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/MemoryBuiltins.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 *visitFree(Instruction &FI);
289 Instruction *visitLoadInst(LoadInst &LI);
290 Instruction *visitStoreInst(StoreInst &SI);
291 Instruction *visitBranchInst(BranchInst &BI);
292 Instruction *visitSwitchInst(SwitchInst &SI);
293 Instruction *visitInsertElementInst(InsertElementInst &IE);
294 Instruction *visitExtractElementInst(ExtractElementInst &EI);
295 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
296 Instruction *visitExtractValueInst(ExtractValueInst &EV);
298 // visitInstruction - Specify what to return for unhandled instructions...
299 Instruction *visitInstruction(Instruction &I) { return 0; }
302 Instruction *visitCallSite(CallSite CS);
303 bool transformConstExprCastCall(CallSite CS);
304 Instruction *transformCallThroughTrampoline(CallSite CS);
305 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
306 bool DoXform = true);
307 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
308 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
312 // InsertNewInstBefore - insert an instruction New before instruction Old
313 // in the program. Add the new instruction to the worklist.
315 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
316 assert(New && New->getParent() == 0 &&
317 "New instruction already inserted into a basic block!");
318 BasicBlock *BB = Old.getParent();
319 BB->getInstList().insert(&Old, New); // Insert inst
324 // ReplaceInstUsesWith - This method is to be used when an instruction is
325 // found to be dead, replacable with another preexisting expression. Here
326 // we add all uses of I to the worklist, replace all uses of I with the new
327 // value, then return I, so that the inst combiner will know that I was
330 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
331 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
333 // If we are replacing the instruction with itself, this must be in a
334 // segment of unreachable code, so just clobber the instruction.
336 V = UndefValue::get(I.getType());
338 I.replaceAllUsesWith(V);
342 // EraseInstFromFunction - When dealing with an instruction that has side
343 // effects or produces a void value, we can't rely on DCE to delete the
344 // instruction. Instead, visit methods should return the value returned by
346 Instruction *EraseInstFromFunction(Instruction &I) {
347 DEBUG(errs() << "IC: ERASE " << I << '\n');
349 assert(I.use_empty() && "Cannot erase instruction that is used!");
350 // Make sure that we reprocess all operands now that we reduced their
352 if (I.getNumOperands() < 8) {
353 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
354 if (Instruction *Op = dyn_cast<Instruction>(*i))
360 return 0; // Don't do anything with FI
363 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
364 APInt &KnownOne, unsigned Depth = 0) const {
365 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
368 bool MaskedValueIsZero(Value *V, const APInt &Mask,
369 unsigned Depth = 0) const {
370 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
372 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
373 return llvm::ComputeNumSignBits(Op, TD, Depth);
378 /// SimplifyCommutative - This performs a few simplifications for
379 /// commutative operators.
380 bool SimplifyCommutative(BinaryOperator &I);
382 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
383 /// most-complex to least-complex order.
384 bool SimplifyCompare(CmpInst &I);
386 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
387 /// based on the demanded bits.
388 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
389 APInt& KnownZero, APInt& KnownOne,
391 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
392 APInt& KnownZero, APInt& KnownOne,
395 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
396 /// SimplifyDemandedBits knows about. See if the instruction has any
397 /// properties that allow us to simplify its operands.
398 bool SimplifyDemandedInstructionBits(Instruction &Inst);
400 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
401 APInt& UndefElts, unsigned Depth = 0);
403 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
404 // which has a PHI node as operand #0, see if we can fold the instruction
405 // into the PHI (which is only possible if all operands to the PHI are
408 // If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
409 // that would normally be unprofitable because they strongly encourage jump
411 Instruction *FoldOpIntoPhi(Instruction &I, bool AllowAggressive = false);
413 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
414 // operator and they all are only used by the PHI, PHI together their
415 // inputs, and do the operation once, to the result of the PHI.
416 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
417 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
418 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
421 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
422 ConstantInt *AndRHS, BinaryOperator &TheAnd);
424 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
425 bool isSub, Instruction &I);
426 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
427 bool isSigned, bool Inside, Instruction &IB);
428 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
429 Instruction *MatchBSwap(BinaryOperator &I);
430 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
431 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
432 Instruction *SimplifyMemSet(MemSetInst *MI);
435 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
437 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
438 unsigned CastOpc, int &NumCastsRemoved);
439 unsigned GetOrEnforceKnownAlignment(Value *V,
440 unsigned PrefAlign = 0);
443 } // end anonymous namespace
445 char InstCombiner::ID = 0;
446 static RegisterPass<InstCombiner>
447 X("instcombine", "Combine redundant instructions");
449 // getComplexity: Assign a complexity or rank value to LLVM Values...
450 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
451 static unsigned getComplexity(Value *V) {
452 if (isa<Instruction>(V)) {
453 if (BinaryOperator::isNeg(V) ||
454 BinaryOperator::isFNeg(V) ||
455 BinaryOperator::isNot(V))
459 if (isa<Argument>(V)) return 3;
460 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
463 // isOnlyUse - Return true if this instruction will be deleted if we stop using
465 static bool isOnlyUse(Value *V) {
466 return V->hasOneUse() || isa<Constant>(V);
469 // getPromotedType - Return the specified type promoted as it would be to pass
470 // though a va_arg area...
471 static const Type *getPromotedType(const Type *Ty) {
472 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
473 if (ITy->getBitWidth() < 32)
474 return Type::getInt32Ty(Ty->getContext());
479 /// getBitCastOperand - If the specified operand is a CastInst, a constant
480 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
481 /// operand value, otherwise return null.
482 static Value *getBitCastOperand(Value *V) {
483 if (Operator *O = dyn_cast<Operator>(V)) {
484 if (O->getOpcode() == Instruction::BitCast)
485 return O->getOperand(0);
486 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
487 if (GEP->hasAllZeroIndices())
488 return GEP->getPointerOperand();
493 /// This function is a wrapper around CastInst::isEliminableCastPair. It
494 /// simply extracts arguments and returns what that function returns.
495 static Instruction::CastOps
496 isEliminableCastPair(
497 const CastInst *CI, ///< The first cast instruction
498 unsigned opcode, ///< The opcode of the second cast instruction
499 const Type *DstTy, ///< The target type for the second cast instruction
500 TargetData *TD ///< The target data for pointer size
503 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
504 const Type *MidTy = CI->getType(); // B from above
506 // Get the opcodes of the two Cast instructions
507 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
508 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
510 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
512 TD ? TD->getIntPtrType(CI->getContext()) : 0);
514 // We don't want to form an inttoptr or ptrtoint that converts to an integer
515 // type that differs from the pointer size.
516 if ((Res == Instruction::IntToPtr &&
517 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
518 (Res == Instruction::PtrToInt &&
519 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
522 return Instruction::CastOps(Res);
525 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
526 /// in any code being generated. It does not require codegen if V is simple
527 /// enough or if the cast can be folded into other casts.
528 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
529 const Type *Ty, TargetData *TD) {
530 if (V->getType() == Ty || isa<Constant>(V)) return false;
532 // If this is another cast that can be eliminated, it isn't codegen either.
533 if (const CastInst *CI = dyn_cast<CastInst>(V))
534 if (isEliminableCastPair(CI, opcode, Ty, TD))
539 // SimplifyCommutative - This performs a few simplifications for commutative
542 // 1. Order operands such that they are listed from right (least complex) to
543 // left (most complex). This puts constants before unary operators before
546 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
547 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
549 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
550 bool Changed = false;
551 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
552 Changed = !I.swapOperands();
554 if (!I.isAssociative()) return Changed;
555 Instruction::BinaryOps Opcode = I.getOpcode();
556 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
557 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
558 if (isa<Constant>(I.getOperand(1))) {
559 Constant *Folded = ConstantExpr::get(I.getOpcode(),
560 cast<Constant>(I.getOperand(1)),
561 cast<Constant>(Op->getOperand(1)));
562 I.setOperand(0, Op->getOperand(0));
563 I.setOperand(1, Folded);
565 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
566 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
567 isOnlyUse(Op) && isOnlyUse(Op1)) {
568 Constant *C1 = cast<Constant>(Op->getOperand(1));
569 Constant *C2 = cast<Constant>(Op1->getOperand(1));
571 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
572 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
573 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
577 I.setOperand(0, New);
578 I.setOperand(1, Folded);
585 /// SimplifyCompare - For a CmpInst this function just orders the operands
586 /// so that theyare listed from right (least complex) to left (most complex).
587 /// This puts constants before unary operators before binary operators.
588 bool InstCombiner::SimplifyCompare(CmpInst &I) {
589 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
592 // Compare instructions are not associative so there's nothing else we can do.
596 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
597 // if the LHS is a constant zero (which is the 'negate' form).
599 static inline Value *dyn_castNegVal(Value *V) {
600 if (BinaryOperator::isNeg(V))
601 return BinaryOperator::getNegArgument(V);
603 // Constants can be considered to be negated values if they can be folded.
604 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
605 return ConstantExpr::getNeg(C);
607 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
608 if (C->getType()->getElementType()->isInteger())
609 return ConstantExpr::getNeg(C);
614 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
615 // instruction if the LHS is a constant negative zero (which is the 'negate'
618 static inline Value *dyn_castFNegVal(Value *V) {
619 if (BinaryOperator::isFNeg(V))
620 return BinaryOperator::getFNegArgument(V);
622 // Constants can be considered to be negated values if they can be folded.
623 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
624 return ConstantExpr::getFNeg(C);
626 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
627 if (C->getType()->getElementType()->isFloatingPoint())
628 return ConstantExpr::getFNeg(C);
633 /// isFreeToInvert - Return true if the specified value is free to invert (apply
634 /// ~ to). This happens in cases where the ~ can be eliminated.
635 static inline bool isFreeToInvert(Value *V) {
637 if (BinaryOperator::isNot(V))
640 // Constants can be considered to be not'ed values.
641 if (isa<ConstantInt>(V))
644 // Compares can be inverted if they have a single use.
645 if (CmpInst *CI = dyn_cast<CmpInst>(V))
646 return CI->hasOneUse();
651 static inline Value *dyn_castNotVal(Value *V) {
652 // If this is not(not(x)) don't return that this is a not: we want the two
653 // not's to be folded first.
654 if (BinaryOperator::isNot(V)) {
655 Value *Operand = BinaryOperator::getNotArgument(V);
656 if (!isFreeToInvert(Operand))
660 // Constants can be considered to be not'ed values...
661 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
662 return ConstantInt::get(C->getType(), ~C->getValue());
668 // dyn_castFoldableMul - If this value is a multiply that can be folded into
669 // other computations (because it has a constant operand), return the
670 // non-constant operand of the multiply, and set CST to point to the multiplier.
671 // Otherwise, return null.
673 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
674 if (V->hasOneUse() && V->getType()->isInteger())
675 if (Instruction *I = dyn_cast<Instruction>(V)) {
676 if (I->getOpcode() == Instruction::Mul)
677 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
678 return I->getOperand(0);
679 if (I->getOpcode() == Instruction::Shl)
680 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
681 // The multiplier is really 1 << CST.
682 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
683 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
684 CST = ConstantInt::get(V->getType()->getContext(),
685 APInt(BitWidth, 1).shl(CSTVal));
686 return I->getOperand(0);
692 /// AddOne - Add one to a ConstantInt
693 static Constant *AddOne(Constant *C) {
694 return ConstantExpr::getAdd(C,
695 ConstantInt::get(C->getType(), 1));
697 /// SubOne - Subtract one from a ConstantInt
698 static Constant *SubOne(ConstantInt *C) {
699 return ConstantExpr::getSub(C,
700 ConstantInt::get(C->getType(), 1));
702 /// MultiplyOverflows - True if the multiply can not be expressed in an int
704 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
705 uint32_t W = C1->getBitWidth();
706 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
715 APInt MulExt = LHSExt * RHSExt;
718 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
719 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
720 return MulExt.slt(Min) || MulExt.sgt(Max);
722 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
726 /// ShrinkDemandedConstant - Check to see if the specified operand of the
727 /// specified instruction is a constant integer. If so, check to see if there
728 /// are any bits set in the constant that are not demanded. If so, shrink the
729 /// constant and return true.
730 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
732 assert(I && "No instruction?");
733 assert(OpNo < I->getNumOperands() && "Operand index too large");
735 // If the operand is not a constant integer, nothing to do.
736 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
737 if (!OpC) return false;
739 // If there are no bits set that aren't demanded, nothing to do.
740 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
741 if ((~Demanded & OpC->getValue()) == 0)
744 // This instruction is producing bits that are not demanded. Shrink the RHS.
745 Demanded &= OpC->getValue();
746 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
750 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
751 // 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 ComputeSignedMinMaxValuesFromKnownBits(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 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
761 APInt UnknownBits = ~(KnownZero|KnownOne);
763 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
764 // bit if it is unknown.
766 Max = KnownOne|UnknownBits;
768 if (UnknownBits.isNegative()) { // Sign bit is unknown
769 Min.set(Min.getBitWidth()-1);
770 Max.clear(Max.getBitWidth()-1);
774 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
775 // a set of known zero and one bits, compute the maximum and minimum values that
776 // could have the specified known zero and known one bits, returning them in
778 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
779 const APInt &KnownOne,
780 APInt &Min, APInt &Max) {
781 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
782 KnownZero.getBitWidth() == Min.getBitWidth() &&
783 KnownZero.getBitWidth() == Max.getBitWidth() &&
784 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
785 APInt UnknownBits = ~(KnownZero|KnownOne);
787 // The minimum value is when the unknown bits are all zeros.
789 // The maximum value is when the unknown bits are all ones.
790 Max = KnownOne|UnknownBits;
793 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
794 /// SimplifyDemandedBits knows about. See if the instruction has any
795 /// properties that allow us to simplify its operands.
796 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
797 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
798 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
799 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
801 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
802 KnownZero, KnownOne, 0);
803 if (V == 0) return false;
804 if (V == &Inst) return true;
805 ReplaceInstUsesWith(Inst, V);
809 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
810 /// specified instruction operand if possible, updating it in place. It returns
811 /// true if it made any change and false otherwise.
812 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
813 APInt &KnownZero, APInt &KnownOne,
815 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
816 KnownZero, KnownOne, Depth);
817 if (NewVal == 0) return false;
823 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
824 /// value based on the demanded bits. When this function is called, it is known
825 /// that only the bits set in DemandedMask of the result of V are ever used
826 /// downstream. Consequently, depending on the mask and V, it may be possible
827 /// to replace V with a constant or one of its operands. In such cases, this
828 /// function does the replacement and returns true. In all other cases, it
829 /// returns false after analyzing the expression and setting KnownOne and known
830 /// to be one in the expression. KnownZero contains all the bits that are known
831 /// to be zero in the expression. These are provided to potentially allow the
832 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
833 /// the expression. KnownOne and KnownZero always follow the invariant that
834 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
835 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
836 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
837 /// and KnownOne must all be the same.
839 /// This returns null if it did not change anything and it permits no
840 /// simplification. This returns V itself if it did some simplification of V's
841 /// operands based on the information about what bits are demanded. This returns
842 /// some other non-null value if it found out that V is equal to another value
843 /// in the context where the specified bits are demanded, but not for all users.
844 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
845 APInt &KnownZero, APInt &KnownOne,
847 assert(V != 0 && "Null pointer of Value???");
848 assert(Depth <= 6 && "Limit Search Depth");
849 uint32_t BitWidth = DemandedMask.getBitWidth();
850 const Type *VTy = V->getType();
851 assert((TD || !isa<PointerType>(VTy)) &&
852 "SimplifyDemandedBits needs to know bit widths!");
853 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
854 (!VTy->isIntOrIntVector() ||
855 VTy->getScalarSizeInBits() == BitWidth) &&
856 KnownZero.getBitWidth() == BitWidth &&
857 KnownOne.getBitWidth() == BitWidth &&
858 "Value *V, DemandedMask, KnownZero and KnownOne "
859 "must have same BitWidth");
860 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
861 // We know all of the bits for a constant!
