSizeInBits()-1) {
2542 // Ok, the transformation is safe. Insert LShr.
2543 return BinaryOperator::CreateLShr(
2544 SI->getOperand(0), CU, SI->getName());
2551 // Try to fold constant sub into select arguments.
2552 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2553 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2556 // C - zext(bool) -> bool ? C - 1 : C
2557 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2558 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2559 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2562 if (I.getType() == Type::getInt1Ty(*Context))
2563 return BinaryOperator::CreateXor(Op0, Op1);
2565 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2566 if (Op1I->getOpcode() == Instruction::Add) {
2567 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2568 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2570 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2571 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2573 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2574 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2575 // C1-(X+C2) --> (C1-C2)-X
2576 return BinaryOperator::CreateSub(
2577 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2581 if (Op1I->hasOneUse()) {
2582 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2583 // is not used by anyone else...
2585 if (Op1I->getOpcode() == Instruction::Sub) {
2586 // Swap the two operands of the subexpr...
2587 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2588 Op1I->setOperand(0, IIOp1);
2589 Op1I->setOperand(1, IIOp0);
2591 // Create the new top level add instruction...
2592 return BinaryOperator::CreateAdd(Op0, Op1);
2595 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2597 if (Op1I->getOpcode() == Instruction::And &&
2598 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2599 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2601 Value *NewNot = Builder->CreateNot(OtherOp, "B.not");
2602 return BinaryOperator::CreateAnd(Op0, NewNot);
2605 // 0 - (X sdiv C) -> (X sdiv -C)
2606 if (Op1I->getOpcode() == Instruction::SDiv)
2607 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2609 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2610 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2611 ConstantExpr::getNeg(DivRHS));
2613 // X - X*C --> X * (1-C)
2614 ConstantInt *C2 = 0;
2615 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2617 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2619 return BinaryOperator::CreateMul(Op0, CP1);
2624 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2625 if (Op0I->getOpcode() == Instruction::Add) {
2626 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2627 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2628 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2629 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2630 } else if (Op0I->getOpcode() == Instruction::Sub) {
2631 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2632 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2638 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2639 if (X == Op1) // X*C - X --> X * (C-1)
2640 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2642 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2643 if (X == dyn_castFoldableMul(Op1, C2))
2644 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2649 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2650 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2652 // If this is a 'B = x-(-A)', change to B = x+A...
2653 if (Value *V = dyn_castFNegVal(Op1))
2654 return BinaryOperator::CreateFAdd(Op0, V);
2656 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2657 if (Op1I->getOpcode() == Instruction::FAdd) {
2658 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2659 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2661 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2662 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2670 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2671 /// comparison only checks the sign bit. If it only checks the sign bit, set
2672 /// TrueIfSigned if the result of the comparison is true when the input value is
2674 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2675 bool &TrueIfSigned) {
2677 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2678 TrueIfSigned = true;
2679 return RHS->isZero();
2680 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2681 TrueIfSigned = true;
2682 return RHS->isAllOnesValue();
2683 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2684 TrueIfSigned = false;
2685 return RHS->isAllOnesValue();
2686 case ICmpInst::ICMP_UGT:
2687 // True if LHS u> RHS and RHS == high-bit-mask - 1
2688 TrueIfSigned = true;
2689 return RHS->getValue() ==
2690 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2691 case ICmpInst::ICMP_UGE:
2692 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2693 TrueIfSigned = true;
2694 return RHS->getValue().isSignBit();
2700 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2701 bool Changed = SimplifyCommutative(I);
2702 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2704 if (isa<UndefValue>(Op1)) // undef * X -> 0
2705 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2707 // Simplify mul instructions with a constant RHS.
2708 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2709 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1C)) {
2711 // ((X << C1)*C2) == (X * (C2 << C1))
2712 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2713 if (SI->getOpcode() == Instruction::Shl)
2714 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2715 return BinaryOperator::CreateMul(SI->getOperand(0),
2716 ConstantExpr::getShl(CI, ShOp));
2719 return ReplaceInstUsesWith(I, Op1C); // X * 0 == 0
2720 if (CI->equalsInt(1)) // X * 1 == X
2721 return ReplaceInstUsesWith(I, Op0);
2722 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2723 return BinaryOperator::CreateNeg(Op0, I.getName());
2725 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2726 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2727 return BinaryOperator::CreateShl(Op0,
2728 ConstantInt::get(Op0->getType(), Val.logBase2()));
2730 } else if (isa<VectorType>(Op1C->getType())) {
2731 if (Op1C->isNullValue())
2732 return ReplaceInstUsesWith(I, Op1C);
2734 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
2735 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2736 return BinaryOperator::CreateNeg(Op0, I.getName());
2738 // As above, vector X*splat(1.0) -> X in all defined cases.
2739 if (Constant *Splat = Op1V->getSplatValue()) {
2740 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2741 if (CI->equalsInt(1))
2742 return ReplaceInstUsesWith(I, Op0);
2747 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2748 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2749 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1C)) {
2750 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2751 Value *Add = Builder->CreateMul(Op0I->getOperand(0), Op1C, "tmp");
2752 Value *C1C2 = Builder->CreateMul(Op1C, Op0I->getOperand(1));
2753 return BinaryOperator::CreateAdd(Add, C1C2);
2757 // Try to fold constant mul into select arguments.
2758 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2759 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2762 if (isa<PHINode>(Op0))
2763 if (Instruction *NV = FoldOpIntoPhi(I))
2767 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2768 if (Value *Op1v = dyn_castNegVal(Op1))
2769 return BinaryOperator::CreateMul(Op0v, Op1v);
2771 // (X / Y) * Y = X - (X % Y)
2772 // (X / Y) * -Y = (X % Y) - X
2775 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2777 (BO->getOpcode() != Instruction::UDiv &&
2778 BO->getOpcode() != Instruction::SDiv)) {
2780 BO = dyn_cast<BinaryOperator>(Op1);
2782 Value *Neg = dyn_castNegVal(Op1C);
2783 if (BO && BO->hasOneUse() &&
2784 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
2785 (BO->getOpcode() == Instruction::UDiv ||
2786 BO->getOpcode() == Instruction::SDiv)) {
2787 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2789 // If the division is exact, X % Y is zero.
2790 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
2791 if (SDiv->isExact()) {
2793 return ReplaceInstUsesWith(I, Op0BO);
2794 return BinaryOperator::CreateNeg(Op0BO);
2798 if (BO->getOpcode() == Instruction::UDiv)
2799 Rem = Builder->CreateURem(Op0BO, Op1BO);
2801 Rem = Builder->CreateSRem(Op0BO, Op1BO);
2805 return BinaryOperator::CreateSub(Op0BO, Rem);
2806 return BinaryOperator::CreateSub(Rem, Op0BO);
2810 /// i1 mul -> i1 and.
2811 if (I.getType() == Type::getInt1Ty(*Context))
2812 return BinaryOperator::CreateAnd(Op0, Op1);
2814 // X*(1 << Y) --> X << Y
2815 // (1 << Y)*X --> X << Y
2818 if (match(Op0, m_Shl(m_One(), m_Value(Y))))
2819 return BinaryOperator::CreateShl(Op1, Y);
2820 if (match(Op1, m_Shl(m_One(), m_Value(Y))))
2821 return BinaryOperator::CreateShl(Op0, Y);
2824 // If one of the operands of the multiply is a cast from a boolean value, then
2825 // we know the bool is either zero or one, so this is a 'masking' multiply.
2826 // X * Y (where Y is 0 or 1) -> X & (0-Y)
2827 if (!isa<VectorType>(I.getType())) {
2828 // -2 is "-1 << 1" so it is all bits set except the low one.
2829 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
2831 Value *BoolCast = 0, *OtherOp = 0;
2832 if (MaskedValueIsZero(Op0, Negative2))
2833 BoolCast = Op0, OtherOp = Op1;
2834 else if (MaskedValueIsZero(Op1, Negative2))
2835 BoolCast = Op1, OtherOp = Op0;
2838 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
2840 return BinaryOperator::CreateAnd(V, OtherOp);
2844 return Changed ? &I : 0;
2847 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2848 bool Changed = SimplifyCommutative(I);
2849 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2851 // Simplify mul instructions with a constant RHS...
2852 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2853 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1C)) {
2854 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2855 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2856 if (Op1F->isExactlyValue(1.0))
2857 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2858 } else if (isa<VectorType>(Op1C->getType())) {
2859 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
2860 // As above, vector X*splat(1.0) -> X in all defined cases.
2861 if (Constant *Splat = Op1V->getSplatValue()) {
2862 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2863 if (F->isExactlyValue(1.0))
2864 return ReplaceInstUsesWith(I, Op0);
2869 // Try to fold constant mul into select arguments.
2870 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2871 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2874 if (isa<PHINode>(Op0))
2875 if (Instruction *NV = FoldOpIntoPhi(I))
2879 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
2880 if (Value *Op1v = dyn_castFNegVal(Op1))
2881 return BinaryOperator::CreateFMul(Op0v, Op1v);
2883 return Changed ? &I : 0;
2886 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2888 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2889 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2891 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2892 int NonNullOperand = -1;
2893 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2894 if (ST->isNullValue())
2896 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2897 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2898 if (ST->isNullValue())
2901 if (NonNullOperand == -1)
2904 Value *SelectCond = SI->getOperand(0);
2906 // Change the div/rem to use 'Y' instead of the select.
2907 I.setOperand(1, SI->getOperand(NonNullOperand));
2909 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2910 // problem. However, the select, or the condition of the select may have
2911 // multiple uses. Based on our knowledge that the operand must be non-zero,
2912 // propagate the known value for the select into other uses of it, and
2913 // propagate a known value of the condition into its other users.
2915 // If the select and condition only have a single use, don't bother with this,
2917 if (SI->use_empty() && SelectCond->hasOneUse())
2920 // Scan the current block backward, looking for other uses of SI.
2921 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2923 while (BBI != BBFront) {
2925 // If we found a call to a function, we can't assume it will return, so
2926 // information from below it cannot be propagated above it.
2927 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2930 // Replace uses of the select or its condition with the known values.
2931 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2934 *I = SI->getOperand(NonNullOperand);
2936 } else if (*I == SelectCond) {
2937 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2938 ConstantInt::getFalse(*Context);
2943 // If we past the instruction, quit looking for it.
2946 if (&*BBI == SelectCond)
2949 // If we ran out of things to eliminate, break out of the loop.
2950 if (SelectCond == 0 && SI == 0)
2958 /// This function implements the transforms on div instructions that work
2959 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2960 /// used by the visitors to those instructions.
2961 /// @brief Transforms common to all three div instructions
2962 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2963 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2965 // undef / X -> 0 for integer.
2966 // undef / X -> undef for FP (the undef could be a snan).
2967 if (isa<UndefValue>(Op0)) {
2968 if (Op0->getType()->isFPOrFPVector())
2969 return ReplaceInstUsesWith(I, Op0);
2970 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2973 // X / undef -> undef
2974 if (isa<UndefValue>(Op1))
2975 return ReplaceInstUsesWith(I, Op1);
2980 /// This function implements the transforms common to both integer division
2981 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2982 /// division instructions.
2983 /// @brief Common integer divide transforms
2984 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2985 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2987 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2989 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2990 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
2991 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2992 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2995 Constant *CI = ConstantInt::get(I.getType(), 1);
2996 return ReplaceInstUsesWith(I, CI);
2999 if (Instruction *Common = commonDivTransforms(I))
3002 // Handle cases involving: [su]div X, (select Cond, Y, Z)
3003 // This does not apply for fdiv.
3004 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3007 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3009 if (RHS->equalsInt(1))
3010 return ReplaceInstUsesWith(I, Op0);
3012 // (X / C1) / C2 -> X / (C1*C2)
3013 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3014 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3015 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3016 if (MultiplyOverflows(RHS, LHSRHS,
3017 I.getOpcode()==Instruction::SDiv))
3018 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3020 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3021 ConstantExpr::getMul(RHS, LHSRHS));
3024 if (!RHS->isZero()) { // avoid X udiv 0
3025 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3026 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3028 if (isa<PHINode>(Op0))
3029 if (Instruction *NV = FoldOpIntoPhi(I))
3034 // 0 / X == 0, we don't need to preserve faults!
3035 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3036 if (LHS->equalsInt(0))
3037 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3039 // It can't be division by zero, hence it must be division by one.
3040 if (I.getType() == Type::getInt1Ty(*Context))
3041 return ReplaceInstUsesWith(I, Op0);
3043 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3044 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3047 return ReplaceInstUsesWith(I, Op0);
3053 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3054 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3056 // Handle the integer div common cases
3057 if (Instruction *Common = commonIDivTransforms(I))
3060 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3061 // X udiv C^2 -> X >> C
3062 // Check to see if this is an unsigned division with an exact power of 2,
3063 // if so, convert to a right shift.
3064 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3065 return BinaryOperator::CreateLShr(Op0,
3066 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3068 // X udiv C, where C >= signbit
3069 if (C->getValue().isNegative()) {
3070 Value *IC = Builder->CreateICmpULT( Op0, C);
3071 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3072 ConstantInt::get(I.getType(), 1));
3076 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3077 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3078 if (RHSI->getOpcode() == Instruction::Shl &&
3079 isa<ConstantInt>(RHSI->getOperand(0))) {
3080 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3081 if (C1.isPowerOf2()) {
3082 Value *N = RHSI->getOperand(1);
3083 const Type *NTy = N->getType();
3084 if (uint32_t C2 = C1.logBase2())
3085 N = Builder->CreateAdd(N, ConstantInt::get(NTy, C2), "tmp");
3086 return BinaryOperator::CreateLShr(Op0, N);
3091 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3092 // where C1&C2 are powers of two.
3093 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3094 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3095 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3096 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3097 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3098 // Compute the shift amounts
3099 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3100 // Construct the "on true" case of the select
3101 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3102 Value *TSI = Builder->CreateLShr(Op0, TC, SI->getName()+".t");
3104 // Construct the "on false" case of the select
3105 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3106 Value *FSI = Builder->CreateLShr(Op0, FC, SI->getName()+".f");
3108 // construct the select instruction and return it.
3109 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3115 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3116 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3118 // Handle the integer div common cases
3119 if (Instruction *Common = commonIDivTransforms(I))
3122 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3124 if (RHS->isAllOnesValue())
3125 return BinaryOperator::CreateNeg(Op0);
3127 // sdiv X, C --> ashr X, log2(C)
3128 if (cast<SDivOperator>(&I)->isExact() &&
3129 RHS->getValue().isNonNegative() &&
3130 RHS->getValue().isPowerOf2()) {
3131 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3132 RHS->getValue().exactLogBase2());
3133 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3136 // -X/C --> X/-C provided the negation doesn't overflow.
3137 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3138 if (isa<Constant>(Sub->getOperand(0)) &&
3139 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3140 Sub->hasNoSignedWrap())
3141 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3142 ConstantExpr::getNeg(RHS));
3145 // If the sign bits of both operands are zero (i.e. we can prove they are
3146 // unsigned inputs), turn this into a udiv.
3147 if (I.getType()->isInteger()) {
3148 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3149 if (MaskedValueIsZero(Op0, Mask)) {
3150 if (MaskedValueIsZero(Op1, Mask)) {
3151 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3152 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3154 ConstantInt *ShiftedInt;
3155 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3156 ShiftedInt->getValue().isPowerOf2()) {
3157 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3158 // Safe because the only negative value (1 << Y) can take on is
3159 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3160 // the sign bit set.
3161 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3169 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3170 return commonDivTransforms(I);
3173 /// This function implements the transforms on rem instructions that work
3174 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3175 /// is used by the visitors to those instructions.
3176 /// @brief Transforms common to all three rem instructions
3177 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3178 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3180 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3181 if (I.getType()->isFPOrFPVector())
3182 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3183 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3185 if (isa<UndefValue>(Op1))
3186 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3188 // Handle cases involving: rem X, (select Cond, Y, Z)
3189 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3195 /// This function implements the transforms common to both integer remainder
3196 /// instructions (urem and srem). It is called by the visitors to those integer
3197 /// remainder instructions.
3198 /// @brief Common integer remainder transforms
3199 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3200 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3202 if (Instruction *common = commonRemTransforms(I))
3205 // 0 % X == 0 for integer, we don't need to preserve faults!
3206 if (Constant *LHS = dyn_cast<Constant>(Op0))
3207 if (LHS->isNullValue())
3208 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3210 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3211 // X % 0 == undef, we don't need to preserve faults!
3212 if (RHS->equalsInt(0))
3213 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3215 if (RHS->equalsInt(1)) // X % 1 == 0
3216 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3218 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3219 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3220 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3222 } else if (isa<PHINode>(Op0I)) {
3223 if (Instruction *NV = FoldOpIntoPhi(I))
3227 // See if we can fold away this rem instruction.
3228 if (SimplifyDemandedInstructionBits(I))
3236 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3237 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3239 if (Instruction *common = commonIRemTransforms(I))
3242 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3243 // X urem C^2 -> X and C
3244 // Check to see if this is an unsigned remainder with an exact power of 2,
3245 // if so, convert to a bitwise and.
3246 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3247 if (C->getValue().isPowerOf2())
3248 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3251 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3252 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3253 if (RHSI->getOpcode() == Instruction::Shl &&
3254 isa<ConstantInt>(RHSI->getOperand(0))) {
3255 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3256 Constant *N1 = Constant::getAllOnesValue(I.getType());
3257 Value *Add = Builder->CreateAdd(RHSI, N1, "tmp");
3258 return BinaryOperator::CreateAnd(Op0, Add);
3263 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3264 // where C1&C2 are powers of two.
3265 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3266 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3267 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3268 // STO == 0 and SFO == 0 handled above.
3269 if ((STO->getValue().isPowerOf2()) &&
3270 (SFO->getValue().isPowerOf2())) {
3271 Value *TrueAnd = Builder->CreateAnd(Op0, SubOne(STO),
3272 SI->getName()+".t");
3273 Value *FalseAnd = Builder->CreateAnd(Op0, SubOne(SFO),
3274 SI->getName()+".f");
3275 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3283 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3284 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3286 // Handle the integer rem common cases
3287 if (Instruction *Common = commonIRemTransforms(I))
3290 if (Value *RHSNeg = dyn_castNegVal(Op1))
3291 if (!isa<Constant>(RHSNeg) ||
3292 (isa<ConstantInt>(RHSNeg) &&
3293 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3295 Worklist.AddValue(I.getOperand(1));
3296 I.setOperand(1, RHSNeg);
3300 // If the sign bits of both operands are zero (i.e. we can prove they are
3301 // unsigned inputs), turn this into a urem.
3302 if (I.getType()->isInteger()) {
3303 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3304 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3305 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3306 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3310 // If it's a constant vector, flip any negative values positive.
3311 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3312 unsigned VWidth = RHSV->getNumOperands();
3314 bool hasNegative = false;
3315 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3316 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3317 if (RHS->getValue().isNegative())
3321 std::vector<Constant *> Elts(VWidth);
3322 for (unsigned i = 0; i != VWidth; ++i) {
3323 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3324 if (RHS->getValue().isNegative())
3325 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3331 Constant *NewRHSV = ConstantVector::get(Elts);
3332 if (NewRHSV != RHSV) {
3333 Worklist.AddValue(I.getOperand(1));
3334 I.setOperand(1, NewRHSV);
3343 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3344 return commonRemTransforms(I);
3347 // isOneBitSet - Return true if there is exactly one bit set in the specified
3349 static bool isOneBitSet(const ConstantInt *CI) {
3350 return CI->getValue().isPowerOf2();
3353 // isHighOnes - Return true if the constant is of the form 1+0+.
3354 // This is the same as lowones(~X).
3355 static bool isHighOnes(const ConstantInt *CI) {
3356 return (~CI->getValue() + 1).isPowerOf2();
3359 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3360 /// are carefully arranged to allow folding of expressions such as:
3362 /// (A < B) | (A > B) --> (A != B)
3364 /// Note that this is only valid if the first and second predicates have the
3365 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3367 /// Three bits are used to represent the condition, as follows:
3372 /// <=> Value Definition
3373 /// 000 0 Always false
3380 /// 111 7 Always true
3382 static unsigned getICmpCode(const ICmpInst *ICI) {
3383 switch (ICI->getPredicate()) {
3385 case ICmpInst::ICMP_UGT: return 1; // 001
3386 case ICmpInst::ICMP_SGT: return 1; // 001
3387 case ICmpInst::ICMP_EQ: return 2; // 010
3388 case ICmpInst::ICMP_UGE: return 3; // 011
3389 case ICmpInst::ICMP_SGE: return 3; // 011
3390 case ICmpInst::ICMP_ULT: return 4; // 100
3391 case ICmpInst::ICMP_SLT: return 4; // 100
3392 case ICmpInst::ICMP_NE: return 5; // 101
3393 case ICmpInst::ICMP_ULE: return 6; // 110
3394 case ICmpInst::ICMP_SLE: return 6; // 110
3397 llvm_unreachable("Invalid ICmp predicate!");
3402 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3403 /// predicate into a three bit mask. It also returns whether it is an ordered
3404 /// predicate by reference.
3405 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3408 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3409 case FCmpInst::FCMP_UNO: return 0; // 000
3410 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3411 case FCmpInst::FCMP_UGT: return 1; // 001
3412 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3413 case FCmpInst::FCMP_UEQ: return 2; // 010
3414 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3415 case FCmpInst::FCMP_UGE: return 3; // 011
3416 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3417 case FCmpInst::FCMP_ULT: return 4; // 100
3418 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3419 case FCmpInst::FCMP_UNE: return 5; // 101
3420 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3421 case FCmpInst::FCMP_ULE: return 6; // 110
3424 // Not expecting FCMP_FALSE and FCMP_TRUE;
3425 llvm_unreachable("Unexpected FCmp predicate!");
3430 /// getICmpValue - This is the complement of getICmpCode, which turns an
3431 /// opcode and two operands into either a constant true or false, or a brand
3432 /// new ICmp instruction. The sign is passed in to determine which kind
3433 /// of predicate to use in the new icmp instruction.
3434 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3435 LLVMContext *Context) {
3437 default: llvm_unreachable("Illegal ICmp code!");
3438 case 0: return ConstantInt::getFalse(*Context);
3441 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3443 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3444 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3447 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3449 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3452 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3454 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3455 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3458 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3460 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3461 case 7: return ConstantInt::getTrue(*Context);
3465 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3466 /// opcode and two operands into either a FCmp instruction. isordered is passed
3467 /// in to determine which kind of predicate to use in the new fcmp instruction.
3468 static Value *getFCmpValue(bool isordered, unsigned code,
3469 Value *LHS, Value *RHS, LLVMContext *Context) {
3471 default: llvm_unreachable("Illegal FCmp code!");
3474 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3476 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3479 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3481 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3484 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3486 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3489 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3491 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3494 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3496 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3499 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3501 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3504 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3506 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3507 case 7: return ConstantInt::getTrue(*Context);
3511 /// PredicatesFoldable - Return true if both predicates match sign or if at
3512 /// least one of them is an equality comparison (which is signless).
3513 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3514 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
3515 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
3516 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
3520 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3521 struct FoldICmpLogical {
3524 ICmpInst::Predicate pred;
3525 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3526 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3527 pred(ICI->getPredicate()) {}
3528 bool shouldApply(Value *V) const {
3529 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3530 if (PredicatesFoldable(pred, ICI->getPredicate()))
3531 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3532 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3535 Instruction *apply(Instruction &Log) const {
3536 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3537 if (ICI->getOperand(0) != LHS) {
3538 assert(ICI->getOperand(1) == LHS);
3539 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3542 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3543 unsigned LHSCode = getICmpCode(ICI);
3544 unsigned RHSCode = getICmpCode(RHSICI);
3546 switch (Log.getOpcode()) {
3547 case Instruction::And: Code = LHSCode & RHSCode; break;
3548 case Instruction::Or: Code = LHSCode | RHSCode; break;
3549 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3550 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3553 bool isSigned = RHSICI->isSigned() || ICI->isSigned();
3554 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3555 if (Instruction *I = dyn_cast<Instruction>(RV))
3557 // Otherwise, it's a constant boolean value...
3558 return IC.ReplaceInstUsesWith(Log, RV);
3561 } // end anonymous namespace
3563 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3564 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3565 // guaranteed to be a binary operator.
3566 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3568 ConstantInt *AndRHS,
3569 BinaryOperator &TheAnd) {
3570 Value *X = Op->getOperand(0);
3571 Constant *Together = 0;
3573 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3575 switch (Op->getOpcode()) {
3576 case Instruction::Xor:
3577 if (Op->hasOneUse()) {
3578 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3579 Value *And = Builder->CreateAnd(X, AndRHS);
3581 return BinaryOperator::CreateXor(And, Together);
3584 case Instruction::Or:
3585 if (Together == AndRHS) // (X | C) & C --> C
3586 return ReplaceInstUsesWith(TheAnd, AndRHS);
3588 if (Op->hasOneUse() && Together != OpRHS) {
3589 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3590 Value *Or = Builder->CreateOr(X, Together);
3592 return BinaryOperator::CreateAnd(Or, AndRHS);
3595 case Instruction::Add:
3596 if (Op->hasOneUse()) {
3597 // Adding a one to a single bit bit-field should be turned into an XOR
3598 // of the bit. First thing to check is to see if this AND is with a
3599 // single bit constant.
3600 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3602 // If there is only one bit set...
3603 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3604 // Ok, at this point, we know that we are masking the result of the
3605 // ADD down to exactly one bit. If the constant we are adding has
3606 // no bits set below this bit, then we can eliminate the ADD.
3607 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3609 // Check to see if any bits below the one bit set in AndRHSV are set.
3610 if ((AddRHS & (AndRHSV-1)) == 0) {
3611 // If not, the only thing that can effect the output of the AND is
3612 // the bit specified by AndRHSV. If that bit is set, the effect of
3613 // the XOR is to toggle the bit. If it is clear, then the ADD has
3615 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3616 TheAnd.setOperand(0, X);
3619 // Pull the XOR out of the AND.
3620 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
3621 NewAnd->takeName(Op);
3622 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3629 case Instruction::Shl: {
3630 // We know that the AND will not produce any of the bits shifted in, so if
3631 // the anded constant includes them, clear them now!
3633 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3634 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3635 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3636 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3638 if (CI->getValue() == ShlMask) {
3639 // Masking out bits that the shift already masks
3640 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3641 } else if (CI != AndRHS) { // Reducing bits set in and.
3642 TheAnd.setOperand(1, CI);
3647 case Instruction::LShr:
3649 // We know that the AND will not produce any of the bits shifted in, so if
3650 // the anded constant includes them, clear them now! This only applies to
3651 // unsigned shifts, because a signed shr may bring in set bits!
3653 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3654 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3655 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3656 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3658 if (CI->getValue() == ShrMask) {
3659 // Masking out bits that the shift already masks.
3660 return ReplaceInstUsesWith(TheAnd, Op);
3661 } else if (CI != AndRHS) {
3662 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3667 case Instruction::AShr:
3669 // See if this is shifting in some sign extension, then masking it out
3671 if (Op->hasOneUse()) {
3672 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3673 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3674 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3675 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3676 if (C == AndRHS) { // Masking out bits shifted in.
3677 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3678 // Make the argument unsigned.
3679 Value *ShVal = Op->getOperand(0);
3680 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
3681 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3690 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3691 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3692 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3693 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3694 /// insert new instructions.
3695 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3696 bool isSigned, bool Inside,
3698 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3699 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3700 "Lo is not <= Hi in range emission code!");
3703 if (Lo == Hi) // Trivially false.
3704 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3706 // V >= Min && V < Hi --> V < Hi
3707 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3708 ICmpInst::Predicate pred = (isSigned ?
3709 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3710 return new ICmpInst(pred, V, Hi);
3713 // Emit V-Lo <u Hi-Lo
3714 Constant *NegLo = ConstantExpr::getNeg(Lo);
3715 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3716 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3717 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3720 if (Lo == Hi) // Trivially true.
3721 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3723 // V < Min || V >= Hi -> V > Hi-1
3724 Hi = SubOne(cast<ConstantInt>(Hi));
3725 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3726 ICmpInst::Predicate pred = (isSigned ?
3727 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3728 return new ICmpInst(pred, V, Hi);
3731 // Emit V-Lo >u Hi-1-Lo
3732 // Note that Hi has already had one subtracted from it, above.
3733 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3734 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3735 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3736 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3739 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3740 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3741 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3742 // not, since all 1s are not contiguous.
3743 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3744 const APInt& V = Val->getValue();
3745 uint32_t BitWidth = Val->getType()->getBitWidth();
3746 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3748 // look for the first zero bit after the run of ones
3749 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3750 // look for the first non-zero bit
3751 ME = V.getActiveBits();
3755 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3756 /// where isSub determines whether the operator is a sub. If we can fold one of
3757 /// the following xforms:
3759 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3760 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3761 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3763 /// return (A +/- B).
3765 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3766 ConstantInt *Mask, bool isSub,
3768 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3769 if (!LHSI || LHSI->getNumOperands() != 2 ||
3770 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3772 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3774 switch (LHSI->getOpcode()) {
3776 case Instruction::And:
3777 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3778 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3779 if ((Mask->getValue().countLeadingZeros() +
3780 Mask->getValue().countPopulation()) ==
3781 Mask->getValue().getBitWidth())
3784 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3785 // part, we don't need any explicit masks to take them out of A. If that
3786 // is all N is, ignore it.
3787 uint32_t MB = 0, ME = 0;
3788 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3789 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3790 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3791 if (MaskedValueIsZero(RHS, Mask))
3796 case Instruction::Or:
3797 case Instruction::Xor:
3798 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3799 if ((Mask->getValue().countLeadingZeros() +
3800 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3801 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3807 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
3808 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
3811 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3812 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3813 ICmpInst *LHS, ICmpInst *RHS) {
3815 ConstantInt *LHSCst, *RHSCst;
3816 ICmpInst::Predicate LHSCC, RHSCC;
3818 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3819 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3820 m_ConstantInt(LHSCst))) ||
3821 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3822 m_ConstantInt(RHSCst))))
3825 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3826 // where C is a power of 2
3827 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3828 LHSCst->getValue().isPowerOf2()) {
3829 Value *NewOr = Builder->CreateOr(Val, Val2);
3830 return new ICmpInst(LHSCC, NewOr, LHSCst);
3833 // From here on, we only handle:
3834 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3835 if (Val != Val2) return 0;
3837 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3838 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3839 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3840 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3841 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3844 // We can't fold (ugt x, C) & (sgt x, C2).
3845 if (!PredicatesFoldable(LHSCC, RHSCC))
3848 // Ensure that the larger constant is on the RHS.
3850 if (CmpInst::isSigned(LHSCC) ||
3851 (ICmpInst::isEquality(LHSCC) &&
3852 CmpInst::isSigned(RHSCC)))
3853 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3855 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3858 std::swap(LHS, RHS);
3859 std::swap(LHSCst, RHSCst);
3860 std::swap(LHSCC, RHSCC);
3863 // At this point, we know we have have two icmp instructions
3864 // comparing a value against two constants and and'ing the result
3865 // together. Because of the above check, we know that we only have
3866 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3867 // (from the FoldICmpLogical check above), that the two constants
3868 // are not equal and that the larger constant is on the RHS
3869 assert(LHSCst != RHSCst && "Compares not folded above?");
3872 default: llvm_unreachable("Unknown integer condition code!");
3873 case ICmpInst::ICMP_EQ:
3875 default: llvm_unreachable("Unknown integer condition code!");
3876 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3877 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3878 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3879 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3880 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3881 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3882 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3883 return ReplaceInstUsesWith(I, LHS);
3885 case ICmpInst::ICMP_NE:
3887 default: llvm_unreachable("Unknown integer condition code!");
3888 case ICmpInst::ICMP_ULT:
3889 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3890 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3891 break; // (X != 13 & X u< 15) -> no change
3892 case ICmpInst::ICMP_SLT:
3893 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3894 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3895 break; // (X != 13 & X s< 15) -> no change
3896 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3897 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3898 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3899 return ReplaceInstUsesWith(I, RHS);
3900 case ICmpInst::ICMP_NE:
3901 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3902 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3903 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
3904 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3905 ConstantInt::get(Add->getType(), 1));
3907 break; // (X != 13 & X != 15) -> no change
3910 case ICmpInst::ICMP_ULT:
3912 default: llvm_unreachable("Unknown integer condition code!");
3913 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3914 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3915 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3916 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3918 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3919 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3920 return ReplaceInstUsesWith(I, LHS);
3921 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3925 case ICmpInst::ICMP_SLT:
3927 default: llvm_unreachable("Unknown integer condition code!");
3928 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3929 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3930 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3931 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3933 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3934 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3935 return ReplaceInstUsesWith(I, LHS);
3936 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3940 case ICmpInst::ICMP_UGT:
3942 default: llvm_unreachable("Unknown integer condition code!");
3943 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3944 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3945 return ReplaceInstUsesWith(I, RHS);
3946 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3948 case ICmpInst::ICMP_NE:
3949 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3950 return new ICmpInst(LHSCC, Val, RHSCst);
3951 break; // (X u> 13 & X != 15) -> no change
3952 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3953 return InsertRangeTest(Val, AddOne(LHSCst),
3954 RHSCst, false, true, I);
3955 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3959 case ICmpInst::ICMP_SGT:
3961 default: llvm_unreachable("Unknown integer condition code!");
3962 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3963 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3964 return ReplaceInstUsesWith(I, RHS);
3965 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3967 case ICmpInst::ICMP_NE:
3968 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3969 return new ICmpInst(LHSCC, Val, RHSCst);
3970 break; // (X s> 13 & X != 15) -> no change
3971 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3972 return InsertRangeTest(Val, AddOne(LHSCst),
3973 RHSCst, true, true, I);
3974 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3983 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
3986 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3987 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3988 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3989 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3990 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3991 // If either of the constants are nans, then the whole thing returns
3993 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3994 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3995 return new FCmpInst(FCmpInst::FCMP_ORD,
3996 LHS->getOperand(0), RHS->getOperand(0));
3999 // Handle vector zeros. This occurs because the canonical form of
4000 // "fcmp ord x,x" is "fcmp ord x, 0".