862 KnownOne = CI->getValue() & DemandedMask;
863 KnownZero = ~KnownOne & DemandedMask;
866 if (isa<ConstantPointerNull>(V)) {
867 // We know all of the bits for a constant!
869 KnownZero = DemandedMask;
875 if (DemandedMask == 0) { // Not demanding any bits from V.
876 if (isa<UndefValue>(V))
878 return UndefValue::get(VTy);
881 if (Depth == 6) // Limit search depth.
884 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
885 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
887 Instruction *I = dyn_cast<Instruction>(V);
889 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
890 return 0; // Only analyze instructions.
893 // If there are multiple uses of this value and we aren't at the root, then
894 // we can't do any simplifications of the operands, because DemandedMask
895 // only reflects the bits demanded by *one* of the users.
896 if (Depth != 0 && !I->hasOneUse()) {
897 // Despite the fact that we can't simplify this instruction in all User's
898 // context, we can at least compute the knownzero/knownone bits, and we can
899 // do simplifications that apply to *just* the one user if we know that
900 // this instruction has a simpler value in that context.
901 if (I->getOpcode() == Instruction::And) {
902 // If either the LHS or the RHS are Zero, the result is zero.
903 ComputeMaskedBits(I->getOperand(1), DemandedMask,
904 RHSKnownZero, RHSKnownOne, Depth+1);
905 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
906 LHSKnownZero, LHSKnownOne, Depth+1);
908 // If all of the demanded bits are known 1 on one side, return the other.
909 // These bits cannot contribute to the result of the 'and' in this
911 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
912 (DemandedMask & ~LHSKnownZero))
913 return I->getOperand(0);
914 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
915 (DemandedMask & ~RHSKnownZero))
916 return I->getOperand(1);
918 // If all of the demanded bits in the inputs are known zeros, return zero.
919 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
920 return Constant::getNullValue(VTy);
922 } else if (I->getOpcode() == Instruction::Or) {
923 // We can simplify (X|Y) -> X or Y in the user's context if we know that
924 // only bits from X or Y are demanded.
926 // If either the LHS or the RHS are One, the result is One.
927 ComputeMaskedBits(I->getOperand(1), DemandedMask,
928 RHSKnownZero, RHSKnownOne, Depth+1);
929 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
930 LHSKnownZero, LHSKnownOne, Depth+1);
932 // If all of the demanded bits are known zero on one side, return the
933 // other. These bits cannot contribute to the result of the 'or' in this
935 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
936 (DemandedMask & ~LHSKnownOne))
937 return I->getOperand(0);
938 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
939 (DemandedMask & ~RHSKnownOne))
940 return I->getOperand(1);
942 // If all of the potentially set bits on one side are known to be set on
943 // the other side, just use the 'other' side.
944 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
945 (DemandedMask & (~RHSKnownZero)))
946 return I->getOperand(0);
947 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
948 (DemandedMask & (~LHSKnownZero)))
949 return I->getOperand(1);
952 // Compute the KnownZero/KnownOne bits to simplify things downstream.
953 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
957 // If this is the root being simplified, allow it to have multiple uses,
958 // just set the DemandedMask to all bits so that we can try to simplify the
959 // operands. This allows visitTruncInst (for example) to simplify the
960 // operand of a trunc without duplicating all the logic below.
961 if (Depth == 0 && !V->hasOneUse())
962 DemandedMask = APInt::getAllOnesValue(BitWidth);
964 switch (I->getOpcode()) {
966 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
968 case Instruction::And:
969 // If either the LHS or the RHS are Zero, the result is zero.
970 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
971 RHSKnownZero, RHSKnownOne, Depth+1) ||
972 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
973 LHSKnownZero, LHSKnownOne, Depth+1))
975 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
976 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
978 // If all of the demanded bits are known 1 on one side, return the other.
979 // These bits cannot contribute to the result of the 'and'.
980 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
981 (DemandedMask & ~LHSKnownZero))
982 return I->getOperand(0);
983 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
984 (DemandedMask & ~RHSKnownZero))
985 return I->getOperand(1);
987 // If all of the demanded bits in the inputs are known zeros, return zero.
988 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
989 return Constant::getNullValue(VTy);
991 // If the RHS is a constant, see if we can simplify it.
992 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
995 // Output known-1 bits are only known if set in both the LHS & RHS.
996 RHSKnownOne &= LHSKnownOne;
997 // Output known-0 are known to be clear if zero in either the LHS | RHS.
998 RHSKnownZero |= LHSKnownZero;
1000 case Instruction::Or:
1001 // If either the LHS or the RHS are One, the result is One.
1002 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1003 RHSKnownZero, RHSKnownOne, Depth+1) ||
1004 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
1005 LHSKnownZero, LHSKnownOne, Depth+1))
1007 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1008 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1010 // If all of the demanded bits are known zero on one side, return the other.
1011 // These bits cannot contribute to the result of the 'or'.
1012 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1013 (DemandedMask & ~LHSKnownOne))
1014 return I->getOperand(0);
1015 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1016 (DemandedMask & ~RHSKnownOne))
1017 return I->getOperand(1);
1019 // If all of the potentially set bits on one side are known to be set on
1020 // the other side, just use the 'other' side.
1021 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1022 (DemandedMask & (~RHSKnownZero)))
1023 return I->getOperand(0);
1024 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1025 (DemandedMask & (~LHSKnownZero)))
1026 return I->getOperand(1);
1028 // If the RHS is a constant, see if we can simplify it.
1029 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1032 // Output known-0 bits are only known if clear in both the LHS & RHS.
1033 RHSKnownZero &= LHSKnownZero;
1034 // Output known-1 are known to be set if set in either the LHS | RHS.
1035 RHSKnownOne |= LHSKnownOne;
1037 case Instruction::Xor: {
1038 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1039 RHSKnownZero, RHSKnownOne, Depth+1) ||
1040 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1041 LHSKnownZero, LHSKnownOne, Depth+1))
1043 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1044 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1046 // If all of the demanded bits are known zero on one side, return the other.
1047 // These bits cannot contribute to the result of the 'xor'.
1048 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1049 return I->getOperand(0);
1050 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1051 return I->getOperand(1);
1053 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1054 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1055 (RHSKnownOne & LHSKnownOne);
1056 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1057 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1058 (RHSKnownOne & LHSKnownZero);
1060 // If all of the demanded bits are known to be zero on one side or the
1061 // other, turn this into an *inclusive* or.
1062 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1063 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1065 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1067 return InsertNewInstBefore(Or, *I);
1070 // If all of the demanded bits on one side are known, and all of the set
1071 // bits on that side are also known to be set on the other side, turn this
1072 // into an AND, as we know the bits will be cleared.
1073 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1074 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1076 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1077 Constant *AndC = Constant::getIntegerValue(VTy,
1078 ~RHSKnownOne & DemandedMask);
1080 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1081 return InsertNewInstBefore(And, *I);
1085 // If the RHS is a constant, see if we can simplify it.
1086 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1087 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1090 // If our LHS is an 'and' and if it has one use, and if any of the bits we
1091 // are flipping are known to be set, then the xor is just resetting those
1092 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
1093 // simplifying both of them.
1094 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0)))
1095 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
1096 isa<ConstantInt>(I->getOperand(1)) &&
1097 isa<ConstantInt>(LHSInst->getOperand(1)) &&
1098 (LHSKnownOne & RHSKnownOne & DemandedMask) != 0) {
1099 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
1100 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
1101 APInt NewMask = ~(LHSKnownOne & RHSKnownOne & DemandedMask);
1104 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
1105 Instruction *NewAnd =
1106 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1107 InsertNewInstBefore(NewAnd, *I);
1110 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
1111 Instruction *NewXor =
1112 BinaryOperator::CreateXor(NewAnd, XorC, "tmp");
1113 return InsertNewInstBefore(NewXor, *I);
1117 RHSKnownZero = KnownZeroOut;
1118 RHSKnownOne = KnownOneOut;
1121 case Instruction::Select:
1122 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1123 RHSKnownZero, RHSKnownOne, Depth+1) ||
1124 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1125 LHSKnownZero, LHSKnownOne, Depth+1))
1127 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1128 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1130 // If the operands are constants, see if we can simplify them.
1131 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1132 ShrinkDemandedConstant(I, 2, DemandedMask))
1135 // Only known if known in both the LHS and RHS.
1136 RHSKnownOne &= LHSKnownOne;
1137 RHSKnownZero &= LHSKnownZero;
1139 case Instruction::Trunc: {
1140 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1141 DemandedMask.zext(truncBf);
1142 RHSKnownZero.zext(truncBf);
1143 RHSKnownOne.zext(truncBf);
1144 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1145 RHSKnownZero, RHSKnownOne, Depth+1))
1147 DemandedMask.trunc(BitWidth);
1148 RHSKnownZero.trunc(BitWidth);
1149 RHSKnownOne.trunc(BitWidth);
1150 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1153 case Instruction::BitCast:
1154 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1155 return false; // vector->int or fp->int?
1157 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1158 if (const VectorType *SrcVTy =
1159 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1160 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1161 // Don't touch a bitcast between vectors of different element counts.
1164 // Don't touch a scalar-to-vector bitcast.
1166 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1167 // Don't touch a vector-to-scalar bitcast.
1170 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1171 RHSKnownZero, RHSKnownOne, Depth+1))
1173 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1175 case Instruction::ZExt: {
1176 // Compute the bits in the result that are not present in the input.
1177 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1179 DemandedMask.trunc(SrcBitWidth);
1180 RHSKnownZero.trunc(SrcBitWidth);
1181 RHSKnownOne.trunc(SrcBitWidth);
1182 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1183 RHSKnownZero, RHSKnownOne, Depth+1))
1185 DemandedMask.zext(BitWidth);
1186 RHSKnownZero.zext(BitWidth);
1187 RHSKnownOne.zext(BitWidth);
1188 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1189 // The top bits are known to be zero.
1190 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1193 case Instruction::SExt: {
1194 // Compute the bits in the result that are not present in the input.
1195 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1197 APInt InputDemandedBits = DemandedMask &
1198 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1200 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1201 // If any of the sign extended bits are demanded, we know that the sign
1203 if ((NewBits & DemandedMask) != 0)
1204 InputDemandedBits.set(SrcBitWidth-1);
1206 InputDemandedBits.trunc(SrcBitWidth);
1207 RHSKnownZero.trunc(SrcBitWidth);
1208 RHSKnownOne.trunc(SrcBitWidth);
1209 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1210 RHSKnownZero, RHSKnownOne, Depth+1))
1212 InputDemandedBits.zext(BitWidth);
1213 RHSKnownZero.zext(BitWidth);
1214 RHSKnownOne.zext(BitWidth);
1215 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1217 // If the sign bit of the input is known set or clear, then we know the
1218 // top bits of the result.
1220 // If the input sign bit is known zero, or if the NewBits are not demanded
1221 // convert this into a zero extension.
1222 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1223 // Convert to ZExt cast
1224 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1225 return InsertNewInstBefore(NewCast, *I);
1226 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1227 RHSKnownOne |= NewBits;
1231 case Instruction::Add: {
1232 // Figure out what the input bits are. If the top bits of the and result
1233 // are not demanded, then the add doesn't demand them from its input
1235 unsigned NLZ = DemandedMask.countLeadingZeros();
1237 // If there is a constant on the RHS, there are a variety of xformations
1239 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1240 // If null, this should be simplified elsewhere. Some of the xforms here
1241 // won't work if the RHS is zero.
1245 // If the top bit of the output is demanded, demand everything from the
1246 // input. Otherwise, we demand all the input bits except NLZ top bits.
1247 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1249 // Find information about known zero/one bits in the input.
1250 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1251 LHSKnownZero, LHSKnownOne, Depth+1))
1254 // If the RHS of the add has bits set that can't affect the input, reduce
1256 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1259 // Avoid excess work.
1260 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1263 // Turn it into OR if input bits are zero.
1264 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1266 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1268 return InsertNewInstBefore(Or, *I);
1271 // We can say something about the output known-zero and known-one bits,
1272 // depending on potential carries from the input constant and the
1273 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1274 // bits set and the RHS constant is 0x01001, then we know we have a known
1275 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1277 // To compute this, we first compute the potential carry bits. These are
1278 // the bits which may be modified. I'm not aware of a better way to do
1280 const APInt &RHSVal = RHS->getValue();
1281 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1283 // Now that we know which bits have carries, compute the known-1/0 sets.
1285 // Bits are known one if they are known zero in one operand and one in the
1286 // other, and there is no input carry.
1287 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1288 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1290 // Bits are known zero if they are known zero in both operands and there
1291 // is no input carry.
1292 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1294 // If the high-bits of this ADD are not demanded, then it does not demand
1295 // the high bits of its LHS or RHS.
1296 if (DemandedMask[BitWidth-1] == 0) {
1297 // Right fill the mask of bits for this ADD to demand the most
1298 // significant bit and all those below it.
1299 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1300 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1301 LHSKnownZero, LHSKnownOne, Depth+1) ||
1302 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1303 LHSKnownZero, LHSKnownOne, Depth+1))
1309 case Instruction::Sub:
1310 // If the high-bits of this SUB are not demanded, then it does not demand
1311 // the high bits of its LHS or RHS.
1312 if (DemandedMask[BitWidth-1] == 0) {
1313 // Right fill the mask of bits for this SUB to demand the most
1314 // significant bit and all those below it.
1315 uint32_t NLZ = DemandedMask.countLeadingZeros();
1316 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1317 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1318 LHSKnownZero, LHSKnownOne, Depth+1) ||
1319 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1320 LHSKnownZero, LHSKnownOne, Depth+1))
1323 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1324 // the known zeros and ones.
1325 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1327 case Instruction::Shl:
1328 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1329 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1330 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1331 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1332 RHSKnownZero, RHSKnownOne, Depth+1))
1334 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1335 RHSKnownZero <<= ShiftAmt;
1336 RHSKnownOne <<= ShiftAmt;
1337 // low bits known zero.
1339 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1342 case Instruction::LShr:
1343 // For a logical shift right
1344 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1345 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1347 // Unsigned shift right.
1348 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1349 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1350 RHSKnownZero, RHSKnownOne, Depth+1))
1352 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1353 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1354 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1356 // Compute the new bits that are at the top now.
1357 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1358 RHSKnownZero |= HighBits; // high bits known zero.
1362 case Instruction::AShr:
1363 // If this is an arithmetic shift right and only the low-bit is set, we can
1364 // always convert this into a logical shr, even if the shift amount is
1365 // variable. The low bit of the shift cannot be an input sign bit unless
1366 // the shift amount is >= the size of the datatype, which is undefined.
1367 if (DemandedMask == 1) {
1368 // Perform the logical shift right.
1369 Instruction *NewVal = BinaryOperator::CreateLShr(
1370 I->getOperand(0), I->getOperand(1), I->getName());
1371 return InsertNewInstBefore(NewVal, *I);
1374 // If the sign bit is the only bit demanded by this ashr, then there is no
1375 // need to do it, the shift doesn't change the high bit.
1376 if (DemandedMask.isSignBit())
1377 return I->getOperand(0);
1379 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1380 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1382 // Signed shift right.
1383 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1384 // If any of the "high bits" are demanded, we should set the sign bit as
1386 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1387 DemandedMaskIn.set(BitWidth-1);
1388 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1389 RHSKnownZero, RHSKnownOne, Depth+1))
1391 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1392 // Compute the new bits that are at the top now.
1393 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1394 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1395 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1397 // Handle the sign bits.
1398 APInt SignBit(APInt::getSignBit(BitWidth));
1399 // Adjust to where it is now in the mask.
1400 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1402 // If the input sign bit is known to be zero, or if none of the top bits
1403 // are demanded, turn this into an unsigned shift right.
1404 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1405 (HighBits & ~DemandedMask) == HighBits) {
1406 // Perform the logical shift right.
1407 Instruction *NewVal = BinaryOperator::CreateLShr(
1408 I->getOperand(0), SA, I->getName());
1409 return InsertNewInstBefore(NewVal, *I);
1410 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1411 RHSKnownOne |= HighBits;
1415 case Instruction::SRem:
1416 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1417 APInt RA = Rem->getValue().abs();
1418 if (RA.isPowerOf2()) {
1419 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1420 return I->getOperand(0);
1422 APInt LowBits = RA - 1;
1423 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1424 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1425 LHSKnownZero, LHSKnownOne, Depth+1))
1428 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1429 LHSKnownZero |= ~LowBits;
1431 KnownZero |= LHSKnownZero & DemandedMask;
1433 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1437 case Instruction::URem: {
1438 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1439 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1440 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1441 KnownZero2, KnownOne2, Depth+1) ||
1442 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1443 KnownZero2, KnownOne2, Depth+1))
1446 unsigned Leaders = KnownZero2.countLeadingOnes();
1447 Leaders = std::max(Leaders,
1448 KnownZero2.countLeadingOnes());
1449 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1452 case Instruction::Call:
1453 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1454 switch (II->getIntrinsicID()) {
1456 case Intrinsic::bswap: {
1457 // If the only bits demanded come from one byte of the bswap result,
1458 // just shift the input byte into position to eliminate the bswap.
1459 unsigned NLZ = DemandedMask.countLeadingZeros();
1460 unsigned NTZ = DemandedMask.countTrailingZeros();
1462 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1463 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1464 // have 14 leading zeros, round to 8.
1467 // If we need exactly one byte, we can do this transformation.
1468 if (BitWidth-NLZ-NTZ == 8) {
1469 unsigned ResultBit = NTZ;
1470 unsigned InputBit = BitWidth-NTZ-8;
1472 // Replace this with either a left or right shift to get the byte into
1474 Instruction *NewVal;
1475 if (InputBit > ResultBit)
1476 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1477 ConstantInt::get(I->getType(), InputBit-ResultBit));
1479 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1480 ConstantInt::get(I->getType(), ResultBit-InputBit));
1481 NewVal->takeName(I);
1482 return InsertNewInstBefore(NewVal, *I);
1485 // TODO: Could compute known zero/one bits based on the input.
1490 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1494 // If the client is only demanding bits that we know, return the known
1496 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1497 return Constant::getIntegerValue(VTy, RHSKnownOne);
1502 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1503 /// any number of elements. DemandedElts contains the set of elements that are
1504 /// actually used by the caller. This method analyzes which elements of the
1505 /// operand are undef and returns that information in UndefElts.
1507 /// If the information about demanded elements can be used to simplify the
1508 /// operation, the operation is simplified, then the resultant value is
1509 /// returned. This returns null if no change was made.
1510 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1513 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1514 APInt EltMask(APInt::getAllOnesValue(VWidth));
1515 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1517 if (isa<UndefValue>(V)) {
1518 // If the entire vector is undefined, just return this info.
1519 UndefElts = EltMask;
1521 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1522 UndefElts = EltMask;
1523 return UndefValue::get(V->getType());
1527 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1528 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1529 Constant *Undef = UndefValue::get(EltTy);
1531 std::vector<Constant*> Elts;
1532 for (unsigned i = 0; i != VWidth; ++i)
1533 if (!DemandedElts[i]) { // If not demanded, set to undef.
1534 Elts.push_back(Undef);
1536 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1537 Elts.push_back(Undef);
1539 } else { // Otherwise, defined.
1540 Elts.push_back(CP->getOperand(i));
1543 // If we changed the constant, return it.
1544 Constant *NewCP = ConstantVector::get(Elts);
1545 return NewCP != CP ? NewCP : 0;
1546 } else if (isa<ConstantAggregateZero>(V)) {
1547 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1550 // Check if this is identity. If so, return 0 since we are not simplifying
1552 if (DemandedElts == ((1ULL << VWidth) -1))
1555 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1556 Constant *Zero = Constant::getNullValue(EltTy);
1557 Constant *Undef = UndefValue::get(EltTy);
1558 std::vector<Constant*> Elts;
1559 for (unsigned i = 0; i != VWidth; ++i) {
1560 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1561 Elts.push_back(Elt);
1563 UndefElts = DemandedElts ^ EltMask;
1564 return ConstantVector::get(Elts);
1567 // Limit search depth.
1571 // If multiple users are using the root value, procede with
1572 // simplification conservatively assuming that all elements
1574 if (!V->hasOneUse()) {
1575 // Quit if we find multiple users of a non-root value though.
1576 // They'll be handled when it's their turn to be visited by
1577 // the main instcombine process.
1579 // TODO: Just compute the UndefElts information recursively.
1582 // Conservatively assume that all elements are needed.
1583 DemandedElts = EltMask;
1586 Instruction *I = dyn_cast<Instruction>(V);
1587 if (!I) return 0; // Only analyze instructions.
1589 bool MadeChange = false;
1590 APInt UndefElts2(VWidth, 0);
1592 switch (I->getOpcode()) {
1595 case Instruction::InsertElement: {
1596 // If this is a variable index, we don't know which element it overwrites.
1597 // demand exactly the same input as we produce.
1598 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1600 // Note that we can't propagate undef elt info, because we don't know
1601 // which elt is getting updated.
1602 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1603 UndefElts2, Depth+1);
1604 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1608 // If this is inserting an element that isn't demanded, remove this
1610 unsigned IdxNo = Idx->getZExtValue();
1611 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1613 return I->getOperand(0);
1616 // Otherwise, the element inserted overwrites whatever was there, so the
1617 // input demanded set is simpler than the output set.
1618 APInt DemandedElts2 = DemandedElts;
1619 DemandedElts2.clear(IdxNo);
1620 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1621 UndefElts, Depth+1);
1622 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1624 // The inserted element is defined.
1625 UndefElts.clear(IdxNo);
1628 case Instruction::ShuffleVector: {
1629 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1630 uint64_t LHSVWidth =
1631 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1632 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1633 for (unsigned i = 0; i < VWidth; i++) {
1634 if (DemandedElts[i]) {
1635 unsigned MaskVal = Shuffle->getMaskValue(i);
1636 if (MaskVal != -1u) {
1637 assert(MaskVal < LHSVWidth * 2 &&
1638 "shufflevector mask index out of range!");
1639 if (MaskVal < LHSVWidth)
1640 LeftDemanded.set(MaskVal);
1642 RightDemanded.set(MaskVal - LHSVWidth);
1647 APInt UndefElts4(LHSVWidth, 0);
1648 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1649 UndefElts4, Depth+1);
1650 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1652 APInt UndefElts3(LHSVWidth, 0);
1653 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1654 UndefElts3, Depth+1);
1655 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1657 bool NewUndefElts = false;
1658 for (unsigned i = 0; i < VWidth; i++) {
1659 unsigned MaskVal = Shuffle->getMaskValue(i);
1660 if (MaskVal == -1u) {
1662 } else if (MaskVal < LHSVWidth) {
1663 if (UndefElts4[MaskVal]) {
1664 NewUndefElts = true;
1668 if (UndefElts3[MaskVal - LHSVWidth]) {
1669 NewUndefElts = true;
1676 // Add additional discovered undefs.
1677 std::vector<Constant*> Elts;
1678 for (unsigned i = 0; i < VWidth; ++i) {
1680 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1682 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1683 Shuffle->getMaskValue(i)));
1685 I->setOperand(2, ConstantVector::get(Elts));
1690 case Instruction::BitCast: {
1691 // Vector->vector casts only.
1692 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1694 unsigned InVWidth = VTy->getNumElements();
1695 APInt InputDemandedElts(InVWidth, 0);
1698 if (VWidth == InVWidth) {
1699 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1700 // elements as are demanded of us.
1702 InputDemandedElts = DemandedElts;
1703 } else if (VWidth > InVWidth) {
1707 // If there are more elements in the result than there are in the source,
1708 // then an input element is live if any of the corresponding output
1709 // elements are live.
1710 Ratio = VWidth/InVWidth;
1711 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1712 if (DemandedElts[OutIdx])
1713 InputDemandedElts.set(OutIdx/Ratio);
1719 // If there are more elements in the source than there are in the result,
1720 // then an input element is live if the corresponding output element is
1722 Ratio = InVWidth/VWidth;
1723 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1724 if (DemandedElts[InIdx/Ratio])
1725 InputDemandedElts.set(InIdx);
1728 // div/rem demand all inputs, because they don't want divide by zero.
1729 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1730 UndefElts2, Depth+1);
1732 I->setOperand(0, TmpV);
1736 UndefElts = UndefElts2;
1737 if (VWidth > InVWidth) {
1738 llvm_unreachable("Unimp");
1739 // If there are more elements in the result than there are in the source,
1740 // then an output element is undef if the corresponding input element is
1742 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1743 if (UndefElts2[OutIdx/Ratio])
1744 UndefElts.set(OutIdx);
1745 } else if (VWidth < InVWidth) {
1746 llvm_unreachable("Unimp");
1747 // If there are more elements in the source than there are in the result,
1748 // then a result element is undef if all of the corresponding input
1749 // elements are undef.
1750 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1751 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1752 if (!UndefElts2[InIdx]) // Not undef?
1753 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1757 case Instruction::And:
1758 case Instruction::Or:
1759 case Instruction::Xor:
1760 case Instruction::Add:
1761 case Instruction::Sub:
1762 case Instruction::Mul:
1763 // div/rem demand all inputs, because they don't want divide by zero.
1764 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1765 UndefElts, Depth+1);
1766 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1767 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1768 UndefElts2, Depth+1);
1769 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1771 // Output elements are undefined if both are undefined. Consider things
1772 // like undef&0. The result is known zero, not undef.
1773 UndefElts &= UndefElts2;
1776 case Instruction::Call: {
1777 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1779 switch (II->getIntrinsicID()) {
1782 // Binary vector operations that work column-wise. A dest element is a
1783 // function of the corresponding input elements from the two inputs.
1784 case Intrinsic::x86_sse_sub_ss:
1785 case Intrinsic::x86_sse_mul_ss:
1786 case Intrinsic::x86_sse_min_ss:
1787 case Intrinsic::x86_sse_max_ss:
1788 case Intrinsic::x86_sse2_sub_sd:
1789 case Intrinsic::x86_sse2_mul_sd:
1790 case Intrinsic::x86_sse2_min_sd:
1791 case Intrinsic::x86_sse2_max_sd:
1792 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1793 UndefElts, Depth+1);
1794 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1795 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1796 UndefElts2, Depth+1);
1797 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1799 // If only the low elt is demanded and this is a scalarizable intrinsic,
1800 // scalarize it now.
1801 if (DemandedElts == 1) {
1802 switch (II->getIntrinsicID()) {
1804 case Intrinsic::x86_sse_sub_ss:
1805 case Intrinsic::x86_sse_mul_ss:
1806 case Intrinsic::x86_sse2_sub_sd:
1807 case Intrinsic::x86_sse2_mul_sd:
1808 // TODO: Lower MIN/MAX/ABS/etc
1809 Value *LHS = II->getOperand(1);
1810 Value *RHS = II->getOperand(2);
1811 // Extract the element as scalars.
1812 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1813 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1814 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1815 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1817 switch (II->getIntrinsicID()) {
1818 default: llvm_unreachable("Case stmts out of sync!");
1819 case Intrinsic::x86_sse_sub_ss:
1820 case Intrinsic::x86_sse2_sub_sd:
1821 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1822 II->getName()), *II);
1824 case Intrinsic::x86_sse_mul_ss:
1825 case Intrinsic::x86_sse2_mul_sd:
1826 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1827 II->getName()), *II);
1832 InsertElementInst::Create(
1833 UndefValue::get(II->getType()), TmpV,
1834 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1835 InsertNewInstBefore(New, *II);
1840 // Output elements are undefined if both are undefined. Consider things
1841 // like undef&0. The result is known zero, not undef.
1842 UndefElts &= UndefElts2;
1848 return MadeChange ? I : 0;
1852 /// AssociativeOpt - Perform an optimization on an associative operator. This
1853 /// function is designed to check a chain of associative operators for a
1854 /// potential to apply a certain optimization. Since the optimization may be
1855 /// applicable if the expression was reassociated, this checks the chain, then
1856 /// reassociates the expression as necessary to expose the optimization
1857 /// opportunity. This makes use of a special Functor, which must define
1858 /// 'shouldApply' and 'apply' methods.
1860 template<typename Functor>
1861 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1862 unsigned Opcode = Root.getOpcode();
1863 Value *LHS = Root.getOperand(0);
1865 // Quick check, see if the immediate LHS matches...
1866 if (F.shouldApply(LHS))
1867 return F.apply(Root);
1869 // Otherwise, if the LHS is not of the same opcode as the root, return.
1870 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1871 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1872 // Should we apply this transform to the RHS?
1873 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1875 // If not to the RHS, check to see if we should apply to the LHS...
1876 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1877 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1881 // If the functor wants to apply the optimization to the RHS of LHSI,
1882 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1884 // Now all of the instructions are in the current basic block, go ahead
1885 // and perform the reassociation.
1886 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1888 // First move the selected RHS to the LHS of the root...
1889 Root.setOperand(0, LHSI->getOperand(1));
1891 // Make what used to be the LHS of the root be the user of the root...
1892 Value *ExtraOperand = TmpLHSI->getOperand(1);
1893 if (&Root == TmpLHSI) {
1894 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1897 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1898 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1899 BasicBlock::iterator ARI = &Root; ++ARI;
1900 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1903 // Now propagate the ExtraOperand down the chain of instructions until we
1905 while (TmpLHSI != LHSI) {
1906 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1907 // Move the instruction to immediately before the chain we are
1908 // constructing to avoid breaking dominance properties.
1909 NextLHSI->moveBefore(ARI);
1912 Value *NextOp = NextLHSI->getOperand(1);
1913 NextLHSI->setOperand(1, ExtraOperand);
1915 ExtraOperand = NextOp;
1918 // Now that the instructions are reassociated, have the functor perform
1919 // the transformation...
1920 return F.apply(Root);
1923 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1930 // AddRHS - Implements: X + X --> X << 1
1933 explicit AddRHS(Value *rhs) : RHS(rhs) {}
1934 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1935 Instruction *apply(BinaryOperator &Add) const {
1936 return BinaryOperator::CreateShl(Add.getOperand(0),
1937 ConstantInt::get(Add.getType(), 1));
1941 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1943 struct AddMaskingAnd {
1945 explicit AddMaskingAnd(Constant *c) : C2(c) {}
1946 bool shouldApply(Value *LHS) const {
1948 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1949 ConstantExpr::getAnd(C1, C2)->isNullValue();
1951 Instruction *apply(BinaryOperator &Add) const {
1952 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1958 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1960 if (CastInst *CI = dyn_cast<CastInst>(&I))
1961 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
1963 // Figure out if the constant is the left or the right argument.
1964 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1965 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1967 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1969 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1970 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1973 Value *Op0 = SO, *Op1 = ConstOperand;
1975 std::swap(Op0, Op1);
1977 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1978 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
1979 SO->getName()+".op");
1980 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
1981 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1982 SO->getName()+".cmp");
1983 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
1984 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1985 SO->getName()+".cmp");
1986 llvm_unreachable("Unknown binary instruction type!");
1989 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1990 // constant as the other operand, try to fold the binary operator into the
1991 // select arguments. This also works for Cast instructions, which obviously do
1992 // not have a second operand.
1993 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1995 // Don't modify shared select instructions
1996 if (!SI->hasOneUse()) return 0;
1997 Value *TV = SI->getOperand(1);
1998 Value *FV = SI->getOperand(2);
2000 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2001 // Bool selects with constant operands can be folded to logical ops.