4001 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4002 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4003 return new FCmpInst(FCmpInst::FCMP_ORD,
4004 LHS->getOperand(0), RHS->getOperand(0));
4008 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4009 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4010 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4013 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4014 // Swap RHS operands to match LHS.
4015 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4016 std::swap(Op1LHS, Op1RHS);
4019 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4020 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4022 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4024 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
4025 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4026 if (Op0CC == FCmpInst::FCMP_TRUE)
4027 return ReplaceInstUsesWith(I, RHS);
4028 if (Op1CC == FCmpInst::FCMP_TRUE)
4029 return ReplaceInstUsesWith(I, LHS);
4033 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4034 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4036 std::swap(LHS, RHS);
4037 std::swap(Op0Pred, Op1Pred);
4038 std::swap(Op0Ordered, Op1Ordered);
4041 // uno && ueq -> uno && (uno || eq) -> ueq
4042 // ord && olt -> ord && (ord && lt) -> olt
4043 if (Op0Ordered == Op1Ordered)
4044 return ReplaceInstUsesWith(I, RHS);
4046 // uno && oeq -> uno && (ord && eq) -> false
4047 // uno && ord -> false
4049 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4050 // ord && ueq -> ord && (uno || eq) -> oeq
4051 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4052 Op0LHS, Op0RHS, Context));
4060 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4061 bool Changed = SimplifyCommutative(I);
4062 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4064 if (isa<UndefValue>(Op1)) // X & undef -> 0
4065 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4069 return ReplaceInstUsesWith(I, Op1);
4071 // See if we can simplify any instructions used by the instruction whose sole
4072 // purpose is to compute bits we don't care about.
4073 if (SimplifyDemandedInstructionBits(I))
4075 if (isa<VectorType>(I.getType())) {
4076 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4077 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4078 return ReplaceInstUsesWith(I, I.getOperand(0));
4079 } else if (isa<ConstantAggregateZero>(Op1)) {
4080 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4084 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4085 const APInt &AndRHSMask = AndRHS->getValue();
4086 APInt NotAndRHS(~AndRHSMask);
4088 // Optimize a variety of ((val OP C1) & C2) combinations...
4089 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4090 Value *Op0LHS = Op0I->getOperand(0);
4091 Value *Op0RHS = Op0I->getOperand(1);
4092 switch (Op0I->getOpcode()) {
4094 case Instruction::Xor:
4095 case Instruction::Or:
4096 // If the mask is only needed on one incoming arm, push it up.
4097 if (!Op0I->hasOneUse()) break;
4099 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4100 // Not masking anything out for the LHS, move to RHS.
4101 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
4102 Op0RHS->getName()+".masked");
4103 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
4105 if (!isa<Constant>(Op0RHS) &&
4106 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4107 // Not masking anything out for the RHS, move to LHS.
4108 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
4109 Op0LHS->getName()+".masked");
4110 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
4114 case Instruction::Add:
4115 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4116 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4117 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4118 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4119 return BinaryOperator::CreateAnd(V, AndRHS);
4120 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4121 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4124 case Instruction::Sub:
4125 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4126 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4127 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4128 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4129 return BinaryOperator::CreateAnd(V, AndRHS);
4131 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4132 // has 1's for all bits that the subtraction with A might affect.
4133 if (Op0I->hasOneUse()) {
4134 uint32_t BitWidth = AndRHSMask.getBitWidth();
4135 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4136 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4138 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4139 if (!(A && A->isZero()) && // avoid infinite recursion.
4140 MaskedValueIsZero(Op0LHS, Mask)) {
4141 Value *NewNeg = Builder->CreateNeg(Op0RHS);
4142 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4147 case Instruction::Shl:
4148 case Instruction::LShr:
4149 // (1 << x) & 1 --> zext(x == 0)
4150 // (1 >> x) & 1 --> zext(x == 0)
4151 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4153 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
4154 return new ZExtInst(NewICmp, I.getType());
4159 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4160 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4162 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4163 // If this is an integer truncation or change from signed-to-unsigned, and
4164 // if the source is an and/or with immediate, transform it. This
4165 // frequently occurs for bitfield accesses.
4166 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4167 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4168 CastOp->getNumOperands() == 2)
4169 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4170 if (CastOp->getOpcode() == Instruction::And) {
4171 // Change: and (cast (and X, C1) to T), C2
4172 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4173 // This will fold the two constants together, which may allow
4174 // other simplifications.
4175 Value *NewCast = Builder->CreateTruncOrBitCast(
4176 CastOp->getOperand(0), I.getType(),
4177 CastOp->getName()+".shrunk");
4178 // trunc_or_bitcast(C1)&C2
4179 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4180 C3 = ConstantExpr::getAnd(C3, AndRHS);
4181 return BinaryOperator::CreateAnd(NewCast, C3);
4182 } else if (CastOp->getOpcode() == Instruction::Or) {
4183 // Change: and (cast (or X, C1) to T), C2
4184 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4185 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4186 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4188 return ReplaceInstUsesWith(I, AndRHS);
4194 // Try to fold constant and into select arguments.
4195 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4196 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4198 if (isa<PHINode>(Op0))
4199 if (Instruction *NV = FoldOpIntoPhi(I))
4203 Value *Op0NotVal = dyn_castNotVal(Op0);
4204 Value *Op1NotVal = dyn_castNotVal(Op1);
4206 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4207 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4209 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4210 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4211 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
4212 I.getName()+".demorgan");
4213 return BinaryOperator::CreateNot(Or);
4217 Value *A = 0, *B = 0, *C = 0, *D = 0;
4218 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4219 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4220 return ReplaceInstUsesWith(I, Op1);
4222 // (A|B) & ~(A&B) -> A^B
4223 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4224 if ((A == C && B == D) || (A == D && B == C))
4225 return BinaryOperator::CreateXor(A, B);
4229 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4230 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4231 return ReplaceInstUsesWith(I, Op0);
4233 // ~(A&B) & (A|B) -> A^B
4234 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4235 if ((A == C && B == D) || (A == D && B == C))
4236 return BinaryOperator::CreateXor(A, B);
4240 if (Op0->hasOneUse() &&
4241 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4242 if (A == Op1) { // (A^B)&A -> A&(A^B)
4243 I.swapOperands(); // Simplify below
4244 std::swap(Op0, Op1);
4245 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4246 cast<BinaryOperator>(Op0)->swapOperands();
4247 I.swapOperands(); // Simplify below
4248 std::swap(Op0, Op1);
4252 if (Op1->hasOneUse() &&
4253 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4254 if (B == Op0) { // B&(A^B) -> B&(B^A)
4255 cast<BinaryOperator>(Op1)->swapOperands();
4258 if (A == Op0) // A&(A^B) -> A & ~B
4259 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
4262 // (A&((~A)|B)) -> A&B
4263 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4264 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4265 return BinaryOperator::CreateAnd(A, Op1);
4266 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4267 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4268 return BinaryOperator::CreateAnd(A, Op0);
4271 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4272 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4273 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4276 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4277 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4281 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4282 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4283 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4284 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4285 const Type *SrcTy = Op0C->getOperand(0)->getType();
4286 if (SrcTy == Op1C->getOperand(0)->getType() &&
4287 SrcTy->isIntOrIntVector() &&
4288 // Only do this if the casts both really cause code to be generated.
4289 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4291 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4293 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
4294 Op1C->getOperand(0), I.getName());
4295 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4299 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4300 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4301 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4302 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4303 SI0->getOperand(1) == SI1->getOperand(1) &&
4304 (SI0->hasOneUse() || SI1->hasOneUse())) {
4306 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
4308 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4309 SI1->getOperand(1));
4313 // If and'ing two fcmp, try combine them into one.
4314 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4315 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4316 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4320 return Changed ? &I : 0;
4323 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4324 /// capable of providing pieces of a bswap. The subexpression provides pieces
4325 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4326 /// the expression came from the corresponding "byte swapped" byte in some other
4327 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4328 /// we know that the expression deposits the low byte of %X into the high byte
4329 /// of the bswap result and that all other bytes are zero. This expression is
4330 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4333 /// This function returns true if the match was unsuccessful and false if so.
4334 /// On entry to the function the "OverallLeftShift" is a signed integer value
4335 /// indicating the number of bytes that the subexpression is later shifted. For
4336 /// example, if the expression is later right shifted by 16 bits, the
4337 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4338 /// byte of ByteValues is actually being set.
4340 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4341 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4342 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4343 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4344 /// always in the local (OverallLeftShift) coordinate space.
4346 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4347 SmallVector<Value*, 8> &ByteValues) {
4348 if (Instruction *I = dyn_cast<Instruction>(V)) {
4349 // If this is an or instruction, it may be an inner node of the bswap.
4350 if (I->getOpcode() == Instruction::Or) {
4351 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4353 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4357 // If this is a logical shift by a constant multiple of 8, recurse with
4358 // OverallLeftShift and ByteMask adjusted.
4359 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4361 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4362 // Ensure the shift amount is defined and of a byte value.
4363 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4366 unsigned ByteShift = ShAmt >> 3;
4367 if (I->getOpcode() == Instruction::Shl) {
4368 // X << 2 -> collect(X, +2)
4369 OverallLeftShift += ByteShift;
4370 ByteMask >>= ByteShift;
4372 // X >>u 2 -> collect(X, -2)
4373 OverallLeftShift -= ByteShift;
4374 ByteMask <<= ByteShift;
4375 ByteMask &= (~0U >> (32-ByteValues.size()));
4378 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4379 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4381 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4385 // If this is a logical 'and' with a mask that clears bytes, clear the
4386 // corresponding bytes in ByteMask.
4387 if (I->getOpcode() == Instruction::And &&
4388 isa<ConstantInt>(I->getOperand(1))) {
4389 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4390 unsigned NumBytes = ByteValues.size();
4391 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4392 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4394 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4395 // If this byte is masked out by a later operation, we don't care what
4397 if ((ByteMask & (1 << i)) == 0)
4400 // If the AndMask is all zeros for this byte, clear the bit.
4401 APInt MaskB = AndMask & Byte;
4403 ByteMask &= ~(1U << i);
4407 // If the AndMask is not all ones for this byte, it's not a bytezap.
4411 // Otherwise, this byte is kept.
4414 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4419 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4420 // the input value to the bswap. Some observations: 1) if more than one byte
4421 // is demanded from this input, then it could not be successfully assembled
4422 // into a byteswap. At least one of the two bytes would not be aligned with
4423 // their ultimate destination.
4424 if (!isPowerOf2_32(ByteMask)) return true;
4425 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4427 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4428 // is demanded, it needs to go into byte 0 of the result. This means that the
4429 // byte needs to be shifted until it lands in the right byte bucket. The
4430 // shift amount depends on the position: if the byte is coming from the high
4431 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4432 // low part, it must be shifted left.
4433 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4434 if (InputByteNo < ByteValues.size()/2) {
4435 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4438 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4442 // If the destination byte value is already defined, the values are or'd
4443 // together, which isn't a bswap (unless it's an or of the same bits).
4444 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4446 ByteValues[DestByteNo] = V;
4450 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4451 /// If so, insert the new bswap intrinsic and return it.
4452 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4453 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4454 if (!ITy || ITy->getBitWidth() % 16 ||
4455 // ByteMask only allows up to 32-byte values.
4456 ITy->getBitWidth() > 32*8)
4457 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4459 /// ByteValues - For each byte of the result, we keep track of which value
4460 /// defines each byte.
4461 SmallVector<Value*, 8> ByteValues;
4462 ByteValues.resize(ITy->getBitWidth()/8);
4464 // Try to find all the pieces corresponding to the bswap.
4465 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4466 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4469 // Check to see if all of the bytes come from the same value.
4470 Value *V = ByteValues[0];
4471 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4473 // Check to make sure that all of the bytes come from the same value.
4474 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4475 if (ByteValues[i] != V)
4477 const Type *Tys[] = { ITy };
4478 Module *M = I.getParent()->getParent()->getParent();
4479 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4480 return CallInst::Create(F, V);
4483 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4484 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4485 /// we can simplify this expression to "cond ? C : D or B".
4486 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4488 LLVMContext *Context) {
4489 // If A is not a select of -1/0, this cannot match.
4491 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4494 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4495 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4496 return SelectInst::Create(Cond, C, B);
4497 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4498 return SelectInst::Create(Cond, C, B);
4499 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4500 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4501 return SelectInst::Create(Cond, C, D);
4502 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4503 return SelectInst::Create(Cond, C, D);
4507 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4508 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4509 ICmpInst *LHS, ICmpInst *RHS) {
4511 ConstantInt *LHSCst, *RHSCst;
4512 ICmpInst::Predicate LHSCC, RHSCC;
4514 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4515 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4516 m_ConstantInt(LHSCst))) ||
4517 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4518 m_ConstantInt(RHSCst))))
4521 // From here on, we only handle:
4522 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4523 if (Val != Val2) return 0;
4525 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4526 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4527 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4528 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4529 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4532 // We can't fold (ugt x, C) | (sgt x, C2).
4533 if (!PredicatesFoldable(LHSCC, RHSCC))
4536 // Ensure that the larger constant is on the RHS.
4538 if (CmpInst::isSigned(LHSCC) ||
4539 (ICmpInst::isEquality(LHSCC) &&
4540 CmpInst::isSigned(RHSCC)))
4541 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4543 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4546 std::swap(LHS, RHS);
4547 std::swap(LHSCst, RHSCst);
4548 std::swap(LHSCC, RHSCC);
4551 // At this point, we know we have have two icmp instructions
4552 // comparing a value against two constants and or'ing the result
4553 // together. Because of the above check, we know that we only have
4554 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4555 // FoldICmpLogical check above), that the two constants are not
4557 assert(LHSCst != RHSCst && "Compares not folded above?");
4560 default: llvm_unreachable("Unknown integer condition code!");
4561 case ICmpInst::ICMP_EQ:
4563 default: llvm_unreachable("Unknown integer condition code!");
4564 case ICmpInst::ICMP_EQ:
4565 if (LHSCst == SubOne(RHSCst)) {
4566 // (X == 13 | X == 14) -> X-13 <u 2
4567 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4568 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4569 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4570 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4572 break; // (X == 13 | X == 15) -> no change
4573 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4574 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4576 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4577 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4578 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4579 return ReplaceInstUsesWith(I, RHS);
4582 case ICmpInst::ICMP_NE:
4584 default: llvm_unreachable("Unknown integer condition code!");
4585 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4586 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4587 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4588 return ReplaceInstUsesWith(I, LHS);
4589 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4590 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4591 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4592 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4595 case ICmpInst::ICMP_ULT:
4597 default: llvm_unreachable("Unknown integer condition code!");
4598 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4600 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4601 // If RHSCst is [us]MAXINT, it is always false. Not handling
4602 // this can cause overflow.
4603 if (RHSCst->isMaxValue(false))
4604 return ReplaceInstUsesWith(I, LHS);
4605 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4607 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4609 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4610 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4611 return ReplaceInstUsesWith(I, RHS);
4612 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4616 case ICmpInst::ICMP_SLT:
4618 default: llvm_unreachable("Unknown integer condition code!");
4619 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4621 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4622 // If RHSCst is [us]MAXINT, it is always false. Not handling
4623 // this can cause overflow.
4624 if (RHSCst->isMaxValue(true))
4625 return ReplaceInstUsesWith(I, LHS);
4626 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4628 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4630 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4631 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4632 return ReplaceInstUsesWith(I, RHS);
4633 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4637 case ICmpInst::ICMP_UGT:
4639 default: llvm_unreachable("Unknown integer condition code!");
4640 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4641 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4642 return ReplaceInstUsesWith(I, LHS);
4643 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4645 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4646 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4647 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4648 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4652 case ICmpInst::ICMP_SGT:
4654 default: llvm_unreachable("Unknown integer condition code!");
4655 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4656 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4657 return ReplaceInstUsesWith(I, LHS);
4658 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4660 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4661 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4662 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4663 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4671 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4673 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4674 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4675 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4676 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4677 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4678 // If either of the constants are nans, then the whole thing returns
4680 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4681 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4683 // Otherwise, no need to compare the two constants, compare the
4685 return new FCmpInst(FCmpInst::FCMP_UNO,
4686 LHS->getOperand(0), RHS->getOperand(0));
4689 // Handle vector zeros. This occurs because the canonical form of
4690 // "fcmp uno x,x" is "fcmp uno x, 0".
4691 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4692 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4693 return new FCmpInst(FCmpInst::FCMP_UNO,
4694 LHS->getOperand(0), RHS->getOperand(0));
4699 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4700 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4701 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4703 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4704 // Swap RHS operands to match LHS.
4705 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4706 std::swap(Op1LHS, Op1RHS);
4708 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4709 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4711 return new FCmpInst((FCmpInst::Predicate)Op0CC,
4713 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4714 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4715 if (Op0CC == FCmpInst::FCMP_FALSE)
4716 return ReplaceInstUsesWith(I, RHS);
4717 if (Op1CC == FCmpInst::FCMP_FALSE)
4718 return ReplaceInstUsesWith(I, LHS);
4721 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4722 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4723 if (Op0Ordered == Op1Ordered) {
4724 // If both are ordered or unordered, return a new fcmp with
4725 // or'ed predicates.
4726 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4727 Op0LHS, Op0RHS, Context);
4728 if (Instruction *I = dyn_cast<Instruction>(RV))
4730 // Otherwise, it's a constant boolean value...
4731 return ReplaceInstUsesWith(I, RV);
4737 /// FoldOrWithConstants - This helper function folds:
4739 /// ((A | B) & C1) | (B & C2)
4745 /// when the XOR of the two constants is "all ones" (-1).
4746 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4747 Value *A, Value *B, Value *C) {
4748 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4752 ConstantInt *CI2 = 0;
4753 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4755 APInt Xor = CI1->getValue() ^ CI2->getValue();
4756 if (!Xor.isAllOnesValue()) return 0;
4758 if (V1 == A || V1 == B) {
4759 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
4760 return BinaryOperator::CreateOr(NewOp, V1);
4766 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4767 bool Changed = SimplifyCommutative(I);
4768 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4770 if (isa<UndefValue>(Op1)) // X | undef -> -1
4771 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4775 return ReplaceInstUsesWith(I, Op0);
4777 // See if we can simplify any instructions used by the instruction whose sole
4778 // purpose is to compute bits we don't care about.
4779 if (SimplifyDemandedInstructionBits(I))
4781 if (isa<VectorType>(I.getType())) {
4782 if (isa<ConstantAggregateZero>(Op1)) {
4783 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4784 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4785 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4786 return ReplaceInstUsesWith(I, I.getOperand(1));
4791 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4792 ConstantInt *C1 = 0; Value *X = 0;
4793 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4794 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
4796 Value *Or = Builder->CreateOr(X, RHS);
4798 return BinaryOperator::CreateAnd(Or,
4799 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4802 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4803 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
4805 Value *Or = Builder->CreateOr(X, RHS);
4807 return BinaryOperator::CreateXor(Or,
4808 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4811 // Try to fold constant and into select arguments.
4812 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4813 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4815 if (isa<PHINode>(Op0))
4816 if (Instruction *NV = FoldOpIntoPhi(I))
4820 Value *A = 0, *B = 0;
4821 ConstantInt *C1 = 0, *C2 = 0;
4823 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4824 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4825 return ReplaceInstUsesWith(I, Op1);
4826 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4827 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4828 return ReplaceInstUsesWith(I, Op0);
4830 // (A | B) | C and A | (B | C) -> bswap if possible.
4831 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4832 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4833 match(Op1, m_Or(m_Value(), m_Value())) ||
4834 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4835 match(Op1, m_Shift(m_Value(), m_Value())))) {
4836 if (Instruction *BSwap = MatchBSwap(I))
4840 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4841 if (Op0->hasOneUse() &&
4842 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4843 MaskedValueIsZero(Op1, C1->getValue())) {
4844 Value *NOr = Builder->CreateOr(A, Op1);
4846 return BinaryOperator::CreateXor(NOr, C1);
4849 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4850 if (Op1->hasOneUse() &&
4851 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4852 MaskedValueIsZero(Op0, C1->getValue())) {
4853 Value *NOr = Builder->CreateOr(A, Op0);
4855 return BinaryOperator::CreateXor(NOr, C1);
4859 Value *C = 0, *D = 0;
4860 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4861 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4862 Value *V1 = 0, *V2 = 0, *V3 = 0;
4863 C1 = dyn_cast<ConstantInt>(C);
4864 C2 = dyn_cast<ConstantInt>(D);
4865 if (C1 && C2) { // (A & C1)|(B & C2)
4866 // If we have: ((V + N) & C1) | (V & C2)
4867 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4868 // replace with V+N.
4869 if (C1->getValue() == ~C2->getValue()) {
4870 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4871 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4872 // Add commutes, try both ways.
4873 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4874 return ReplaceInstUsesWith(I, A);
4875 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4876 return ReplaceInstUsesWith(I, A);
4878 // Or commutes, try both ways.
4879 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4880 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4881 // Add commutes, try both ways.
4882 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4883 return ReplaceInstUsesWith(I, B);
4884 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4885 return ReplaceInstUsesWith(I, B);
4888 V1 = 0; V2 = 0; V3 = 0;
4891 // Check to see if we have any common things being and'ed. If so, find the
4892 // terms for V1 & (V2|V3).
4893 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4894 if (A == B) // (A & C)|(A & D) == A & (C|D)
4895 V1 = A, V2 = C, V3 = D;
4896 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4897 V1 = A, V2 = B, V3 = C;
4898 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4899 V1 = C, V2 = A, V3 = D;
4900 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4901 V1 = C, V2 = A, V3 = B;
4904 Value *Or = Builder->CreateOr(V2, V3, "tmp");
4905 return BinaryOperator::CreateAnd(V1, Or);
4909 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4910 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4912 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4914 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4916 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4919 // ((A&~B)|(~A&B)) -> A^B
4920 if ((match(C, m_Not(m_Specific(D))) &&
4921 match(B, m_Not(m_Specific(A)))))
4922 return BinaryOperator::CreateXor(A, D);
4923 // ((~B&A)|(~A&B)) -> A^B
4924 if ((match(A, m_Not(m_Specific(D))) &&
4925 match(B, m_Not(m_Specific(C)))))
4926 return BinaryOperator::CreateXor(C, D);
4927 // ((A&~B)|(B&~A)) -> A^B
4928 if ((match(C, m_Not(m_Specific(B))) &&
4929 match(D, m_Not(m_Specific(A)))))
4930 return BinaryOperator::CreateXor(A, B);
4931 // ((~B&A)|(B&~A)) -> A^B
4932 if ((match(A, m_Not(m_Specific(B))) &&
4933 match(D, m_Not(m_Specific(C)))))
4934 return BinaryOperator::CreateXor(C, B);
4937 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4938 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4939 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4940 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4941 SI0->getOperand(1) == SI1->getOperand(1) &&
4942 (SI0->hasOneUse() || SI1->hasOneUse())) {
4943 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
4945 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4946 SI1->getOperand(1));
4950 // ((A|B)&1)|(B&-2) -> (A&1) | B
4951 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4952 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4953 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4954 if (Ret) return Ret;
4956 // (B&-2)|((A|B)&1) -> (A&1) | B
4957 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4958 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4959 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4960 if (Ret) return Ret;
4963 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4964 if (A == Op1) // ~A | A == -1
4965 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4969 // Note, A is still live here!
4970 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4972 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4974 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4975 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4976 Value *And = Builder->CreateAnd(A, B, I.getName()+".demorgan");
4977 return BinaryOperator::CreateNot(And);
4981 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4982 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4983 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4986 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4987 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4991 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4992 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4993 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4994 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4995 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4996 !isa<ICmpInst>(Op1C->getOperand(0))) {
4997 const Type *SrcTy = Op0C->getOperand(0)->getType();
4998 if (SrcTy == Op1C->getOperand(0)->getType() &&
4999 SrcTy->isIntOrIntVector() &&
5000 // Only do this if the casts both really cause code to be
5002 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5004 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5006 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
5007 Op1C->getOperand(0), I.getName());
5008 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5015 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
5016 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
5017 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
5018 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5022 return Changed ? &I : 0;
5027 // XorSelf - Implements: X ^ X --> 0
5030 XorSelf(Value *rhs) : RHS(rhs) {}
5031 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5032 Instruction *apply(BinaryOperator &Xor) const {
5039 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5040 bool Changed = SimplifyCommutative(I);
5041 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5043 if (isa<UndefValue>(Op1)) {
5044 if (isa<UndefValue>(Op0))
5045 // Handle undef ^ undef -> 0 special case. This is a common
5047 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5048 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5051 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5052 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5053 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5054 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5057 // See if we can simplify any instructions used by the instruction whose sole
5058 // purpose is to compute bits we don't care about.
5059 if (SimplifyDemandedInstructionBits(I))
5061 if (isa<VectorType>(I.getType()))
5062 if (isa<ConstantAggregateZero>(Op1))
5063 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5065 // Is this a ~ operation?
5066 if (Value *NotOp = dyn_castNotVal(&I)) {
5067 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5068 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5069 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5070 if (Op0I->getOpcode() == Instruction::And ||
5071 Op0I->getOpcode() == Instruction::Or) {
5072 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
5073 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5075 Builder->CreateNot(Op0I->getOperand(1),
5076 Op0I->getOperand(1)->getName()+".not");
5077 if (Op0I->getOpcode() == Instruction::And)
5078 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5079 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5086 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5087 if (RHS->isOne() && Op0->hasOneUse()) {
5088 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5089 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5090 return new ICmpInst(ICI->getInversePredicate(),
5091 ICI->getOperand(0), ICI->getOperand(1));
5093 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5094 return new FCmpInst(FCI->getInversePredicate(),
5095 FCI->getOperand(0), FCI->getOperand(1));
5098 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5099 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5100 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5101 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5102 Instruction::CastOps Opcode = Op0C->getOpcode();
5103 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
5104 (RHS == ConstantExpr::getCast(Opcode,
5105 ConstantInt::getTrue(*Context),
5106 Op0C->getDestTy()))) {
5107 CI->setPredicate(CI->getInversePredicate());
5108 return CastInst::Create(Opcode, CI, Op0C->getType());
5114 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5115 // ~(c-X) == X-c-1 == X+(-c-1)
5116 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5117 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5118 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5119 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5120 ConstantInt::get(I.getType(), 1));
5121 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5124 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5125 if (Op0I->getOpcode() == Instruction::Add) {
5126 // ~(X-c) --> (-c-1)-X
5127 if (RHS->isAllOnesValue()) {
5128 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5129 return BinaryOperator::CreateSub(
5130 ConstantExpr::getSub(NegOp0CI,
5131 ConstantInt::get(I.getType(), 1)),
5132 Op0I->getOperand(0));
5133 } else if (RHS->getValue().isSignBit()) {
5134 // (X + C) ^ signbit -> (X + C + signbit)
5135 Constant *C = ConstantInt::get(*Context,
5136 RHS->getValue() + Op0CI->getValue());
5137 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5140 } else if (Op0I->getOpcode() == Instruction::Or) {
5141 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5142 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5143 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5144 // Anything in both C1 and C2 is known to be zero, remove it from
5146 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5147 NewRHS = ConstantExpr::getAnd(NewRHS,
5148 ConstantExpr::getNot(CommonBits));
5150 I.setOperand(0, Op0I->getOperand(0));
5151 I.setOperand(1, NewRHS);
5158 // Try to fold constant and into select arguments.
5159 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5160 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5162 if (isa<PHINode>(Op0))
5163 if (Instruction *NV = FoldOpIntoPhi(I))
5167 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5169 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5171 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5173 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5176 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5179 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5180 if (A == Op0) { // B^(B|A) == (A|B)^B
5181 Op1I->swapOperands();
5183 std::swap(Op0, Op1);
5184 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5185 I.swapOperands(); // Simplified below.
5186 std::swap(Op0, Op1);
5188 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5189 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5190 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5191 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5192 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5194 if (A == Op0) { // A^(A&B) -> A^(B&A)
5195 Op1I->swapOperands();
5198 if (B == Op0) { // A^(B&A) -> (B&A)^A
5199 I.swapOperands(); // Simplified below.
5200 std::swap(Op0, Op1);
5205 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5208 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5209 Op0I->hasOneUse()) {
5210 if (A == Op1) // (B|A)^B == (A|B)^B
5212 if (B == Op1) // (A|B)^B == A & ~B
5213 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
5214 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5215 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5216 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5217 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5218 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5220 if (A == Op1) // (A&B)^A -> (B&A)^A
5222 if (B == Op1 && // (B&A)^A == ~B & A
5223 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5224 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
5229 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5230 if (Op0I && Op1I && Op0I->isShift() &&
5231 Op0I->getOpcode() == Op1I->getOpcode() &&
5232 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5233 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5235 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
5237 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5238 Op1I->getOperand(1));
5242 Value *A, *B, *C, *D;
5243 // (A & B)^(A | B) -> A ^ B
5244 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5245 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5246 if ((A == C && B == D) || (A == D && B == C))
5247 return BinaryOperator::CreateXor(A, B);
5249 // (A | B)^(A & B) -> A ^ B
5250 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5251 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5252 if ((A == C && B == D) || (A == D && B == C))
5253 return BinaryOperator::CreateXor(A, B);
5257 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5258 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5259 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5260 // (X & Y)^(X & Y) -> (Y^Z) & X
5261 Value *X = 0, *Y = 0, *Z = 0;
5263 X = A, Y = B, Z = D;
5265 X = A, Y = B, Z = C;
5267 X = B, Y = A, Z = D;
5269 X = B, Y = A, Z = C;
5272 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
5273 return BinaryOperator::CreateAnd(NewOp, X);
5278 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5279 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5280 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5283 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5284 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5285 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5286 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5287 const Type *SrcTy = Op0C->getOperand(0)->getType();
5288 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5289 // Only do this if the casts both really cause code to be generated.
5290 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5292 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5294 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
5295 Op1C->getOperand(0), I.getName());
5296 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5301 return Changed ? &I : 0;
5304 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5305 LLVMContext *Context) {
5306 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5309 static bool HasAddOverflow(ConstantInt *Result,
5310 ConstantInt *In1, ConstantInt *In2,
5313 if (In2->getValue().isNegative())
5314 return Result->getValue().sgt(In1->getValue());
5316 return Result->getValue().slt(In1->getValue());
5318 return Result->getValue().ult(In1->getValue());
5321 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5322 /// overflowed for this type.
5323 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5324 Constant *In2, LLVMContext *Context,
5325 bool IsSigned = false) {
5326 Result = ConstantExpr::getAdd(In1, In2);
5328 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5329 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5330 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5331 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5332 ExtractElement(In1, Idx, Context),
5333 ExtractElement(In2, Idx, Context),
5340 return HasAddOverflow(cast<ConstantInt>(Result),
5341 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5345 static bool HasSubOverflow(ConstantInt *Result,
5346 ConstantInt *In1, ConstantInt *In2,
5349 if (In2->getValue().isNegative())
5350 return Result->getValue().slt(In1->getValue());
5352 return Result->getValue().sgt(In1->getValue());
5354 return Result->getValue().ugt(In1->getValue());
5357 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5358 /// overflowed for this type.