2002 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
2004 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2005 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2007 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
2014 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
2015 /// has a PHI node as operand #0, see if we can fold the instruction into the
2016 /// PHI (which is only possible if all operands to the PHI are constants).
2018 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
2019 /// that would normally be unprofitable because they strongly encourage jump
2021 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
2022 bool AllowAggressive) {
2023 AllowAggressive = false;
2024 PHINode *PN = cast<PHINode>(I.getOperand(0));
2025 unsigned NumPHIValues = PN->getNumIncomingValues();
2026 if (NumPHIValues == 0 ||
2027 // We normally only transform phis with a single use, unless we're trying
2028 // hard to make jump threading happen.
2029 (!PN->hasOneUse() && !AllowAggressive))
2033 // Check to see if all of the operands of the PHI are simple constants
2034 // (constantint/constantfp/undef). If there is one non-constant value,
2035 // remember the BB it is in. If there is more than one or if *it* is a PHI,
2036 // bail out. We don't do arbitrary constant expressions here because moving
2037 // their computation can be expensive without a cost model.
2038 BasicBlock *NonConstBB = 0;
2039 for (unsigned i = 0; i != NumPHIValues; ++i)
2040 if (!isa<Constant>(PN->getIncomingValue(i)) ||
2041 isa<ConstantExpr>(PN->getIncomingValue(i))) {
2042 if (NonConstBB) return 0; // More than one non-const value.
2043 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2044 NonConstBB = PN->getIncomingBlock(i);
2046 // If the incoming non-constant value is in I's block, we have an infinite
2048 if (NonConstBB == I.getParent())
2052 // If there is exactly one non-constant value, we can insert a copy of the
2053 // operation in that block. However, if this is a critical edge, we would be
2054 // inserting the computation one some other paths (e.g. inside a loop). Only
2055 // do this if the pred block is unconditionally branching into the phi block.
2056 if (NonConstBB != 0 && !AllowAggressive) {
2057 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2058 if (!BI || !BI->isUnconditional()) return 0;
2061 // Okay, we can do the transformation: create the new PHI node.
2062 PHINode *NewPN = PHINode::Create(I.getType(), "");
2063 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2064 InsertNewInstBefore(NewPN, *PN);
2065 NewPN->takeName(PN);
2067 // Next, add all of the operands to the PHI.
2068 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
2069 // We only currently try to fold the condition of a select when it is a phi,
2070 // not the true/false values.
2071 Value *TrueV = SI->getTrueValue();
2072 Value *FalseV = SI->getFalseValue();
2073 BasicBlock *PhiTransBB = PN->getParent();
2074 for (unsigned i = 0; i != NumPHIValues; ++i) {
2075 BasicBlock *ThisBB = PN->getIncomingBlock(i);
2076 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
2077 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
2079 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2080 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
2082 assert(PN->getIncomingBlock(i) == NonConstBB);
2083 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
2085 "phitmp", NonConstBB->getTerminator());
2086 Worklist.Add(cast<Instruction>(InV));
2088 NewPN->addIncoming(InV, ThisBB);
2090 } else if (I.getNumOperands() == 2) {
2091 Constant *C = cast<Constant>(I.getOperand(1));
2092 for (unsigned i = 0; i != NumPHIValues; ++i) {
2094 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2095 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2096 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2098 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2100 assert(PN->getIncomingBlock(i) == NonConstBB);
2101 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2102 InV = BinaryOperator::Create(BO->getOpcode(),
2103 PN->getIncomingValue(i), C, "phitmp",
2104 NonConstBB->getTerminator());
2105 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2106 InV = CmpInst::Create(CI->getOpcode(),
2108 PN->getIncomingValue(i), C, "phitmp",
2109 NonConstBB->getTerminator());
2111 llvm_unreachable("Unknown binop!");
2113 Worklist.Add(cast<Instruction>(InV));
2115 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2118 CastInst *CI = cast<CastInst>(&I);
2119 const Type *RetTy = CI->getType();
2120 for (unsigned i = 0; i != NumPHIValues; ++i) {
2122 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2123 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2125 assert(PN->getIncomingBlock(i) == NonConstBB);
2126 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2127 I.getType(), "phitmp",
2128 NonConstBB->getTerminator());
2129 Worklist.Add(cast<Instruction>(InV));
2131 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2134 return ReplaceInstUsesWith(I, NewPN);
2138 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2139 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2140 /// This basically requires proving that the add in the original type would not
2141 /// overflow to change the sign bit or have a carry out.
2142 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2143 // There are different heuristics we can use for this. Here are some simple
2146 // Add has the property that adding any two 2's complement numbers can only
2147 // have one carry bit which can change a sign. As such, if LHS and RHS each
2148 // have at least two sign bits, we know that the addition of the two values will
2149 // sign extend fine.
2150 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2154 // If one of the operands only has one non-zero bit, and if the other operand
2155 // has a known-zero bit in a more significant place than it (not including the
2156 // sign bit) the ripple may go up to and fill the zero, but won't change the
2157 // sign. For example, (X & ~4) + 1.
2165 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2166 bool Changed = SimplifyCommutative(I);
2167 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2169 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2170 // X + undef -> undef
2171 if (isa<UndefValue>(RHS))
2172 return ReplaceInstUsesWith(I, RHS);
2175 if (RHSC->isNullValue())
2176 return ReplaceInstUsesWith(I, LHS);
2178 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2179 // X + (signbit) --> X ^ signbit
2180 const APInt& Val = CI->getValue();
2181 uint32_t BitWidth = Val.getBitWidth();
2182 if (Val == APInt::getSignBit(BitWidth))
2183 return BinaryOperator::CreateXor(LHS, RHS);
2185 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2186 // (X & 254)+1 -> (X&254)|1
2187 if (SimplifyDemandedInstructionBits(I))
2190 // zext(bool) + C -> bool ? C + 1 : C
2191 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2192 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2193 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2196 if (isa<PHINode>(LHS))
2197 if (Instruction *NV = FoldOpIntoPhi(I))
2200 ConstantInt *XorRHS = 0;
2202 if (isa<ConstantInt>(RHSC) &&
2203 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2204 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2205 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2207 uint32_t Size = TySizeBits / 2;
2208 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2209 APInt CFF80Val(-C0080Val);
2211 if (TySizeBits > Size) {
2212 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2213 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2214 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2215 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2216 // This is a sign extend if the top bits are known zero.
2217 if (!MaskedValueIsZero(XorLHS,
2218 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2219 Size = 0; // Not a sign ext, but can't be any others either.
2224 C0080Val = APIntOps::lshr(C0080Val, Size);
2225 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2226 } while (Size >= 1);
2228 // FIXME: This shouldn't be necessary. When the backends can handle types
2229 // with funny bit widths then this switch statement should be removed. It
2230 // is just here to get the size of the "middle" type back up to something
2231 // that the back ends can handle.
2232 const Type *MiddleType = 0;
2235 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2236 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2237 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2240 Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext");
2241 return new SExtInst(NewTrunc, I.getType(), I.getName());
2246 if (I.getType() == Type::getInt1Ty(*Context))
2247 return BinaryOperator::CreateXor(LHS, RHS);
2250 if (I.getType()->isInteger()) {
2251 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2254 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2255 if (RHSI->getOpcode() == Instruction::Sub)
2256 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2257 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2259 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2260 if (LHSI->getOpcode() == Instruction::Sub)
2261 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2262 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2267 // -A + -B --> -(A + B)
2268 if (Value *LHSV = dyn_castNegVal(LHS)) {
2269 if (LHS->getType()->isIntOrIntVector()) {
2270 if (Value *RHSV = dyn_castNegVal(RHS)) {
2271 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
2272 return BinaryOperator::CreateNeg(NewAdd);
2276 return BinaryOperator::CreateSub(RHS, LHSV);
2280 if (!isa<Constant>(RHS))
2281 if (Value *V = dyn_castNegVal(RHS))
2282 return BinaryOperator::CreateSub(LHS, V);
2286 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2287 if (X == RHS) // X*C + X --> X * (C+1)
2288 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2290 // X*C1 + X*C2 --> X * (C1+C2)
2292 if (X == dyn_castFoldableMul(RHS, C1))
2293 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2296 // X + X*C --> X * (C+1)
2297 if (dyn_castFoldableMul(RHS, C2) == LHS)
2298 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2300 // X + ~X --> -1 since ~X = -X-1
2301 if (dyn_castNotVal(LHS) == RHS ||
2302 dyn_castNotVal(RHS) == LHS)
2303 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2306 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2307 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2308 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2311 // A+B --> A|B iff A and B have no bits set in common.
2312 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2313 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2314 APInt LHSKnownOne(IT->getBitWidth(), 0);
2315 APInt LHSKnownZero(IT->getBitWidth(), 0);
2316 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2317 if (LHSKnownZero != 0) {
2318 APInt RHSKnownOne(IT->getBitWidth(), 0);
2319 APInt RHSKnownZero(IT->getBitWidth(), 0);
2320 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2322 // No bits in common -> bitwise or.
2323 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2324 return BinaryOperator::CreateOr(LHS, RHS);
2328 // W*X + Y*Z --> W * (X+Z) iff W == Y
2329 if (I.getType()->isIntOrIntVector()) {
2330 Value *W, *X, *Y, *Z;
2331 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2332 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2336 } else if (Y == X) {
2338 } else if (X == Z) {
2345 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
2346 return BinaryOperator::CreateMul(W, NewAdd);
2351 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2353 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2354 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2356 // (X & FF00) + xx00 -> (X+xx00) & FF00
2357 if (LHS->hasOneUse() &&
2358 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2359 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2360 if (Anded == CRHS) {
2361 // See if all bits from the first bit set in the Add RHS up are included
2362 // in the mask. First, get the rightmost bit.
2363 const APInt& AddRHSV = CRHS->getValue();
2365 // Form a mask of all bits from the lowest bit added through the top.
2366 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2368 // See if the and mask includes all of these bits.
2369 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2371 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2372 // Okay, the xform is safe. Insert the new add pronto.
2373 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
2374 return BinaryOperator::CreateAnd(NewAdd, C2);
2379 // Try to fold constant add into select arguments.
2380 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2381 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2385 // add (select X 0 (sub n A)) A --> select X A n
2387 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2390 SI = dyn_cast<SelectInst>(RHS);
2393 if (SI && SI->hasOneUse()) {
2394 Value *TV = SI->getTrueValue();
2395 Value *FV = SI->getFalseValue();
2398 // Can we fold the add into the argument of the select?
2399 // We check both true and false select arguments for a matching subtract.
2400 if (match(FV, m_Zero()) &&
2401 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2402 // Fold the add into the true select value.
2403 return SelectInst::Create(SI->getCondition(), N, A);
2404 if (match(TV, m_Zero()) &&
2405 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2406 // Fold the add into the false select value.
2407 return SelectInst::Create(SI->getCondition(), A, N);
2411 // Check for (add (sext x), y), see if we can merge this into an
2412 // integer add followed by a sext.
2413 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2414 // (add (sext x), cst) --> (sext (add x, cst'))
2415 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2417 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2418 if (LHSConv->hasOneUse() &&
2419 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2420 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2421 // Insert the new, smaller add.
2422 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2424 return new SExtInst(NewAdd, I.getType());
2428 // (add (sext x), (sext y)) --> (sext (add int x, y))
2429 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2430 // Only do this if x/y have the same type, if at last one of them has a
2431 // single use (so we don't increase the number of sexts), and if the
2432 // integer add will not overflow.
2433 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2434 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2435 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2436 RHSConv->getOperand(0))) {
2437 // Insert the new integer add.
2438 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2439 RHSConv->getOperand(0), "addconv");
2440 return new SExtInst(NewAdd, I.getType());
2445 return Changed ? &I : 0;
2448 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2449 bool Changed = SimplifyCommutative(I);
2450 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2452 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2454 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2455 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2456 (I.getType())->getValueAPF()))
2457 return ReplaceInstUsesWith(I, LHS);
2460 if (isa<PHINode>(LHS))
2461 if (Instruction *NV = FoldOpIntoPhi(I))
2466 // -A + -B --> -(A + B)
2467 if (Value *LHSV = dyn_castFNegVal(LHS))
2468 return BinaryOperator::CreateFSub(RHS, LHSV);
2471 if (!isa<Constant>(RHS))
2472 if (Value *V = dyn_castFNegVal(RHS))
2473 return BinaryOperator::CreateFSub(LHS, V);
2475 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2476 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2477 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2478 return ReplaceInstUsesWith(I, LHS);
2480 // Check for (add double (sitofp x), y), see if we can merge this into an
2481 // integer add followed by a promotion.
2482 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2483 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2484 // ... if the constant fits in the integer value. This is useful for things
2485 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2486 // requires a constant pool load, and generally allows the add to be better
2488 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2490 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2491 if (LHSConv->hasOneUse() &&
2492 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2493 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2494 // Insert the new integer add.
2495 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2497 return new SIToFPInst(NewAdd, I.getType());
2501 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2502 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2503 // Only do this if x/y have the same type, if at last one of them has a
2504 // single use (so we don't increase the number of int->fp conversions),
2505 // and if the integer add will not overflow.
2506 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2507 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2508 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2509 RHSConv->getOperand(0))) {
2510 // Insert the new integer add.
2511 Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
2512 RHSConv->getOperand(0), "addconv");
2513 return new SIToFPInst(NewAdd, I.getType());
2518 return Changed ? &I : 0;
2521 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2522 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2524 if (Op0 == Op1) // sub X, X -> 0
2525 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2527 // If this is a 'B = x-(-A)', change to B = x+A...
2528 if (Value *V = dyn_castNegVal(Op1))
2529 return BinaryOperator::CreateAdd(Op0, V);
2531 if (isa<UndefValue>(Op0))
2532 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2533 if (isa<UndefValue>(Op1))
2534 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2536 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2537 // Replace (-1 - A) with (~A)...
2538 if (C->isAllOnesValue())
2539 return BinaryOperator::CreateNot(Op1);
2541 // C - ~X == X + (1+C)
2543 if (match(Op1, m_Not(m_Value(X))))
2544 return BinaryOperator::CreateAdd(X, AddOne(C));
2546 // -(X >>u 31) -> (X >>s 31)
2547 // -(X >>s 31) -> (X >>u 31)
2549 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2550 if (SI->getOpcode() == Instruction::LShr) {
2551 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2552 // Check to see if we are shifting out everything but the sign bit.
2553 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2554 SI->getType()->getPrimitiveSizeInBits()-1) {
2555 // Ok, the transformation is safe. Insert AShr.
2556 return BinaryOperator::Create(Instruction::AShr,
2557 SI->getOperand(0), CU, SI->getName());
2561 else if (SI->getOpcode() == Instruction::AShr) {
2562 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2563 // Check to see if we are shifting out everything but the sign bit.
2564 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2565 SI->getType()->getPrimitiveSizeInBits()-1) {
2566 // Ok, the transformation is safe. Insert LShr.
2567 return BinaryOperator::CreateLShr(
2568 SI->getOperand(0), CU, SI->getName());
2575 // Try to fold constant sub into select arguments.
2576 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2577 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2580 // C - zext(bool) -> bool ? C - 1 : C
2581 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2582 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2583 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2586 if (I.getType() == Type::getInt1Ty(*Context))
2587 return BinaryOperator::CreateXor(Op0, Op1);
2589 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2590 if (Op1I->getOpcode() == Instruction::Add) {
2591 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2592 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2594 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2595 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2597 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2598 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2599 // C1-(X+C2) --> (C1-C2)-X
2600 return BinaryOperator::CreateSub(
2601 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2605 if (Op1I->hasOneUse()) {
2606 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2607 // is not used by anyone else...
2609 if (Op1I->getOpcode() == Instruction::Sub) {
2610 // Swap the two operands of the subexpr...
2611 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2612 Op1I->setOperand(0, IIOp1);
2613 Op1I->setOperand(1, IIOp0);
2615 // Create the new top level add instruction...
2616 return BinaryOperator::CreateAdd(Op0, Op1);
2619 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2621 if (Op1I->getOpcode() == Instruction::And &&
2622 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2623 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2625 Value *NewNot = Builder->CreateNot(OtherOp, "B.not");
2626 return BinaryOperator::CreateAnd(Op0, NewNot);
2629 // 0 - (X sdiv C) -> (X sdiv -C)
2630 if (Op1I->getOpcode() == Instruction::SDiv)
2631 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2633 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2634 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2635 ConstantExpr::getNeg(DivRHS));
2637 // X - X*C --> X * (1-C)
2638 ConstantInt *C2 = 0;
2639 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2641 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2643 return BinaryOperator::CreateMul(Op0, CP1);
2648 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2649 if (Op0I->getOpcode() == Instruction::Add) {
2650 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2651 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2652 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2653 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2654 } else if (Op0I->getOpcode() == Instruction::Sub) {
2655 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2656 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2662 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2663 if (X == Op1) // X*C - X --> X * (C-1)
2664 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2666 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2667 if (X == dyn_castFoldableMul(Op1, C2))
2668 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2673 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2674 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2676 // If this is a 'B = x-(-A)', change to B = x+A...
2677 if (Value *V = dyn_castFNegVal(Op1))
2678 return BinaryOperator::CreateFAdd(Op0, V);
2680 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2681 if (Op1I->getOpcode() == Instruction::FAdd) {
2682 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2683 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2685 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2686 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2694 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2695 /// comparison only checks the sign bit. If it only checks the sign bit, set
2696 /// TrueIfSigned if the result of the comparison is true when the input value is
2698 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2699 bool &TrueIfSigned) {
2701 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2702 TrueIfSigned = true;
2703 return RHS->isZero();
2704 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2705 TrueIfSigned = true;
2706 return RHS->isAllOnesValue();
2707 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2708 TrueIfSigned = false;
2709 return RHS->isAllOnesValue();
2710 case ICmpInst::ICMP_UGT:
2711 // True if LHS u> RHS and RHS == high-bit-mask - 1
2712 TrueIfSigned = true;
2713 return RHS->getValue() ==
2714 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2715 case ICmpInst::ICMP_UGE:
2716 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2717 TrueIfSigned = true;
2718 return RHS->getValue().isSignBit();
2724 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2725 bool Changed = SimplifyCommutative(I);
2726 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2728 if (isa<UndefValue>(Op1)) // undef * X -> 0
2729 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2731 // Simplify mul instructions with a constant RHS.
2732 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2733 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1C)) {
2735 // ((X << C1)*C2) == (X * (C2 << C1))
2736 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2737 if (SI->getOpcode() == Instruction::Shl)
2738 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2739 return BinaryOperator::CreateMul(SI->getOperand(0),
2740 ConstantExpr::getShl(CI, ShOp));
2743 return ReplaceInstUsesWith(I, Op1C); // X * 0 == 0
2744 if (CI->equalsInt(1)) // X * 1 == X
2745 return ReplaceInstUsesWith(I, Op0);
2746 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2747 return BinaryOperator::CreateNeg(Op0, I.getName());
2749 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2750 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2751 return BinaryOperator::CreateShl(Op0,
2752 ConstantInt::get(Op0->getType(), Val.logBase2()));
2754 } else if (isa<VectorType>(Op1C->getType())) {
2755 if (Op1C->isNullValue())
2756 return ReplaceInstUsesWith(I, Op1C);
2758 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
2759 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2760 return BinaryOperator::CreateNeg(Op0, I.getName());
2762 // As above, vector X*splat(1.0) -> X in all defined cases.
2763 if (Constant *Splat = Op1V->getSplatValue()) {
2764 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2765 if (CI->equalsInt(1))
2766 return ReplaceInstUsesWith(I, Op0);
2771 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2772 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2773 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1C)) {
2774 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2775 Value *Add = Builder->CreateMul(Op0I->getOperand(0), Op1C, "tmp");
2776 Value *C1C2 = Builder->CreateMul(Op1C, Op0I->getOperand(1));
2777 return BinaryOperator::CreateAdd(Add, C1C2);
2781 // Try to fold constant mul into select arguments.
2782 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2783 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2786 if (isa<PHINode>(Op0))
2787 if (Instruction *NV = FoldOpIntoPhi(I))
2791 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2792 if (Value *Op1v = dyn_castNegVal(Op1))
2793 return BinaryOperator::CreateMul(Op0v, Op1v);
2795 // (X / Y) * Y = X - (X % Y)
2796 // (X / Y) * -Y = (X % Y) - X
2799 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2801 (BO->getOpcode() != Instruction::UDiv &&
2802 BO->getOpcode() != Instruction::SDiv)) {
2804 BO = dyn_cast<BinaryOperator>(Op1);
2806 Value *Neg = dyn_castNegVal(Op1C);
2807 if (BO && BO->hasOneUse() &&
2808 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
2809 (BO->getOpcode() == Instruction::UDiv ||
2810 BO->getOpcode() == Instruction::SDiv)) {
2811 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2813 // If the division is exact, X % Y is zero.
2814 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
2815 if (SDiv->isExact()) {
2817 return ReplaceInstUsesWith(I, Op0BO);
2818 return BinaryOperator::CreateNeg(Op0BO);
2822 if (BO->getOpcode() == Instruction::UDiv)
2823 Rem = Builder->CreateURem(Op0BO, Op1BO);
2825 Rem = Builder->CreateSRem(Op0BO, Op1BO);
2829 return BinaryOperator::CreateSub(Op0BO, Rem);
2830 return BinaryOperator::CreateSub(Rem, Op0BO);
2834 /// i1 mul -> i1 and.
2835 if (I.getType() == Type::getInt1Ty(*Context))
2836 return BinaryOperator::CreateAnd(Op0, Op1);
2838 // X*(1 << Y) --> X << Y
2839 // (1 << Y)*X --> X << Y
2842 if (match(Op0, m_Shl(m_One(), m_Value(Y))))
2843 return BinaryOperator::CreateShl(Op1, Y);
2844 if (match(Op1, m_Shl(m_One(), m_Value(Y))))
2845 return BinaryOperator::CreateShl(Op0, Y);
2848 // If one of the operands of the multiply is a cast from a boolean value, then
2849 // we know the bool is either zero or one, so this is a 'masking' multiply.
2850 // X * Y (where Y is 0 or 1) -> X & (0-Y)
2851 if (!isa<VectorType>(I.getType())) {
2852 // -2 is "-1 << 1" so it is all bits set except the low one.
2853 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
2855 Value *BoolCast = 0, *OtherOp = 0;
2856 if (MaskedValueIsZero(Op0, Negative2))
2857 BoolCast = Op0, OtherOp = Op1;
2858 else if (MaskedValueIsZero(Op1, Negative2))
2859 BoolCast = Op1, OtherOp = Op0;
2862 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
2864 return BinaryOperator::CreateAnd(V, OtherOp);
2868 return Changed ? &I : 0;
2871 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2872 bool Changed = SimplifyCommutative(I);
2873 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2875 // Simplify mul instructions with a constant RHS...
2876 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2877 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1C)) {
2878 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2879 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2880 if (Op1F->isExactlyValue(1.0))
2881 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2882 } else if (isa<VectorType>(Op1C->getType())) {
2883 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
2884 // As above, vector X*splat(1.0) -> X in all defined cases.
2885 if (Constant *Splat = Op1V->getSplatValue()) {
2886 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2887 if (F->isExactlyValue(1.0))
2888 return ReplaceInstUsesWith(I, Op0);
2893 // Try to fold constant mul into select arguments.
2894 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2895 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2898 if (isa<PHINode>(Op0))
2899 if (Instruction *NV = FoldOpIntoPhi(I))
2903 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
2904 if (Value *Op1v = dyn_castFNegVal(Op1))
2905 return BinaryOperator::CreateFMul(Op0v, Op1v);
2907 return Changed ? &I : 0;
2910 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2912 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2913 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2915 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2916 int NonNullOperand = -1;
2917 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2918 if (ST->isNullValue())
2920 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2921 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2922 if (ST->isNullValue())
2925 if (NonNullOperand == -1)
2928 Value *SelectCond = SI->getOperand(0);
2930 // Change the div/rem to use 'Y' instead of the select.
2931 I.setOperand(1, SI->getOperand(NonNullOperand));
2933 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2934 // problem. However, the select, or the condition of the select may have
2935 // multiple uses. Based on our knowledge that the operand must be non-zero,
2936 // propagate the known value for the select into other uses of it, and
2937 // propagate a known value of the condition into its other users.
2939 // If the select and condition only have a single use, don't bother with this,
2941 if (SI->use_empty() && SelectCond->hasOneUse())
2944 // Scan the current block backward, looking for other uses of SI.
2945 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2947 while (BBI != BBFront) {
2949 // If we found a call to a function, we can't assume it will return, so
2950 // information from below it cannot be propagated above it.
2951 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2954 // Replace uses of the select or its condition with the known values.
2955 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2958 *I = SI->getOperand(NonNullOperand);
2960 } else if (*I == SelectCond) {
2961 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2962 ConstantInt::getFalse(*Context);
2967 // If we past the instruction, quit looking for it.
2970 if (&*BBI == SelectCond)
2973 // If we ran out of things to eliminate, break out of the loop.
2974 if (SelectCond == 0 && SI == 0)
2982 /// This function implements the transforms on div instructions that work
2983 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2984 /// used by the visitors to those instructions.
2985 /// @brief Transforms common to all three div instructions
2986 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2987 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2989 // undef / X -> 0 for integer.
2990 // undef / X -> undef for FP (the undef could be a snan).
2991 if (isa<UndefValue>(Op0)) {
2992 if (Op0->getType()->isFPOrFPVector())
2993 return ReplaceInstUsesWith(I, Op0);
2994 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2997 // X / undef -> undef
2998 if (isa<UndefValue>(Op1))
2999 return ReplaceInstUsesWith(I, Op1);
3004 /// This function implements the transforms common to both integer division
3005 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3006 /// division instructions.
3007 /// @brief Common integer divide transforms
3008 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3009 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3011 // (sdiv X, X) --> 1 (udiv X, X) --> 1
3013 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
3014 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
3015 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
3016 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
3019 Constant *CI = ConstantInt::get(I.getType(), 1);
3020 return ReplaceInstUsesWith(I, CI);
3023 if (Instruction *Common = commonDivTransforms(I))
3026 // Handle cases involving: [su]div X, (select Cond, Y, Z)
3027 // This does not apply for fdiv.
3028 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3031 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3033 if (RHS->equalsInt(1))
3034 return ReplaceInstUsesWith(I, Op0);
3036 // (X / C1) / C2 -> X / (C1*C2)
3037 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3038 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3039 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3040 if (MultiplyOverflows(RHS, LHSRHS,
3041 I.getOpcode()==Instruction::SDiv))
3042 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3044 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3045 ConstantExpr::getMul(RHS, LHSRHS));
3048 if (!RHS->isZero()) { // avoid X udiv 0
3049 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3050 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3052 if (isa<PHINode>(Op0))
3053 if (Instruction *NV = FoldOpIntoPhi(I))
3058 // 0 / X == 0, we don't need to preserve faults!
3059 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3060 if (LHS->equalsInt(0))
3061 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3063 // It can't be division by zero, hence it must be division by one.
3064 if (I.getType() == Type::getInt1Ty(*Context))
3065 return ReplaceInstUsesWith(I, Op0);
3067 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3068 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3071 return ReplaceInstUsesWith(I, Op0);
3077 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3078 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3080 // Handle the integer div common cases
3081 if (Instruction *Common = commonIDivTransforms(I))
3084 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3085 // X udiv C^2 -> X >> C
3086 // Check to see if this is an unsigned division with an exact power of 2,
3087 // if so, convert to a right shift.
3088 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3089 return BinaryOperator::CreateLShr(Op0,
3090 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3092 // X udiv C, where C >= signbit
3093 if (C->getValue().isNegative()) {
3094 Value *IC = Builder->CreateICmpULT( Op0, C);
3095 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3096 ConstantInt::get(I.getType(), 1));
3100 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3101 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3102 if (RHSI->getOpcode() == Instruction::Shl &&
3103 isa<ConstantInt>(RHSI->getOperand(0))) {
3104 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3105 if (C1.isPowerOf2()) {
3106 Value *N = RHSI->getOperand(1);
3107 const Type *NTy = N->getType();
3108 if (uint32_t C2 = C1.logBase2())
3109 N = Builder->CreateAdd(N, ConstantInt::get(NTy, C2), "tmp");
3110 return BinaryOperator::CreateLShr(Op0, N);
3115 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3116 // where C1&C2 are powers of two.
3117 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3118 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3119 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3120 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3121 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3122 // Compute the shift amounts
3123 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3124 // Construct the "on true" case of the select
3125 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3126 Value *TSI = Builder->CreateLShr(Op0, TC, SI->getName()+".t");
3128 // Construct the "on false" case of the select
3129 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3130 Value *FSI = Builder->CreateLShr(Op0, FC, SI->getName()+".f");
3132 // construct the select instruction and return it.
3133 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3139 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3140 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3142 // Handle the integer div common cases
3143 if (Instruction *Common = commonIDivTransforms(I))
3146 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3148 if (RHS->isAllOnesValue())
3149 return BinaryOperator::CreateNeg(Op0);
3151 // sdiv X, C --> ashr X, log2(C)
3152 if (cast<SDivOperator>(&I)->isExact() &&
3153 RHS->getValue().isNonNegative() &&
3154 RHS->getValue().isPowerOf2()) {
3155 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3156 RHS->getValue().exactLogBase2());
3157 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3160 // -X/C --> X/-C provided the negation doesn't overflow.
3161 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3162 if (isa<Constant>(Sub->getOperand(0)) &&
3163 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3164 Sub->hasNoSignedWrap())
3165 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3166 ConstantExpr::getNeg(RHS));
3169 // If the sign bits of both operands are zero (i.e. we can prove they are
3170 // unsigned inputs), turn this into a udiv.
3171 if (I.getType()->isInteger()) {
3172 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3173 if (MaskedValueIsZero(Op0, Mask)) {
3174 if (MaskedValueIsZero(Op1, Mask)) {
3175 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3176 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3178 ConstantInt *ShiftedInt;
3179 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3180 ShiftedInt->getValue().isPowerOf2()) {
3181 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3182 // Safe because the only negative value (1 << Y) can take on is
3183 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3184 // the sign bit set.
3185 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3193 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3194 return commonDivTransforms(I);
3197 /// This function implements the transforms on rem instructions that work
3198 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3199 /// is used by the visitors to those instructions.
3200 /// @brief Transforms common to all three rem instructions
3201 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3202 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3204 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3205 if (I.getType()->isFPOrFPVector())
3206 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3207 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3209 if (isa<UndefValue>(Op1))
3210 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3212 // Handle cases involving: rem X, (select Cond, Y, Z)
3213 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3219 /// This function implements the transforms common to both integer remainder
3220 /// instructions (urem and srem). It is called by the visitors to those integer
3221 /// remainder instructions.
3222 /// @brief Common integer remainder transforms
3223 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3224 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3226 if (Instruction *common = commonRemTransforms(I))
3229 // 0 % X == 0 for integer, we don't need to preserve faults!
3230 if (Constant *LHS = dyn_cast<Constant>(Op0))
3231 if (LHS->isNullValue())
3232 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3234 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3235 // X % 0 == undef, we don't need to preserve faults!
3236 if (RHS->equalsInt(0))
3237 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3239 if (RHS->equalsInt(1)) // X % 1 == 0
3240 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3242 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3243 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3244 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3246 } else if (isa<PHINode>(Op0I)) {
3247 if (Instruction *NV = FoldOpIntoPhi(I))
3251 // See if we can fold away this rem instruction.