5359 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5360 Constant *In2, LLVMContext *Context,
5361 bool IsSigned = false) {
5362 Result = ConstantExpr::getSub(In1, In2);
5364 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5365 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5366 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5367 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5368 ExtractElement(In1, Idx, Context),
5369 ExtractElement(In2, Idx, Context),
5376 return HasSubOverflow(cast<ConstantInt>(Result),
5377 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5381 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5382 /// code necessary to compute the offset from the base pointer (without adding
5383 /// in the base pointer). Return the result as a signed integer of intptr size.
5384 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5385 TargetData &TD = *IC.getTargetData();
5386 gep_type_iterator GTI = gep_type_begin(GEP);
5387 const Type *IntPtrTy = TD.getIntPtrType(I.getContext());
5388 Value *Result = Constant::getNullValue(IntPtrTy);
5390 // Build a mask for high order bits.
5391 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5392 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5394 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5397 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5398 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5399 if (OpC->isZero()) continue;
5401 // Handle a struct index, which adds its field offset to the pointer.
5402 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5403 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5405 Result = IC.Builder->CreateAdd(Result,
5406 ConstantInt::get(IntPtrTy, Size),
5407 GEP->getName()+".offs");
5411 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5413 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5414 Scale = ConstantExpr::getMul(OC, Scale);
5415 // Emit an add instruction.
5416 Result = IC.Builder->CreateAdd(Result, Scale, GEP->getName()+".offs");
5419 // Convert to correct type.
5420 if (Op->getType() != IntPtrTy)
5421 Op = IC.Builder->CreateIntCast(Op, IntPtrTy, true, Op->getName()+".c");
5423 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5424 // We'll let instcombine(mul) convert this to a shl if possible.
5425 Op = IC.Builder->CreateMul(Op, Scale, GEP->getName()+".idx");
5428 // Emit an add instruction.
5429 Result = IC.Builder->CreateAdd(Op, Result, GEP->getName()+".offs");
5435 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5436 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5437 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5438 /// be complex, and scales are involved. The above expression would also be
5439 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5440 /// This later form is less amenable to optimization though, and we are allowed
5441 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5443 /// If we can't emit an optimized form for this expression, this returns null.
5445 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5447 TargetData &TD = *IC.getTargetData();
5448 gep_type_iterator GTI = gep_type_begin(GEP);
5450 // Check to see if this gep only has a single variable index. If so, and if
5451 // any constant indices are a multiple of its scale, then we can compute this
5452 // in terms of the scale of the variable index. For example, if the GEP
5453 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5454 // because the expression will cross zero at the same point.
5455 unsigned i, e = GEP->getNumOperands();
5457 for (i = 1; i != e; ++i, ++GTI) {
5458 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5459 // Compute the aggregate offset of constant indices.
5460 if (CI->isZero()) continue;
5462 // Handle a struct index, which adds its field offset to the pointer.
5463 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5464 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5466 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5467 Offset += Size*CI->getSExtValue();
5470 // Found our variable index.
5475 // If there are no variable indices, we must have a constant offset, just
5476 // evaluate it the general way.
5477 if (i == e) return 0;
5479 Value *VariableIdx = GEP->getOperand(i);
5480 // Determine the scale factor of the variable element. For example, this is
5481 // 4 if the variable index is into an array of i32.
5482 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5484 // Verify that there are no other variable indices. If so, emit the hard way.
5485 for (++i, ++GTI; i != e; ++i, ++GTI) {
5486 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5489 // Compute the aggregate offset of constant indices.
5490 if (CI->isZero()) continue;
5492 // Handle a struct index, which adds its field offset to the pointer.
5493 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5494 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5496 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5497 Offset += Size*CI->getSExtValue();
5501 // Okay, we know we have a single variable index, which must be a
5502 // pointer/array/vector index. If there is no offset, life is simple, return
5504 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5506 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5507 // we don't need to bother extending: the extension won't affect where the
5508 // computation crosses zero.
5509 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5510 VariableIdx = new TruncInst(VariableIdx,
5511 TD.getIntPtrType(VariableIdx->getContext()),
5512 VariableIdx->getName(), &I);
5516 // Otherwise, there is an index. The computation we will do will be modulo
5517 // the pointer size, so get it.
5518 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5520 Offset &= PtrSizeMask;
5521 VariableScale &= PtrSizeMask;
5523 // To do this transformation, any constant index must be a multiple of the
5524 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5525 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5526 // multiple of the variable scale.
5527 int64_t NewOffs = Offset / (int64_t)VariableScale;
5528 if (Offset != NewOffs*(int64_t)VariableScale)
5531 // Okay, we can do this evaluation. Start by converting the index to intptr.
5532 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
5533 if (VariableIdx->getType() != IntPtrTy)
5534 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5536 VariableIdx->getName(), &I);
5537 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5538 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5542 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5543 /// else. At this point we know that the GEP is on the LHS of the comparison.
5544 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5545 ICmpInst::Predicate Cond,
5547 // Look through bitcasts.
5548 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5549 RHS = BCI->getOperand(0);
5551 Value *PtrBase = GEPLHS->getOperand(0);
5552 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5553 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5554 // This transformation (ignoring the base and scales) is valid because we
5555 // know pointers can't overflow since the gep is inbounds. See if we can
5556 // output an optimized form.
5557 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5559 // If not, synthesize the offset the hard way.
5561 Offset = EmitGEPOffset(GEPLHS, I, *this);
5562 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5563 Constant::getNullValue(Offset->getType()));
5564 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5565 // If the base pointers are different, but the indices are the same, just
5566 // compare the base pointer.
5567 if (PtrBase != GEPRHS->getOperand(0)) {
5568 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5569 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5570 GEPRHS->getOperand(0)->getType();
5572 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5573 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5574 IndicesTheSame = false;
5578 // If all indices are the same, just compare the base pointers.
5580 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5581 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5583 // Otherwise, the base pointers are different and the indices are
5584 // different, bail out.
5588 // If one of the GEPs has all zero indices, recurse.
5589 bool AllZeros = true;
5590 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5591 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5592 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5597 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5598 ICmpInst::getSwappedPredicate(Cond), I);
5600 // If the other GEP has all zero indices, recurse.
5602 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5603 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5604 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5609 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5611 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5612 // If the GEPs only differ by one index, compare it.
5613 unsigned NumDifferences = 0; // Keep track of # differences.
5614 unsigned DiffOperand = 0; // The operand that differs.
5615 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5616 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5617 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5618 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5619 // Irreconcilable differences.
5623 if (NumDifferences++) break;
5628 if (NumDifferences == 0) // SAME GEP?
5629 return ReplaceInstUsesWith(I, // No comparison is needed here.
5630 ConstantInt::get(Type::getInt1Ty(*Context),
5631 ICmpInst::isTrueWhenEqual(Cond)));
5633 else if (NumDifferences == 1) {
5634 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5635 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5636 // Make sure we do a signed comparison here.
5637 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5641 // Only lower this if the icmp is the only user of the GEP or if we expect
5642 // the result to fold to a constant!
5644 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5645 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5646 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5647 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5648 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5649 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5655 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5657 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5660 if (!isa<ConstantFP>(RHSC)) return 0;
5661 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5663 // Get the width of the mantissa. We don't want to hack on conversions that
5664 // might lose information from the integer, e.g. "i64 -> float"
5665 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5666 if (MantissaWidth == -1) return 0; // Unknown.
5668 // Check to see that the input is converted from an integer type that is small
5669 // enough that preserves all bits. TODO: check here for "known" sign bits.
5670 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5671 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5673 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5674 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5678 // If the conversion would lose info, don't hack on this.
5679 if ((int)InputSize > MantissaWidth)
5682 // Otherwise, we can potentially simplify the comparison. We know that it
5683 // will always come through as an integer value and we know the constant is
5684 // not a NAN (it would have been previously simplified).
5685 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5687 ICmpInst::Predicate Pred;
5688 switch (I.getPredicate()) {
5689 default: llvm_unreachable("Unexpected predicate!");
5690 case FCmpInst::FCMP_UEQ:
5691 case FCmpInst::FCMP_OEQ:
5692 Pred = ICmpInst::ICMP_EQ;
5694 case FCmpInst::FCMP_UGT:
5695 case FCmpInst::FCMP_OGT:
5696 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5698 case FCmpInst::FCMP_UGE:
5699 case FCmpInst::FCMP_OGE:
5700 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5702 case FCmpInst::FCMP_ULT:
5703 case FCmpInst::FCMP_OLT:
5704 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5706 case FCmpInst::FCMP_ULE:
5707 case FCmpInst::FCMP_OLE:
5708 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5710 case FCmpInst::FCMP_UNE:
5711 case FCmpInst::FCMP_ONE:
5712 Pred = ICmpInst::ICMP_NE;
5714 case FCmpInst::FCMP_ORD:
5715 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5716 case FCmpInst::FCMP_UNO:
5717 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5720 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5722 // Now we know that the APFloat is a normal number, zero or inf.
5724 // See if the FP constant is too large for the integer. For example,
5725 // comparing an i8 to 300.0.
5726 unsigned IntWidth = IntTy->getScalarSizeInBits();
5729 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5730 // and large values.
5731 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5732 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5733 APFloat::rmNearestTiesToEven);
5734 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5735 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5736 Pred == ICmpInst::ICMP_SLE)
5737 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5738 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5741 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5742 // +INF and large values.
5743 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5744 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5745 APFloat::rmNearestTiesToEven);
5746 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5747 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5748 Pred == ICmpInst::ICMP_ULE)
5749 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5750 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5755 // See if the RHS value is < SignedMin.
5756 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5757 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5758 APFloat::rmNearestTiesToEven);
5759 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5760 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5761 Pred == ICmpInst::ICMP_SGE)
5762 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5763 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5767 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5768 // [0, UMAX], but it may still be fractional. See if it is fractional by
5769 // casting the FP value to the integer value and back, checking for equality.
5770 // Don't do this for zero, because -0.0 is not fractional.
5771 Constant *RHSInt = LHSUnsigned
5772 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5773 : ConstantExpr::getFPToSI(RHSC, IntTy);
5774 if (!RHS.isZero()) {
5775 bool Equal = LHSUnsigned
5776 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5777 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5779 // If we had a comparison against a fractional value, we have to adjust
5780 // the compare predicate and sometimes the value. RHSC is rounded towards
5781 // zero at this point.
5783 default: llvm_unreachable("Unexpected integer comparison!");
5784 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5785 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5786 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5787 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5788 case ICmpInst::ICMP_ULE:
5789 // (float)int <= 4.4 --> int <= 4
5790 // (float)int <= -4.4 --> false
5791 if (RHS.isNegative())
5792 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5794 case ICmpInst::ICMP_SLE:
5795 // (float)int <= 4.4 --> int <= 4
5796 // (float)int <= -4.4 --> int < -4
5797 if (RHS.isNegative())
5798 Pred = ICmpInst::ICMP_SLT;
5800 case ICmpInst::ICMP_ULT:
5801 // (float)int < -4.4 --> false
5802 // (float)int < 4.4 --> int <= 4
5803 if (RHS.isNegative())
5804 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5805 Pred = ICmpInst::ICMP_ULE;
5807 case ICmpInst::ICMP_SLT:
5808 // (float)int < -4.4 --> int < -4
5809 // (float)int < 4.4 --> int <= 4
5810 if (!RHS.isNegative())
5811 Pred = ICmpInst::ICMP_SLE;
5813 case ICmpInst::ICMP_UGT:
5814 // (float)int > 4.4 --> int > 4
5815 // (float)int > -4.4 --> true
5816 if (RHS.isNegative())
5817 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5819 case ICmpInst::ICMP_SGT:
5820 // (float)int > 4.4 --> int > 4
5821 // (float)int > -4.4 --> int >= -4
5822 if (RHS.isNegative())
5823 Pred = ICmpInst::ICMP_SGE;
5825 case ICmpInst::ICMP_UGE:
5826 // (float)int >= -4.4 --> true
5827 // (float)int >= 4.4 --> int > 4
5828 if (!RHS.isNegative())
5829 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5830 Pred = ICmpInst::ICMP_UGT;
5832 case ICmpInst::ICMP_SGE:
5833 // (float)int >= -4.4 --> int >= -4
5834 // (float)int >= 4.4 --> int > 4
5835 if (!RHS.isNegative())
5836 Pred = ICmpInst::ICMP_SGT;
5842 // Lower this FP comparison into an appropriate integer version of the
5844 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5847 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5848 bool Changed = SimplifyCompare(I);
5849 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5851 // Fold trivial predicates.
5852 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5853 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 0));
5854 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5855 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 1));
5857 // Simplify 'fcmp pred X, X'
5859 switch (I.getPredicate()) {
5860 default: llvm_unreachable("Unknown predicate!");
5861 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5862 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5863 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5864 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 1));
5865 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5866 case FCmpInst::FCMP_OLT: // True if ordered and less than
5867 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5868 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 0));
5870 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5871 case FCmpInst::FCMP_ULT: // True if unordered or less than
5872 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5873 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5874 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5875 I.setPredicate(FCmpInst::FCMP_UNO);
5876 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5879 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5880 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5881 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5882 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5883 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5884 I.setPredicate(FCmpInst::FCMP_ORD);
5885 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5890 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5891 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
5893 // Handle fcmp with constant RHS
5894 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5895 // If the constant is a nan, see if we can fold the comparison based on it.
5896 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5897 if (CFP->getValueAPF().isNaN()) {
5898 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5899 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5900 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5901 "Comparison must be either ordered or unordered!");
5902 // True if unordered.
5903 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5907 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5908 switch (LHSI->getOpcode()) {
5909 case Instruction::PHI:
5910 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5911 // block. If in the same block, we're encouraging jump threading. If
5912 // not, we are just pessimizing the code by making an i1 phi.
5913 if (LHSI->getParent() == I.getParent())
5914 if (Instruction *NV = FoldOpIntoPhi(I, true))
5917 case Instruction::SIToFP:
5918 case Instruction::UIToFP:
5919 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5922 case Instruction::Select:
5923 // If either operand of the select is a constant, we can fold the
5924 // comparison into the select arms, which will cause one to be
5925 // constant folded and the select turned into a bitwise or.
5926 Value *Op1 = 0, *Op2 = 0;
5927 if (LHSI->hasOneUse()) {
5928 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5929 // Fold the known value into the constant operand.
5930 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5931 // Insert a new FCmp of the other select operand.
5932 Op2 = Builder->CreateFCmp(I.getPredicate(),
5933 LHSI->getOperand(2), RHSC, I.getName());
5934 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5935 // Fold the known value into the constant operand.
5936 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5937 // Insert a new FCmp of the other select operand.
5938 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
5944 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5949 return Changed ? &I : 0;
5952 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5953 bool Changed = SimplifyCompare(I);
5954 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5955 const Type *Ty = Op0->getType();
5959 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(),
5960 I.isTrueWhenEqual()));
5962 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5963 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
5965 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5966 // addresses never equal each other! We already know that Op0 != Op1.
5967 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5968 isa<ConstantPointerNull>(Op0)) &&
5969 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5970 isa<ConstantPointerNull>(Op1)))
5971 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
5972 !I.isTrueWhenEqual()));
5974 // icmp's with boolean values can always be turned into bitwise operations
5975 if (Ty == Type::getInt1Ty(*Context)) {
5976 switch (I.getPredicate()) {
5977 default: llvm_unreachable("Invalid icmp instruction!");
5978 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5979 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
5980 return BinaryOperator::CreateNot(Xor);
5982 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5983 return BinaryOperator::CreateXor(Op0, Op1);
5985 case ICmpInst::ICMP_UGT:
5986 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5988 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5989 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
5990 return BinaryOperator::CreateAnd(Not, Op1);
5992 case ICmpInst::ICMP_SGT:
5993 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5995 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5996 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
5997 return BinaryOperator::CreateAnd(Not, Op0);
5999 case ICmpInst::ICMP_UGE:
6000 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6002 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6003 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6004 return BinaryOperator::CreateOr(Not, Op1);
6006 case ICmpInst::ICMP_SGE:
6007 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6009 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6010 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6011 return BinaryOperator::CreateOr(Not, Op0);
6016 unsigned BitWidth = 0;
6018 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6019 else if (Ty->isIntOrIntVector())
6020 BitWidth = Ty->getScalarSizeInBits();
6022 bool isSignBit = false;
6024 // See if we are doing a comparison with a constant.
6025 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6026 Value *A = 0, *B = 0;
6028 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6029 if (I.isEquality() && CI->isNullValue() &&
6030 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6031 // (icmp cond A B) if cond is equality
6032 return new ICmpInst(I.getPredicate(), A, B);
6035 // If we have an icmp le or icmp ge instruction, turn it into the
6036 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6037 // them being folded in the code below.
6038 switch (I.getPredicate()) {
6040 case ICmpInst::ICMP_ULE:
6041 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6042 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6043 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
6045 case ICmpInst::ICMP_SLE:
6046 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6047 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6048 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6050 case ICmpInst::ICMP_UGE:
6051 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6052 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6053 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
6055 case ICmpInst::ICMP_SGE:
6056 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6057 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6058 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6062 // If this comparison is a normal comparison, it demands all
6063 // bits, if it is a sign bit comparison, it only demands the sign bit.
6065 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6068 // See if we can fold the comparison based on range information we can get
6069 // by checking whether bits are known to be zero or one in the input.
6070 if (BitWidth != 0) {
6071 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6072 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6074 if (SimplifyDemandedBits(I.getOperandUse(0),
6075 isSignBit ? APInt::getSignBit(BitWidth)
6076 : APInt::getAllOnesValue(BitWidth),
6077 Op0KnownZero, Op0KnownOne, 0))
6079 if (SimplifyDemandedBits(I.getOperandUse(1),
6080 APInt::getAllOnesValue(BitWidth),
6081 Op1KnownZero, Op1KnownOne, 0))
6084 // Given the known and unknown bits, compute a range that the LHS could be
6085 // in. Compute the Min, Max and RHS values based on the known bits. For the
6086 // EQ and NE we use unsigned values.
6087 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6088 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6090 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6092 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6095 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6097 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6101 // If Min and Max are known to be the same, then SimplifyDemandedBits
6102 // figured out that the LHS is a constant. Just constant fold this now so
6103 // that code below can assume that Min != Max.
6104 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6105 return new ICmpInst(I.getPredicate(),
6106 ConstantInt::get(*Context, Op0Min), Op1);
6107 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6108 return new ICmpInst(I.getPredicate(), Op0,
6109 ConstantInt::get(*Context, Op1Min));
6111 // Based on the range information we know about the LHS, see if we can
6112 // simplify this comparison. For example, (x&4) < 8 is always true.
6113 switch (I.getPredicate()) {
6114 default: llvm_unreachable("Unknown icmp opcode!");
6115 case ICmpInst::ICMP_EQ:
6116 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6117 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6119 case ICmpInst::ICMP_NE:
6120 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6121 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6123 case ICmpInst::ICMP_ULT:
6124 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6125 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6126 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6127 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6128 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6129 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6130 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6131 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6132 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6135 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6136 if (CI->isMinValue(true))
6137 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6138 Constant::getAllOnesValue(Op0->getType()));
6141 case ICmpInst::ICMP_UGT:
6142 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6143 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6144 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6145 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6147 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6148 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6149 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6150 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6151 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6154 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6155 if (CI->isMaxValue(true))
6156 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6157 Constant::getNullValue(Op0->getType()));
6160 case ICmpInst::ICMP_SLT:
6161 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6162 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6163 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6164 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6165 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6166 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6167 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6168 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6169 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6173 case ICmpInst::ICMP_SGT:
6174 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6175 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6176 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6177 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6179 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6180 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6181 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6182 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6183 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6187 case ICmpInst::ICMP_SGE:
6188 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6189 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6190 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6191 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6192 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6194 case ICmpInst::ICMP_SLE:
6195 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6196 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6197 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6198 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6199 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6201 case ICmpInst::ICMP_UGE:
6202 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6203 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6204 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6205 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6206 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6208 case ICmpInst::ICMP_ULE:
6209 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6210 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6211 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6212 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6213 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6217 // Turn a signed comparison into an unsigned one if both operands
6218 // are known to have the same sign.
6220 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6221 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6222 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6225 // Test if the ICmpInst instruction is used exclusively by a select as
6226 // part of a minimum or maximum operation. If so, refrain from doing
6227 // any other folding. This helps out other analyses which understand
6228 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6229 // and CodeGen. And in this case, at least one of the comparison
6230 // operands has at least one user besides the compare (the select),
6231 // which would often largely negate the benefit of folding anyway.
6233 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6234 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6235 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6238 // See if we are doing a comparison between a constant and an instruction that
6239 // can be folded into the comparison.
6240 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6241 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6242 // instruction, see if that instruction also has constants so that the
6243 // instruction can be folded into the icmp
6244 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6245 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6249 // Handle icmp with constant (but not simple integer constant) RHS
6250 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6251 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6252 switch (LHSI->getOpcode()) {
6253 case Instruction::GetElementPtr:
6254 if (RHSC->isNullValue()) {
6255 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6256 bool isAllZeros = true;
6257 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6258 if (!isa<Constant>(LHSI->getOperand(i)) ||
6259 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6264 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6265 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6269 case Instruction::PHI:
6270 // Only fold icmp into the PHI if the phi and icmp are in the same
6271 // block. If in the same block, we're encouraging jump threading. If
6272 // not, we are just pessimizing the code by making an i1 phi.
6273 if (LHSI->getParent() == I.getParent())
6274 if (Instruction *NV = FoldOpIntoPhi(I, true))
6277 case Instruction::Select: {
6278 // If either operand of the select is a constant, we can fold the
6279 // comparison into the select arms, which will cause one to be
6280 // constant folded and the select turned into a bitwise or.
6281 Value *Op1 = 0, *Op2 = 0;
6282 if (LHSI->hasOneUse()) {
6283 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6284 // Fold the known value into the constant operand.
6285 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6286 // Insert a new ICmp of the other select operand.
6287 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
6289 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6290 // Fold the known value into the constant operand.
6291 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6292 // Insert a new ICmp of the other select operand.
6293 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
6299 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6302 case Instruction::Call:
6303 // If we have (malloc != null), and if the malloc has a single use, we
6304 // can assume it is successful and remove the malloc.
6305 if (isMalloc(LHSI) && LHSI->hasOneUse() &&
6306 isa<ConstantPointerNull>(RHSC)) {
6307 // Need to explicitly erase malloc call here, instead of adding it to
6308 // Worklist, because it won't get DCE'd from the Worklist since
6309 // isInstructionTriviallyDead() returns false for function calls.
6310 // It is OK to replace LHSI/MallocCall with Undef because the
6311 // instruction that uses it will be erased via Worklist.
6312 if (extractMallocCall(LHSI)) {
6313 LHSI->replaceAllUsesWith(UndefValue::get(LHSI->getType()));
6314 EraseInstFromFunction(*LHSI);
6315 return ReplaceInstUsesWith(I,
6316 ConstantInt::get(Type::getInt1Ty(*Context),
6317 !I.isTrueWhenEqual()));
6319 if (CallInst* MallocCall = extractMallocCallFromBitCast(LHSI))
6320 if (MallocCall->hasOneUse()) {
6321 MallocCall->replaceAllUsesWith(
6322 UndefValue::get(MallocCall->getType()));
6323 EraseInstFromFunction(*MallocCall);
6324 Worklist.Add(LHSI); // The malloc's bitcast use.
6325 return ReplaceInstUsesWith(I,
6326 ConstantInt::get(Type::getInt1Ty(*Context),
6327 !I.isTrueWhenEqual()));
6334 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6335 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6336 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6338 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6339 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6340 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6343 // Test to see if the operands of the icmp are casted versions of other
6344 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6346 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6347 if (isa<PointerType>(Op0->getType()) &&
6348 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6349 // We keep moving the cast from the left operand over to the right
6350 // operand, where it can often be eliminated completely.
6351 Op0 = CI->getOperand(0);
6353 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6354 // so eliminate it as well.
6355 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6356 Op1 = CI2->getOperand(0);
6358 // If Op1 is a constant, we can fold the cast into the constant.
6359 if (Op0->getType() != Op1->getType()) {
6360 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6361 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6363 // Otherwise, cast the RHS right before the icmp
6364 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
6367 return new ICmpInst(I.getPredicate(), Op0, Op1);
6371 if (isa<CastInst>(Op0)) {
6372 // Handle the special case of: icmp (cast bool to X), <cst>
6373 // This comes up when you have code like
6376 // For generality, we handle any zero-extension of any operand comparison
6377 // with a constant or another cast from the same type.
6378 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6379 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6383 // See if it's the same type of instruction on the left and right.
6384 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6385 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6386 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6387 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6388 switch (Op0I->getOpcode()) {
6390 case Instruction::Add:
6391 case Instruction::Sub:
6392 case Instruction::Xor:
6393 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6394 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6395 Op1I->getOperand(0));
6396 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6397 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6398 if (CI->getValue().isSignBit()) {
6399 ICmpInst::Predicate Pred = I.isSigned()
6400 ? I.getUnsignedPredicate()
6401 : I.getSignedPredicate();
6402 return new ICmpInst(Pred, Op0I->getOperand(0),
6403 Op1I->getOperand(0));
6406 if (CI->getValue().isMaxSignedValue()) {
6407 ICmpInst::Predicate Pred = I.isSigned()
6408 ? I.getUnsignedPredicate()
6409 : I.getSignedPredicate();
6410 Pred = I.getSwappedPredicate(Pred);
6411 return new ICmpInst(Pred, Op0I->getOperand(0),
6412 Op1I->getOperand(0));
6416 case Instruction::Mul:
6417 if (!I.isEquality())
6420 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6421 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6422 // Mask = -1 >> count-trailing-zeros(Cst).
6423 if (!CI->isZero() && !CI->isOne()) {
6424 const APInt &AP = CI->getValue();
6425 ConstantInt *Mask = ConstantInt::get(*Context,
6426 APInt::getLowBitsSet(AP.getBitWidth(),
6428 AP.countTrailingZeros()));
6429 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
6430 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
6431 return new ICmpInst(I.getPredicate(), And1, And2);
6440 // ~x < ~y --> y < x
6442 if (match(Op0, m_Not(m_Value(A))) &&
6443 match(Op1, m_Not(m_Value(B))))
6444 return new ICmpInst(I.getPredicate(), B, A);
6447 if (I.isEquality()) {
6448 Value *A, *B, *C, *D;
6450 // -x == -y --> x == y
6451 if (match(Op0, m_Neg(m_Value(A))) &&
6452 match(Op1, m_Neg(m_Value(B))))
6453 return new ICmpInst(I.getPredicate(), A, B);
6455 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6456 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6457 Value *OtherVal = A == Op1 ? B : A;
6458 return new ICmpInst(I.getPredicate(), OtherVal,
6459 Constant::getNullValue(A->getType()));
6462 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6463 // A^c1 == C^c2 --> A == C^(c1^c2)
6464 ConstantInt *C1, *C2;
6465 if (match(B, m_ConstantInt(C1)) &&
6466 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6468 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6469 Value *Xor = Builder->CreateXor(C, NC, "tmp");
6470 return new ICmpInst(I.getPredicate(), A, Xor);
6473 // A^B == A^D -> B == D
6474 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6475 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6476 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6477 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6481 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6482 (A == Op0 || B == Op0)) {
6483 // A == (A^B) -> B == 0
6484 Value *OtherVal = A == Op0 ? B : A;
6485 return new ICmpInst(I.getPredicate(), OtherVal,
6486 Constant::getNullValue(A->getType()));
6489 // (A-B) == A -> B == 0
6490 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6491 return new ICmpInst(I.getPredicate(), B,
6492 Constant::getNullValue(B->getType()));
6494 // A == (A-B) -> B == 0
6495 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6496 return new ICmpInst(I.getPredicate(), B,
6497 Constant::getNullValue(B->getType()));
6499 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6500 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6501 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6502 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6503 Value *X = 0, *Y = 0, *Z = 0;
6506 X = B; Y = D; Z = A;
6507 } else if (A == D) {
6508 X = B; Y = C; Z = A;
6509 } else if (B == C) {
6510 X = A; Y = D; Z = B;
6511 } else if (B == D) {
6512 X = A; Y = C; Z = B;
6515 if (X) { // Build (X^Y) & Z
6516 Op1 = Builder->CreateXor(X, Y, "tmp");
6517 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
6518 I.setOperand(0, Op1);
6519 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6524 return Changed ? &I : 0;
6528 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6529 /// and CmpRHS are both known to be integer constants.
6530 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6531 ConstantInt *DivRHS) {
6532 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6533 const APInt &CmpRHSV = CmpRHS->getValue();
6535 // FIXME: If the operand types don't match the type of the divide
6536 // then don't attempt this transform. The code below doesn't have the
6537 // logic to deal with a signed divide and an unsigned compare (and
6538 // vice versa). This is because (x /s C1) <s C2 produces different
6539 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6540 // (x /u C1) <u C2. Simply casting the operands and result won't
6541 // work. :( The if statement below tests that condition and bails
6543 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6544 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
6546 if (DivRHS->isZero())
6547 return 0; // The ProdOV computation fails on divide by zero.
6548 if (DivIsSigned && DivRHS->isAllOnesValue())
6549 return 0; // The overflow computation also screws up here
6550 if (DivRHS->isOne())
6551 return 0; // Not worth bothering, and eliminates some funny cases
6554 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6555 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6556 // C2 (CI). By solving for X we can turn this into a range check
6557 // instead of computing a divide.
6558 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6560 // Determine if the product overflows by seeing if the product is
6561 // not equal to the divide. Make sure we do the same kind of divide
6562 // as in the LHS instruction that we're folding.
6563 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6564 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6566 // Get the ICmp opcode
6567 ICmpInst::Predicate Pred = ICI.getPredicate();
6569 // Figure out the interval that is being checked. For example, a comparison
6570 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6571 // Compute this interval based on the constants involved and the signedness of
6572 // the compare/divide. This computes a half-open interval, keeping track of
6573 // whether either value in the interval overflows. After analysis each
6574 // overflow variable is set to 0 if it's corresponding bound variable is valid
6575 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6576 int LoOverflow = 0, HiOverflow = 0;
6577 Constant *LoBound = 0, *HiBound = 0;
6579 if (!DivIsSigned) { // udiv
6580 // e.g. X/5 op 3 --> [15, 20)
6582 HiOverflow = LoOverflow = ProdOV;
6584 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6585 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6586 if (CmpRHSV == 0) { // (X / pos) op 0
6587 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6588 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6590 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6591 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6592 HiOverflow = LoOverflow = ProdOV;
6594 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6595 } else { // (X / pos) op neg
6596 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6597 HiBound = AddOne(Prod);
6598 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6600 ConstantInt* DivNeg =
6601 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6602 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6606 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6607 if (CmpRHSV == 0) { // (X / neg) op 0
6608 // e.g. X/-5 op 0 --> [-4, 5)
6609 LoBound = AddOne(DivRHS);
6610 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6611 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6612 HiOverflow = 1; // [INTMIN+1, overflow)
6613 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6615 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6616 // e.g. X/-5 op 3 --> [-19, -14)
6617 HiBound = AddOne(Prod);
6618 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6620 LoOverflow = AddWithOverflow(LoBound, HiBound,
6621 DivRHS, Context, true) ? -1 : 0;
6622 } else { // (X / neg) op neg
6623 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6624 LoOverflow = HiOverflow = ProdOV;
6626 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6629 // Dividing by a negative swaps the condition. LT <-> GT
6630 Pred = ICmpInst::getSwappedPredicate(Pred);
6633 Value *X = DivI->getOperand(0);
6635 default: llvm_unreachable("Unhandled icmp opcode!");
6636 case ICmpInst::ICMP_EQ:
6637 if (LoOverflow && HiOverflow)
6638 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6639 else if (HiOverflow)
6640 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6641 ICmpInst::ICMP_UGE, X, LoBound);
6642 else if (LoOverflow)
6643 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6644 ICmpInst::ICMP_ULT, X, HiBound);
6646 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6647 case ICmpInst::ICMP_NE:
6648 if (LoOverflow && HiOverflow)
6649 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6650 else if (HiOverflow)
6651 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6652 ICmpInst::ICMP_ULT, X, LoBound);
6653 else if (LoOverflow)
6654 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6655 ICmpInst::ICMP_UGE, X, HiBound);
6657 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6658 case ICmpInst::ICMP_ULT:
6659 case ICmpInst::ICMP_SLT:
6660 if (LoOverflow == +1) // Low bound is greater than input range.