3252 if (SimplifyDemandedInstructionBits(I))
3260 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3261 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3263 if (Instruction *common = commonIRemTransforms(I))
3266 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3267 // X urem C^2 -> X and C
3268 // Check to see if this is an unsigned remainder with an exact power of 2,
3269 // if so, convert to a bitwise and.
3270 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3271 if (C->getValue().isPowerOf2())
3272 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3275 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3276 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3277 if (RHSI->getOpcode() == Instruction::Shl &&
3278 isa<ConstantInt>(RHSI->getOperand(0))) {
3279 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3280 Constant *N1 = Constant::getAllOnesValue(I.getType());
3281 Value *Add = Builder->CreateAdd(RHSI, N1, "tmp");
3282 return BinaryOperator::CreateAnd(Op0, Add);
3287 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3288 // where C1&C2 are powers of two.
3289 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3290 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3291 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3292 // STO == 0 and SFO == 0 handled above.
3293 if ((STO->getValue().isPowerOf2()) &&
3294 (SFO->getValue().isPowerOf2())) {
3295 Value *TrueAnd = Builder->CreateAnd(Op0, SubOne(STO),
3296 SI->getName()+".t");
3297 Value *FalseAnd = Builder->CreateAnd(Op0, SubOne(SFO),
3298 SI->getName()+".f");
3299 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3307 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3308 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3310 // Handle the integer rem common cases
3311 if (Instruction *Common = commonIRemTransforms(I))
3314 if (Value *RHSNeg = dyn_castNegVal(Op1))
3315 if (!isa<Constant>(RHSNeg) ||
3316 (isa<ConstantInt>(RHSNeg) &&
3317 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3319 Worklist.AddValue(I.getOperand(1));
3320 I.setOperand(1, RHSNeg);
3324 // If the sign bits of both operands are zero (i.e. we can prove they are
3325 // unsigned inputs), turn this into a urem.
3326 if (I.getType()->isInteger()) {
3327 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3328 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3329 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3330 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3334 // If it's a constant vector, flip any negative values positive.
3335 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3336 unsigned VWidth = RHSV->getNumOperands();
3338 bool hasNegative = false;
3339 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3340 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3341 if (RHS->getValue().isNegative())
3345 std::vector<Constant *> Elts(VWidth);
3346 for (unsigned i = 0; i != VWidth; ++i) {
3347 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3348 if (RHS->getValue().isNegative())
3349 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3355 Constant *NewRHSV = ConstantVector::get(Elts);
3356 if (NewRHSV != RHSV) {
3357 Worklist.AddValue(I.getOperand(1));
3358 I.setOperand(1, NewRHSV);
3367 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3368 return commonRemTransforms(I);
3371 // isOneBitSet - Return true if there is exactly one bit set in the specified
3373 static bool isOneBitSet(const ConstantInt *CI) {
3374 return CI->getValue().isPowerOf2();
3377 // isHighOnes - Return true if the constant is of the form 1+0+.
3378 // This is the same as lowones(~X).
3379 static bool isHighOnes(const ConstantInt *CI) {
3380 return (~CI->getValue() + 1).isPowerOf2();
3383 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3384 /// are carefully arranged to allow folding of expressions such as:
3386 /// (A < B) | (A > B) --> (A != B)
3388 /// Note that this is only valid if the first and second predicates have the
3389 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3391 /// Three bits are used to represent the condition, as follows:
3396 /// <=> Value Definition
3397 /// 000 0 Always false
3404 /// 111 7 Always true
3406 static unsigned getICmpCode(const ICmpInst *ICI) {
3407 switch (ICI->getPredicate()) {
3409 case ICmpInst::ICMP_UGT: return 1; // 001
3410 case ICmpInst::ICMP_SGT: return 1; // 001
3411 case ICmpInst::ICMP_EQ: return 2; // 010
3412 case ICmpInst::ICMP_UGE: return 3; // 011
3413 case ICmpInst::ICMP_SGE: return 3; // 011
3414 case ICmpInst::ICMP_ULT: return 4; // 100
3415 case ICmpInst::ICMP_SLT: return 4; // 100
3416 case ICmpInst::ICMP_NE: return 5; // 101
3417 case ICmpInst::ICMP_ULE: return 6; // 110
3418 case ICmpInst::ICMP_SLE: return 6; // 110
3421 llvm_unreachable("Invalid ICmp predicate!");
3426 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3427 /// predicate into a three bit mask. It also returns whether it is an ordered
3428 /// predicate by reference.
3429 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3432 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3433 case FCmpInst::FCMP_UNO: return 0; // 000
3434 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3435 case FCmpInst::FCMP_UGT: return 1; // 001
3436 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3437 case FCmpInst::FCMP_UEQ: return 2; // 010
3438 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3439 case FCmpInst::FCMP_UGE: return 3; // 011
3440 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3441 case FCmpInst::FCMP_ULT: return 4; // 100
3442 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3443 case FCmpInst::FCMP_UNE: return 5; // 101
3444 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3445 case FCmpInst::FCMP_ULE: return 6; // 110
3448 // Not expecting FCMP_FALSE and FCMP_TRUE;
3449 llvm_unreachable("Unexpected FCmp predicate!");
3454 /// getICmpValue - This is the complement of getICmpCode, which turns an
3455 /// opcode and two operands into either a constant true or false, or a brand
3456 /// new ICmp instruction. The sign is passed in to determine which kind
3457 /// of predicate to use in the new icmp instruction.
3458 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3459 LLVMContext *Context) {
3461 default: llvm_unreachable("Illegal ICmp code!");
3462 case 0: return ConstantInt::getFalse(*Context);
3465 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3467 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3468 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3471 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3473 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3476 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3478 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3479 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3482 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3484 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3485 case 7: return ConstantInt::getTrue(*Context);
3489 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3490 /// opcode and two operands into either a FCmp instruction. isordered is passed
3491 /// in to determine which kind of predicate to use in the new fcmp instruction.
3492 static Value *getFCmpValue(bool isordered, unsigned code,
3493 Value *LHS, Value *RHS, LLVMContext *Context) {
3495 default: llvm_unreachable("Illegal FCmp code!");
3498 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3500 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3503 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3505 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3508 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3510 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3513 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3515 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3518 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3520 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3523 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3525 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3528 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3530 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3531 case 7: return ConstantInt::getTrue(*Context);
3535 /// PredicatesFoldable - Return true if both predicates match sign or if at
3536 /// least one of them is an equality comparison (which is signless).
3537 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3538 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
3539 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
3540 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
3544 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3545 struct FoldICmpLogical {
3548 ICmpInst::Predicate pred;
3549 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3550 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3551 pred(ICI->getPredicate()) {}
3552 bool shouldApply(Value *V) const {
3553 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3554 if (PredicatesFoldable(pred, ICI->getPredicate()))
3555 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3556 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3559 Instruction *apply(Instruction &Log) const {
3560 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3561 if (ICI->getOperand(0) != LHS) {
3562 assert(ICI->getOperand(1) == LHS);
3563 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3566 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3567 unsigned LHSCode = getICmpCode(ICI);
3568 unsigned RHSCode = getICmpCode(RHSICI);
3570 switch (Log.getOpcode()) {
3571 case Instruction::And: Code = LHSCode & RHSCode; break;
3572 case Instruction::Or: Code = LHSCode | RHSCode; break;
3573 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3574 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3577 bool isSigned = RHSICI->isSigned() || ICI->isSigned();
3578 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3579 if (Instruction *I = dyn_cast<Instruction>(RV))
3581 // Otherwise, it's a constant boolean value...
3582 return IC.ReplaceInstUsesWith(Log, RV);
3585 } // end anonymous namespace
3587 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3588 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3589 // guaranteed to be a binary operator.
3590 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3592 ConstantInt *AndRHS,
3593 BinaryOperator &TheAnd) {
3594 Value *X = Op->getOperand(0);
3595 Constant *Together = 0;
3597 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3599 switch (Op->getOpcode()) {
3600 case Instruction::Xor:
3601 if (Op->hasOneUse()) {
3602 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3603 Value *And = Builder->CreateAnd(X, AndRHS);
3605 return BinaryOperator::CreateXor(And, Together);
3608 case Instruction::Or:
3609 if (Together == AndRHS) // (X | C) & C --> C
3610 return ReplaceInstUsesWith(TheAnd, AndRHS);
3612 if (Op->hasOneUse() && Together != OpRHS) {
3613 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3614 Value *Or = Builder->CreateOr(X, Together);
3616 return BinaryOperator::CreateAnd(Or, AndRHS);
3619 case Instruction::Add:
3620 if (Op->hasOneUse()) {
3621 // Adding a one to a single bit bit-field should be turned into an XOR
3622 // of the bit. First thing to check is to see if this AND is with a
3623 // single bit constant.
3624 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3626 // If there is only one bit set...
3627 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3628 // Ok, at this point, we know that we are masking the result of the
3629 // ADD down to exactly one bit. If the constant we are adding has
3630 // no bits set below this bit, then we can eliminate the ADD.
3631 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3633 // Check to see if any bits below the one bit set in AndRHSV are set.
3634 if ((AddRHS & (AndRHSV-1)) == 0) {
3635 // If not, the only thing that can effect the output of the AND is
3636 // the bit specified by AndRHSV. If that bit is set, the effect of
3637 // the XOR is to toggle the bit. If it is clear, then the ADD has
3639 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3640 TheAnd.setOperand(0, X);
3643 // Pull the XOR out of the AND.
3644 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
3645 NewAnd->takeName(Op);
3646 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3653 case Instruction::Shl: {
3654 // We know that the AND will not produce any of the bits shifted in, so if
3655 // the anded constant includes them, clear them now!
3657 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3658 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3659 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3660 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3662 if (CI->getValue() == ShlMask) {
3663 // Masking out bits that the shift already masks
3664 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3665 } else if (CI != AndRHS) { // Reducing bits set in and.
3666 TheAnd.setOperand(1, CI);
3671 case Instruction::LShr:
3673 // We know that the AND will not produce any of the bits shifted in, so if
3674 // the anded constant includes them, clear them now! This only applies to
3675 // unsigned shifts, because a signed shr may bring in set bits!
3677 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3678 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3679 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3680 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3682 if (CI->getValue() == ShrMask) {
3683 // Masking out bits that the shift already masks.
3684 return ReplaceInstUsesWith(TheAnd, Op);
3685 } else if (CI != AndRHS) {
3686 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3691 case Instruction::AShr:
3693 // See if this is shifting in some sign extension, then masking it out
3695 if (Op->hasOneUse()) {
3696 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3697 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3698 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3699 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3700 if (C == AndRHS) { // Masking out bits shifted in.
3701 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3702 // Make the argument unsigned.
3703 Value *ShVal = Op->getOperand(0);
3704 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
3705 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3714 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3715 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3716 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3717 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3718 /// insert new instructions.
3719 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3720 bool isSigned, bool Inside,
3722 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3723 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3724 "Lo is not <= Hi in range emission code!");
3727 if (Lo == Hi) // Trivially false.
3728 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3730 // V >= Min && V < Hi --> V < Hi
3731 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3732 ICmpInst::Predicate pred = (isSigned ?
3733 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3734 return new ICmpInst(pred, V, Hi);
3737 // Emit V-Lo <u Hi-Lo
3738 Constant *NegLo = ConstantExpr::getNeg(Lo);
3739 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3740 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3741 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3744 if (Lo == Hi) // Trivially true.
3745 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3747 // V < Min || V >= Hi -> V > Hi-1
3748 Hi = SubOne(cast<ConstantInt>(Hi));
3749 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3750 ICmpInst::Predicate pred = (isSigned ?
3751 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3752 return new ICmpInst(pred, V, Hi);
3755 // Emit V-Lo >u Hi-1-Lo
3756 // Note that Hi has already had one subtracted from it, above.
3757 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3758 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3759 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3760 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3763 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3764 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3765 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3766 // not, since all 1s are not contiguous.
3767 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3768 const APInt& V = Val->getValue();
3769 uint32_t BitWidth = Val->getType()->getBitWidth();
3770 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3772 // look for the first zero bit after the run of ones
3773 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3774 // look for the first non-zero bit
3775 ME = V.getActiveBits();
3779 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3780 /// where isSub determines whether the operator is a sub. If we can fold one of
3781 /// the following xforms:
3783 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3784 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3785 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3787 /// return (A +/- B).
3789 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3790 ConstantInt *Mask, bool isSub,
3792 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3793 if (!LHSI || LHSI->getNumOperands() != 2 ||
3794 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3796 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3798 switch (LHSI->getOpcode()) {
3800 case Instruction::And:
3801 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3802 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3803 if ((Mask->getValue().countLeadingZeros() +
3804 Mask->getValue().countPopulation()) ==
3805 Mask->getValue().getBitWidth())
3808 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3809 // part, we don't need any explicit masks to take them out of A. If that
3810 // is all N is, ignore it.
3811 uint32_t MB = 0, ME = 0;
3812 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3813 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3814 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3815 if (MaskedValueIsZero(RHS, Mask))
3820 case Instruction::Or:
3821 case Instruction::Xor:
3822 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3823 if ((Mask->getValue().countLeadingZeros() +
3824 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3825 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3831 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
3832 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
3835 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3836 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3837 ICmpInst *LHS, ICmpInst *RHS) {
3839 ConstantInt *LHSCst, *RHSCst;
3840 ICmpInst::Predicate LHSCC, RHSCC;
3842 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3843 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3844 m_ConstantInt(LHSCst))) ||
3845 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3846 m_ConstantInt(RHSCst))))
3849 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3850 // where C is a power of 2
3851 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3852 LHSCst->getValue().isPowerOf2()) {
3853 Value *NewOr = Builder->CreateOr(Val, Val2);
3854 return new ICmpInst(LHSCC, NewOr, LHSCst);
3857 // From here on, we only handle:
3858 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3859 if (Val != Val2) return 0;
3861 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3862 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3863 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3864 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3865 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3868 // We can't fold (ugt x, C) & (sgt x, C2).
3869 if (!PredicatesFoldable(LHSCC, RHSCC))
3872 // Ensure that the larger constant is on the RHS.