6661 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6662 if (LoOverflow == -1) // Low bound is less than input range.
6663 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6664 return new ICmpInst(Pred, X, LoBound);
6665 case ICmpInst::ICMP_UGT:
6666 case ICmpInst::ICMP_SGT:
6667 if (HiOverflow == +1) // High bound greater than input range.
6668 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6669 else if (HiOverflow == -1) // High bound less than input range.
6670 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6671 if (Pred == ICmpInst::ICMP_UGT)
6672 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6674 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6679 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6681 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6684 const APInt &RHSV = RHS->getValue();
6686 switch (LHSI->getOpcode()) {
6687 case Instruction::Trunc:
6688 if (ICI.isEquality() && LHSI->hasOneUse()) {
6689 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6690 // of the high bits truncated out of x are known.
6691 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6692 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6693 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6694 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6695 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6697 // If all the high bits are known, we can do this xform.
6698 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6699 // Pull in the high bits from known-ones set.
6700 APInt NewRHS(RHS->getValue());
6701 NewRHS.zext(SrcBits);
6703 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6704 ConstantInt::get(*Context, NewRHS));
6709 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6710 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6711 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6713 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6714 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6715 Value *CompareVal = LHSI->getOperand(0);
6717 // If the sign bit of the XorCST is not set, there is no change to
6718 // the operation, just stop using the Xor.
6719 if (!XorCST->getValue().isNegative()) {
6720 ICI.setOperand(0, CompareVal);
6725 // Was the old condition true if the operand is positive?
6726 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6728 // If so, the new one isn't.
6729 isTrueIfPositive ^= true;
6731 if (isTrueIfPositive)
6732 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
6735 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
6739 if (LHSI->hasOneUse()) {
6740 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6741 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6742 const APInt &SignBit = XorCST->getValue();
6743 ICmpInst::Predicate Pred = ICI.isSigned()
6744 ? ICI.getUnsignedPredicate()
6745 : ICI.getSignedPredicate();
6746 return new ICmpInst(Pred, LHSI->getOperand(0),
6747 ConstantInt::get(*Context, RHSV ^ SignBit));
6750 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6751 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6752 const APInt &NotSignBit = XorCST->getValue();
6753 ICmpInst::Predicate Pred = ICI.isSigned()
6754 ? ICI.getUnsignedPredicate()
6755 : ICI.getSignedPredicate();
6756 Pred = ICI.getSwappedPredicate(Pred);
6757 return new ICmpInst(Pred, LHSI->getOperand(0),
6758 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6763 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6764 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6765 LHSI->getOperand(0)->hasOneUse()) {
6766 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6768 // If the LHS is an AND of a truncating cast, we can widen the
6769 // and/compare to be the input width without changing the value
6770 // produced, eliminating a cast.
6771 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6772 // We can do this transformation if either the AND constant does not
6773 // have its sign bit set or if it is an equality comparison.
6774 // Extending a relational comparison when we're checking the sign
6775 // bit would not work.
6776 if (Cast->hasOneUse() &&
6777 (ICI.isEquality() ||
6778 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6780 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6781 APInt NewCST = AndCST->getValue();
6782 NewCST.zext(BitWidth);
6784 NewCI.zext(BitWidth);
6786 Builder->CreateAnd(Cast->getOperand(0),
6787 ConstantInt::get(*Context, NewCST), LHSI->getName());
6788 return new ICmpInst(ICI.getPredicate(), NewAnd,
6789 ConstantInt::get(*Context, NewCI));
6793 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6794 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6795 // happens a LOT in code produced by the C front-end, for bitfield
6797 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6798 if (Shift && !Shift->isShift())
6802 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6803 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6804 const Type *AndTy = AndCST->getType(); // Type of the and.
6806 // We can fold this as long as we can't shift unknown bits
6807 // into the mask. This can only happen with signed shift
6808 // rights, as they sign-extend.
6810 bool CanFold = Shift->isLogicalShift();
6812 // To test for the bad case of the signed shr, see if any
6813 // of the bits shifted in could be tested after the mask.
6814 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6815 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6817 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6818 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6819 AndCST->getValue()) == 0)
6825 if (Shift->getOpcode() == Instruction::Shl)
6826 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6828 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6830 // Check to see if we are shifting out any of the bits being
6832 if (ConstantExpr::get(Shift->getOpcode(),
6833 NewCst, ShAmt) != RHS) {
6834 // If we shifted bits out, the fold is not going to work out.
6835 // As a special case, check to see if this means that the
6836 // result is always true or false now.
6837 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6838 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6839 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6840 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6842 ICI.setOperand(1, NewCst);
6843 Constant *NewAndCST;
6844 if (Shift->getOpcode() == Instruction::Shl)
6845 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6847 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6848 LHSI->setOperand(1, NewAndCST);
6849 LHSI->setOperand(0, Shift->getOperand(0));
6850 Worklist.Add(Shift); // Shift is dead.
6856 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6857 // preferable because it allows the C<<Y expression to be hoisted out
6858 // of a loop if Y is invariant and X is not.
6859 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6860 ICI.isEquality() && !Shift->isArithmeticShift() &&
6861 !isa<Constant>(Shift->getOperand(0))) {
6864 if (Shift->getOpcode() == Instruction::LShr) {
6865 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
6867 // Insert a logical shift.
6868 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
6871 // Compute X & (C << Y).
6873 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6875 ICI.setOperand(0, NewAnd);
6881 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6882 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6885 uint32_t TypeBits = RHSV.getBitWidth();
6887 // Check that the shift amount is in range. If not, don't perform
6888 // undefined shifts. When the shift is visited it will be
6890 if (ShAmt->uge(TypeBits))
6893 if (ICI.isEquality()) {
6894 // If we are comparing against bits always shifted out, the
6895 // comparison cannot succeed.
6897 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6899 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6900 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6901 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6902 return ReplaceInstUsesWith(ICI, Cst);
6905 if (LHSI->hasOneUse()) {
6906 // Otherwise strength reduce the shift into an and.
6907 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6909 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6910 TypeBits-ShAmtVal));
6913 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
6914 return new ICmpInst(ICI.getPredicate(), And,
6915 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6919 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6920 bool TrueIfSigned = false;
6921 if (LHSI->hasOneUse() &&
6922 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6923 // (X << 31) <s 0 --> (X&1) != 0
6924 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6925 (TypeBits-ShAmt->getZExtValue()-1));
6927 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
6928 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6929 And, Constant::getNullValue(And->getType()));
6934 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6935 case Instruction::AShr: {
6936 // Only handle equality comparisons of shift-by-constant.
6937 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6938 if (!ShAmt || !ICI.isEquality()) break;
6940 // Check that the shift amount is in range. If not, don't perform
6941 // undefined shifts. When the shift is visited it will be
6943 uint32_t TypeBits = RHSV.getBitWidth();
6944 if (ShAmt->uge(TypeBits))
6947 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6949 // If we are comparing against bits always shifted out, the
6950 // comparison cannot succeed.
6951 APInt Comp = RHSV << ShAmtVal;
6952 if (LHSI->getOpcode() == Instruction::LShr)
6953 Comp = Comp.lshr(ShAmtVal);
6955 Comp = Comp.ashr(ShAmtVal);
6957 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6958 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6959 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6960 return ReplaceInstUsesWith(ICI, Cst);
6963 // Otherwise, check to see if the bits shifted out are known to be zero.
6964 // If so, we can compare against the unshifted value:
6965 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6966 if (LHSI->hasOneUse() &&
6967 MaskedValueIsZero(LHSI->getOperand(0),
6968 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6969 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6970 ConstantExpr::getShl(RHS, ShAmt));
6973 if (LHSI->hasOneUse()) {
6974 // Otherwise strength reduce the shift into an and.
6975 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6976 Constant *Mask = ConstantInt::get(*Context, Val);
6978 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
6979 Mask, LHSI->getName()+".mask");
6980 return new ICmpInst(ICI.getPredicate(), And,
6981 ConstantExpr::getShl(RHS, ShAmt));
6986 case Instruction::SDiv:
6987 case Instruction::UDiv:
6988 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6989 // Fold this div into the comparison, producing a range check.
6990 // Determine, based on the divide type, what the range is being
6991 // checked. If there is an overflow on the low or high side, remember
6992 // it, otherwise compute the range [low, hi) bounding the new value.
6993 // See: InsertRangeTest above for the kinds of replacements possible.
6994 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6995 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7000 case Instruction::Add:
7001 // Fold: icmp pred (add, X, C1), C2
7003 if (!ICI.isEquality()) {
7004 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7006 const APInt &LHSV = LHSC->getValue();
7008 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7011 if (ICI.isSigned()) {
7012 if (CR.getLower().isSignBit()) {
7013 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7014 ConstantInt::get(*Context, CR.getUpper()));
7015 } else if (CR.getUpper().isSignBit()) {
7016 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7017 ConstantInt::get(*Context, CR.getLower()));
7020 if (CR.getLower().isMinValue()) {
7021 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7022 ConstantInt::get(*Context, CR.getUpper()));
7023 } else if (CR.getUpper().isMinValue()) {
7024 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7025 ConstantInt::get(*Context, CR.getLower()));
7032 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7033 if (ICI.isEquality()) {
7034 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7036 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7037 // the second operand is a constant, simplify a bit.
7038 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7039 switch (BO->getOpcode()) {
7040 case Instruction::SRem:
7041 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7042 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7043 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7044 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7046 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
7048 return new ICmpInst(ICI.getPredicate(), NewRem,
7049 Constant::getNullValue(BO->getType()));
7053 case Instruction::Add:
7054 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7055 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7056 if (BO->hasOneUse())
7057 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7058 ConstantExpr::getSub(RHS, BOp1C));
7059 } else if (RHSV == 0) {
7060 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7061 // efficiently invertible, or if the add has just this one use.
7062 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7064 if (Value *NegVal = dyn_castNegVal(BOp1))
7065 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
7066 else if (Value *NegVal = dyn_castNegVal(BOp0))
7067 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
7068 else if (BO->hasOneUse()) {
7069 Value *Neg = Builder->CreateNeg(BOp1);
7071 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
7075 case Instruction::Xor:
7076 // For the xor case, we can xor two constants together, eliminating
7077 // the explicit xor.
7078 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7079 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7080 ConstantExpr::getXor(RHS, BOC));
7083 case Instruction::Sub:
7084 // Replace (([sub|xor] A, B) != 0) with (A != B)
7086 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7090 case Instruction::Or:
7091 // If bits are being or'd in that are not present in the constant we
7092 // are comparing against, then the comparison could never succeed!
7093 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7094 Constant *NotCI = ConstantExpr::getNot(RHS);
7095 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7096 return ReplaceInstUsesWith(ICI,
7097 ConstantInt::get(Type::getInt1Ty(*Context),
7102 case Instruction::And:
7103 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7104 // If bits are being compared against that are and'd out, then the
7105 // comparison can never succeed!
7106 if ((RHSV & ~BOC->getValue()) != 0)
7107 return ReplaceInstUsesWith(ICI,
7108 ConstantInt::get(Type::getInt1Ty(*Context),
7111 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7112 if (RHS == BOC && RHSV.isPowerOf2())
7113 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7114 ICmpInst::ICMP_NE, LHSI,
7115 Constant::getNullValue(RHS->getType()));
7117 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7118 if (BOC->getValue().isSignBit()) {
7119 Value *X = BO->getOperand(0);
7120 Constant *Zero = Constant::getNullValue(X->getType());
7121 ICmpInst::Predicate pred = isICMP_NE ?
7122 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7123 return new ICmpInst(pred, X, Zero);
7126 // ((X & ~7) == 0) --> X < 8
7127 if (RHSV == 0 && isHighOnes(BOC)) {
7128 Value *X = BO->getOperand(0);
7129 Constant *NegX = ConstantExpr::getNeg(BOC);
7130 ICmpInst::Predicate pred = isICMP_NE ?
7131 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7132 return new ICmpInst(pred, X, NegX);
7137 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7138 // Handle icmp {eq|ne} <intrinsic>, intcst.
7139 if (II->getIntrinsicID() == Intrinsic::bswap) {
7141 ICI.setOperand(0, II->getOperand(1));
7142 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7150 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7151 /// We only handle extending casts so far.
7153 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7154 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7155 Value *LHSCIOp = LHSCI->getOperand(0);
7156 const Type *SrcTy = LHSCIOp->getType();
7157 const Type *DestTy = LHSCI->getType();
7160 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7161 // integer type is the same size as the pointer type.
7162 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7163 TD->getPointerSizeInBits() ==
7164 cast<IntegerType>(DestTy)->getBitWidth()) {
7166 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7167 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7168 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7169 RHSOp = RHSC->getOperand(0);
7170 // If the pointer types don't match, insert a bitcast.
7171 if (LHSCIOp->getType() != RHSOp->getType())
7172 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
7176 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7179 // The code below only handles extension cast instructions, so far.
7181 if (LHSCI->getOpcode() != Instruction::ZExt &&
7182 LHSCI->getOpcode() != Instruction::SExt)
7185 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7186 bool isSignedCmp = ICI.isSigned();
7188 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7189 // Not an extension from the same type?
7190 RHSCIOp = CI->getOperand(0);
7191 if (RHSCIOp->getType() != LHSCIOp->getType())
7194 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7195 // and the other is a zext), then we can't handle this.
7196 if (CI->getOpcode() != LHSCI->getOpcode())
7199 // Deal with equality cases early.
7200 if (ICI.isEquality())
7201 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7203 // A signed comparison of sign extended values simplifies into a
7204 // signed comparison.
7205 if (isSignedCmp && isSignedExt)
7206 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7208 // The other three cases all fold into an unsigned comparison.
7209 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7212 // If we aren't dealing with a constant on the RHS, exit early
7213 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7217 // Compute the constant that would happen if we truncated to SrcTy then
7218 // reextended to DestTy.
7219 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7220 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7223 // If the re-extended constant didn't change...
7225 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7226 // For example, we might have:
7227 // %A = sext i16 %X to i32
7228 // %B = icmp ugt i32 %A, 1330
7229 // It is incorrect to transform this into
7230 // %B = icmp ugt i16 %X, 1330
7231 // because %A may have negative value.
7233 // However, we allow this when the compare is EQ/NE, because they are
7235 if (isSignedExt == isSignedCmp || ICI.isEquality())
7236 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7240 // The re-extended constant changed so the constant cannot be represented
7241 // in the shorter type. Consequently, we cannot emit a simple comparison.
7243 // First, handle some easy cases. We know the result cannot be equal at this
7244 // point so handle the ICI.isEquality() cases
7245 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7246 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7247 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7248 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7250 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7251 // should have been folded away previously and not enter in here.
7254 // We're performing a signed comparison.
7255 if (cast<ConstantInt>(CI)->getValue().isNegative())
7256 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7258 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7260 // We're performing an unsigned comparison.
7262 // We're performing an unsigned comp with a sign extended value.
7263 // This is true if the input is >= 0. [aka >s -1]
7264 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7265 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
7267 // Unsigned extend & unsigned compare -> always true.
7268 Result = ConstantInt::getTrue(*Context);
7272 // Finally, return the value computed.
7273 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7274 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7275 return ReplaceInstUsesWith(ICI, Result);
7277 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7278 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7279 "ICmp should be folded!");
7280 if (Constant *CI = dyn_cast<Constant>(Result))
7281 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7282 return BinaryOperator::CreateNot(Result);
7285 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7286 return commonShiftTransforms(I);
7289 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7290 return commonShiftTransforms(I);
7293 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7294 if (Instruction *R = commonShiftTransforms(I))
7297 Value *Op0 = I.getOperand(0);
7299 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7300 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7301 if (CSI->isAllOnesValue())
7302 return ReplaceInstUsesWith(I, CSI);
7304 // See if we can turn a signed shr into an unsigned shr.
7305 if (MaskedValueIsZero(Op0,
7306 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7307 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7309 // Arithmetic shifting an all-sign-bit value is a no-op.
7310 unsigned NumSignBits = ComputeNumSignBits(Op0);
7311 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7312 return ReplaceInstUsesWith(I, Op0);
7317 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7318 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7319 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7321 // shl X, 0 == X and shr X, 0 == X
7322 // shl 0, X == 0 and shr 0, X == 0
7323 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7324 Op0 == Constant::getNullValue(Op0->getType()))
7325 return ReplaceInstUsesWith(I, Op0);
7327 if (isa<UndefValue>(Op0)) {
7328 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7329 return ReplaceInstUsesWith(I, Op0);
7330 else // undef << X -> 0, undef >>u X -> 0
7331 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7333 if (isa<UndefValue>(Op1)) {
7334 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7335 return ReplaceInstUsesWith(I, Op0);
7336 else // X << undef, X >>u undef -> 0
7337 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7340 // See if we can fold away this shift.
7341 if (SimplifyDemandedInstructionBits(I))
7344 // Try to fold constant and into select arguments.
7345 if (isa<Constant>(Op0))
7346 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7347 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7350 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7351 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7356 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7357 BinaryOperator &I) {
7358 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7360 // See if we can simplify any instructions used by the instruction whose sole
7361 // purpose is to compute bits we don't care about.
7362 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7364 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7367 if (Op1->uge(TypeBits)) {
7368 if (I.getOpcode() != Instruction::AShr)
7369 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7371 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7376 // ((X*C1) << C2) == (X * (C1 << C2))
7377 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7378 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7379 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7380 return BinaryOperator::CreateMul(BO->getOperand(0),
7381 ConstantExpr::getShl(BOOp, Op1));
7383 // Try to fold constant and into select arguments.
7384 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7385 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7387 if (isa<PHINode>(Op0))
7388 if (Instruction *NV = FoldOpIntoPhi(I))
7391 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7392 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7393 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7394 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7395 // place. Don't try to do this transformation in this case. Also, we
7396 // require that the input operand is a shift-by-constant so that we have
7397 // confidence that the shifts will get folded together. We could do this
7398 // xform in more cases, but it is unlikely to be profitable.
7399 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7400 isa<ConstantInt>(TrOp->getOperand(1))) {
7401 // Okay, we'll do this xform. Make the shift of shift.
7402 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7403 // (shift2 (shift1 & 0x00FF), c2)
7404 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
7406 // For logical shifts, the truncation has the effect of making the high
7407 // part of the register be zeros. Emulate this by inserting an AND to
7408 // clear the top bits as needed. This 'and' will usually be zapped by
7409 // other xforms later if dead.
7410 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7411 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7412 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7414 // The mask we constructed says what the trunc would do if occurring
7415 // between the shifts. We want to know the effect *after* the second
7416 // shift. We know that it is a logical shift by a constant, so adjust the
7417 // mask as appropriate.
7418 if (I.getOpcode() == Instruction::Shl)
7419 MaskV <<= Op1->getZExtValue();
7421 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7422 MaskV = MaskV.lshr(Op1->getZExtValue());
7426 Value *And = Builder->CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7429 // Return the value truncated to the interesting size.
7430 return new TruncInst(And, I.getType());
7434 if (Op0->hasOneUse()) {
7435 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7436 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7439 switch (Op0BO->getOpcode()) {
7441 case Instruction::Add:
7442 case Instruction::And:
7443 case Instruction::Or:
7444 case Instruction::Xor: {
7445 // These operators commute.
7446 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7447 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7448 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7449 m_Specific(Op1)))) {
7450 Value *YS = // (Y << C)
7451 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
7453 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
7454 Op0BO->getOperand(1)->getName());
7455 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7456 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7457 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7460 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7461 Value *Op0BOOp1 = Op0BO->getOperand(1);
7462 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7464 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7465 m_ConstantInt(CC))) &&
7466 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7467 Value *YS = // (Y << C)
7468 Builder->CreateShl(Op0BO->getOperand(0), Op1,
7471 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7472 V1->getName()+".mask");
7473 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7478 case Instruction::Sub: {
7479 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7480 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7481 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7482 m_Specific(Op1)))) {
7483 Value *YS = // (Y << C)
7484 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7486 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
7487 Op0BO->getOperand(0)->getName());
7488 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7489 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7490 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7493 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7494 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7495 match(Op0BO->getOperand(0),
7496 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7497 m_ConstantInt(CC))) && V2 == Op1 &&
7498 cast<BinaryOperator>(Op0BO->getOperand(0))
7499 ->getOperand(0)->hasOneUse()) {
7500 Value *YS = // (Y << C)
7501 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7503 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7504 V1->getName()+".mask");
7506 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7514 // If the operand is an bitwise operator with a constant RHS, and the
7515 // shift is the only use, we can pull it out of the shift.
7516 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7517 bool isValid = true; // Valid only for And, Or, Xor
7518 bool highBitSet = false; // Transform if high bit of constant set?
7520 switch (Op0BO->getOpcode()) {
7521 default: isValid = false; break; // Do not perform transform!
7522 case Instruction::Add:
7523 isValid = isLeftShift;
7525 case Instruction::Or:
7526 case Instruction::Xor:
7529 case Instruction::And:
7534 // If this is a signed shift right, and the high bit is modified
7535 // by the logical operation, do not perform the transformation.
7536 // The highBitSet boolean indicates the value of the high bit of
7537 // the constant which would cause it to be modified for this
7540 if (isValid && I.getOpcode() == Instruction::AShr)
7541 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7544 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7547 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
7548 NewShift->takeName(Op0BO);
7550 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7557 // Find out if this is a shift of a shift by a constant.
7558 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7559 if (ShiftOp && !ShiftOp->isShift())
7562 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7563 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7564 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7565 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7566 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7567 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7568 Value *X = ShiftOp->getOperand(0);
7570 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7572 const IntegerType *Ty = cast<IntegerType>(I.getType());
7574 // Check for (X << c1) << c2 and (X >> c1) >> c2
7575 if (I.getOpcode() == ShiftOp->getOpcode()) {
7576 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7578 if (AmtSum >= TypeBits) {
7579 if (I.getOpcode() != Instruction::AShr)
7580 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7581 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7584 return BinaryOperator::Create(I.getOpcode(), X,
7585 ConstantInt::get(Ty, AmtSum));
7588 if (ShiftOp->getOpcode() == Instruction::LShr &&
7589 I.getOpcode() == Instruction::AShr) {
7590 if (AmtSum >= TypeBits)
7591 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7593 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7594 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7597 if (ShiftOp->getOpcode() == Instruction::AShr &&
7598 I.getOpcode() == Instruction::LShr) {
7599 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7600 if (AmtSum >= TypeBits)
7601 AmtSum = TypeBits-1;
7603 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7605 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7606 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7609 // Okay, if we get here, one shift must be left, and the other shift must be
7610 // right. See if the amounts are equal.
7611 if (ShiftAmt1 == ShiftAmt2) {
7612 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7613 if (I.getOpcode() == Instruction::Shl) {
7614 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7615 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7617 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7618 if (I.getOpcode() == Instruction::LShr) {
7619 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7620 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7622 // We can simplify ((X << C) >>s C) into a trunc + sext.
7623 // NOTE: we could do this for any C, but that would make 'unusual' integer
7624 // types. For now, just stick to ones well-supported by the code
7626 const Type *SExtType = 0;
7627 switch (Ty->getBitWidth() - ShiftAmt1) {
7634 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
7639 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
7640 // Otherwise, we can't handle it yet.
7641 } else if (ShiftAmt1 < ShiftAmt2) {
7642 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7644 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7645 if (I.getOpcode() == Instruction::Shl) {
7646 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7647 ShiftOp->getOpcode() == Instruction::AShr);
7648 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7650 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7651 return BinaryOperator::CreateAnd(Shift,
7652 ConstantInt::get(*Context, Mask));
7655 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7656 if (I.getOpcode() == Instruction::LShr) {
7657 assert(ShiftOp->getOpcode() == Instruction::Shl);
7658 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7660 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7661 return BinaryOperator::CreateAnd(Shift,
7662 ConstantInt::get(*Context, Mask));
7665 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7667 assert(ShiftAmt2 < ShiftAmt1);
7668 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7670 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7671 if (I.getOpcode() == Instruction::Shl) {
7672 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7673 ShiftOp->getOpcode() == Instruction::AShr);
7674 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
7675 ConstantInt::get(Ty, ShiftDiff));
7677 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7678 return BinaryOperator::CreateAnd(Shift,
7679 ConstantInt::get(*Context, Mask));
7682 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7683 if (I.getOpcode() == Instruction::LShr) {
7684 assert(ShiftOp->getOpcode() == Instruction::Shl);
7685 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7687 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7688 return BinaryOperator::CreateAnd(Shift,
7689 ConstantInt::get(*Context, Mask));
7692 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7699 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7700 /// expression. If so, decompose it, returning some value X, such that Val is
7703 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7704 int &Offset, LLVMContext *Context) {
7705 assert(Val->getType() == Type::getInt32Ty(*Context) &&
7706 "Unexpected allocation size type!");
7707 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7708 Offset = CI->getZExtValue();
7710 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
7711 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7712 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7713 if (I->getOpcode() == Instruction::Shl) {
7714 // This is a value scaled by '1 << the shift amt'.
7715 Scale = 1U << RHS->getZExtValue();
7717 return I->getOperand(0);
7718 } else if (I->getOpcode() == Instruction::Mul) {
7719 // This value is scaled by 'RHS'.
7720 Scale = RHS->getZExtValue();
7722 return I->getOperand(0);
7723 } else if (I->getOpcode() == Instruction::Add) {
7724 // We have X+C. Check to see if we really have (X*C2)+C1,
7725 // where C1 is divisible by C2.
7728 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7730 Offset += RHS->getZExtValue();
7737 // Otherwise, we can't look past this.
7744 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7745 /// try to eliminate the cast by moving the type information into the alloc.
7746 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7748 const PointerType *PTy = cast<PointerType>(CI.getType());
7750 BuilderTy AllocaBuilder(*Builder);
7751 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
7753 // Remove any uses of AI that are dead.
7754 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7756 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7757 Instruction *User = cast<Instruction>(*UI++);
7758 if (isInstructionTriviallyDead(User)) {
7759 while (UI != E && *UI == User)
7760 ++UI; // If this instruction uses AI more than once, don't break UI.
7763 DEBUG(errs() << "IC: DCE: " << *User << '\n');
7764 EraseInstFromFunction(*User);
7768 // This requires TargetData to get the alloca alignment and size information.
7771 // Get the type really allocated and the type casted to.
7772 const Type *AllocElTy = AI.getAllocatedType();
7773 const Type *CastElTy = PTy->getElementType();
7774 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7776 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7777 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7778 if (CastElTyAlign < AllocElTyAlign) return 0;
7780 // If the allocation has multiple uses, only promote it if we are strictly
7781 // increasing the alignment of the resultant allocation. If we keep it the
7782 // same, we open the door to infinite loops of various kinds. (A reference
7783 // from a dbg.declare doesn't count as a use for this purpose.)
7784 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7785 CastElTyAlign == AllocElTyAlign) return 0;
7787 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7788 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7789 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7791 // See if we can satisfy the modulus by pulling a scale out of the array
7793 unsigned ArraySizeScale;
7795 Value *NumElements = // See if the array size is a decomposable linear expr.
7796 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7797 ArrayOffset, Context);
7799 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7801 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7802 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7804 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7809 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
7810 // Insert before the alloca, not before the cast.
7811 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
7814 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7815 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
7816 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
7819 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
7820 New->setAlignment(AI.getAlignment());
7823 // If the allocation has one real use plus a dbg.declare, just remove the
7825 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7826 EraseInstFromFunction(*DI);
7828 // If the allocation has multiple real uses, insert a cast and change all
7829 // things that used it to use the new cast. This will also hack on CI, but it
7831 else if (!AI.hasOneUse()) {
7832 // New is the allocation instruction, pointer typed. AI is the original
7833 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7834 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
7835 AI.replaceAllUsesWith(NewCast);
7837 return ReplaceInstUsesWith(CI, New);
7840 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7841 /// and return it as type Ty without inserting any new casts and without
7842 /// changing the computed value. This is used by code that tries to decide
7843 /// whether promoting or shrinking integer operations to wider or smaller types
7844 /// will allow us to eliminate a truncate or extend.
7846 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7847 /// extension operation if Ty is larger.
7849 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7850 /// should return true if trunc(V) can be computed by computing V in the smaller
7851 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7852 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7853 /// efficiently truncated.
7855 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7856 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7857 /// the final result.
7858 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7860 int &NumCastsRemoved){
7861 // We can always evaluate constants in another type.
7862 if (isa<Constant>(V))
7865 Instruction *I = dyn_cast<Instruction>(V);
7866 if (!I) return false;
7868 const Type *OrigTy = V->getType();
7870 // If this is an extension or truncate, we can often eliminate it.
7871 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7872 // If this is a cast from the destination type, we can trivially eliminate
7873 // it, and this will remove a cast overall.
7874 if (I->getOperand(0)->getType() == Ty) {
7875 // If the first operand is itself a cast, and is eliminable, do not count
7876 // this as an eliminable cast. We would prefer to eliminate those two
7878 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7884 // We can't extend or shrink something that has multiple uses: doing so would
7885 // require duplicating the instruction in general, which isn't profitable.
7886 if (!I->hasOneUse()) return false;
7888 unsigned Opc = I->getOpcode();
7890 case Instruction::Add:
7891 case Instruction::Sub:
7892 case Instruction::Mul:
7893 case Instruction::And:
7894 case Instruction::Or:
7895 case Instruction::Xor:
7896 // These operators can all arbitrarily be extended or truncated.
7897 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7899 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7902 case Instruction::UDiv:
7903 case Instruction::URem: {
7904 // UDiv and URem can be truncated if all the truncated bits are zero.
7905 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7906 uint32_t BitWidth = Ty->getScalarSizeInBits();
7907 if (BitWidth < OrigBitWidth) {
7908 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7909 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7910 MaskedValueIsZero(I->getOperand(1), Mask)) {
7911 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7913 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7919 case Instruction::Shl:
7920 // If we are truncating the result of this SHL, and if it's a shift of a
7921 // constant amount, we can always perform a SHL in a smaller type.
7922 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7923 uint32_t BitWidth = Ty->getScalarSizeInBits();
7924 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7925 CI->getLimitedValue(BitWidth) < BitWidth)
7926 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7930 case Instruction::LShr:
7931 // If this is a truncate of a logical shr, we can truncate it to a smaller
7932 // lshr iff we know that the bits we would otherwise be shifting in are
7934 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7935 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7936 uint32_t BitWidth = Ty->getScalarSizeInBits();
7937 if (BitWidth < OrigBitWidth &&
7938 MaskedValueIsZero(I->getOperand(0),
7939 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7940 CI->getLimitedValue(BitWidth) < BitWidth) {
7941 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7946 case Instruction::ZExt:
7947 case Instruction::SExt:
7948 case Instruction::Trunc:
7949 // If this is the same kind of case as our original (e.g. zext+zext), we
7950 // can safely replace it. Note that replacing it does not reduce the number
7951 // of casts in the input.
7955 // sext (zext ty1), ty2 -> zext ty2
7956 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7959 case Instruction::Select: {
7960 SelectInst *SI = cast<SelectInst>(I);
7961 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7963 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7966 case Instruction::PHI: {
7967 // We can change a phi if we can change all operands.
7968 PHINode *PN = cast<PHINode>(I);
7969 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7970 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7976 // TODO: Can handle more cases here.
7983 /// EvaluateInDifferentType - Given an expression that
7984 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7985 /// evaluate the expression.
7986 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7988 if (Constant *C = dyn_cast<Constant>(V))
7989 return ConstantExpr::getIntegerCast(C, Ty,
7990 isSigned /*Sext or ZExt*/);
7992 // Otherwise, it must be an instruction.
7993 Instruction *I = cast<Instruction>(V);
7994 Instruction *Res = 0;
7995 unsigned Opc = I->getOpcode();
7997 case Instruction::Add:
7998 case Instruction::Sub:
7999 case Instruction::Mul:
8000 case Instruction::And:
8001 case Instruction::Or:
8002 case Instruction::Xor:
8003 case Instruction::AShr:
8004 case Instruction::LShr:
8005 case Instruction::Shl:
8006 case Instruction::UDiv:
8007 case Instruction::URem: {
8008 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8009 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8010 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8013 case Instruction::Trunc:
8014 case Instruction::ZExt:
8015 case Instruction::SExt:
8016 // If the source type of the cast is the type we're trying for then we can
8017 // just return the source. There's no need to insert it because it is not
8019 if (I->getOperand(0)->getType() == Ty)
8020 return I->getOperand(0);
8022 // Otherwise, must be the same type of cast, so just reinsert a new one.
8023 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8026 case Instruction::Select: {
8027 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8028 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8029 Res = SelectInst::Create(I->getOperand(0), True, False);
8032 case Instruction::PHI: {
8033 PHINode *OPN = cast<PHINode>(I);
8034 PHINode *NPN = PHINode::Create(Ty);
8035 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8036 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8037 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8043 // TODO: Can handle more cases here.
8044 llvm_unreachable("Unreachable!");
8049 return InsertNewInstBefore(Res, *I);
8052 /// @brief Implement the transforms common to all CastInst visitors.
8053 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8054 Value *Src = CI.getOperand(0);
8056 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8057 // eliminate it now.
8058 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8059 if (Instruction::CastOps opc =
8060 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8061 // The first cast (CSrc) is eliminable so we need to fix up or replace
8062 // the second cast (CI). CSrc will then have a good chance of being dead.
8063 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8067 // If we are casting a select then fold the cast into the select
8068 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8069 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8072 // If we are casting a PHI then fold the cast into the PHI
8073 if (isa<PHINode>(Src))
8074 if (Instruction *NV = FoldOpIntoPhi(CI))
8080 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8081 /// or not there is a sequence of GEP indices into the type that will land us at
8082 /// the specified offset. If so, fill them into NewIndices and return the
8083 /// resultant element type, otherwise return null.
8084 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8085 SmallVectorImpl<Value*> &NewIndices,
8086 const TargetData *TD,
8087 LLVMContext *Context) {
8089 if (!Ty->isSized()) return 0;
8091 // Start with the index over the outer type. Note that the type size
8092 // might be zero (even if the offset isn't zero) if the indexed type
8093 // is something like [0 x {int, int}]
8094 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8095 int64_t FirstIdx = 0;
8096 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8097 FirstIdx = Offset/TySize;
8098 Offset -= FirstIdx*TySize;
8100 // Handle hosts where % returns negative instead of values [0..TySize).
8104 assert(Offset >= 0);
8106 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8109 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8111 // Index into the types. If we fail, set OrigBase to null.
8113 // Indexing into tail padding between struct/array elements.
8114 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8117 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8118 const StructLayout *SL = TD->getStructLayout(STy);
8119 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8120 "Offset must stay within the indexed type");
8122 unsigned Elt = SL->getElementContainingOffset(Offset);
8123 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8125 Offset -= SL->getElementOffset(Elt);
8126 Ty = STy->getElementType(Elt);
8127 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8128 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8129 assert(EltSize && "Cannot index into a zero-sized array");
8130 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8132 Ty = AT->getElementType();
8134 // Otherwise, we can't index into the middle of this atomic type, bail.
8142 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8143 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8144 Value *Src = CI.getOperand(0);
8146 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8147 // If casting the result of a getelementptr instruction with no offset, turn
8148 // this into a cast of the original pointer!
8149 if (GEP->hasAllZeroIndices()) {
8150 // Changing the cast operand is usually not a good idea but it is safe
8151 // here because the pointer operand is being replaced with another
8152 // pointer operand so the opcode doesn't need to change.
8154 CI.setOperand(0, GEP->getOperand(0));
8158 // If the GEP has a single use, and the base pointer is a bitcast, and the
8159 // GEP computes a constant offset, see if we can convert these three
8160 // instructions into fewer. This typically happens with unions and other
8161 // non-type-safe code.
8162 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8163 if (GEP->hasAllConstantIndices()) {
8164 // We are guaranteed to get a constant from EmitGEPOffset.
8165 ConstantInt *OffsetV =
8166 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8167 int64_t Offset = OffsetV->getSExtValue();
8169 // Get the base pointer input of the bitcast, and the type it points to.
8170 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8171 const Type *GEPIdxTy =
8172 cast<PointerType>(OrigBase->getType())->getElementType();
8173 SmallVector<Value*, 8> NewIndices;
8174 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8175 // If we were able to index down into an element, create the GEP
8176 // and bitcast the result. This eliminates one bitcast, potentially
8178 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
8179 Builder->CreateInBoundsGEP(OrigBase,
8180 NewIndices.begin(), NewIndices.end()) :
8181 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
8182 NGEP->takeName(GEP);
8184 if (isa<BitCastInst>(CI))
8185 return new BitCastInst(NGEP, CI.getType());
8186 assert(isa<PtrToIntInst>(CI));
8187 return new PtrToIntInst(NGEP, CI.getType());
8193 return commonCastTransforms(CI);
8196 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8197 /// type like i42. We don't want to introduce operations on random non-legal
8198 /// integer types where they don't already exist in the code. In the future,
8199 /// we should consider making this based off target-data, so that 32-bit targets
8200 /// won't get i64 operations etc.
8201 static bool isSafeIntegerType(const Type *Ty) {
8202 switch (Ty->getPrimitiveSizeInBits()) {
8213 /// commonIntCastTransforms - This function implements the common transforms
8214 /// for trunc, zext, and sext.
8215 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8216 if (Instruction *Result = commonCastTransforms(CI))
8219 Value *Src = CI.getOperand(0);
8220 const Type *SrcTy = Src->getType();
8221 const Type *DestTy = CI.getType();
8222 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8223 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8225 // See if we can simplify any instructions used by the LHS whose sole
8226 // purpose is to compute bits we don't care about.
8227 if (SimplifyDemandedInstructionBits(CI))
8230 // If the source isn't an instruction or has more than one use then we
8231 // can't do anything more.
8232 Instruction *SrcI = dyn_cast<Instruction>(Src);
8233 if (!SrcI || !Src->hasOneUse())
8236 // Attempt to propagate the cast into the instruction for int->int casts.
8237 int NumCastsRemoved = 0;
8238 // Only do this if the dest type is a simple type, don't convert the
8239 // expression tree to something weird like i93 unless the source is also
8241 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8242 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8243 CanEvaluateInDifferentType(SrcI, DestTy,
8244 CI.getOpcode(), NumCastsRemoved)) {
8245 // If this cast is a truncate, evaluting in a different type always
8246 // eliminates the cast, so it is always a win. If this is a zero-extension,
8247 // we need to do an AND to maintain the clear top-part of the computation,
8248 // so we require that the input have eliminated at least one cast. If this
8249 // is a sign extension, we insert two new casts (to do the extension) so we
8250 // require that two casts have been eliminated.
8251 bool DoXForm = false;
8252 bool JustReplace = false;
8253 switch (CI.getOpcode()) {
8255 // All the others use floating point so we shouldn't actually
8256 // get here because of the check above.
8257 llvm_unreachable("Unknown cast type");
8258 case Instruction::Trunc:
8261 case Instruction::ZExt: {
8262 DoXForm = NumCastsRemoved >= 1;
8263 if (!DoXForm && 0) {
8264 // If it's unnecessary to issue an AND to clear the high bits, it's
8265 // always profitable to do this xform.
8266 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8267 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8268 if (MaskedValueIsZero(TryRes, Mask))
8269 return ReplaceInstUsesWith(CI, TryRes);
8271 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8272 if (TryI->use_empty())
8273 EraseInstFromFunction(*TryI);
8277 case Instruction::SExt: {
8278 DoXForm = NumCastsRemoved >= 2;
8279 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8280 // If we do not have to emit the truncate + sext pair, then it's always
8281 // profitable to do this xform.
8283 // It's not safe to eliminate the trunc + sext pair if one of the
8284 // eliminated cast is a truncate. e.g.
8285 // t2 = trunc i32 t1 to i16
8286 // t3 = sext i16 t2 to i32
8289 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8290 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8291 if (NumSignBits > (DestBitSize - SrcBitSize))
8292 return ReplaceInstUsesWith(CI, TryRes);
8294 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8295 if (TryI->use_empty())
8296 EraseInstFromFunction(*TryI);
8303 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8304 " to avoid cast: " << CI);
8305 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8306 CI.getOpcode() == Instruction::SExt);
8308 // Just replace this cast with the result.
8309 return ReplaceInstUsesWith(CI, Res);
8311 assert(Res->getType() == DestTy);
8312 switch (CI.getOpcode()) {
8313 default: llvm_unreachable("Unknown cast type!");
8314 case Instruction::Trunc:
8315 // Just replace this cast with the result.
8316 return ReplaceInstUsesWith(CI, Res);
8317 case Instruction::ZExt: {
8318 assert(SrcBitSize < DestBitSize && "Not a zext?");
8320 // If the high bits are already zero, just replace this cast with the
8322 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8323 if (MaskedValueIsZero(Res, Mask))
8324 return ReplaceInstUsesWith(CI, Res);
8326 // We need to emit an AND to clear the high bits.
8327 Constant *C = ConstantInt::get(*Context,
8328 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8329 return BinaryOperator::CreateAnd(Res, C);
8331 case Instruction::SExt: {
8332 // If the high bits are already filled with sign bit, just replace this
8333 // cast with the result.
8334 unsigned NumSignBits = ComputeNumSignBits(Res);
8335 if (NumSignBits > (DestBitSize - SrcBitSize))
8336 return ReplaceInstUsesWith(CI, Res);
8338 // We need to emit a cast to truncate, then a cast to sext.
8339 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
8345 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8346 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8348 switch (SrcI->getOpcode()) {
8349 case Instruction::Add:
8350 case Instruction::Mul:
8351 case Instruction::And:
8352 case Instruction::Or:
8353 case Instruction::Xor:
8354 // If we are discarding information, rewrite.
8355 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8356 // Don't insert two casts unless at least one can be eliminated.
8357 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8358 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8359 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8360 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8361 return BinaryOperator::Create(
8362 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8366 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8367 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8368 SrcI->getOpcode() == Instruction::Xor &&
8369 Op1 == ConstantInt::getTrue(*Context) &&
8370 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8371 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName());
8372 return BinaryOperator::CreateXor(New,
8373 ConstantInt::get(CI.getType(), 1));
8377 case Instruction::Shl: {
8378 // Canonicalize trunc inside shl, if we can.
8379 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8380 if (CI && DestBitSize < SrcBitSize &&
8381 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8382 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8383 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8384 return BinaryOperator::CreateShl(Op0c, Op1c);
8392 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8393 if (Instruction *Result = commonIntCastTransforms(CI))
8396 Value *Src = CI.getOperand(0);
8397 const Type *Ty = CI.getType();
8398 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8399 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8401 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8402 if (DestBitWidth == 1) {
8403 Constant *One = ConstantInt::get(Src->getType(), 1);
8404 Src = Builder->CreateAnd(Src, One, "tmp");
8405 Value *Zero = Constant::getNullValue(Src->getType());
8406 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8409 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8410 ConstantInt *ShAmtV = 0;
8412 if (Src->hasOneUse() &&
8413 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8414 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8416 // Get a mask for the bits shifting in.
8417 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8418 if (MaskedValueIsZero(ShiftOp, Mask)) {
8419 if (ShAmt >= DestBitWidth) // All zeros.
8420 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8422 // Okay, we can shrink this. Truncate the input, then return a new
8424 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName());
8425 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8426 return BinaryOperator::CreateLShr(V1, V2);
8433 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8434 /// in order to eliminate the icmp.
8435 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8437 // If we are just checking for a icmp eq of a single bit and zext'ing it
8438 // to an integer, then shift the bit to the appropriate place and then
8439 // cast to integer to avoid the comparison.
8440 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8441 const APInt &Op1CV = Op1C->getValue();
8443 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8444 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8445 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8446 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8447 if (!DoXform) return ICI;
8449 Value *In = ICI->getOperand(0);
8450 Value *Sh = ConstantInt::get(In->getType(),
8451 In->getType()->getScalarSizeInBits()-1);
8452 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
8453 if (In->getType() != CI.getType())
8454 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
8456 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8457 Constant *One = ConstantInt::get(In->getType(), 1);
8458 In = Builder->CreateXor(In, One, In->getName()+".not");
8461 return ReplaceInstUsesWith(CI, In);
8466 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8467 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8468 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8469 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8470 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8471 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8472 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8473 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8474 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8475 // This only works for EQ and NE
8476 ICI->isEquality()) {
8477 // If Op1C some other power of two, convert:
8478 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8479 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8480 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8481 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8483 APInt KnownZeroMask(~KnownZero);
8484 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8485 if (!DoXform) return ICI;
8487 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8488 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8489 // (X&4) == 2 --> false
8490 // (X&4) != 2 --> true
8491 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8492 Res = ConstantExpr::getZExt(Res, CI.getType());
8493 return ReplaceInstUsesWith(CI, Res);
8496 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8497 Value *In = ICI->getOperand(0);
8499 // Perform a logical shr by shiftamt.
8500 // Insert the shift to put the result in the low bit.
8501 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
8502 In->getName()+".lobit");
8505 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8506 Constant *One = ConstantInt::get(In->getType(), 1);
8507 In = Builder->CreateXor(In, One, "tmp");
8510 if (CI.getType() == In->getType())
8511 return ReplaceInstUsesWith(CI, In);
8513 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8521 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8522 // If one of the common conversion will work ..
8523 if (Instruction *Result = commonIntCastTransforms(CI))
8526 Value *Src = CI.getOperand(0);
8528 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8529 // types and if the sizes are just right we can convert this into a logical
8530 // 'and' which will be much cheaper than the pair of casts.
8531 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8532 // Get the sizes of the types involved. We know that the intermediate type
8533 // will be smaller than A or C, but don't know the relation between A and C.
8534 Value *A = CSrc->getOperand(0);
8535 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8536 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8537 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8538 // If we're actually extending zero bits, then if
8539 // SrcSize < DstSize: zext(a & mask)
8540 // SrcSize == DstSize: a & mask
8541 // SrcSize > DstSize: trunc(a) & mask
8542 if (SrcSize < DstSize) {
8543 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8544 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8545 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
8546 return new ZExtInst(And, CI.getType());
8549 if (SrcSize == DstSize) {
8550 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8551 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8554 if (SrcSize > DstSize) {
8555 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
8556 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8557 return BinaryOperator::CreateAnd(Trunc,
8558 ConstantInt::get(Trunc->getType(),
8563 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8564 return transformZExtICmp(ICI, CI);
8566 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8567 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8568 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8569 // of the (zext icmp) will be transformed.
8570 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8571 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8572 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8573 (transformZExtICmp(LHS, CI, false) ||
8574 transformZExtICmp(RHS, CI, false))) {
8575 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
8576 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
8577 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8581 // zext(trunc(t) & C) -> (t & zext(C)).
8582 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8583 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8584 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8585 Value *TI0 = TI->getOperand(0);
8586 if (TI0->getType() == CI.getType())
8588 BinaryOperator::CreateAnd(TI0,
8589 ConstantExpr::getZExt(C, CI.getType()));
8592 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8593 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8594 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8595 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8596 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8597 And->getOperand(1) == C)
8598 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8599 Value *TI0 = TI->getOperand(0);
8600 if (TI0->getType() == CI.getType()) {
8601 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8602 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
8603 return BinaryOperator::CreateXor(NewAnd, ZC);
8610 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8611 if (Instruction *I = commonIntCastTransforms(CI))
8614 Value *Src = CI.getOperand(0);
8616 // Canonicalize sign-extend from i1 to a select.
8617 if (Src->getType() == Type::getInt1Ty(*Context))
8618 return SelectInst::Create(Src,
8619 Constant::getAllOnesValue(CI.getType()),
8620 Constant::getNullValue(CI.getType()));
8622 // See if the value being truncated is already sign extended. If so, just
8623 // eliminate the trunc/sext pair.
8624 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8625 Value *Op = cast<User>(Src)->getOperand(0);
8626 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8627 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8628 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8629 unsigned NumSignBits = ComputeNumSignBits(Op);
8631 if (OpBits == DestBits) {
8632 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8633 // bits, it is already ready.
8634 if (NumSignBits > DestBits-MidBits)
8635 return ReplaceInstUsesWith(CI, Op);
8636 } else if (OpBits < DestBits) {
8637 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8638 // bits, just sext from i32.
8639 if (NumSignBits > OpBits-MidBits)
8640 return new SExtInst(Op, CI.getType(), "tmp");
8642 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8643 // bits, just truncate to i32.
8644 if (NumSignBits > OpBits-MidBits)
8645 return new TruncInst(Op, CI.getType(), "tmp");
8649 // If the input is a shl/ashr pair of a same constant, then this is a sign
8650 // extension from a smaller value. If we could trust arbitrary bitwidth
8651 // integers, we could turn this into a truncate to the smaller bit and then
8652 // use a sext for the whole extension. Since we don't, look deeper and check
8653 // for a truncate. If the source and dest are the same type, eliminate the
8654 // trunc and extend and just do shifts. For example, turn:
8655 // %a = trunc i32 %i to i8
8656 // %b = shl i8 %a, 6
8657 // %c = ashr i8 %b, 6
8658 // %d = sext i8 %c to i32
8660 // %a = shl i32 %i, 30
8661 // %d = ashr i32 %a, 30
8663 ConstantInt *BA = 0, *CA = 0;
8664 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8665 m_ConstantInt(CA))) &&
8666 BA == CA && isa<TruncInst>(A)) {
8667 Value *I = cast<TruncInst>(A)->getOperand(0);
8668 if (I->getType() == CI.getType()) {
8669 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8670 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8671 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8672 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8673 I = Builder->CreateShl(I, ShAmtV, CI.getName());
8674 return BinaryOperator::CreateAShr(I, ShAmtV);
8681 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8682 /// in the specified FP type without changing its value.
8683 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8684 LLVMContext *Context) {
8686 APFloat F = CFP->getValueAPF();
8687 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8689 return ConstantFP::get(*Context, F);
8693 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8694 /// through it until we get the source value.
8695 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8696 if (Instruction *I = dyn_cast<Instruction>(V))
8697 if (I->getOpcode() == Instruction::FPExt)
8698 return LookThroughFPExtensions(I->getOperand(0), Context);
8700 // If this value is a constant, return the constant in the smallest FP type
8701 // that can accurately represent it. This allows us to turn
8702 // (float)((double)X+2.0) into x+2.0f.
8703 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8704 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
8705 return V; // No constant folding of this.
8706 // See if the value can be truncated to float and then reextended.
8707 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8709 if (CFP->getType() == Type::getDoubleTy(*Context))
8710 return V; // Won't shrink.
8711 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8713 // Don't try to shrink to various long double types.
8719 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8720 if (Instruction *I = commonCastTransforms(CI))
8723 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8724 // smaller than the destination type, we can eliminate the truncate by doing
8725 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8726 // many builtins (sqrt, etc).
8727 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8728 if (OpI && OpI->hasOneUse()) {
8729 switch (OpI->getOpcode()) {
8731 case Instruction::FAdd:
8732 case Instruction::FSub:
8733 case Instruction::FMul:
8734 case Instruction::FDiv:
8735 case Instruction::FRem:
8736 const Type *SrcTy = OpI->getType();
8737 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8738 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8739 if (LHSTrunc->getType() != SrcTy &&
8740 RHSTrunc->getType() != SrcTy) {
8741 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8742 // If the source types were both smaller than the destination type of
8743 // the cast, do this xform.
8744 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8745 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8746 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
8747 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
8748 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8757 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8758 return commonCastTransforms(CI);
8761 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8762 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8764 return commonCastTransforms(FI);
8766 // fptoui(uitofp(X)) --> X
8767 // fptoui(sitofp(X)) --> X
8768 // This is safe if the intermediate type has enough bits in its mantissa to
8769 // accurately represent all values of X. For example, do not do this with
8770 // i64->float->i64. This is also safe for sitofp case, because any negative
8771 // 'X' value would cause an undefined result for the fptoui.
8772 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8773 OpI->getOperand(0)->getType() == FI.getType() &&
8774 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8775 OpI->getType()->getFPMantissaWidth())
8776 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8778 return commonCastTransforms(FI);
8781 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8782 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8784 return commonCastTransforms(FI);
8786 // fptosi(sitofp(X)) --> X
8787 // fptosi(uitofp(X)) --> X
8788 // This is safe if the intermediate type has enough bits in its mantissa to
8789 // accurately represent all values of X. For example, do not do this with
8790 // i64->float->i64. This is also safe for sitofp case, because any negative
8791 // 'X' value would cause an undefined result for the fptoui.
8792 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8793 OpI->getOperand(0)->getType() == FI.getType() &&
8794 (int)FI.getType()->getScalarSizeInBits() <=
8795 OpI->getType()->getFPMantissaWidth())
8796 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8798 return commonCastTransforms(FI);
8801 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8802 return commonCastTransforms(CI);
8805 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8806 return commonCastTransforms(CI);
8809 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8810 // If the destination integer type is smaller than the intptr_t type for
8811 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8812 // trunc to be exposed to other transforms. Don't do this for extending
8813 // ptrtoint's, because we don't know if the target sign or zero extends its
8816 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8817 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
8818 TD->getIntPtrType(CI.getContext()),
8820 return new TruncInst(P, CI.getType());
8823 return commonPointerCastTransforms(CI);
8826 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8827 // If the source integer type is larger than the intptr_t type for
8828 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8829 // allows the trunc to be exposed to other transforms. Don't do this for
8830 // extending inttoptr's, because we don't know if the target sign or zero
8831 // extends to pointers.
8832 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
8833 TD->getPointerSizeInBits()) {
8834 Value *P = Builder->CreateTrunc(CI.getOperand(0),
8835 TD->getIntPtrType(CI.getContext()), "tmp");
8836 return new IntToPtrInst(P, CI.getType());
8839 if (Instruction *I = commonCastTransforms(CI))
8845 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8846 // If the operands are integer typed then apply the integer transforms,
8847 // otherwise just apply the common ones.
8848 Value *Src = CI.getOperand(0);
8849 const Type *SrcTy = Src->getType();
8850 const Type *DestTy = CI.getType();
8852 if (isa<PointerType>(SrcTy)) {
8853 if (Instruction *I = commonPointerCastTransforms(CI))
8856 if (Instruction *Result = commonCastTransforms(CI))
8861 // Get rid of casts from one type to the same type. These are useless and can
8862 // be replaced by the operand.
8863 if (DestTy == Src->getType())
8864 return ReplaceInstUsesWith(CI, Src);
8866 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8867 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8868 const Type *DstElTy = DstPTy->getElementType();
8869 const Type *SrcElTy = SrcPTy->getElementType();
8871 // If the address spaces don't match, don't eliminate the bitcast, which is
8872 // required for changing types.
8873 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8876 // If we are casting a alloca to a pointer to a type of the same
8877 // size, rewrite the allocation instruction to allocate the "right" type.
8878 // There is no need to modify malloc calls because it is their bitcast that
8879 // needs to be cleaned up.
8880 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
8881 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8884 // If the source and destination are pointers, and this cast is equivalent
8885 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8886 // This can enhance SROA and other transforms that want type-safe pointers.
8887 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
8888 unsigned NumZeros = 0;
8889 while (SrcElTy != DstElTy &&
8890 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8891 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8892 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8896 // If we found a path from the src to dest, create the getelementptr now.
8897 if (SrcElTy == DstElTy) {
8898 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8899 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(), "",
8900 ((Instruction*) NULL));
8904 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
8905 if (DestVTy->getNumElements() == 1) {
8906 if (!isa<VectorType>(SrcTy)) {
8907 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
8908 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
8909 Constant::getNullValue(Type::getInt32Ty(*Context)));
8911 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
8915 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
8916 if (SrcVTy->getNumElements() == 1) {
8917 if (!isa<VectorType>(DestTy)) {
8919 Builder->CreateExtractElement(Src,
8920 Constant::getNullValue(Type::getInt32Ty(*Context)));
8921 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
8926 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8927 if (SVI->hasOneUse()) {
8928 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8929 // a bitconvert to a vector with the same # elts.
8930 if (isa<VectorType>(DestTy) &&
8931 cast<VectorType>(DestTy)->getNumElements() ==
8932 SVI->getType()->getNumElements() &&
8933 SVI->getType()->getNumElements() ==
8934 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8936 // If either of the operands is a cast from CI.getType(), then
8937 // evaluating the shuffle in the casted destination's type will allow
8938 // us to eliminate at least one cast.
8939 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8940 Tmp->getOperand(0)->getType() == DestTy) ||
8941 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8942 Tmp->getOperand(0)->getType() == DestTy)) {
8943 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
8944 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
8945 // Return a new shuffle vector. Use the same element ID's, as we
8946 // know the vector types match #elts.
8947 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8955 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8957 /// %D = select %cond, %C, %A
8959 /// %C = select %cond, %B, 0
8962 /// Assuming that the specified instruction is an operand to the select, return
8963 /// a bitmask indicating which operands of this instruction are foldable if they
8964 /// equal the other incoming value of the select.
8966 static unsigned GetSelectFoldableOperands(Instruction *I) {
8967 switch (I->getOpcode()) {
8968 case Instruction::Add:
8969 case Instruction::Mul:
8970 case Instruction::And:
8971 case Instruction::Or:
8972 case Instruction::Xor:
8973 return 3; // Can fold through either operand.
8974 case Instruction::Sub: // Can only fold on the amount subtracted.
8975 case Instruction::Shl: // Can only fold on the shift amount.
8976 case Instruction::LShr:
8977 case Instruction::AShr:
8980 return 0; // Cannot fold
8984 /// GetSelectFoldableConstant - For the same transformation as the previous
8985 /// function, return the identity constant that goes into the select.
8986 static Constant *GetSelectFoldableConstant(Instruction *I,
8987 LLVMContext *Context) {
8988 switch (I->getOpcode()) {
8989 default: llvm_unreachable("This cannot happen!");
8990 case Instruction::Add:
8991 case Instruction::Sub:
8992 case Instruction::Or:
8993 case Instruction::Xor:
8994 case Instruction::Shl:
8995 case Instruction::LShr:
8996 case Instruction::AShr:
8997 return Constant::getNullValue(I->getType());
8998 case Instruction::And:
8999 return Constant::getAllOnesValue(I->getType());
9000 case Instruction::Mul:
9001 return ConstantInt::get(I->getType(), 1);
9005 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9006 /// have the same opcode and only one use each. Try to simplify this.
9007 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9009 if (TI->getNumOperands() == 1) {
9010 // If this is a non-volatile load or a cast from the same type,
9013 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9016 return 0; // unknown unary op.
9019 // Fold this by inserting a select from the input values.
9020 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9021 FI->getOperand(0), SI.getName()+".v");
9022 InsertNewInstBefore(NewSI, SI);
9023 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9027 // Only handle binary operators here.
9028 if (!isa<BinaryOperator>(TI))
9031 // Figure out if the operations have any operands in common.
9032 Value *MatchOp, *OtherOpT, *OtherOpF;
9034 if (TI->getOperand(0) == FI->getOperand(0)) {
9035 MatchOp = TI->getOperand(0);
9036 OtherOpT = TI->getOperand(1);
9037 OtherOpF = FI->getOperand(1);
9038 MatchIsOpZero = true;
9039 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9040 MatchOp = TI->getOperand(1);
9041 OtherOpT = TI->getOperand(0);
9042 OtherOpF = FI->getOperand(0);
9043 MatchIsOpZero = false;
9044 } else if (!TI->isCommutative()) {
9046 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9047 MatchOp = TI->getOperand(0);
9048 OtherOpT = TI->getOperand(1);
9049 OtherOpF = FI->getOperand(0);
9050 MatchIsOpZero = true;
9051 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9052 MatchOp = TI->getOperand(1);
9053 OtherOpT = TI->getOperand(0);
9054 OtherOpF = FI->getOperand(1);
9055 MatchIsOpZero = true;
9060 // If we reach here, they do have operations in common.
9061 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9062 OtherOpF, SI.getName()+".v");
9063 InsertNewInstBefore(NewSI, SI);
9065 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9067 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9069 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9071 llvm_unreachable("Shouldn't get here");
9075 static bool isSelect01(Constant *C1, Constant *C2) {
9076 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9079 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9082 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9085 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9086 /// facilitate further optimization.
9087 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9089 // See the comment above GetSelectFoldableOperands for a description of the
9090 // transformation we are doing here.
9091 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9092 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9093 !isa<Constant>(FalseVal)) {
9094 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9095 unsigned OpToFold = 0;
9096 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9098 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9103 Constant *C = GetSelectFoldableConstant(TVI, Context);
9104 Value *OOp = TVI->getOperand(2-OpToFold);
9105 // Avoid creating select between 2 constants unless it's selecting
9107 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9108 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9109 InsertNewInstBefore(NewSel, SI);
9110 NewSel->takeName(TVI);
9111 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9112 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9113 llvm_unreachable("Unknown instruction!!");
9120 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9121 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9122 !isa<Constant>(TrueVal)) {
9123 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9124 unsigned OpToFold = 0;
9125 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9127 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9132 Constant *C = GetSelectFoldableConstant(FVI, Context);
9133 Value *OOp = FVI->getOperand(2-OpToFold);
9134 // Avoid creating select between 2 constants unless it's selecting
9136 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9137 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9138 InsertNewInstBefore(NewSel, SI);
9139 NewSel->takeName(FVI);
9140 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9141 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9142 llvm_unreachable("Unknown instruction!!");
9152 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9153 /// ICmpInst as its first operand.
9155 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9157 bool Changed = false;
9158 ICmpInst::Predicate Pred = ICI->getPredicate();
9159 Value *CmpLHS = ICI->getOperand(0);
9160 Value *CmpRHS = ICI->getOperand(1);
9161 Value *TrueVal = SI.getTrueValue();
9162 Value *FalseVal = SI.getFalseValue();
9164 // Check cases where the comparison is with a constant that
9165 // can be adjusted to fit the min/max idiom. We may edit ICI in
9166 // place here, so make sure the select is the only user.
9167 if (ICI->hasOneUse())
9168 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9171 case ICmpInst::ICMP_ULT:
9172 case ICmpInst::ICMP_SLT: {
9173 // X < MIN ? T : F --> F
9174 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9175 return ReplaceInstUsesWith(SI, FalseVal);
9176 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9177 Constant *AdjustedRHS = SubOne(CI);
9178 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9179 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9180 Pred = ICmpInst::getSwappedPredicate(Pred);
9181 CmpRHS = AdjustedRHS;
9182 std::swap(FalseVal, TrueVal);
9183 ICI->setPredicate(Pred);
9184 ICI->setOperand(1, CmpRHS);
9185 SI.setOperand(1, TrueVal);
9186 SI.setOperand(2, FalseVal);
9191 case ICmpInst::ICMP_UGT:
9192 case ICmpInst::ICMP_SGT: {
9193 // X > MAX ? T : F --> F
9194 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9195 return ReplaceInstUsesWith(SI, FalseVal);
9196 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9197 Constant *AdjustedRHS = AddOne(CI);
9198 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9199 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9200 Pred = ICmpInst::getSwappedPredicate(Pred);
9201 CmpRHS = AdjustedRHS;
9202 std::swap(FalseVal, TrueVal);
9203 ICI->setPredicate(Pred);
9204 ICI->setOperand(1, CmpRHS);
9205 SI.setOperand(1, TrueVal);
9206 SI.setOperand(2, FalseVal);
9213 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9214 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9215 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9216 if (match(TrueVal, m_ConstantInt<-1>()) &&
9217 match(FalseVal, m_ConstantInt<0>()))
9218 Pred = ICI->getPredicate();
9219 else if (match(TrueVal, m_ConstantInt<0>()) &&
9220 match(FalseVal, m_ConstantInt<-1>()))
9221 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9223 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9224 // If we are just checking for a icmp eq of a single bit and zext'ing it
9225 // to an integer, then shift the bit to the appropriate place and then
9226 // cast to integer to avoid the comparison.
9227 const APInt &Op1CV = CI->getValue();
9229 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9230 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9231 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9232 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9233 Value *In = ICI->getOperand(0);
9234 Value *Sh = ConstantInt::get(In->getType(),
9235 In->getType()->getScalarSizeInBits()-1);
9236 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9237 In->getName()+".lobit"),
9239 if (In->getType() != SI.getType())
9240 In = CastInst::CreateIntegerCast(In, SI.getType(),
9241 true/*SExt*/, "tmp", ICI);
9243 if (Pred == ICmpInst::ICMP_SGT)
9244 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9245 In->getName()+".not"), *ICI);
9247 return ReplaceInstUsesWith(SI, In);
9252 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9253 // Transform (X == Y) ? X : Y -> Y
9254 if (Pred == ICmpInst::ICMP_EQ)
9255 return ReplaceInstUsesWith(SI, FalseVal);
9256 // Transform (X != Y) ? X : Y -> X
9257 if (Pred == ICmpInst::ICMP_NE)
9258 return ReplaceInstUsesWith(SI, TrueVal);
9259 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9261 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9262 // Transform (X == Y) ? Y : X -> X
9263 if (Pred == ICmpInst::ICMP_EQ)
9264 return ReplaceInstUsesWith(SI, FalseVal);
9265 // Transform (X != Y) ? Y : X -> Y
9266 if (Pred == ICmpInst::ICMP_NE)
9267 return ReplaceInstUsesWith(SI, TrueVal);
9268 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9271 /// NOTE: if we wanted to, this is where to detect integer ABS
9273 return Changed ? &SI : 0;
9277 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a
9278 /// PHI node (but the two may be in different blocks). See if the true/false
9279 /// values (V) are live in all of the predecessor blocks of the PHI. For
9280 /// example, cases like this cannot be mapped:
9282 /// X = phi [ C1, BB1], [C2, BB2]
9284 /// Z = select X, Y, 0
9286 /// because Y is not live in BB1/BB2.