3874 if (CmpInst::isSigned(LHSCC) ||
3875 (ICmpInst::isEquality(LHSCC) &&
3876 CmpInst::isSigned(RHSCC)))
3877 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3879 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3882 std::swap(LHS, RHS);
3883 std::swap(LHSCst, RHSCst);
3884 std::swap(LHSCC, RHSCC);
3887 // At this point, we know we have have two icmp instructions
3888 // comparing a value against two constants and and'ing the result
3889 // together. Because of the above check, we know that we only have
3890 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3891 // (from the FoldICmpLogical check above), that the two constants
3892 // are not equal and that the larger constant is on the RHS
3893 assert(LHSCst != RHSCst && "Compares not folded above?");
3896 default: llvm_unreachable("Unknown integer condition code!");
3897 case ICmpInst::ICMP_EQ:
3899 default: llvm_unreachable("Unknown integer condition code!");
3900 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3901 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3902 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3903 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3904 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3905 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3906 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3907 return ReplaceInstUsesWith(I, LHS);
3909 case ICmpInst::ICMP_NE:
3911 default: llvm_unreachable("Unknown integer condition code!");
3912 case ICmpInst::ICMP_ULT:
3913 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3914 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3915 break; // (X != 13 & X u< 15) -> no change
3916 case ICmpInst::ICMP_SLT:
3917 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3918 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3919 break; // (X != 13 & X s< 15) -> no change
3920 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3921 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3922 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3923 return ReplaceInstUsesWith(I, RHS);
3924 case ICmpInst::ICMP_NE:
3925 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3926 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3927 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
3928 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3929 ConstantInt::get(Add->getType(), 1));
3931 break; // (X != 13 & X != 15) -> no change
3934 case ICmpInst::ICMP_ULT:
3936 default: llvm_unreachable("Unknown integer condition code!");
3937 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3938 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3939 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3940 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3942 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3943 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3944 return ReplaceInstUsesWith(I, LHS);
3945 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3949 case ICmpInst::ICMP_SLT:
3951 default: llvm_unreachable("Unknown integer condition code!");
3952 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3953 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3954 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3955 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3957 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3958 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3959 return ReplaceInstUsesWith(I, LHS);
3960 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3964 case ICmpInst::ICMP_UGT:
3966 default: llvm_unreachable("Unknown integer condition code!");
3967 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3968 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3969 return ReplaceInstUsesWith(I, RHS);
3970 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3972 case ICmpInst::ICMP_NE:
3973 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3974 return new ICmpInst(LHSCC, Val, RHSCst);
3975 break; // (X u> 13 & X != 15) -> no change
3976 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3977 return InsertRangeTest(Val, AddOne(LHSCst),
3978 RHSCst, false, true, I);
3979 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3983 case ICmpInst::ICMP_SGT:
3985 default: llvm_unreachable("Unknown integer condition code!");
3986 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3987 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3988 return ReplaceInstUsesWith(I, RHS);
3989 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3991 case ICmpInst::ICMP_NE:
3992 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3993 return new ICmpInst(LHSCC, Val, RHSCst);
3994 break; // (X s> 13 & X != 15) -> no change
3995 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3996 return InsertRangeTest(Val, AddOne(LHSCst),
3997 RHSCst, true, true, I);
3998 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4007 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
4010 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4011 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4012 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4013 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4014 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4015 // If either of the constants are nans, then the whole thing returns
4017 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4018 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4019 return new FCmpInst(FCmpInst::FCMP_ORD,
4020 LHS->getOperand(0), RHS->getOperand(0));
4023 // Handle vector zeros. This occurs because the canonical form of
4024 // "fcmp ord x,x" is "fcmp ord x, 0".
4025 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4026 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4027 return new FCmpInst(FCmpInst::FCMP_ORD,
4028 LHS->getOperand(0), RHS->getOperand(0));
4032 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4033 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4034 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4037 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4038 // Swap RHS operands to match LHS.
4039 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4040 std::swap(Op1LHS, Op1RHS);
4043 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4044 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4046 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4048 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
4049 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4050 if (Op0CC == FCmpInst::FCMP_TRUE)
4051 return ReplaceInstUsesWith(I, RHS);
4052 if (Op1CC == FCmpInst::FCMP_TRUE)
4053 return ReplaceInstUsesWith(I, LHS);
4057 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4058 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4060 std::swap(LHS, RHS);
4061 std::swap(Op0Pred, Op1Pred);
4062 std::swap(Op0Ordered, Op1Ordered);
4065 // uno && ueq -> uno && (uno || eq) -> ueq
4066 // ord && olt -> ord && (ord && lt) -> olt
4067 if (Op0Ordered == Op1Ordered)
4068 return ReplaceInstUsesWith(I, RHS);
4070 // uno && oeq -> uno && (ord && eq) -> false
4071 // uno && ord -> false
4073 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4074 // ord && ueq -> ord && (uno || eq) -> oeq
4075 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4076 Op0LHS, Op0RHS, Context));
4084 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4085 bool Changed = SimplifyCommutative(I);
4086 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4088 if (isa<UndefValue>(Op1)) // X & undef -> 0
4089 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4093 return ReplaceInstUsesWith(I, Op1);
4095 // See if we can simplify any instructions used by the instruction whose sole
4096 // purpose is to compute bits we don't care about.
4097 if (SimplifyDemandedInstructionBits(I))
4099 if (isa<VectorType>(I.getType())) {
4100 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4101 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4102 return ReplaceInstUsesWith(I, I.getOperand(0));
4103 } else if (isa<ConstantAggregateZero>(Op1)) {
4104 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4108 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4109 const APInt &AndRHSMask = AndRHS->getValue();
4110 APInt NotAndRHS(~AndRHSMask);
4112 // Optimize a variety of ((val OP C1) & C2) combinations...
4113 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4114 Value *Op0LHS = Op0I->getOperand(0);
4115 Value *Op0RHS = Op0I->getOperand(1);
4116 switch (Op0I->getOpcode()) {
4118 case Instruction::Xor:
4119 case Instruction::Or:
4120 // If the mask is only needed on one incoming arm, push it up.
4121 if (!Op0I->hasOneUse()) break;
4123 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4124 // Not masking anything out for the LHS, move to RHS.
4125 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
4126 Op0RHS->getName()+".masked");
4127 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
4129 if (!isa<Constant>(Op0RHS) &&
4130 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4131 // Not masking anything out for the RHS, move to LHS.
4132 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
4133 Op0LHS->getName()+".masked");
4134 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
4138 case Instruction::Add:
4139 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4140 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4141 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4142 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4143 return BinaryOperator::CreateAnd(V, AndRHS);
4144 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4145 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4148 case Instruction::Sub:
4149 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4150 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4151 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4152 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4153 return BinaryOperator::CreateAnd(V, AndRHS);
4155 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4156 // has 1's for all bits that the subtraction with A might affect.
4157 if (Op0I->hasOneUse()) {
4158 uint32_t BitWidth = AndRHSMask.getBitWidth();
4159 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4160 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4162 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4163 if (!(A && A->isZero()) && // avoid infinite recursion.
4164 MaskedValueIsZero(Op0LHS, Mask)) {
4165 Value *NewNeg = Builder->CreateNeg(Op0RHS);
4166 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4171 case Instruction::Shl:
4172 case Instruction::LShr:
4173 // (1 << x) & 1 --> zext(x == 0)
4174 // (1 >> x) & 1 --> zext(x == 0)
4175 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4177 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
4178 return new ZExtInst(NewICmp, I.getType());
4183 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4184 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4186 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4187 // If this is an integer truncation or change from signed-to-unsigned, and
4188 // if the source is an and/or with immediate, transform it. This
4189 // frequently occurs for bitfield accesses.
4190 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4191 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4192 CastOp->getNumOperands() == 2)
4193 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
4194 if (CastOp->getOpcode() == Instruction::And) {
4195 // Change: and (cast (and X, C1) to T), C2
4196 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4197 // This will fold the two constants together, which may allow
4198 // other simplifications.
4199 Value *NewCast = Builder->CreateTruncOrBitCast(
4200 CastOp->getOperand(0), I.getType(),
4201 CastOp->getName()+".shrunk");
4202 // trunc_or_bitcast(C1)&C2
4203 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4204 C3 = ConstantExpr::getAnd(C3, AndRHS);
4205 return BinaryOperator::CreateAnd(NewCast, C3);
4206 } else if (CastOp->getOpcode() == Instruction::Or) {
4207 // Change: and (cast (or X, C1) to T), C2
4208 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4209 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4210 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4212 return ReplaceInstUsesWith(I, AndRHS);
4218 // Try to fold constant and into select arguments.
4219 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4220 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4222 if (isa<PHINode>(Op0))
4223 if (Instruction *NV = FoldOpIntoPhi(I))
4227 Value *Op0NotVal = dyn_castNotVal(Op0);
4228 Value *Op1NotVal = dyn_castNotVal(Op1);
4230 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4231 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4233 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4234 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4235 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
4236 I.getName()+".demorgan");
4237 return BinaryOperator::CreateNot(Or);
4241 Value *A = 0, *B = 0, *C = 0, *D = 0;
4242 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4243 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4244 return ReplaceInstUsesWith(I, Op1);
4246 // (A|B) & ~(A&B) -> A^B
4247 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4248 if ((A == C && B == D) || (A == D && B == C))
4249 return BinaryOperator::CreateXor(A, B);
4253 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4254 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4255 return ReplaceInstUsesWith(I, Op0);
4257 // ~(A&B) & (A|B) -> A^B
4258 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4259 if ((A == C && B == D) || (A == D && B == C))
4260 return BinaryOperator::CreateXor(A, B);
4264 if (Op0->hasOneUse() &&
4265 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4266 if (A == Op1) { // (A^B)&A -> A&(A^B)
4267 I.swapOperands(); // Simplify below
4268 std::swap(Op0, Op1);
4269 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4270 cast<BinaryOperator>(Op0)->swapOperands();
4271 I.swapOperands(); // Simplify below
4272 std::swap(Op0, Op1);
4276 if (Op1->hasOneUse() &&
4277 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4278 if (B == Op0) { // B&(A^B) -> B&(B^A)
4279 cast<BinaryOperator>(Op1)->swapOperands();
4282 if (A == Op0) // A&(A^B) -> A & ~B
4283 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
4286 // (A&((~A)|B)) -> A&B
4287 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4288 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4289 return BinaryOperator::CreateAnd(A, Op1);
4290 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4291 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4292 return BinaryOperator::CreateAnd(A, Op0);
4295 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4296 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4297 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4300 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4301 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4305 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4306 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4307 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4308 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4309 const Type *SrcTy = Op0C->getOperand(0)->getType();
4310 if (SrcTy == Op1C->getOperand(0)->getType() &&
4311 SrcTy->isIntOrIntVector() &&
4312 // Only do this if the casts both really cause code to be generated.
4313 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4315 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4317 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
4318 Op1C->getOperand(0), I.getName());
4319 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4323 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4324 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4325 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4326 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4327 SI0->getOperand(1) == SI1->getOperand(1) &&
4328 (SI0->hasOneUse() || SI1->hasOneUse())) {
4330 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
4332 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4333 SI1->getOperand(1));
4337 // If and'ing two fcmp, try combine them into one.
4338 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4339 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4340 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4344 return Changed ? &I : 0;
4347 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4348 /// capable of providing pieces of a bswap. The subexpression provides pieces
4349 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4350 /// the expression came from the corresponding "byte swapped" byte in some other
4351 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4352 /// we know that the expression deposits the low byte of %X into the high byte
4353 /// of the bswap result and that all other bytes are zero. This expression is
4354 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4357 /// This function returns true if the match was unsuccessful and false if so.
4358 /// On entry to the function the "OverallLeftShift" is a signed integer value
4359 /// indicating the number of bytes that the subexpression is later shifted. For
4360 /// example, if the expression is later right shifted by 16 bits, the
4361 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4362 /// byte of ByteValues is actually being set.
4364 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4365 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4366 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4367 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4368 /// always in the local (OverallLeftShift) coordinate space.
4370 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4371 SmallVector<Value*, 8> &ByteValues) {
4372 if (Instruction *I = dyn_cast<Instruction>(V)) {
4373 // If this is an or instruction, it may be an inner node of the bswap.
4374 if (I->getOpcode() == Instruction::Or) {
4375 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4377 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4381 // If this is a logical shift by a constant multiple of 8, recurse with
4382 // OverallLeftShift and ByteMask adjusted.
4383 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4385 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4386 // Ensure the shift amount is defined and of a byte value.
4387 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4390 unsigned ByteShift = ShAmt >> 3;
4391 if (I->getOpcode() == Instruction::Shl) {
4392 // X << 2 -> collect(X, +2)
4393 OverallLeftShift += ByteShift;
4394 ByteMask >>= ByteShift;
4396 // X >>u 2 -> collect(X, -2)
4397 OverallLeftShift -= ByteShift;
4398 ByteMask <<= ByteShift;
4399 ByteMask &= (~0U >> (32-ByteValues.size()));
4402 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4403 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4405 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4409 // If this is a logical 'and' with a mask that clears bytes, clear the
4410 // corresponding bytes in ByteMask.
4411 if (I->getOpcode() == Instruction::And &&
4412 isa<ConstantInt>(I->getOperand(1))) {
4413 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4414 unsigned NumBytes = ByteValues.size();
4415 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4416 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4418 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4419 // If this byte is masked out by a later operation, we don't care what
4421 if ((ByteMask & (1 << i)) == 0)
4424 // If the AndMask is all zeros for this byte, clear the bit.
4425 APInt MaskB = AndMask & Byte;
4427 ByteMask &= ~(1U << i);
4431 // If the AndMask is not all ones for this byte, it's not a bytezap.
4435 // Otherwise, this byte is kept.
4438 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4443 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4444 // the input value to the bswap. Some observations: 1) if more than one byte
4445 // is demanded from this input, then it could not be successfully assembled
4446 // into a byteswap. At least one of the two bytes would not be aligned with
4447 // their ultimate destination.
4448 if (!isPowerOf2_32(ByteMask)) return true;
4449 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4451 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4452 // is demanded, it needs to go into byte 0 of the result. This means that the
4453 // byte needs to be shifted until it lands in the right byte bucket. The
4454 // shift amount depends on the position: if the byte is coming from the high
4455 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4456 // low part, it must be shifted left.
4457 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4458 if (InputByteNo < ByteValues.size()/2) {
4459 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4462 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4466 // If the destination byte value is already defined, the values are or'd
4467 // together, which isn't a bswap (unless it's an or of the same bits).
4468 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4470 ByteValues[DestByteNo] = V;
4474 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4475 /// If so, insert the new bswap intrinsic and return it.
4476 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4477 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4478 if (!ITy || ITy->getBitWidth() % 16 ||
4479 // ByteMask only allows up to 32-byte values.
4480 ITy->getBitWidth() > 32*8)
4481 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4483 /// ByteValues - For each byte of the result, we keep track of which value
4484 /// defines each byte.
4485 SmallVector<Value*, 8> ByteValues;
4486 ByteValues.resize(ITy->getBitWidth()/8);
4488 // Try to find all the pieces corresponding to the bswap.
4489 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4490 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4493 // Check to see if all of the bytes come from the same value.
4494 Value *V = ByteValues[0];
4495 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4497 // Check to make sure that all of the bytes come from the same value.
4498 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4499 if (ByteValues[i] != V)
4501 const Type *Tys[] = { ITy };
4502 Module *M = I.getParent()->getParent()->getParent();
4503 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4504 return CallInst::Create(F, V);
4507 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4508 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4509 /// we can simplify this expression to "cond ? C : D or B".
4510 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4512 LLVMContext *Context) {
4513 // If A is not a select of -1/0, this cannot match.
4515 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4518 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4519 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4520 return SelectInst::Create(Cond, C, B);
4521 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4522 return SelectInst::Create(Cond, C, B);
4523 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4524 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4525 return SelectInst::Create(Cond, C, D);
4526 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4527 return SelectInst::Create(Cond, C, D);
4531 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4532 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4533 ICmpInst *LHS, ICmpInst *RHS) {
4535 ConstantInt *LHSCst, *RHSCst;
4536 ICmpInst::Predicate LHSCC, RHSCC;
4538 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4539 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4540 m_ConstantInt(LHSCst))) ||
4541 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4542 m_ConstantInt(RHSCst))))
4545 // From here on, we only handle:
4546 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4547 if (Val != Val2) return 0;
4549 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4550 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4551 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4552 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4553 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4556 // We can't fold (ugt x, C) | (sgt x, C2).
4557 if (!PredicatesFoldable(LHSCC, RHSCC))
4560 // Ensure that the larger constant is on the RHS.