9288 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V,
9289 const SelectInst &SI) {
9290 // If the value is a non-instruction value like a constant or argument, it
9291 // can always be mapped.
9292 const Instruction *I = dyn_cast<Instruction>(V);
9293 if (I == 0) return true;
9295 // If V is a PHI node defined in the same block as the condition PHI, we can
9296 // map the arguments.
9297 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
9299 if (const PHINode *VP = dyn_cast<PHINode>(I))
9300 if (VP->getParent() == CondPHI->getParent())
9303 // Otherwise, if the PHI and select are defined in the same block and if V is
9304 // defined in a different block, then we can transform it.
9305 if (SI.getParent() == CondPHI->getParent() &&
9306 I->getParent() != CondPHI->getParent())
9309 // Otherwise we have a 'hard' case and we can't tell without doing more
9310 // detailed dominator based analysis, punt.
9314 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9315 Value *CondVal = SI.getCondition();
9316 Value *TrueVal = SI.getTrueValue();
9317 Value *FalseVal = SI.getFalseValue();
9319 // select true, X, Y -> X
9320 // select false, X, Y -> Y
9321 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9322 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9324 // select C, X, X -> X
9325 if (TrueVal == FalseVal)
9326 return ReplaceInstUsesWith(SI, TrueVal);
9328 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9329 return ReplaceInstUsesWith(SI, FalseVal);
9330 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9331 return ReplaceInstUsesWith(SI, TrueVal);
9332 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9333 if (isa<Constant>(TrueVal))
9334 return ReplaceInstUsesWith(SI, TrueVal);
9336 return ReplaceInstUsesWith(SI, FalseVal);
9339 if (SI.getType() == Type::getInt1Ty(*Context)) {
9340 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9341 if (C->getZExtValue()) {
9342 // Change: A = select B, true, C --> A = or B, C
9343 return BinaryOperator::CreateOr(CondVal, FalseVal);
9345 // Change: A = select B, false, C --> A = and !B, C
9347 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9348 "not."+CondVal->getName()), SI);
9349 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9351 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9352 if (C->getZExtValue() == false) {
9353 // Change: A = select B, C, false --> A = and B, C
9354 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9356 // Change: A = select B, C, true --> A = or !B, C
9358 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9359 "not."+CondVal->getName()), SI);
9360 return BinaryOperator::CreateOr(NotCond, TrueVal);
9364 // select a, b, a -> a&b
9365 // select a, a, b -> a|b
9366 if (CondVal == TrueVal)
9367 return BinaryOperator::CreateOr(CondVal, FalseVal);
9368 else if (CondVal == FalseVal)
9369 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9372 // Selecting between two integer constants?
9373 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9374 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9375 // select C, 1, 0 -> zext C to int
9376 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9377 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9378 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9379 // select C, 0, 1 -> zext !C to int
9381 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9382 "not."+CondVal->getName()), SI);
9383 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9386 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9387 // If one of the constants is zero (we know they can't both be) and we
9388 // have an icmp instruction with zero, and we have an 'and' with the
9389 // non-constant value, eliminate this whole mess. This corresponds to
9390 // cases like this: ((X & 27) ? 27 : 0)
9391 if (TrueValC->isZero() || FalseValC->isZero())
9392 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9393 cast<Constant>(IC->getOperand(1))->isNullValue())
9394 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9395 if (ICA->getOpcode() == Instruction::And &&
9396 isa<ConstantInt>(ICA->getOperand(1)) &&
9397 (ICA->getOperand(1) == TrueValC ||
9398 ICA->getOperand(1) == FalseValC) &&
9399 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9400 // Okay, now we know that everything is set up, we just don't
9401 // know whether we have a icmp_ne or icmp_eq and whether the
9402 // true or false val is the zero.
9403 bool ShouldNotVal = !TrueValC->isZero();
9404 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9407 V = InsertNewInstBefore(BinaryOperator::Create(
9408 Instruction::Xor, V, ICA->getOperand(1)), SI);
9409 return ReplaceInstUsesWith(SI, V);
9414 // See if we are selecting two values based on a comparison of the two values.
9415 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9416 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9417 // Transform (X == Y) ? X : Y -> Y
9418 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9419 // This is not safe in general for floating point:
9420 // consider X== -0, Y== +0.
9421 // It becomes safe if either operand is a nonzero constant.
9422 ConstantFP *CFPt, *CFPf;
9423 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9424 !CFPt->getValueAPF().isZero()) ||
9425 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9426 !CFPf->getValueAPF().isZero()))
9427 return ReplaceInstUsesWith(SI, FalseVal);
9429 // Transform (X != Y) ? X : Y -> X
9430 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9431 return ReplaceInstUsesWith(SI, TrueVal);
9432 // NOTE: if we wanted to, this is where to detect MIN/MAX
9434 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9435 // Transform (X == Y) ? Y : X -> X
9436 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9437 // This is not safe in general for floating point:
9438 // consider X== -0, Y== +0.
9439 // It becomes safe if either operand is a nonzero constant.
9440 ConstantFP *CFPt, *CFPf;
9441 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9442 !CFPt->getValueAPF().isZero()) ||
9443 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9444 !CFPf->getValueAPF().isZero()))
9445 return ReplaceInstUsesWith(SI, FalseVal);
9447 // Transform (X != Y) ? Y : X -> Y
9448 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9449 return ReplaceInstUsesWith(SI, TrueVal);
9450 // NOTE: if we wanted to, this is where to detect MIN/MAX
9452 // NOTE: if we wanted to, this is where to detect ABS
9455 // See if we are selecting two values based on a comparison of the two values.
9456 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9457 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9460 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9461 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9462 if (TI->hasOneUse() && FI->hasOneUse()) {
9463 Instruction *AddOp = 0, *SubOp = 0;
9465 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9466 if (TI->getOpcode() == FI->getOpcode())
9467 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9470 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9471 // even legal for FP.
9472 if ((TI->getOpcode() == Instruction::Sub &&
9473 FI->getOpcode() == Instruction::Add) ||
9474 (TI->getOpcode() == Instruction::FSub &&
9475 FI->getOpcode() == Instruction::FAdd)) {
9476 AddOp = FI; SubOp = TI;
9477 } else if ((FI->getOpcode() == Instruction::Sub &&
9478 TI->getOpcode() == Instruction::Add) ||
9479 (FI->getOpcode() == Instruction::FSub &&
9480 TI->getOpcode() == Instruction::FAdd)) {
9481 AddOp = TI; SubOp = FI;
9485 Value *OtherAddOp = 0;
9486 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9487 OtherAddOp = AddOp->getOperand(1);
9488 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9489 OtherAddOp = AddOp->getOperand(0);
9493 // So at this point we know we have (Y -> OtherAddOp):
9494 // select C, (add X, Y), (sub X, Z)
9495 Value *NegVal; // Compute -Z
9496 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9497 NegVal = ConstantExpr::getNeg(C);
9499 NegVal = InsertNewInstBefore(
9500 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9504 Value *NewTrueOp = OtherAddOp;
9505 Value *NewFalseOp = NegVal;
9507 std::swap(NewTrueOp, NewFalseOp);
9508 Instruction *NewSel =
9509 SelectInst::Create(CondVal, NewTrueOp,
9510 NewFalseOp, SI.getName() + ".p");
9512 NewSel = InsertNewInstBefore(NewSel, SI);
9513 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9518 // See if we can fold the select into one of our operands.
9519 if (SI.getType()->isInteger()) {
9520 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9525 // See if we can fold the select into a phi node if the condition is a select.
9526 if (isa<PHINode>(SI.getCondition()))
9527 // The true/false values have to be live in the PHI predecessor's blocks.
9528 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
9529 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
9530 if (Instruction *NV = FoldOpIntoPhi(SI))
9533 if (BinaryOperator::isNot(CondVal)) {
9534 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9535 SI.setOperand(1, FalseVal);
9536 SI.setOperand(2, TrueVal);
9543 /// EnforceKnownAlignment - If the specified pointer points to an object that
9544 /// we control, modify the object's alignment to PrefAlign. This isn't
9545 /// often possible though. If alignment is important, a more reliable approach
9546 /// is to simply align all global variables and allocation instructions to
9547 /// their preferred alignment from the beginning.
9549 static unsigned EnforceKnownAlignment(Value *V,
9550 unsigned Align, unsigned PrefAlign) {
9552 User *U = dyn_cast<User>(V);
9553 if (!U) return Align;
9555 switch (Operator::getOpcode(U)) {
9557 case Instruction::BitCast:
9558 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9559 case Instruction::GetElementPtr: {
9560 // If all indexes are zero, it is just the alignment of the base pointer.
9561 bool AllZeroOperands = true;
9562 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9563 if (!isa<Constant>(*i) ||
9564 !cast<Constant>(*i)->isNullValue()) {
9565 AllZeroOperands = false;
9569 if (AllZeroOperands) {
9570 // Treat this like a bitcast.
9571 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9577 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9578 // If there is a large requested alignment and we can, bump up the alignment
9580 if (!GV->isDeclaration()) {
9581 if (GV->getAlignment() >= PrefAlign)
9582 Align = GV->getAlignment();
9584 GV->setAlignment(PrefAlign);
9588 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
9589 // If there is a requested alignment and if this is an alloca, round up.
9590 if (AI->getAlignment() >= PrefAlign)
9591 Align = AI->getAlignment();
9593 AI->setAlignment(PrefAlign);
9601 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9602 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9603 /// and it is more than the alignment of the ultimate object, see if we can
9604 /// increase the alignment of the ultimate object, making this check succeed.
9605 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9606 unsigned PrefAlign) {
9607 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9608 sizeof(PrefAlign) * CHAR_BIT;
9609 APInt Mask = APInt::getAllOnesValue(BitWidth);
9610 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9611 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9612 unsigned TrailZ = KnownZero.countTrailingOnes();
9613 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9615 if (PrefAlign > Align)
9616 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9618 // We don't need to make any adjustment.
9622 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9623 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9624 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9625 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9626 unsigned CopyAlign = MI->getAlignment();
9628 if (CopyAlign < MinAlign) {
9629 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9634 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9636 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9637 if (MemOpLength == 0) return 0;
9639 // Source and destination pointer types are always "i8*" for intrinsic. See
9640 // if the size is something we can handle with a single primitive load/store.
9641 // A single load+store correctly handles overlapping memory in the memmove
9643 unsigned Size = MemOpLength->getZExtValue();
9644 if (Size == 0) return MI; // Delete this mem transfer.
9646 if (Size > 8 || (Size&(Size-1)))
9647 return 0; // If not 1/2/4/8 bytes, exit.
9649 // Use an integer load+store unless we can find something better.
9651 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
9653 // Memcpy forces the use of i8* for the source and destination. That means
9654 // that if you're using memcpy to move one double around, you'll get a cast
9655 // from double* to i8*. We'd much rather use a double load+store rather than
9656 // an i64 load+store, here because this improves the odds that the source or
9657 // dest address will be promotable. See if we can find a better type than the
9658 // integer datatype.
9659 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9660 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9661 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9662 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9663 // down through these levels if so.
9664 while (!SrcETy->isSingleValueType()) {
9665 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9666 if (STy->getNumElements() == 1)
9667 SrcETy = STy->getElementType(0);
9670 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9671 if (ATy->getNumElements() == 1)
9672 SrcETy = ATy->getElementType();
9679 if (SrcETy->isSingleValueType())
9680 NewPtrTy = PointerType::getUnqual(SrcETy);
9685 // If the memcpy/memmove provides better alignment info than we can
9687 SrcAlign = std::max(SrcAlign, CopyAlign);
9688 DstAlign = std::max(DstAlign, CopyAlign);
9690 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy);
9691 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy);
9692 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9693 InsertNewInstBefore(L, *MI);
9694 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9696 // Set the size of the copy to 0, it will be deleted on the next iteration.
9697 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9701 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9702 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9703 if (MI->getAlignment() < Alignment) {
9704 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9709 // Extract the length and alignment and fill if they are constant.
9710 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9711 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9712 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
9714 uint64_t Len = LenC->getZExtValue();
9715 Alignment = MI->getAlignment();
9717 // If the length is zero, this is a no-op
9718 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9720 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9721 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9722 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
9724 Value *Dest = MI->getDest();
9725 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
9727 // Alignment 0 is identity for alignment 1 for memset, but not store.
9728 if (Alignment == 0) Alignment = 1;
9730 // Extract the fill value and store.
9731 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9732 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9733 Dest, false, Alignment), *MI);
9735 // Set the size of the copy to 0, it will be deleted on the next iteration.
9736 MI->setLength(Constant::getNullValue(LenC->getType()));
9744 /// visitCallInst - CallInst simplification. This mostly only handles folding
9745 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9746 /// the heavy lifting.
9748 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9749 if (isFreeCall(&CI))
9750 return visitFree(CI);
9752 // If the caller function is nounwind, mark the call as nounwind, even if the
9754 if (CI.getParent()->getParent()->doesNotThrow() &&
9755 !CI.doesNotThrow()) {
9756 CI.setDoesNotThrow();
9760 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9761 if (!II) return visitCallSite(&CI);
9763 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9765 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9766 bool Changed = false;
9768 // memmove/cpy/set of zero bytes is a noop.
9769 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9770 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9772 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9773 if (CI->getZExtValue() == 1) {
9774 // Replace the instruction with just byte operations. We would
9775 // transform other cases to loads/stores, but we don't know if
9776 // alignment is sufficient.
9780 // If we have a memmove and the source operation is a constant global,
9781 // then the source and dest pointers can't alias, so we can change this
9782 // into a call to memcpy.
9783 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9784 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9785 if (GVSrc->isConstant()) {
9786 Module *M = CI.getParent()->getParent()->getParent();
9787 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9789 Tys[0] = CI.getOperand(3)->getType();
9791 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9795 // memmove(x,x,size) -> noop.
9796 if (MMI->getSource() == MMI->getDest())
9797 return EraseInstFromFunction(CI);
9800 // If we can determine a pointer alignment that is bigger than currently
9801 // set, update the alignment.
9802 if (isa<MemTransferInst>(MI)) {
9803 if (Instruction *I = SimplifyMemTransfer(MI))
9805 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9806 if (Instruction *I = SimplifyMemSet(MSI))
9810 if (Changed) return II;
9813 switch (II->getIntrinsicID()) {
9815 case Intrinsic::bswap:
9816 // bswap(bswap(x)) -> x
9817 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9818 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9819 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9821 case Intrinsic::ppc_altivec_lvx:
9822 case Intrinsic::ppc_altivec_lvxl:
9823 case Intrinsic::x86_sse_loadu_ps:
9824 case Intrinsic::x86_sse2_loadu_pd:
9825 case Intrinsic::x86_sse2_loadu_dq:
9826 // Turn PPC lvx -> load if the pointer is known aligned.
9827 // Turn X86 loadups -> load if the pointer is known aligned.
9828 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9829 Value *Ptr = Builder->CreateBitCast(II->getOperand(1),
9830 PointerType::getUnqual(II->getType()));
9831 return new LoadInst(Ptr);
9834 case Intrinsic::ppc_altivec_stvx:
9835 case Intrinsic::ppc_altivec_stvxl:
9836 // Turn stvx -> store if the pointer is known aligned.
9837 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9838 const Type *OpPtrTy =
9839 PointerType::getUnqual(II->getOperand(1)->getType());
9840 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy);
9841 return new StoreInst(II->getOperand(1), Ptr);
9844 case Intrinsic::x86_sse_storeu_ps:
9845 case Intrinsic::x86_sse2_storeu_pd:
9846 case Intrinsic::x86_sse2_storeu_dq:
9847 // Turn X86 storeu -> store if the pointer is known aligned.
9848 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9849 const Type *OpPtrTy =
9850 PointerType::getUnqual(II->getOperand(2)->getType());
9851 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
9852 return new StoreInst(II->getOperand(2), Ptr);
9856 case Intrinsic::x86_sse_cvttss2si: {
9857 // These intrinsics only demands the 0th element of its input vector. If
9858 // we can simplify the input based on that, do so now.
9860 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9861 APInt DemandedElts(VWidth, 1);
9862 APInt UndefElts(VWidth, 0);
9863 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9865 II->setOperand(1, V);
9871 case Intrinsic::ppc_altivec_vperm:
9872 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9873 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9874 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9876 // Check that all of the elements are integer constants or undefs.
9877 bool AllEltsOk = true;
9878 for (unsigned i = 0; i != 16; ++i) {
9879 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9880 !isa<UndefValue>(Mask->getOperand(i))) {
9887 // Cast the input vectors to byte vectors.
9888 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
9889 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType());
9890 Value *Result = UndefValue::get(Op0->getType());
9892 // Only extract each element once.
9893 Value *ExtractedElts[32];
9894 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9896 for (unsigned i = 0; i != 16; ++i) {
9897 if (isa<UndefValue>(Mask->getOperand(i)))
9899 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9900 Idx &= 31; // Match the hardware behavior.
9902 if (ExtractedElts[Idx] == 0) {
9903 ExtractedElts[Idx] =
9904 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
9905 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false),
9909 // Insert this value into the result vector.
9910 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
9911 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
9914 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9919 case Intrinsic::stackrestore: {
9920 // If the save is right next to the restore, remove the restore. This can
9921 // happen when variable allocas are DCE'd.
9922 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9923 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9924 BasicBlock::iterator BI = SS;
9926 return EraseInstFromFunction(CI);
9930 // Scan down this block to see if there is another stack restore in the
9931 // same block without an intervening call/alloca.
9932 BasicBlock::iterator BI = II;
9933 TerminatorInst *TI = II->getParent()->getTerminator();
9934 bool CannotRemove = false;
9935 for (++BI; &*BI != TI; ++BI) {
9936 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
9937 CannotRemove = true;
9940 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9941 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9942 // If there is a stackrestore below this one, remove this one.
9943 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9944 return EraseInstFromFunction(CI);
9945 // Otherwise, ignore the intrinsic.
9947 // If we found a non-intrinsic call, we can't remove the stack
9949 CannotRemove = true;
9955 // If the stack restore is in a return/unwind block and if there are no
9956 // allocas or calls between the restore and the return, nuke the restore.
9957 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9958 return EraseInstFromFunction(CI);
9963 return visitCallSite(II);
9966 // InvokeInst simplification
9968 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9969 return visitCallSite(&II);
9972 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9973 /// passed through the varargs area, we can eliminate the use of the cast.
9974 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9975 const CastInst * const CI,
9976 const TargetData * const TD,
9978 if (!CI->isLosslessCast())
9981 // The size of ByVal arguments is derived from the type, so we
9982 // can't change to a type with a different size. If the size were
9983 // passed explicitly we could avoid this check.
9984 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9988 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9989 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9990 if (!SrcTy->isSized() || !DstTy->isSized())
9992 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
9997 // visitCallSite - Improvements for call and invoke instructions.
9999 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10000 bool Changed = false;
10002 // If the callee is a constexpr cast of a function, attempt to move the cast
10003 // to the arguments of the call/invoke.
10004 if (transformConstExprCastCall(CS)) return 0;
10006 Value *Callee = CS.getCalledValue();
10008 if (Function *CalleeF = dyn_cast<Function>(Callee))
10009 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10010 Instruction *OldCall = CS.getInstruction();
10011 // If the call and callee calling conventions don't match, this call must
10012 // be unreachable, as the call is undefined.
10013 new StoreInst(ConstantInt::getTrue(*Context),
10014 UndefValue::get(Type::getInt1PtrTy(*Context)),
10016 // If OldCall dues not return void then replaceAllUsesWith undef.
10017 // This allows ValueHandlers and custom metadata to adjust itself.
10018 if (!OldCall->getType()->isVoidTy())
10019 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10020 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10021 return EraseInstFromFunction(*OldCall);
10025 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10026 // This instruction is not reachable, just remove it. We insert a store to
10027 // undef so that we know that this code is not reachable, despite the fact
10028 // that we can't modify the CFG here.
10029 new StoreInst(ConstantInt::getTrue(*Context),
10030 UndefValue::get(Type::getInt1PtrTy(*Context)),
10031 CS.getInstruction());
10033 // If CS dues not return void then replaceAllUsesWith undef.
10034 // This allows ValueHandlers and custom metadata to adjust itself.
10035 if (!CS.getInstruction()->getType()->isVoidTy())
10036 CS.getInstruction()->
10037 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10039 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10040 // Don't break the CFG, insert a dummy cond branch.
10041 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10042 ConstantInt::getTrue(*Context), II);
10044 return EraseInstFromFunction(*CS.getInstruction());
10047 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10048 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10049 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10050 return transformCallThroughTrampoline(CS);
10052 const PointerType *PTy = cast<PointerType>(Callee->getType());
10053 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10054 if (FTy->isVarArg()) {
10055 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10056 // See if we can optimize any arguments passed through the varargs area of
10058 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10059 E = CS.arg_end(); I != E; ++I, ++ix) {
10060 CastInst *CI = dyn_cast<CastInst>(*I);
10061 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10062 *I = CI->getOperand(0);
10068 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10069 // Inline asm calls cannot throw - mark them 'nounwind'.
10070 CS.setDoesNotThrow();
10074 return Changed ? CS.getInstruction() : 0;
10077 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10078 // attempt to move the cast to the arguments of the call/invoke.
10080 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10081 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10082 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10083 if (CE->getOpcode() != Instruction::BitCast ||
10084 !isa<Function>(CE->getOperand(0)))
10086 Function *Callee = cast<Function>(CE->getOperand(0));
10087 Instruction *Caller = CS.getInstruction();
10088 const AttrListPtr &CallerPAL = CS.getAttributes();
10090 // Okay, this is a cast from a function to a different type. Unless doing so
10091 // would cause a type conversion of one of our arguments, change this call to
10092 // be a direct call with arguments casted to the appropriate types.
10094 const FunctionType *FT = Callee->getFunctionType();
10095 const Type *OldRetTy = Caller->getType();
10096 const Type *NewRetTy = FT->getReturnType();
10098 if (isa<StructType>(NewRetTy))
10099 return false; // TODO: Handle multiple return values.
10101 // Check to see if we are changing the return type...
10102 if (OldRetTy != NewRetTy) {
10103 if (Callee->isDeclaration() &&
10104 // Conversion is ok if changing from one pointer type to another or from
10105 // a pointer to an integer of the same size.
10106 !((isa<PointerType>(OldRetTy) || !TD ||
10107 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
10108 (isa<PointerType>(NewRetTy) || !TD ||
10109 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
10110 return false; // Cannot transform this return value.
10112 if (!Caller->use_empty() &&
10113 // void -> non-void is handled specially
10114 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
10115 return false; // Cannot transform this return value.
10117 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10118 Attributes RAttrs = CallerPAL.getRetAttributes();
10119 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10120 return false; // Attribute not compatible with transformed value.
10123 // If the callsite is an invoke instruction, and the return value is used by
10124 // a PHI node in a successor, we cannot change the return type of the call
10125 // because there is no place to put the cast instruction (without breaking
10126 // the critical edge). Bail out in this case.
10127 if (!Caller->use_empty())
10128 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10129 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10131 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10132 if (PN->getParent() == II->getNormalDest() ||
10133 PN->getParent() == II->getUnwindDest())
10137 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10138 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10140 CallSite::arg_iterator AI = CS.arg_begin();
10141 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10142 const Type *ParamTy = FT->getParamType(i);
10143 const Type *ActTy = (*AI)->getType();
10145 if (!CastInst::isCastable(ActTy, ParamTy))
10146 return false; // Cannot transform this parameter value.
10148 if (CallerPAL.getParamAttributes(i + 1)
10149 & Attribute::typeIncompatible(ParamTy))
10150 return false; // Attribute not compatible with transformed value.
10152 // Converting from one pointer type to another or between a pointer and an
10153 // integer of the same size is safe even if we do not have a body.
10154 bool isConvertible = ActTy == ParamTy ||
10155 (TD && ((isa<PointerType>(ParamTy) ||
10156 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10157 (isa<PointerType>(ActTy) ||
10158 ActTy == TD->getIntPtrType(Caller->getContext()))));
10159 if (Callee->isDeclaration() && !isConvertible) return false;
10162 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10163 Callee->isDeclaration())
10164 return false; // Do not delete arguments unless we have a function body.
10166 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10167 !CallerPAL.isEmpty())
10168 // In this case we have more arguments than the new function type, but we
10169 // won't be dropping them. Check that these extra arguments have attributes
10170 // that are compatible with being a vararg call argument.
10171 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10172 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10174 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10175 if (PAttrs & Attribute::VarArgsIncompatible)
10179 // Okay, we decided that this is a safe thing to do: go ahead and start
10180 // inserting cast instructions as necessary...
10181 std::vector<Value*> Args;
10182 Args.reserve(NumActualArgs);
10183 SmallVector<AttributeWithIndex, 8> attrVec;
10184 attrVec.reserve(NumCommonArgs);
10186 // Get any return attributes.
10187 Attributes RAttrs = CallerPAL.getRetAttributes();
10189 // If the return value is not being used, the type may not be compatible
10190 // with the existing attributes. Wipe out any problematic attributes.
10191 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10193 // Add the new return attributes.
10195 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10197 AI = CS.arg_begin();
10198 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10199 const Type *ParamTy = FT->getParamType(i);
10200 if ((*AI)->getType() == ParamTy) {
10201 Args.push_back(*AI);
10203 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10204 false, ParamTy, false);
10205 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
10208 // Add any parameter attributes.
10209 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10210 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10213 // If the function takes more arguments than the call was taking, add them
10215 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10216 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10218 // If we are removing arguments to the function, emit an obnoxious warning.
10219 if (FT->getNumParams() < NumActualArgs) {
10220 if (!FT->isVarArg()) {
10221 errs() << "WARNING: While resolving call to function '"
10222 << Callee->getName() << "' arguments were dropped!\n";
10224 // Add all of the arguments in their promoted form to the arg list.
10225 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10226 const Type *PTy = getPromotedType((*AI)->getType());
10227 if (PTy != (*AI)->getType()) {
10228 // Must promote to pass through va_arg area!
10229 Instruction::CastOps opcode =
10230 CastInst::getCastOpcode(*AI, false, PTy, false);
10231 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
10233 Args.push_back(*AI);
10236 // Add any parameter attributes.
10237 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10238 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10243 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10244 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10246 if (NewRetTy->isVoidTy())
10247 Caller->setName(""); // Void type should not have a name.
10249 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10253 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10254 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10255 Args.begin(), Args.end(),
10256 Caller->getName(), Caller);
10257 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10258 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10260 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10261 Caller->getName(), Caller);
10262 CallInst *CI = cast<CallInst>(Caller);
10263 if (CI->isTailCall())
10264 cast<CallInst>(NC)->setTailCall();
10265 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10266 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10269 // Insert a cast of the return type as necessary.
10271 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10272 if (!NV->getType()->isVoidTy()) {
10273 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10275 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10277 // If this is an invoke instruction, we should insert it after the first
10278 // non-phi, instruction in the normal successor block.
10279 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10280 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10281 InsertNewInstBefore(NC, *I);
10283 // Otherwise, it's a call, just insert cast right after the call instr
10284 InsertNewInstBefore(NC, *Caller);
10286 Worklist.AddUsersToWorkList(*Caller);
10288 NV = UndefValue::get(Caller->getType());
10293 if (!Caller->use_empty())
10294 Caller->replaceAllUsesWith(NV);
10296 EraseInstFromFunction(*Caller);
10300 // transformCallThroughTrampoline - Turn a call to a function created by the
10301 // init_trampoline intrinsic into a direct call to the underlying function.
10303 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10304 Value *Callee = CS.getCalledValue();
10305 const PointerType *PTy = cast<PointerType>(Callee->getType());
10306 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10307 const AttrListPtr &Attrs = CS.getAttributes();
10309 // If the call already has the 'nest' attribute somewhere then give up -
10310 // otherwise 'nest' would occur twice after splicing in the chain.
10311 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10314 IntrinsicInst *Tramp =
10315 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10317 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10318 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10319 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10321 const AttrListPtr &NestAttrs = NestF->getAttributes();
10322 if (!NestAttrs.isEmpty()) {
10323 unsigned NestIdx = 1;
10324 const Type *NestTy = 0;
10325 Attributes NestAttr = Attribute::None;
10327 // Look for a parameter marked with the 'nest' attribute.
10328 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10329 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10330 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10331 // Record the parameter type and any other attributes.
10333 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10338 Instruction *Caller = CS.getInstruction();
10339 std::vector<Value*> NewArgs;
10340 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10342 SmallVector<AttributeWithIndex, 8> NewAttrs;
10343 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10345 // Insert the nest argument into the call argument list, which may
10346 // mean appending it. Likewise for attributes.
10348 // Add any result attributes.
10349 if (Attributes Attr = Attrs.getRetAttributes())
10350 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10354 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10356 if (Idx == NestIdx) {
10357 // Add the chain argument and attributes.
10358 Value *NestVal = Tramp->getOperand(3);
10359 if (NestVal->getType() != NestTy)
10360 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10361 NewArgs.push_back(NestVal);
10362 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10368 // Add the original argument and attributes.
10369 NewArgs.push_back(*I);
10370 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10372 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10378 // Add any function attributes.
10379 if (Attributes Attr = Attrs.getFnAttributes())
10380 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10382 // The trampoline may have been bitcast to a bogus type (FTy).
10383 // Handle this by synthesizing a new function type, equal to FTy
10384 // with the chain parameter inserted.
10386 std::vector<const Type*> NewTypes;
10387 NewTypes.reserve(FTy->getNumParams()+1);
10389 // Insert the chain's type into the list of parameter types, which may
10390 // mean appending it.
10393 FunctionType::param_iterator I = FTy->param_begin(),
10394 E = FTy->param_end();
10397 if (Idx == NestIdx)
10398 // Add the chain's type.
10399 NewTypes.push_back(NestTy);
10404 // Add the original type.
10405 NewTypes.push_back(*I);
10411 // Replace the trampoline call with a direct call. Let the generic
10412 // code sort out any function type mismatches.
10413 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10415 Constant *NewCallee =
10416 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10417 NestF : ConstantExpr::getBitCast(NestF,
10418 PointerType::getUnqual(NewFTy));
10419 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10422 Instruction *NewCaller;
10423 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10424 NewCaller = InvokeInst::Create(NewCallee,
10425 II->getNormalDest(), II->getUnwindDest(),
10426 NewArgs.begin(), NewArgs.end(),
10427 Caller->getName(), Caller);
10428 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10429 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10431 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10432 Caller->getName(), Caller);
10433 if (cast<CallInst>(Caller)->isTailCall())
10434 cast<CallInst>(NewCaller)->setTailCall();
10435 cast<CallInst>(NewCaller)->
10436 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10437 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10439 if (!Caller->getType()->isVoidTy())
10440 Caller->replaceAllUsesWith(NewCaller);
10441 Caller->eraseFromParent();
10442 Worklist.Remove(Caller);
10447 // Replace the trampoline call with a direct call. Since there is no 'nest'
10448 // parameter, there is no need to adjust the argument list. Let the generic
10449 // code sort out any function type mismatches.
10450 Constant *NewCallee =
10451 NestF->getType() == PTy ? NestF :
10452 ConstantExpr::getBitCast(NestF, PTy);
10453 CS.setCalledFunction(NewCallee);
10454 return CS.getInstruction();
10457 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
10458 /// and if a/b/c and the add's all have a single use, turn this into a phi
10459 /// and a single binop.
10460 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10461 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10462 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10463 unsigned Opc = FirstInst->getOpcode();
10464 Value *LHSVal = FirstInst->getOperand(0);
10465 Value *RHSVal = FirstInst->getOperand(1);
10467 const Type *LHSType = LHSVal->getType();
10468 const Type *RHSType = RHSVal->getType();
10470 // Scan to see if all operands are the same opcode, and all have one use.
10471 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10472 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10473 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10474 // Verify type of the LHS matches so we don't fold cmp's of different
10475 // types or GEP's with different index types.
10476 I->getOperand(0)->getType() != LHSType ||
10477 I->getOperand(1)->getType() != RHSType)
10480 // If they are CmpInst instructions, check their predicates
10481 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10482 if (cast<CmpInst>(I)->getPredicate() !=
10483 cast<CmpInst>(FirstInst)->getPredicate())
10486 // Keep track of which operand needs a phi node.
10487 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10488 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10491 // If both LHS and RHS would need a PHI, don't do this transformation,
10492 // because it would increase the number of PHIs entering the block,
10493 // which leads to higher register pressure. This is especially
10494 // bad when the PHIs are in the header of a loop.
10495 if (!LHSVal && !RHSVal)
10498 // Otherwise, this is safe to transform!
10500 Value *InLHS = FirstInst->getOperand(0);
10501 Value *InRHS = FirstInst->getOperand(1);
10502 PHINode *NewLHS = 0, *NewRHS = 0;
10504 NewLHS = PHINode::Create(LHSType,
10505 FirstInst->getOperand(0)->getName() + ".pn");
10506 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10507 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10508 InsertNewInstBefore(NewLHS, PN);
10513 NewRHS = PHINode::Create(RHSType,
10514 FirstInst->getOperand(1)->getName() + ".pn");
10515 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10516 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10517 InsertNewInstBefore(NewRHS, PN);
10521 // Add all operands to the new PHIs.
10522 if (NewLHS || NewRHS) {
10523 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10524 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10526 Value *NewInLHS = InInst->getOperand(0);
10527 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10530 Value *NewInRHS = InInst->getOperand(1);
10531 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10536 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10537 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10538 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10539 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10543 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10544 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10546 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10547 FirstInst->op_end());
10548 // This is true if all GEP bases are allocas and if all indices into them are
10550 bool AllBasePointersAreAllocas = true;
10552 // We don't want to replace this phi if the replacement would require
10553 // more than one phi, which leads to higher register pressure. This is
10554 // especially bad when the PHIs are in the header of a loop.
10555 bool NeededPhi = false;
10557 // Scan to see if all operands are the same opcode, and all have one use.
10558 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10559 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10560 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10561 GEP->getNumOperands() != FirstInst->getNumOperands())
10564 // Keep track of whether or not all GEPs are of alloca pointers.
10565 if (AllBasePointersAreAllocas &&
10566 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10567 !GEP->hasAllConstantIndices()))
10568 AllBasePointersAreAllocas = false;
10570 // Compare the operand lists.
10571 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10572 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10575 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10576 // if one of the PHIs has a constant for the index. The index may be
10577 // substantially cheaper to compute for the constants, so making it a
10578 // variable index could pessimize the path. This also handles the case
10579 // for struct indices, which must always be constant.
10580 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10581 isa<ConstantInt>(GEP->getOperand(op)))
10584 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10587 // If we already needed a PHI for an earlier operand, and another operand
10588 // also requires a PHI, we'd be introducing more PHIs than we're
10589 // eliminating, which increases register pressure on entry to the PHI's
10594 FixedOperands[op] = 0; // Needs a PHI.
10599 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10600 // bother doing this transformation. At best, this will just save a bit of
10601 // offset calculation, but all the predecessors will have to materialize the
10602 // stack address into a register anyway. We'd actually rather *clone* the
10603 // load up into the predecessors so that we have a load of a gep of an alloca,
10604 // which can usually all be folded into the load.
10605 if (AllBasePointersAreAllocas)
10608 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10609 // that is variable.
10610 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10612 bool HasAnyPHIs = false;
10613 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10614 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10615 Value *FirstOp = FirstInst->getOperand(i);
10616 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10617 FirstOp->getName()+".pn");
10618 InsertNewInstBefore(NewPN, PN);
10620 NewPN->reserveOperandSpace(e);
10621 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10622 OperandPhis[i] = NewPN;
10623 FixedOperands[i] = NewPN;
10628 // Add all operands to the new PHIs.
10630 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10631 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10632 BasicBlock *InBB = PN.getIncomingBlock(i);
10634 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10635 if (PHINode *OpPhi = OperandPhis[op])
10636 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10640 Value *Base = FixedOperands[0];
10641 return cast<GEPOperator>(FirstInst)->isInBounds() ?
10642 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
10643 FixedOperands.end()) :
10644 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10645 FixedOperands.end());
10649 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10650 /// sink the load out of the block that defines it. This means that it must be
10651 /// obvious the value of the load is not changed from the point of the load to
10652 /// the end of the block it is in.
10654 /// Finally, it is safe, but not profitable, to sink a load targetting a
10655 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10657 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10658 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10660 for (++BBI; BBI != E; ++BBI)
10661 if (BBI->mayWriteToMemory())
10664 // Check for non-address taken alloca. If not address-taken already, it isn't
10665 // profitable to do this xform.
10666 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10667 bool isAddressTaken = false;
10668 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10670 if (isa<LoadInst>(UI)) continue;
10671 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10672 // If storing TO the alloca, then the address isn't taken.
10673 if (SI->getOperand(1) == AI) continue;
10675 isAddressTaken = true;
10679 if (!isAddressTaken && AI->isStaticAlloca())
10683 // If this load is a load from a GEP with a constant offset from an alloca,
10684 // then we don't want to sink it. In its present form, it will be
10685 // load [constant stack offset]. Sinking it will cause us to have to
10686 // materialize the stack addresses in each predecessor in a register only to
10687 // do a shared load from register in the successor.
10688 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10689 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10690 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10697 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10698 // operator and they all are only used by the PHI, PHI together their
10699 // inputs, and do the operation once, to the result of the PHI.
10700 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10701 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10703 // Scan the instruction, looking for input operations that can be folded away.
10704 // If all input operands to the phi are the same instruction (e.g. a cast from
10705 // the same type or "+42") we can pull the operation through the PHI, reducing
10706 // code size and simplifying code.
10707 Constant *ConstantOp = 0;
10708 const Type *CastSrcTy = 0;
10709 bool isVolatile = false;
10710 if (isa<CastInst>(FirstInst)) {
10711 CastSrcTy = FirstInst->getOperand(0)->getType();
10712 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10713 // Can fold binop, compare or shift here if the RHS is a constant,
10714 // otherwise call FoldPHIArgBinOpIntoPHI.
10715 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10716 if (ConstantOp == 0)
10717 return FoldPHIArgBinOpIntoPHI(PN);
10718 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10719 isVolatile = LI->isVolatile();
10720 // We can't sink the load if the loaded value could be modified between the
10721 // load and the PHI.
10722 if (LI->getParent() != PN.getIncomingBlock(0) ||
10723 !isSafeAndProfitableToSinkLoad(LI))
10726 // If the PHI is of volatile loads and the load block has multiple
10727 // successors, sinking it would remove a load of the volatile value from
10728 // the path through the other successor.
10730 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10733 } else if (isa<GetElementPtrInst>(FirstInst)) {
10734 return FoldPHIArgGEPIntoPHI(PN);
10736 return 0; // Cannot fold this operation.
10739 // Check to see if all arguments are the same operation.
10740 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10741 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10742 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10743 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10746 if (I->getOperand(0)->getType() != CastSrcTy)
10747 return 0; // Cast operation must match.
10748 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10749 // We can't sink the load if the loaded value could be modified between
10750 // the load and the PHI.
10751 if (LI->isVolatile() != isVolatile ||
10752 LI->getParent() != PN.getIncomingBlock(i) ||
10753 !isSafeAndProfitableToSinkLoad(LI))
10756 // If the PHI is of volatile loads and the load block has multiple
10757 // successors, sinking it would remove a load of the volatile value from
10758 // the path through the other successor.
10760 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10763 } else if (I->getOperand(1) != ConstantOp) {
10768 // Okay, they are all the same operation. Create a new PHI node of the
10769 // correct type, and PHI together all of the LHS's of the instructions.
10770 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10771 PN.getName()+".in");
10772 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10774 Value *InVal = FirstInst->getOperand(0);
10775 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10777 // Add all operands to the new PHI.
10778 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10779 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10780 if (NewInVal != InVal)
10782 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10787 // The new PHI unions all of the same values together. This is really
10788 // common, so we handle it intelligently here for compile-time speed.
10792 InsertNewInstBefore(NewPN, PN);
10796 // Insert and return the new operation.
10797 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10798 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10799 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10800 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10801 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10802 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10803 PhiVal, ConstantOp);
10804 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10806 // If this was a volatile load that we are merging, make sure to loop through
10807 // and mark all the input loads as non-volatile. If we don't do this, we will
10808 // insert a new volatile load and the old ones will not be deletable.
10810 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10811 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10813 return new LoadInst(PhiVal, "", isVolatile);
10816 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10818 static bool DeadPHICycle(PHINode *PN,
10819 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10820 if (PN->use_empty()) return true;
10821 if (!PN->hasOneUse()) return false;
10823 // Remember this node, and if we find the cycle, return.
10824 if (!PotentiallyDeadPHIs.insert(PN))
10827 // Don't scan crazily complex things.
10828 if (PotentiallyDeadPHIs.size() == 16)
10831 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10832 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10837 /// PHIsEqualValue - Return true if this phi node is always equal to
10838 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10839 /// z = some value; x = phi (y, z); y = phi (x, z)
10840 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10841 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10842 // See if we already saw this PHI node.
10843 if (!ValueEqualPHIs.insert(PN))
10846 // Don't scan crazily complex things.
10847 if (ValueEqualPHIs.size() == 16)
10850 // Scan the operands to see if they are either phi nodes or are equal to
10852 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10853 Value *Op = PN->getIncomingValue(i);
10854 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10855 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10857 } else if (Op != NonPhiInVal)
10865 // PHINode simplification
10867 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10868 // If LCSSA is around, don't mess with Phi nodes
10869 if (MustPreserveLCSSA) return 0;
10871 if (Value *V = PN.hasConstantValue())
10872 return ReplaceInstUsesWith(PN, V);
10874 // If all PHI operands are the same operation, pull them through the PHI,
10875 // reducing code size.
10876 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10877 isa<Instruction>(PN.getIncomingValue(1)) &&
10878 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10879 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10880 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10881 // than themselves more than once.
10882 PN.getIncomingValue(0)->hasOneUse())
10883 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10886 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10887 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10888 // PHI)... break the cycle.
10889 if (PN.hasOneUse()) {
10890 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10891 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10892 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10893 PotentiallyDeadPHIs.insert(&PN);
10894 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10895 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10898 // If this phi has a single use, and if that use just computes a value for
10899 // the next iteration of a loop, delete the phi. This occurs with unused
10900 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10901 // common case here is good because the only other things that catch this
10902 // are induction variable analysis (sometimes) and ADCE, which is only run
10904 if (PHIUser->hasOneUse() &&
10905 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10906 PHIUser->use_back() == &PN) {
10907 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10911 // We sometimes end up with phi cycles that non-obviously end up being the
10912 // same value, for example:
10913 // z = some value; x = phi (y, z); y = phi (x, z)
10914 // where the phi nodes don't necessarily need to be in the same block. Do a
10915 // quick check to see if the PHI node only contains a single non-phi value, if
10916 // so, scan to see if the phi cycle is actually equal to that value.
10918 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10919 // Scan for the first non-phi operand.
10920 while (InValNo != NumOperandVals &&
10921 isa<PHINode>(PN.getIncomingValue(InValNo)))
10924 if (InValNo != NumOperandVals) {
10925 Value *NonPhiInVal = PN.getOperand(InValNo);
10927 // Scan the rest of the operands to see if there are any conflicts, if so
10928 // there is no need to recursively scan other phis.
10929 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10930 Value *OpVal = PN.getIncomingValue(InValNo);
10931 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10935 // If we scanned over all operands, then we have one unique value plus
10936 // phi values. Scan PHI nodes to see if they all merge in each other or
10938 if (InValNo == NumOperandVals) {
10939 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10940 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10941 return ReplaceInstUsesWith(PN, NonPhiInVal);
10948 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10949 Value *PtrOp = GEP.getOperand(0);
10950 // Eliminate 'getelementptr %P, i32 0' and 'getelementptr %P', they are noops.
10951 if (GEP.getNumOperands() == 1)
10952 return ReplaceInstUsesWith(GEP, PtrOp);
10954 if (isa<UndefValue>(GEP.getOperand(0)))
10955 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10957 bool HasZeroPointerIndex = false;
10958 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10959 HasZeroPointerIndex = C->isNullValue();
10961 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10962 return ReplaceInstUsesWith(GEP, PtrOp);
10964 // Eliminate unneeded casts for indices.
10966 bool MadeChange = false;
10967 unsigned PtrSize = TD->getPointerSizeInBits();
10969 gep_type_iterator GTI = gep_type_begin(GEP);
10970 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
10971 I != E; ++I, ++GTI) {
10972 if (!isa<SequentialType>(*GTI)) continue;
10974 // If we are using a wider index than needed for this platform, shrink it
10975 // to what we need. If narrower, sign-extend it to what we need. This
10976 // explicit cast can make subsequent optimizations more obvious.
10977 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
10978 if (OpBits == PtrSize)
10981 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
10984 if (MadeChange) return &GEP;
10987 // Combine Indices - If the source pointer to this getelementptr instruction
10988 // is a getelementptr instruction, combine the indices of the two
10989 // getelementptr instructions into a single instruction.
10991 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
10992 // Note that if our source is a gep chain itself that we wait for that
10993 // chain to be resolved before we perform this transformation. This
10994 // avoids us creating a TON of code in some cases.
10996 if (GetElementPtrInst *SrcGEP =
10997 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
10998 if (SrcGEP->getNumOperands() == 2)
10999 return 0; // Wait until our source is folded to completion.
11001 SmallVector<Value*, 8> Indices;
11003 // Find out whether the last index in the source GEP is a sequential idx.
11004 bool EndsWithSequential = false;
11005 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
11007 EndsWithSequential = !isa<StructType>(*I);
11009 // Can we combine the two pointer arithmetics offsets?
11010 if (EndsWithSequential) {
11011 // Replace: gep (gep %P, long B), long A, ...
11012 // With: T = long A+B; gep %P, T, ...
11015 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
11016 Value *GO1 = GEP.getOperand(1);
11017 if (SO1 == Constant::getNullValue(SO1->getType())) {
11019 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
11022 // If they aren't the same type, then the input hasn't been processed
11023 // by the loop above yet (which canonicalizes sequential index types to
11024 // intptr_t). Just avoid transforming this until the input has been
11026 if (SO1->getType() != GO1->getType())
11028 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11031 // Update the GEP in place if possible.
11032 if (Src->getNumOperands() == 2) {
11033 GEP.setOperand(0, Src->getOperand(0));
11034 GEP.setOperand(1, Sum);
11037 Indices.append(Src->op_begin()+1, Src->op_end()-1);
11038 Indices.push_back(Sum);
11039 Indices.append(GEP.op_begin()+2, GEP.op_end());
11040 } else if (isa<Constant>(*GEP.idx_begin()) &&
11041 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11042 Src->getNumOperands() != 1) {
11043 // Otherwise we can do the fold if the first index of the GEP is a zero
11044 Indices.append(Src->op_begin()+1, Src->op_end());
11045 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
11048 if (!Indices.empty())
11049 return (cast<GEPOperator>(&GEP)->isInBounds() &&
11050 Src->isInBounds()) ?
11051 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
11052 Indices.end(), GEP.getName()) :
11053 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
11054 Indices.end(), GEP.getName());
11057 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
11058 if (Value *X = getBitCastOperand(PtrOp)) {
11059 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
11061 // If the input bitcast is actually "bitcast(bitcast(x))", then we don't
11062 // want to change the gep until the bitcasts are eliminated.
11063 if (getBitCastOperand(X)) {
11064 Worklist.AddValue(PtrOp);
11068 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11069 // into : GEP [10 x i8]* X, i32 0, ...
11071 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11072 // into : GEP i8* X, ...
11074 // This occurs when the program declares an array extern like "int X[];"
11075 if (HasZeroPointerIndex) {
11076 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11077 const PointerType *XTy = cast<PointerType>(X->getType());
11078 if (const ArrayType *CATy =
11079 dyn_cast<ArrayType>(CPTy->getElementType())) {
11080 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11081 if (CATy->getElementType() == XTy->getElementType()) {
11082 // -> GEP i8* X, ...
11083 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11084 return cast<GEPOperator>(&GEP)->isInBounds() ?
11085 GetElementPtrInst::CreateInBounds(X, Indices.begin(), Indices.end(),
11087 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11091 if (const ArrayType *XATy = dyn_cast<ArrayType>(XTy->getElementType())){
11092 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11093 if (CATy->getElementType() == XATy->getElementType()) {
11094 // -> GEP [10 x i8]* X, i32 0, ...
11095 // At this point, we know that the cast source type is a pointer
11096 // to an array of the same type as the destination pointer
11097 // array. Because the array type is never stepped over (there
11098 // is a leading zero) we can fold the cast into this GEP.
11099 GEP.setOperand(0, X);
11104 } else if (GEP.getNumOperands() == 2) {
11105 // Transform things like:
11106 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11107 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11108 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11109 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11110 if (TD && isa<ArrayType>(SrcElTy) &&
11111 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11112 TD->getTypeAllocSize(ResElTy)) {
11114 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11115 Idx[1] = GEP.getOperand(1);
11116 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11117 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11118 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11119 // V and GEP are both pointer types --> BitCast
11120 return new BitCastInst(NewGEP, GEP.getType());
11123 // Transform things like:
11124 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11125 // (where tmp = 8*tmp2) into:
11126 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11128 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
11129 uint64_t ArrayEltSize =
11130 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11132 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11133 // allow either a mul, shift, or constant here.
11135 ConstantInt *Scale = 0;
11136 if (ArrayEltSize == 1) {
11137 NewIdx = GEP.getOperand(1);
11138 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11139 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11140 NewIdx = ConstantInt::get(CI->getType(), 1);
11142 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11143 if (Inst->getOpcode() == Instruction::Shl &&
11144 isa<ConstantInt>(Inst->getOperand(1))) {
11145 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11146 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11147 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11149 NewIdx = Inst->getOperand(0);
11150 } else if (Inst->getOpcode() == Instruction::Mul &&
11151 isa<ConstantInt>(Inst->getOperand(1))) {
11152 Scale = cast<ConstantInt>(Inst->getOperand(1));
11153 NewIdx = Inst->getOperand(0);
11157 // If the index will be to exactly the right offset with the scale taken
11158 // out, perform the transformation. Note, we don't know whether Scale is
11159 // signed or not. We'll use unsigned version of division/modulo
11160 // operation after making sure Scale doesn't have the sign bit set.
11161 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11162 Scale->getZExtValue() % ArrayEltSize == 0) {
11163 Scale = ConstantInt::get(Scale->getType(),
11164 Scale->getZExtValue() / ArrayEltSize);
11165 if (Scale->getZExtValue() != 1) {
11166 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11168 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
11171 // Insert the new GEP instruction.
11173 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11175 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11176 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11177 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11178 // The NewGEP must be pointer typed, so must the old one -> BitCast
11179 return new BitCastInst(NewGEP, GEP.getType());
11185 /// See if we can simplify:
11186 /// X = bitcast A* to B*
11187 /// Y = gep X, <...constant indices...>
11188 /// into a gep of the original struct. This is important for SROA and alias
11189 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11190 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11192 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11193 // Determine how much the GEP moves the pointer. We are guaranteed to get
11194 // a constant back from EmitGEPOffset.
11195 ConstantInt *OffsetV =
11196 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11197 int64_t Offset = OffsetV->getSExtValue();
11199 // If this GEP instruction doesn't move the pointer, just replace the GEP
11200 // with a bitcast of the real input to the dest type.
11202 // If the bitcast is of an allocation, and the allocation will be
11203 // converted to match the type of the cast, don't touch this.
11204 if (isa<AllocaInst>(BCI->getOperand(0)) ||
11205 isMalloc(BCI->getOperand(0))) {
11206 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11207 if (Instruction *I = visitBitCast(*BCI)) {
11210 BCI->getParent()->getInstList().insert(BCI, I);
11211 ReplaceInstUsesWith(*BCI, I);
11216 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11219 // Otherwise, if the offset is non-zero, we need to find out if there is a
11220 // field at Offset in 'A's type. If so, we can pull the cast through the
11222 SmallVector<Value*, 8> NewIndices;
11224 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11225 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11226 Value *NGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11227 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
11228 NewIndices.end()) :
11229 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
11232 if (NGEP->getType() == GEP.getType())
11233 return ReplaceInstUsesWith(GEP, NGEP);
11234 NGEP->takeName(&GEP);
11235 return new BitCastInst(NGEP, GEP.getType());
11243 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
11244 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11245 if (AI.isArrayAllocation()) { // Check C != 1
11246 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11247 const Type *NewTy =
11248 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11249 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11250 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
11251 New->setAlignment(AI.getAlignment());
11253 // Scan to the end of the allocation instructions, to skip over a block of
11254 // allocas if possible...also skip interleaved debug info
11256 BasicBlock::iterator It = New;
11257 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11259 // Now that I is pointing to the first non-allocation-inst in the block,
11260 // insert our getelementptr instruction...
11262 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
11266 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
11267 New->getName()+".sub", It);
11269 // Now make everything use the getelementptr instead of the original
11271 return ReplaceInstUsesWith(AI, V);
11272 } else if (isa<UndefValue>(AI.getArraySize())) {
11273 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11277 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11278 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11279 // Note that we only do this for alloca's, because malloc should allocate
11280 // and return a unique pointer, even for a zero byte allocation.
11281 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11282 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11284 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11285 if (AI.getAlignment() == 0)
11286 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11292 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11293 Value *Op = FI.getOperand(0);
11295 // free undef -> unreachable.
11296 if (isa<UndefValue>(Op)) {
11297 // Insert a new store to null because we cannot modify the CFG here.
11298 new StoreInst(ConstantInt::getTrue(*Context),
11299 UndefValue::get(Type::getInt1PtrTy(*Context)), &FI);
11300 return EraseInstFromFunction(FI);
11303 // If we have 'free null' delete the instruction. This can happen in stl code
11304 // when lots of inlining happens.
11305 if (isa<ConstantPointerNull>(Op))
11306 return EraseInstFromFunction(FI);
11308 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11309 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11310 FI.setOperand(0, CI->getOperand(0));
11314 // Change free (gep X, 0,0,0,0) into free(X)
11315 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11316 if (GEPI->hasAllZeroIndices()) {
11317 Worklist.Add(GEPI);
11318 FI.setOperand(0, GEPI->getOperand(0));
11323 if (isMalloc(Op)) {
11324 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
11325 if (Op->hasOneUse() && CI->hasOneUse()) {
11326 EraseInstFromFunction(FI);
11327 EraseInstFromFunction(*CI);
11328 return EraseInstFromFunction(*cast<Instruction>(Op));
11331 // Op is a call to malloc
11332 if (Op->hasOneUse()) {
11333 EraseInstFromFunction(FI);
11334 return EraseInstFromFunction(*cast<Instruction>(Op));
11342 Instruction *InstCombiner::visitFree(Instruction &FI) {
11343 Value *Op = FI.getOperand(1);
11345 // free undef -> unreachable.
11346 if (isa<UndefValue>(Op)) {
11347 // Insert a new store to null because we cannot modify the CFG here.
11348 new StoreInst(ConstantInt::getTrue(*Context),
11349 UndefValue::get(Type::getInt1PtrTy(*Context)), &FI);
11350 return EraseInstFromFunction(FI);
11353 // If we have 'free null' delete the instruction. This can happen in stl code
11354 // when lots of inlining happens.
11355 if (isa<ConstantPointerNull>(Op))
11356 return EraseInstFromFunction(FI);
11358 // FIXME: Bring back free (gep X, 0,0,0,0) into free(X) transform
11360 if (isMalloc(Op)) {
11361 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
11362 if (Op->hasOneUse() && CI->hasOneUse()) {
11363 EraseInstFromFunction(FI);
11364 EraseInstFromFunction(*CI);
11365 return EraseInstFromFunction(*cast<Instruction>(Op));
11368 // Op is a call to malloc
11369 if (Op->hasOneUse()) {
11370 EraseInstFromFunction(FI);
11371 return EraseInstFromFunction(*cast<Instruction>(Op));
11379 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11380 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11381 const TargetData *TD) {
11382 User *CI = cast<User>(LI.getOperand(0));
11383 Value *CastOp = CI->getOperand(0);
11384 LLVMContext *Context = IC.getContext();
11386 const PointerType *DestTy = cast<PointerType>(CI->getType());
11387 const Type *DestPTy = DestTy->getElementType();
11388 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11390 // If the address spaces don't match, don't eliminate the cast.
11391 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11394 const Type *SrcPTy = SrcTy->getElementType();
11396 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11397 isa<VectorType>(DestPTy)) {
11398 // If the source is an array, the code below will not succeed. Check to
11399 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11401 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11402 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11403 if (ASrcTy->getNumElements() != 0) {
11405 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11407 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11408 SrcTy = cast<PointerType>(CastOp->getType());
11409 SrcPTy = SrcTy->getElementType();
11412 if (IC.getTargetData() &&
11413 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11414 isa<VectorType>(SrcPTy)) &&
11415 // Do not allow turning this into a load of an integer, which is then
11416 // casted to a pointer, this pessimizes pointer analysis a lot.
11417 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11418 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11419 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11421 // Okay, we are casting from one integer or pointer type to another of
11422 // the same size. Instead of casting the pointer before the load, cast
11423 // the result of the loaded value.
11425 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
11426 // Now cast the result of the load.
11427 return new BitCastInst(NewLoad, LI.getType());
11434 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11435 Value *Op = LI.getOperand(0);
11437 // Attempt to improve the alignment.
11439 unsigned KnownAlign =
11440 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11442 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11443 LI.getAlignment()))
11444 LI.setAlignment(KnownAlign);
11447 // load (cast X) --> cast (load X) iff safe.
11448 if (isa<CastInst>(Op))
11449 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11452 // None of the following transforms are legal for volatile loads.
11453 if (LI.isVolatile()) return 0;
11455 // Do really simple store-to-load forwarding and load CSE, to catch cases
11456 // where there are several consequtive memory accesses to the same location,
11457 // separated by a few arithmetic operations.
11458 BasicBlock::iterator BBI = &LI;
11459 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11460 return ReplaceInstUsesWith(LI, AvailableVal);
11462 // load(gep null, ...) -> unreachable
11463 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11464 const Value *GEPI0 = GEPI->getOperand(0);
11465 // TODO: Consider a target hook for valid address spaces for this xform.
11466 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
11467 // Insert a new store to null instruction before the load to indicate
11468 // that this code is not reachable. We do this instead of inserting
11469 // an unreachable instruction directly because we cannot modify the
11471 new StoreInst(UndefValue::get(LI.getType()),
11472 Constant::getNullValue(Op->getType()), &LI);
11473 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11477 // load null/undef -> unreachable
11478 // TODO: Consider a target hook for valid address spaces for this xform.
11479 if (isa<UndefValue>(Op) ||
11480 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
11481 // Insert a new store to null instruction before the load to indicate that
11482 // this code is not reachable. We do this instead of inserting an
11483 // unreachable instruction directly because we cannot modify the CFG.
11484 new StoreInst(UndefValue::get(LI.getType()),
11485 Constant::getNullValue(Op->getType()), &LI);
11486 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11489 // Instcombine load (constantexpr_cast global) -> cast (load global)
11490 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
11492 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11495 if (Op->hasOneUse()) {
11496 // Change select and PHI nodes to select values instead of addresses: this
11497 // helps alias analysis out a lot, allows many others simplifications, and
11498 // exposes redundancy in the code.
11500 // Note that we cannot do the transformation unless we know that the
11501 // introduced loads cannot trap! Something like this is valid as long as
11502 // the condition is always false: load (select bool %C, int* null, int* %G),
11503 // but it would not be valid if we transformed it to load from null
11504 // unconditionally.
11506 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11507 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11508 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11509 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11510 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
11511 SI->getOperand(1)->getName()+".val");
11512 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
11513 SI->getOperand(2)->getName()+".val");
11514 return SelectInst::Create(SI->getCondition(), V1, V2);
11517 // load (select (cond, null, P)) -> load P
11518 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11519 if (C->isNullValue()) {
11520 LI.setOperand(0, SI->getOperand(2));
11524 // load (select (cond, P, null)) -> load P
11525 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11526 if (C->isNullValue()) {
11527 LI.setOperand(0, SI->getOperand(1));
11535 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11536 /// when possible. This makes it generally easy to do alias analysis and/or
11537 /// SROA/mem2reg of the memory object.
11538 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11539 User *CI = cast<User>(SI.getOperand(1));
11540 Value *CastOp = CI->getOperand(0);
11542 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11543 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11544 if (SrcTy == 0) return 0;
11546 const Type *SrcPTy = SrcTy->getElementType();
11548 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11551 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11552 /// to its first element. This allows us to handle things like:
11553 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11554 /// on 32-bit hosts.
11555 SmallVector<Value*, 4> NewGEPIndices;
11557 // If the source is an array, the code below will not succeed. Check to
11558 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11560 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11561 // Index through pointer.
11562 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
11563 NewGEPIndices.push_back(Zero);
11566 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11567 if (!STy->getNumElements()) /* Struct can be empty {} */
11569 NewGEPIndices.push_back(Zero);
11570 SrcPTy = STy->getElementType(0);
11571 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11572 NewGEPIndices.push_back(Zero);
11573 SrcPTy = ATy->getElementType();
11579 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11582 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11585 // If the pointers point into different address spaces or if they point to
11586 // values with different sizes, we can't do the transformation.
11587 if (!IC.getTargetData() ||
11588 SrcTy->getAddressSpace() !=
11589 cast<PointerType>(CI->getType())->getAddressSpace() ||
11590 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11591 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11594 // Okay, we are casting from one integer or pointer type to another of
11595 // the same size. Instead of casting the pointer before
11596 // the store, cast the value to be stored.
11598 Value *SIOp0 = SI.getOperand(0);
11599 Instruction::CastOps opcode = Instruction::BitCast;
11600 const Type* CastSrcTy = SIOp0->getType();
11601 const Type* CastDstTy = SrcPTy;
11602 if (isa<PointerType>(CastDstTy)) {
11603 if (CastSrcTy->isInteger())
11604 opcode = Instruction::IntToPtr;
11605 } else if (isa<IntegerType>(CastDstTy)) {
11606 if (isa<PointerType>(SIOp0->getType()))
11607 opcode = Instruction::PtrToInt;
11610 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11611 // emit a GEP to index into its first field.
11612 if (!NewGEPIndices.empty())
11613 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
11614 NewGEPIndices.end());
11616 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
11617 SIOp0->getName()+".c");
11618 return new StoreInst(NewCast, CastOp);
11621 /// equivalentAddressValues - Test if A and B will obviously have the same
11622 /// value. This includes recognizing that %t0 and %t1 will have the same
11623 /// value in code like this:
11624 /// %t0 = getelementptr \@a, 0, 3
11625 /// store i32 0, i32* %t0
11626 /// %t1 = getelementptr \@a, 0, 3
11627 /// %t2 = load i32* %t1
11629 static bool equivalentAddressValues(Value *A, Value *B) {
11630 // Test if the values are trivially equivalent.
11631 if (A == B) return true;
11633 // Test if the values come form identical arithmetic instructions.
11634 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
11635 // its only used to compare two uses within the same basic block, which
11636 // means that they'll always either have the same value or one of them
11637 // will have an undefined value.
11638 if (isa<BinaryOperator>(A) ||
11639 isa<CastInst>(A) ||
11641 isa<GetElementPtrInst>(A))
11642 if (Instruction *BI = dyn_cast<Instruction>(B))
11643 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
11646 // Otherwise they may not be equivalent.
11650 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11651 // return the llvm.dbg.declare.
11652 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11653 if (!V->hasNUses(2))
11655 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11657 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11659 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11660 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11667 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11668 Value *Val = SI.getOperand(0);
11669 Value *Ptr = SI.getOperand(1);
11671 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11672 EraseInstFromFunction(SI);
11677 // If the RHS is an alloca with a single use, zapify the store, making the
11679 // If the RHS is an alloca with a two uses, the other one being a
11680 // llvm.dbg.declare, zapify the store and the declare, making the
11681 // alloca dead. We must do this to prevent declare's from affecting
11683 if (!SI.isVolatile()) {
11684 if (Ptr->hasOneUse()) {
11685 if (isa<AllocaInst>(Ptr)) {
11686 EraseInstFromFunction(SI);
11690 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11691 if (isa<AllocaInst>(GEP->getOperand(0))) {
11692 if (GEP->getOperand(0)->hasOneUse()) {
11693 EraseInstFromFunction(SI);
11697 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11698 EraseInstFromFunction(*DI);
11699 EraseInstFromFunction(SI);
11706 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11707 EraseInstFromFunction(*DI);
11708 EraseInstFromFunction(SI);
11714 // Attempt to improve the alignment.
11716 unsigned KnownAlign =
11717 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11719 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11720 SI.getAlignment()))
11721 SI.setAlignment(KnownAlign);
11724 // Do really simple DSE, to catch cases where there are several consecutive
11725 // stores to the same location, separated by a few arithmetic operations. This
11726 // situation often occurs with bitfield accesses.
11727 BasicBlock::iterator BBI = &SI;
11728 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11731 // Don't count debug info directives, lest they affect codegen,
11732 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11733 // It is necessary for correctness to skip those that feed into a
11734 // llvm.dbg.declare, as these are not present when debugging is off.
11735 if (isa<DbgInfoIntrinsic>(BBI) ||
11736 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11741 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11742 // Prev store isn't volatile, and stores to the same location?
11743 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11744 SI.getOperand(1))) {
11747 EraseInstFromFunction(*PrevSI);
11753 // If this is a load, we have to stop. However, if the loaded value is from
11754 // the pointer we're loading and is producing the pointer we're storing,
11755 // then *this* store is dead (X = load P; store X -> P).
11756 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11757 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11758 !SI.isVolatile()) {
11759 EraseInstFromFunction(SI);
11763 // Otherwise, this is a load from some other location. Stores before it
11764 // may not be dead.
11768 // Don't skip over loads or things that can modify memory.
11769 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11774 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11776 // store X, null -> turns into 'unreachable' in SimplifyCFG
11777 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
11778 if (!isa<UndefValue>(Val)) {
11779 SI.setOperand(0, UndefValue::get(Val->getType()));
11780 if (Instruction *U = dyn_cast<Instruction>(Val))
11781 Worklist.Add(U); // Dropped a use.
11784 return 0; // Do not modify these!
11787 // store undef, Ptr -> noop
11788 if (isa<UndefValue>(Val)) {
11789 EraseInstFromFunction(SI);
11794 // If the pointer destination is a cast, see if we can fold the cast into the
11796 if (isa<CastInst>(Ptr))
11797 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11799 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11801 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11805 // If this store is the last instruction in the basic block (possibly
11806 // excepting debug info instructions and the pointer bitcasts that feed
11807 // into them), and if the block ends with an unconditional branch, try
11808 // to move it to the successor block.
11812 } while (isa<DbgInfoIntrinsic>(BBI) ||
11813 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11814 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11815 if (BI->isUnconditional())
11816 if (SimplifyStoreAtEndOfBlock(SI))
11817 return 0; // xform done!
11822 /// SimplifyStoreAtEndOfBlock - Turn things like:
11823 /// if () { *P = v1; } else { *P = v2 }
11824 /// into a phi node with a store in the successor.
11826 /// Simplify things like:
11827 /// *P = v1; if () { *P = v2; }
11828 /// into a phi node with a store in the successor.
11830 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11831 BasicBlock *StoreBB = SI.getParent();
11833 // Check to see if the successor block has exactly two incoming edges. If
11834 // so, see if the other predecessor contains a store to the same location.
11835 // if so, insert a PHI node (if needed) and move the stores down.
11836 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11838 // Determine whether Dest has exactly two predecessors and, if so, compute
11839 // the other predecessor.
11840 pred_iterator PI = pred_begin(DestBB);
11841 BasicBlock *OtherBB = 0;
11842 if (*PI != StoreBB)
11845 if (PI == pred_end(DestBB))
11848 if (*PI != StoreBB) {
11853 if (++PI != pred_end(DestBB))
11856 // Bail out if all the relevant blocks aren't distinct (this can happen,
11857 // for example, if SI is in an infinite loop)
11858 if (StoreBB == DestBB || OtherBB == DestBB)
11861 // Verify that the other block ends in a branch and is not otherwise empty.
11862 BasicBlock::iterator BBI = OtherBB->getTerminator();
11863 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11864 if (!OtherBr || BBI == OtherBB->begin())
11867 // If the other block ends in an unconditional branch, check for the 'if then
11868 // else' case. there is an instruction before the branch.
11869 StoreInst *OtherStore = 0;
11870 if (OtherBr->isUnconditional()) {
11872 // Skip over debugging info.
11873 while (isa<DbgInfoIntrinsic>(BBI) ||
11874 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11875 if (BBI==OtherBB->begin())
11879 // If this isn't a store, or isn't a store to the same location, bail out.
11880 OtherStore = dyn_cast<StoreInst>(BBI);
11881 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11884 // Otherwise, the other block ended with a conditional branch. If one of the
11885 // destinations is StoreBB, then we have the if/then case.
11886 if (OtherBr->getSuccessor(0) != StoreBB &&
11887 OtherBr->getSuccessor(1) != StoreBB)
11890 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11891 // if/then triangle. See if there is a store to the same ptr as SI that
11892 // lives in OtherBB.
11894 // Check to see if we find the matching store.
11895 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11896 if (OtherStore->getOperand(1) != SI.getOperand(1))
11900 // If we find something that may be using or overwriting the stored
11901 // value, or if we run out of instructions, we can't do the xform.
11902 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11903 BBI == OtherBB->begin())
11907 // In order to eliminate the store in OtherBr, we have to
11908 // make sure nothing reads or overwrites the stored value in
11910 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11911 // FIXME: This should really be AA driven.
11912 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11917 // Insert a PHI node now if we need it.
11918 Value *MergedVal = OtherStore->getOperand(0);
11919 if (MergedVal != SI.getOperand(0)) {
11920 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11921 PN->reserveOperandSpace(2);
11922 PN->addIncoming(SI.getOperand(0), SI.getParent());
11923 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11924 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11927 // Advance to a place where it is safe to insert the new store and
11929 BBI = DestBB->getFirstNonPHI();
11930 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11931 OtherStore->isVolatile()), *BBI);
11933 // Nuke the old stores.
11934 EraseInstFromFunction(SI);
11935 EraseInstFromFunction(*OtherStore);
11941 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11942 // Change br (not X), label True, label False to: br X, label False, True
11944 BasicBlock *TrueDest;
11945 BasicBlock *FalseDest;
11946 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11947 !isa<Constant>(X)) {
11948 // Swap Destinations and condition...
11949 BI.setCondition(X);
11950 BI.setSuccessor(0, FalseDest);
11951 BI.setSuccessor(1, TrueDest);
11955 // Cannonicalize fcmp_one -> fcmp_oeq
11956 FCmpInst::Predicate FPred; Value *Y;
11957 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11958 TrueDest, FalseDest)) &&
11959 BI.getCondition()->hasOneUse())
11960 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11961 FPred == FCmpInst::FCMP_OGE) {
11962 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
11963 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
11965 // Swap Destinations and condition.
11966 BI.setSuccessor(0, FalseDest);
11967 BI.setSuccessor(1, TrueDest);
11968 Worklist.Add(Cond);
11972 // Cannonicalize icmp_ne -> icmp_eq
11973 ICmpInst::Predicate IPred;
11974 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11975 TrueDest, FalseDest)) &&
11976 BI.getCondition()->hasOneUse())
11977 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11978 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11979 IPred == ICmpInst::ICMP_SGE) {
11980 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
11981 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
11982 // Swap Destinations and condition.
11983 BI.setSuccessor(0, FalseDest);
11984 BI.setSuccessor(1, TrueDest);
11985 Worklist.Add(Cond);
11992 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11993 Value *Cond = SI.getCondition();
11994 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11995 if (I->getOpcode() == Instruction::Add)
11996 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11997 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11998 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12000 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12002 SI.setOperand(0, I->getOperand(0));
12010 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12011 Value *Agg = EV.getAggregateOperand();
12013 if (!EV.hasIndices())
12014 return ReplaceInstUsesWith(EV, Agg);
12016 if (Constant *C = dyn_cast<Constant>(Agg)) {
12017 if (isa<UndefValue>(C))
12018 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12020 if (isa<ConstantAggregateZero>(C))
12021 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12023 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12024 // Extract the element indexed by the first index out of the constant
12025 Value *V = C->getOperand(*EV.idx_begin());
12026 if (EV.getNumIndices() > 1)
12027 // Extract the remaining indices out of the constant indexed by the
12029 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12031 return ReplaceInstUsesWith(EV, V);
12033 return 0; // Can't handle other constants
12035 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12036 // We're extracting from an insertvalue instruction, compare the indices
12037 const unsigned *exti, *exte, *insi, *inse;
12038 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12039 exte = EV.idx_end(), inse = IV->idx_end();
12040 exti != exte && insi != inse;
12042 if (*insi != *exti)
12043 // The insert and extract both reference distinctly different elements.
12044 // This means the extract is not influenced by the insert, and we can
12045 // replace the aggregate operand of the extract with the aggregate
12046 // operand of the insert. i.e., replace
12047 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12048 // %E = extractvalue { i32, { i32 } } %I, 0
12050 // %E = extractvalue { i32, { i32 } } %A, 0
12051 return ExtractValueInst::Create(IV->getAggregateOperand(),
12052 EV.idx_begin(), EV.idx_end());
12054 if (exti == exte && insi == inse)
12055 // Both iterators are at the end: Index lists are identical. Replace
12056 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12057 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12059 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12060 if (exti == exte) {
12061 // The extract list is a prefix of the insert list. i.e. replace
12062 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12063 // %E = extractvalue { i32, { i32 } } %I, 1
12065 // %X = extractvalue { i32, { i32 } } %A, 1
12066 // %E = insertvalue { i32 } %X, i32 42, 0
12067 // by switching the order of the insert and extract (though the
12068 // insertvalue should be left in, since it may have other uses).
12069 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
12070 EV.idx_begin(), EV.idx_end());
12071 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12075 // The insert list is a prefix of the extract list
12076 // We can simply remove the common indices from the extract and make it
12077 // operate on the inserted value instead of the insertvalue result.
12079 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12080 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12082 // %E extractvalue { i32 } { i32 42 }, 0
12083 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12086 // Can't simplify extracts from other values. Note that nested extracts are
12087 // already simplified implicitely by the above (extract ( extract (insert) )
12088 // will be translated into extract ( insert ( extract ) ) first and then just
12089 // the value inserted, if appropriate).
12093 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12094 /// is to leave as a vector operation.
12095 static bool CheapToScalarize(Value *V, bool isConstant) {
12096 if (isa<ConstantAggregateZero>(V))
12098 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12099 if (isConstant) return true;
12100 // If all elts are the same, we can extract.
12101 Constant *Op0 = C->getOperand(0);
12102 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12103 if (C->getOperand(i) != Op0)
12107 Instruction *I = dyn_cast<Instruction>(V);
12108 if (!I) return false;
12110 // Insert element gets simplified to the inserted element or is deleted if
12111 // this is constant idx extract element and its a constant idx insertelt.
12112 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12113 isa<ConstantInt>(I->getOperand(2)))
12115 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12117 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12118 if (BO->hasOneUse() &&
12119 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12120 CheapToScalarize(BO->getOperand(1), isConstant)))
12122 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12123 if (CI->hasOneUse() &&
12124 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12125 CheapToScalarize(CI->getOperand(1), isConstant)))
12131 /// Read and decode a shufflevector mask.
12133 /// It turns undef elements into values that are larger than the number of
12134 /// elements in the input.
12135 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12136 unsigned NElts = SVI->getType()->getNumElements();
12137 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12138 return std::vector<unsigned>(NElts, 0);
12139 if (isa<UndefValue>(SVI->getOperand(2)))
12140 return std::vector<unsigned>(NElts, 2*NElts);
12142 std::vector<unsigned> Result;
12143 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12144 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12145 if (isa<UndefValue>(*i))
12146 Result.push_back(NElts*2); // undef -> 8
12148 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12152 /// FindScalarElement - Given a vector and an element number, see if the scalar
12153 /// value is already around as a register, for example if it were inserted then
12154 /// extracted from the vector.
12155 static Value *FindScalarElement(Value *V, unsigned EltNo,
12156 LLVMContext *Context) {
12157 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12158 const VectorType *PTy = cast<VectorType>(V->getType());
12159 unsigned Width = PTy->getNumElements();
12160 if (EltNo >= Width) // Out of range access.
12161 return UndefValue::get(PTy->getElementType());
12163 if (isa<UndefValue>(V))
12164 return UndefValue::get(PTy->getElementType());
12165 else if (isa<ConstantAggregateZero>(V))
12166 return Constant::getNullValue(PTy->getElementType());
12167 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12168 return CP->getOperand(EltNo);
12169 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12170 // If this is an insert to a variable element, we don't know what it is.
12171 if (!isa<ConstantInt>(III->getOperand(2)))
12173 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12175 // If this is an insert to the element we are looking for, return the
12177 if (EltNo == IIElt)
12178 return III->getOperand(1);
12180 // Otherwise, the insertelement doesn't modify the value, recurse on its
12182 return FindScalarElement(III->getOperand(0), EltNo, Context);
12183 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12184 unsigned LHSWidth =
12185 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12186 unsigned InEl = getShuffleMask(SVI)[EltNo];
12187 if (InEl < LHSWidth)
12188 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12189 else if (InEl < LHSWidth*2)
12190 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12192 return UndefValue::get(PTy->getElementType());
12195 // Otherwise, we don't know.
12199 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12200 // If vector val is undef, replace extract with scalar undef.
12201 if (isa<UndefValue>(EI.getOperand(0)))
12202 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12204 // If vector val is constant 0, replace extract with scalar 0.
12205 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12206 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12208 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12209 // If vector val is constant with all elements the same, replace EI with
12210 // that element. When the elements are not identical, we cannot replace yet
12211 // (we do that below, but only when the index is constant).
12212 Constant *op0 = C->getOperand(0);
12213 for (unsigned i = 1; i != C->getNumOperands(); ++i)
12214 if (C->getOperand(i) != op0) {
12219 return ReplaceInstUsesWith(EI, op0);
12222 // If extracting a specified index from the vector, see if we can recursively
12223 // find a previously computed scalar that was inserted into the vector.
12224 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12225 unsigned IndexVal = IdxC->getZExtValue();
12226 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
12228 // If this is extracting an invalid index, turn this into undef, to avoid
12229 // crashing the code below.
12230 if (IndexVal >= VectorWidth)
12231 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12233 // This instruction only demands the single element from the input vector.
12234 // If the input vector has a single use, simplify it based on this use
12236 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12237 APInt UndefElts(VectorWidth, 0);
12238 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12239 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12240 DemandedMask, UndefElts)) {
12241 EI.setOperand(0, V);
12246 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12247 return ReplaceInstUsesWith(EI, Elt);
12249 // If the this extractelement is directly using a bitcast from a vector of
12250 // the same number of elements, see if we can find the source element from
12251 // it. In this case, we will end up needing to bitcast the scalars.
12252 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12253 if (const VectorType *VT =
12254 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12255 if (VT->getNumElements() == VectorWidth)
12256 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12257 IndexVal, Context))
12258 return new BitCastInst(Elt, EI.getType());
12262 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12263 // Push extractelement into predecessor operation if legal and
12264 // profitable to do so
12265 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12266 if (I->hasOneUse() &&
12267 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
12269 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
12270 EI.getName()+".lhs");
12272 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
12273 EI.getName()+".rhs");
12274 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12276 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12277 // Extracting the inserted element?
12278 if (IE->getOperand(2) == EI.getOperand(1))
12279 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12280 // If the inserted and extracted elements are constants, they must not
12281 // be the same value, extract from the pre-inserted value instead.
12282 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
12283 Worklist.AddValue(EI.getOperand(0));
12284 EI.setOperand(0, IE->getOperand(0));
12287 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12288 // If this is extracting an element from a shufflevector, figure out where
12289 // it came from and extract from the appropriate input element instead.
12290 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12291 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12293 unsigned LHSWidth =
12294 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12296 if (SrcIdx < LHSWidth)
12297 Src = SVI->getOperand(0);
12298 else if (SrcIdx < LHSWidth*2) {
12299 SrcIdx -= LHSWidth;
12300 Src = SVI->getOperand(1);
12302 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12304 return ExtractElementInst::Create(Src,
12305 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx,
12309 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12314 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12315 /// elements from either LHS or RHS, return the shuffle mask and true.
12316 /// Otherwise, return false.
12317 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12318 std::vector<Constant*> &Mask,
12319 LLVMContext *Context) {
12320 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12321 "Invalid CollectSingleShuffleElements");
12322 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12324 if (isa<UndefValue>(V)) {
12325 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12327 } else if (V == LHS) {
12328 for (unsigned i = 0; i != NumElts; ++i)
12329 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12331 } else if (V == RHS) {
12332 for (unsigned i = 0; i != NumElts; ++i)
12333 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
12335 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12336 // If this is an insert of an extract from some other vector, include it.
12337 Value *VecOp = IEI->getOperand(0);
12338 Value *ScalarOp = IEI->getOperand(1);
12339 Value *IdxOp = IEI->getOperand(2);
12341 if (!isa<ConstantInt>(IdxOp))
12343 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12345 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12346 // Okay, we can handle this if the vector we are insertinting into is
12347 // transitively ok.
12348 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12349 // If so, update the mask to reflect the inserted undef.
12350 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
12353 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12354 if (isa<ConstantInt>(EI->getOperand(1)) &&
12355 EI->getOperand(0)->getType() == V->getType()) {
12356 unsigned ExtractedIdx =
12357 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12359 // This must be extracting from either LHS or RHS.
12360 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12361 // Okay, we can handle this if the vector we are insertinting into is
12362 // transitively ok.
12363 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12364 // If so, update the mask to reflect the inserted value.
12365 if (EI->getOperand(0) == LHS) {
12366 Mask[InsertedIdx % NumElts] =
12367 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12369 assert(EI->getOperand(0) == RHS);
12370 Mask[InsertedIdx % NumElts] =
12371 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
12380 // TODO: Handle shufflevector here!
12385 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12386 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12387 /// that computes V and the LHS value of the shuffle.
12388 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12389 Value *&RHS, LLVMContext *Context) {
12390 assert(isa<VectorType>(V->getType()) &&
12391 (RHS == 0 || V->getType() == RHS->getType()) &&
12392 "Invalid shuffle!");
12393 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12395 if (isa<UndefValue>(V)) {
12396 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12398 } else if (isa<ConstantAggregateZero>(V)) {
12399 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
12401 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12402 // If this is an insert of an extract from some other vector, include it.
12403 Value *VecOp = IEI->getOperand(0);
12404 Value *ScalarOp = IEI->getOperand(1);
12405 Value *IdxOp = IEI->getOperand(2);
12407 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12408 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12409 EI->getOperand(0)->getType() == V->getType()) {
12410 unsigned ExtractedIdx =
12411 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12412 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12414 // Either the extracted from or inserted into vector must be RHSVec,
12415 // otherwise we'd end up with a shuffle of three inputs.
12416 if (EI->getOperand(0) == RHS || RHS == 0) {
12417 RHS = EI->getOperand(0);
12418 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12419 Mask[InsertedIdx % NumElts] =
12420 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
12424 if (VecOp == RHS) {
12425 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12427 // Everything but the extracted element is replaced with the RHS.
12428 for (unsigned i = 0; i != NumElts; ++i) {
12429 if (i != InsertedIdx)
12430 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
12435 // If this insertelement is a chain that comes from exactly these two
12436 // vectors, return the vector and the effective shuffle.
12437 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12439 return EI->getOperand(0);
12444 // TODO: Handle shufflevector here!
12446 // Otherwise, can't do anything fancy. Return an identity vector.
12447 for (unsigned i = 0; i != NumElts; ++i)
12448 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12452 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12453 Value *VecOp = IE.getOperand(0);
12454 Value *ScalarOp = IE.getOperand(1);
12455 Value *IdxOp = IE.getOperand(2);
12457 // Inserting an undef or into an undefined place, remove this.
12458 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12459 ReplaceInstUsesWith(IE, VecOp);
12461 // If the inserted element was extracted from some other vector, and if the
12462 // indexes are constant, try to turn this into a shufflevector operation.
12463 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12464 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12465 EI->getOperand(0)->getType() == IE.getType()) {
12466 unsigned NumVectorElts = IE.getType()->getNumElements();
12467 unsigned ExtractedIdx =
12468 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12469 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12471 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12472 return ReplaceInstUsesWith(IE, VecOp);
12474 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12475 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12477 // If we are extracting a value from a vector, then inserting it right
12478 // back into the same place, just use the input vector.
12479 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12480 return ReplaceInstUsesWith(IE, VecOp);
12482 // If this insertelement isn't used by some other insertelement, turn it
12483 // (and any insertelements it points to), into one big shuffle.
12484 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12485 std::vector<Constant*> Mask;
12487 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12488 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12489 // We now have a shuffle of LHS, RHS, Mask.
12490 return new ShuffleVectorInst(LHS, RHS,
12491 ConstantVector::get(Mask));
12496 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12497 APInt UndefElts(VWidth, 0);
12498 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12499 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12506 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12507 Value *LHS = SVI.getOperand(0);
12508 Value *RHS = SVI.getOperand(1);
12509 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12511 bool MadeChange = false;
12513 // Undefined shuffle mask -> undefined value.
12514 if (isa<UndefValue>(SVI.getOperand(2)))
12515 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12517 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12519 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12522 APInt UndefElts(VWidth, 0);
12523 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12524 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12525 LHS = SVI.getOperand(0);
12526 RHS = SVI.getOperand(1);
12530 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12531 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12532 if (LHS == RHS || isa<UndefValue>(LHS)) {
12533 if (isa<UndefValue>(LHS) && LHS == RHS) {
12534 // shuffle(undef,undef,mask) -> undef.
12535 return ReplaceInstUsesWith(SVI, LHS);
12538 // Remap any references to RHS to use LHS.
12539 std::vector<Constant*> Elts;
12540 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12541 if (Mask[i] >= 2*e)
12542 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12544 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12545 (Mask[i] < e && isa<UndefValue>(LHS))) {
12546 Mask[i] = 2*e; // Turn into undef.
12547 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12549 Mask[i] = Mask[i] % e; // Force to LHS.
12550 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
12554 SVI.setOperand(0, SVI.getOperand(1));
12555 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12556 SVI.setOperand(2, ConstantVector::get(Elts));
12557 LHS = SVI.getOperand(0);
12558 RHS = SVI.getOperand(1);
12562 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12563 bool isLHSID = true, isRHSID = true;
12565 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12566 if (Mask[i] >= e*2) continue; // Ignore undef values.
12567 // Is this an identity shuffle of the LHS value?
12568 isLHSID &= (Mask[i] == i);
12570 // Is this an identity shuffle of the RHS value?
12571 isRHSID &= (Mask[i]-e == i);
12574 // Eliminate identity shuffles.
12575 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12576 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12578 // If the LHS is a shufflevector itself, see if we can combine it with this
12579 // one without producing an unusual shuffle. Here we are really conservative:
12580 // we are absolutely afraid of producing a shuffle mask not in the input
12581 // program, because the code gen may not be smart enough to turn a merged
12582 // shuffle into two specific shuffles: it may produce worse code. As such,
12583 // we only merge two shuffles if the result is one of the two input shuffle
12584 // masks. In this case, merging the shuffles just removes one instruction,
12585 // which we know is safe. This is good for things like turning:
12586 // (splat(splat)) -> splat.
12587 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12588 if (isa<UndefValue>(RHS)) {
12589 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12591 std::vector<unsigned> NewMask;
12592 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12593 if (Mask[i] >= 2*e)
12594 NewMask.push_back(2*e);
12596 NewMask.push_back(LHSMask[Mask[i]]);
12598 // If the result mask is equal to the src shuffle or this shuffle mask, do
12599 // the replacement.
12600 if (NewMask == LHSMask || NewMask == Mask) {
12601 unsigned LHSInNElts =
12602 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12603 std::vector<Constant*> Elts;
12604 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12605 if (NewMask[i] >= LHSInNElts*2) {
12606 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12608 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), NewMask[i]));
12611 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12612 LHSSVI->getOperand(1),
12613 ConstantVector::get(Elts));
12618 return MadeChange ? &SVI : 0;
12624 /// TryToSinkInstruction - Try to move the specified instruction from its
12625 /// current block into the beginning of DestBlock, which can only happen if it's
12626 /// safe to move the instruction past all of the instructions between it and the
12627 /// end of its block.
12628 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12629 assert(I->hasOneUse() && "Invariants didn't hold!");
12631 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12632 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12635 // Do not sink alloca instructions out of the entry block.
12636 if (isa<AllocaInst>(I) && I->getParent() ==
12637 &DestBlock->getParent()->getEntryBlock())
12640 // We can only sink load instructions if there is nothing between the load and
12641 // the end of block that could change the value.
12642 if (I->mayReadFromMemory()) {
12643 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12645 if (Scan->mayWriteToMemory())
12649 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12651 CopyPrecedingStopPoint(I, InsertPos);
12652 I->moveBefore(InsertPos);
12658 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12659 /// all reachable code to the worklist.
12661 /// This has a couple of tricks to make the code faster and more powerful. In
12662 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12663 /// them to the worklist (this significantly speeds up instcombine on code where
12664 /// many instructions are dead or constant). Additionally, if we find a branch
12665 /// whose condition is a known constant, we only visit the reachable successors.
12667 static bool AddReachableCodeToWorklist(BasicBlock *BB,
12668 SmallPtrSet<BasicBlock*, 64> &Visited,
12670 const TargetData *TD) {
12671 bool MadeIRChange = false;
12672 SmallVector<BasicBlock*, 256> Worklist;
12673 Worklist.push_back(BB);
12675 std::vector<Instruction*> InstrsForInstCombineWorklist;
12676 InstrsForInstCombineWorklist.reserve(128);
12678 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
12680 while (!Worklist.empty()) {
12681 BB = Worklist.back();
12682 Worklist.pop_back();
12684 // We have now visited this block! If we've already been here, ignore it.
12685 if (!Visited.insert(BB)) continue;
12687 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12688 Instruction *Inst = BBI++;
12690 // DCE instruction if trivially dead.
12691 if (isInstructionTriviallyDead(Inst)) {
12693 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
12694 Inst->eraseFromParent();
12698 // ConstantProp instruction if trivially constant.
12699 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
12700 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12701 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
12703 Inst->replaceAllUsesWith(C);
12705 Inst->eraseFromParent();
12712 // See if we can constant fold its operands.
12713 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
12715 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
12716 if (CE == 0) continue;
12718 // If we already folded this constant, don't try again.
12719 if (!FoldedConstants.insert(CE))
12723 ConstantFoldConstantExpression(CE, BB->getContext(), TD);
12724 if (NewC && NewC != CE) {
12726 MadeIRChange = true;
12732 InstrsForInstCombineWorklist.push_back(Inst);
12735 // Recursively visit successors. If this is a branch or switch on a
12736 // constant, only visit the reachable successor.
12737 TerminatorInst *TI = BB->getTerminator();
12738 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12739 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12740 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12741 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12742 Worklist.push_back(ReachableBB);
12745 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12746 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12747 // See if this is an explicit destination.
12748 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12749 if (SI->getCaseValue(i) == Cond) {
12750 BasicBlock *ReachableBB = SI->getSuccessor(i);
12751 Worklist.push_back(ReachableBB);
12755 // Otherwise it is the default destination.
12756 Worklist.push_back(SI->getSuccessor(0));
12761 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12762 Worklist.push_back(TI->getSuccessor(i));
12765 // Once we've found all of the instructions to add to instcombine's worklist,
12766 // add them in reverse order. This way instcombine will visit from the top
12767 // of the function down. This jives well with the way that it adds all uses
12768 // of instructions to the worklist after doing a transformation, thus avoiding
12769 // some N^2 behavior in pathological cases.
12770 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
12771 InstrsForInstCombineWorklist.size());
12773 return MadeIRChange;
12776 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12777 MadeIRChange = false;
12779 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12780 << F.getNameStr() << "\n");
12783 // Do a depth-first traversal of the function, populate the worklist with
12784 // the reachable instructions. Ignore blocks that are not reachable. Keep
12785 // track of which blocks we visit.
12786 SmallPtrSet<BasicBlock*, 64> Visited;
12787 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12789 // Do a quick scan over the function. If we find any blocks that are
12790 // unreachable, remove any instructions inside of them. This prevents
12791 // the instcombine code from having to deal with some bad special cases.
12792 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12793 if (!Visited.count(BB)) {
12794 Instruction *Term = BB->getTerminator();
12795 while (Term != BB->begin()) { // Remove instrs bottom-up
12796 BasicBlock::iterator I = Term; --I;
12798 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12799 // A debug intrinsic shouldn't force another iteration if we weren't
12800 // going to do one without it.
12801 if (!isa<DbgInfoIntrinsic>(I)) {
12803 MadeIRChange = true;
12806 // If I is not void type then replaceAllUsesWith undef.
12807 // This allows ValueHandlers and custom metadata to adjust itself.
12808 if (!I->getType()->isVoidTy())
12809 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12810 I->eraseFromParent();
12815 while (!Worklist.isEmpty()) {
12816 Instruction *I = Worklist.RemoveOne();
12817 if (I == 0) continue; // skip null values.
12819 // Check to see if we can DCE the instruction.
12820 if (isInstructionTriviallyDead(I)) {
12821 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12822 EraseInstFromFunction(*I);
12824 MadeIRChange = true;
12828 // Instruction isn't dead, see if we can constant propagate it.
12829 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
12830 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
12831 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
12833 // Add operands to the worklist.
12834 ReplaceInstUsesWith(*I, C);
12836 EraseInstFromFunction(*I);
12837 MadeIRChange = true;
12841 // See if we can trivially sink this instruction to a successor basic block.
12842 if (I->hasOneUse()) {
12843 BasicBlock *BB = I->getParent();
12844 Instruction *UserInst = cast<Instruction>(I->use_back());
12845 BasicBlock *UserParent;
12847 // Get the block the use occurs in.
12848 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
12849 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
12851 UserParent = UserInst->getParent();
12853 if (UserParent != BB) {
12854 bool UserIsSuccessor = false;
12855 // See if the user is one of our successors.
12856 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12857 if (*SI == UserParent) {
12858 UserIsSuccessor = true;
12862 // If the user is one of our immediate successors, and if that successor
12863 // only has us as a predecessors (we'd have to split the critical edge
12864 // otherwise), we can keep going.
12865 if (UserIsSuccessor && UserParent->getSinglePredecessor())
12866 // Okay, the CFG is simple enough, try to sink this instruction.
12867 MadeIRChange |= TryToSinkInstruction(I, UserParent);
12871 // Now that we have an instruction, try combining it to simplify it.
12872 Builder->SetInsertPoint(I->getParent(), I);
12877 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
12878 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
12880 if (Instruction *Result = visit(*I)) {
12882 // Should we replace the old instruction with a new one?
12884 DEBUG(errs() << "IC: Old = " << *I << '\n'
12885 << " New = " << *Result << '\n');
12887 // Everything uses the new instruction now.
12888 I->replaceAllUsesWith(Result);
12890 // Push the new instruction and any users onto the worklist.
12891 Worklist.Add(Result);
12892 Worklist.AddUsersToWorkList(*Result);
12894 // Move the name to the new instruction first.
12895 Result->takeName(I);
12897 // Insert the new instruction into the basic block...
12898 BasicBlock *InstParent = I->getParent();
12899 BasicBlock::iterator InsertPos = I;
12901 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12902 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12905 InstParent->getInstList().insert(InsertPos, Result);
12907 EraseInstFromFunction(*I);
12910 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
12911 << " New = " << *I << '\n');
12914 // If the instruction was modified, it's possible that it is now dead.
12915 // if so, remove it.
12916 if (isInstructionTriviallyDead(I)) {
12917 EraseInstFromFunction(*I);
12920 Worklist.AddUsersToWorkList(*I);
12923 MadeIRChange = true;
12928 return MadeIRChange;
12932 bool InstCombiner::runOnFunction(Function &F) {
12933 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12934 Context = &F.getContext();
12935 TD = getAnalysisIfAvailable<TargetData>();
12938 /// Builder - This is an IRBuilder that automatically inserts new
12939 /// instructions into the worklist when they are created.
12940 IRBuilder<true, TargetFolder, InstCombineIRInserter>
12941 TheBuilder(F.getContext(), TargetFolder(TD, F.getContext()),
12942 InstCombineIRInserter(Worklist));
12943 Builder = &TheBuilder;
12945 bool EverMadeChange = false;
12947 // Iterate while there is work to do.
12948 unsigned Iteration = 0;
12949 while (DoOneIteration(F, Iteration++))
12950 EverMadeChange = true;
12953 return EverMadeChange;
12956 FunctionPass *llvm::createInstructionCombiningPass() {
12957 return new InstCombiner();