4562 if (CmpInst::isSigned(LHSCC) ||
4563 (ICmpInst::isEquality(LHSCC) &&
4564 CmpInst::isSigned(RHSCC)))
4565 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4567 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4570 std::swap(LHS, RHS);
4571 std::swap(LHSCst, RHSCst);
4572 std::swap(LHSCC, RHSCC);
4575 // At this point, we know we have have two icmp instructions
4576 // comparing a value against two constants and or'ing the result
4577 // together. Because of the above check, we know that we only have
4578 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4579 // FoldICmpLogical check above), that the two constants are not
4581 assert(LHSCst != RHSCst && "Compares not folded above?");
4584 default: llvm_unreachable("Unknown integer condition code!");
4585 case ICmpInst::ICMP_EQ:
4587 default: llvm_unreachable("Unknown integer condition code!");
4588 case ICmpInst::ICMP_EQ:
4589 if (LHSCst == SubOne(RHSCst)) {
4590 // (X == 13 | X == 14) -> X-13 <u 2
4591 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4592 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4593 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4594 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4596 break; // (X == 13 | X == 15) -> no change
4597 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4598 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4600 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4601 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4602 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4603 return ReplaceInstUsesWith(I, RHS);
4606 case ICmpInst::ICMP_NE:
4608 default: llvm_unreachable("Unknown integer condition code!");
4609 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4610 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4611 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4612 return ReplaceInstUsesWith(I, LHS);
4613 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4614 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4615 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4616 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4619 case ICmpInst::ICMP_ULT:
4621 default: llvm_unreachable("Unknown integer condition code!");
4622 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4624 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4625 // If RHSCst is [us]MAXINT, it is always false. Not handling
4626 // this can cause overflow.
4627 if (RHSCst->isMaxValue(false))
4628 return ReplaceInstUsesWith(I, LHS);
4629 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4631 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4633 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4634 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4635 return ReplaceInstUsesWith(I, RHS);
4636 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4640 case ICmpInst::ICMP_SLT:
4642 default: llvm_unreachable("Unknown integer condition code!");
4643 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4645 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4646 // If RHSCst is [us]MAXINT, it is always false. Not handling
4647 // this can cause overflow.
4648 if (RHSCst->isMaxValue(true))
4649 return ReplaceInstUsesWith(I, LHS);
4650 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4652 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4654 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4655 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4656 return ReplaceInstUsesWith(I, RHS);
4657 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4661 case ICmpInst::ICMP_UGT:
4663 default: llvm_unreachable("Unknown integer condition code!");
4664 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4665 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4666 return ReplaceInstUsesWith(I, LHS);
4667 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4669 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4670 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4671 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4672 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4676 case ICmpInst::ICMP_SGT:
4678 default: llvm_unreachable("Unknown integer condition code!");
4679 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4680 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4681 return ReplaceInstUsesWith(I, LHS);
4682 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4684 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4685 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4686 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4687 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4695 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4697 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4698 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4699 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4700 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4701 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4702 // If either of the constants are nans, then the whole thing returns
4704 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4705 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4707 // Otherwise, no need to compare the two constants, compare the
4709 return new FCmpInst(FCmpInst::FCMP_UNO,
4710 LHS->getOperand(0), RHS->getOperand(0));
4713 // Handle vector zeros. This occurs because the canonical form of
4714 // "fcmp uno x,x" is "fcmp uno x, 0".
4715 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4716 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4717 return new FCmpInst(FCmpInst::FCMP_UNO,
4718 LHS->getOperand(0), RHS->getOperand(0));
4723 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4724 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4725 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4727 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4728 // Swap RHS operands to match LHS.
4729 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4730 std::swap(Op1LHS, Op1RHS);
4732 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4733 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4735 return new FCmpInst((FCmpInst::Predicate)Op0CC,
4737 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4738 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4739 if (Op0CC == FCmpInst::FCMP_FALSE)
4740 return ReplaceInstUsesWith(I, RHS);
4741 if (Op1CC == FCmpInst::FCMP_FALSE)
4742 return ReplaceInstUsesWith(I, LHS);
4745 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4746 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4747 if (Op0Ordered == Op1Ordered) {
4748 // If both are ordered or unordered, return a new fcmp with
4749 // or'ed predicates.
4750 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4751 Op0LHS, Op0RHS, Context);
4752 if (Instruction *I = dyn_cast<Instruction>(RV))
4754 // Otherwise, it's a constant boolean value...
4755 return ReplaceInstUsesWith(I, RV);
4761 /// FoldOrWithConstants - This helper function folds:
4763 /// ((A | B) & C1) | (B & C2)
4769 /// when the XOR of the two constants is "all ones" (-1).
4770 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4771 Value *A, Value *B, Value *C) {
4772 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4776 ConstantInt *CI2 = 0;
4777 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4779 APInt Xor = CI1->getValue() ^ CI2->getValue();
4780 if (!Xor.isAllOnesValue()) return 0;
4782 if (V1 == A || V1 == B) {
4783 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
4784 return BinaryOperator::CreateOr(NewOp, V1);
4790 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4791 bool Changed = SimplifyCommutative(I);
4792 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4794 if (isa<UndefValue>(Op1)) // X | undef -> -1
4795 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4799 return ReplaceInstUsesWith(I, Op0);
4801 // See if we can simplify any instructions used by the instruction whose sole
4802 // purpose is to compute bits we don't care about.
4803 if (SimplifyDemandedInstructionBits(I))
4805 if (isa<VectorType>(I.getType())) {
4806 if (isa<ConstantAggregateZero>(Op1)) {
4807 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4808 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4809 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4810 return ReplaceInstUsesWith(I, I.getOperand(1));
4815 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4816 ConstantInt *C1 = 0; Value *X = 0;
4817 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4818 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
4820 Value *Or = Builder->CreateOr(X, RHS);
4822 return BinaryOperator::CreateAnd(Or,
4823 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4826 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4827 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
4829 Value *Or = Builder->CreateOr(X, RHS);
4831 return BinaryOperator::CreateXor(Or,
4832 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4835 // Try to fold constant and into select arguments.
4836 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4837 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4839 if (isa<PHINode>(Op0))
4840 if (Instruction *NV = FoldOpIntoPhi(I))
4844 Value *A = 0, *B = 0;
4845 ConstantInt *C1 = 0, *C2 = 0;
4847 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4848 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4849 return ReplaceInstUsesWith(I, Op1);
4850 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4851 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4852 return ReplaceInstUsesWith(I, Op0);
4854 // (A | B) | C and A | (B | C) -> bswap if possible.
4855 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4856 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4857 match(Op1, m_Or(m_Value(), m_Value())) ||
4858 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4859 match(Op1, m_Shift(m_Value(), m_Value())))) {
4860 if (Instruction *BSwap = MatchBSwap(I))
4864 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4865 if (Op0->hasOneUse() &&
4866 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4867 MaskedValueIsZero(Op1, C1->getValue())) {
4868 Value *NOr = Builder->CreateOr(A, Op1);
4870 return BinaryOperator::CreateXor(NOr, C1);
4873 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4874 if (Op1->hasOneUse() &&
4875 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4876 MaskedValueIsZero(Op0, C1->getValue())) {
4877 Value *NOr = Builder->CreateOr(A, Op0);
4879 return BinaryOperator::CreateXor(NOr, C1);
4883 Value *C = 0, *D = 0;
4884 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4885 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4886 Value *V1 = 0, *V2 = 0, *V3 = 0;
4887 C1 = dyn_cast<ConstantInt>(C);
4888 C2 = dyn_cast<ConstantInt>(D);
4889 if (C1 && C2) { // (A & C1)|(B & C2)
4890 // If we have: ((V + N) & C1) | (V & C2)
4891 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4892 // replace with V+N.
4893 if (C1->getValue() == ~C2->getValue()) {
4894 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4895 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4896 // Add commutes, try both ways.
4897 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4898 return ReplaceInstUsesWith(I, A);
4899 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4900 return ReplaceInstUsesWith(I, A);
4902 // Or commutes, try both ways.
4903 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4904 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4905 // Add commutes, try both ways.
4906 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4907 return ReplaceInstUsesWith(I, B);
4908 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4909 return ReplaceInstUsesWith(I, B);
4912 V1 = 0; V2 = 0; V3 = 0;
4915 // Check to see if we have any common things being and'ed. If so, find the
4916 // terms for V1 & (V2|V3).
4917 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4918 if (A == B) // (A & C)|(A & D) == A & (C|D)
4919 V1 = A, V2 = C, V3 = D;
4920 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4921 V1 = A, V2 = B, V3 = C;
4922 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4923 V1 = C, V2 = A, V3 = D;
4924 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4925 V1 = C, V2 = A, V3 = B;
4928 Value *Or = Builder->CreateOr(V2, V3, "tmp");
4929 return BinaryOperator::CreateAnd(V1, Or);
4933 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4934 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4936 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4938 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4940 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4943 // ((A&~B)|(~A&B)) -> A^B
4944 if ((match(C, m_Not(m_Specific(D))) &&
4945 match(B, m_Not(m_Specific(A)))))
4946 return BinaryOperator::CreateXor(A, D);
4947 // ((~B&A)|(~A&B)) -> A^B
4948 if ((match(A, m_Not(m_Specific(D))) &&
4949 match(B, m_Not(m_Specific(C)))))
4950 return BinaryOperator::CreateXor(C, D);
4951 // ((A&~B)|(B&~A)) -> A^B
4952 if ((match(C, m_Not(m_Specific(B))) &&
4953 match(D, m_Not(m_Specific(A)))))
4954 return BinaryOperator::CreateXor(A, B);
4955 // ((~B&A)|(B&~A)) -> A^B
4956 if ((match(A, m_Not(m_Specific(B))) &&
4957 match(D, m_Not(m_Specific(C)))))
4958 return BinaryOperator::CreateXor(C, B);
4961 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4962 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4963 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4964 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4965 SI0->getOperand(1) == SI1->getOperand(1) &&
4966 (SI0->hasOneUse() || SI1->hasOneUse())) {
4967 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
4969 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4970 SI1->getOperand(1));
4974 // ((A|B)&1)|(B&-2) -> (A&1) | B
4975 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4976 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4977 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4978 if (Ret) return Ret;
4980 // (B&-2)|((A|B)&1) -> (A&1) | B
4981 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4982 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4983 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4984 if (Ret) return Ret;
4987 if ((A = dyn_castNotVal(Op0))) { // ~A | Op1
4988 if (A == Op1) // ~A | A == -1
4989 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4993 // Note, A is still live here!
4994 if ((B = dyn_castNotVal(Op1))) { // Op0 | ~B
4996 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4998 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4999 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
5000 Value *And = Builder->CreateAnd(A, B, I.getName()+".demorgan");
5001 return BinaryOperator::CreateNot(And);
5005 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
5006 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
5007 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5010 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
5011 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
5015 // fold (or (cast A), (cast B)) -> (cast (or A, B))
5016 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5017 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5018 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
5019 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
5020 !isa<ICmpInst>(Op1C->getOperand(0))) {
5021 const Type *SrcTy = Op0C->getOperand(0)->getType();
5022 if (SrcTy == Op1C->getOperand(0)->getType() &&
5023 SrcTy->isIntOrIntVector() &&
5024 // Only do this if the casts both really cause code to be
5026 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5028 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5030 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
5031 Op1C->getOperand(0), I.getName());
5032 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5039 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
5040 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
5041 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
5042 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5046 return Changed ? &I : 0;
5051 // XorSelf - Implements: X ^ X --> 0
5054 XorSelf(Value *rhs) : RHS(rhs) {}
5055 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5056 Instruction *apply(BinaryOperator &Xor) const {
5063 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5064 bool Changed = SimplifyCommutative(I);
5065 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5067 if (isa<UndefValue>(Op1)) {
5068 if (isa<UndefValue>(Op0))
5069 // Handle undef ^ undef -> 0 special case. This is a common
5071 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5072 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5075 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5076 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5077 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5078 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5081 // See if we can simplify any instructions used by the instruction whose sole
5082 // purpose is to compute bits we don't care about.
5083 if (SimplifyDemandedInstructionBits(I))
5085 if (isa<VectorType>(I.getType()))
5086 if (isa<ConstantAggregateZero>(Op1))
5087 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5089 // Is this a ~ operation?
5090 if (Value *NotOp = dyn_castNotVal(&I)) {
5091 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5092 if (Op0I->getOpcode() == Instruction::And ||
5093 Op0I->getOpcode() == Instruction::Or) {
5094 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5095 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5096 if (dyn_castNotVal(Op0I->getOperand(1)))
5097 Op0I->swapOperands();
5098 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5100 Builder->CreateNot(Op0I->getOperand(1),
5101 Op0I->getOperand(1)->getName()+".not");
5102 if (Op0I->getOpcode() == Instruction::And)
5103 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5104 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5107 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
5108 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
5109 if (isFreeToInvert(Op0I->getOperand(0)) &&
5110 isFreeToInvert(Op0I->getOperand(1))) {
5112 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
5114 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
5115 if (Op0I->getOpcode() == Instruction::And)
5116 return BinaryOperator::CreateOr(NotX, NotY);
5117 return BinaryOperator::CreateAnd(NotX, NotY);
5124 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5125 if (RHS->isOne() && Op0->hasOneUse()) {
5126 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5127 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5128 return new ICmpInst(ICI->getInversePredicate(),
5129 ICI->getOperand(0), ICI->getOperand(1));
5131 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5132 return new FCmpInst(FCI->getInversePredicate(),
5133 FCI->getOperand(0), FCI->getOperand(1));
5136 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5137 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5138 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5139 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5140 Instruction::CastOps Opcode = Op0C->getOpcode();
5141 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
5142 (RHS == ConstantExpr::getCast(Opcode,
5143 ConstantInt::getTrue(*Context),
5144 Op0C->getDestTy()))) {
5145 CI->setPredicate(CI->getInversePredicate());
5146 return CastInst::Create(Opcode, CI, Op0C->getType());
5152 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5153 // ~(c-X) == X-c-1 == X+(-c-1)
5154 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5155 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5156 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5157 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5158 ConstantInt::get(I.getType(), 1));
5159 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5162 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5163 if (Op0I->getOpcode() == Instruction::Add) {
5164 // ~(X-c) --> (-c-1)-X
5165 if (RHS->isAllOnesValue()) {
5166 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5167 return BinaryOperator::CreateSub(
5168 ConstantExpr::getSub(NegOp0CI,
5169 ConstantInt::get(I.getType(), 1)),
5170 Op0I->getOperand(0));
5171 } else if (RHS->getValue().isSignBit()) {
5172 // (X + C) ^ signbit -> (X + C + signbit)
5173 Constant *C = ConstantInt::get(*Context,
5174 RHS->getValue() + Op0CI->getValue());
5175 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5178 } else if (Op0I->getOpcode() == Instruction::Or) {
5179 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5180 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5181 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5182 // Anything in both C1 and C2 is known to be zero, remove it from
5184 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5185 NewRHS = ConstantExpr::getAnd(NewRHS,
5186 ConstantExpr::getNot(CommonBits));
5188 I.setOperand(0, Op0I->getOperand(0));
5189 I.setOperand(1, NewRHS);
5196 // Try to fold constant and into select arguments.
5197 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5198 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5200 if (isa<PHINode>(Op0))
5201 if (Instruction *NV = FoldOpIntoPhi(I))
5205 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5207 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5209 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5211 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5214 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5217 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5218 if (A == Op0) { // B^(B|A) == (A|B)^B
5219 Op1I->swapOperands();
5221 std::swap(Op0, Op1);
5222 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5223 I.swapOperands(); // Simplified below.
5224 std::swap(Op0, Op1);
5226 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5227 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5228 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5229 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5230 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5232 if (A == Op0) { // A^(A&B) -> A^(B&A)
5233 Op1I->swapOperands();
5236 if (B == Op0) { // A^(B&A) -> (B&A)^A
5237 I.swapOperands(); // Simplified below.
5238 std::swap(Op0, Op1);
5243 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5246 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5247 Op0I->hasOneUse()) {
5248 if (A == Op1) // (B|A)^B == (A|B)^B
5250 if (B == Op1) // (A|B)^B == A & ~B
5251 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
5252 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5253 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5254 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5255 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5256 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5258 if (A == Op1) // (A&B)^A -> (B&A)^A
5260 if (B == Op1 && // (B&A)^A == ~B & A
5261 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5262 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
5267 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5268 if (Op0I && Op1I && Op0I->isShift() &&
5269 Op0I->getOpcode() == Op1I->getOpcode() &&
5270 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5271 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5273 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
5275 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5276 Op1I->getOperand(1));
5280 Value *A, *B, *C, *D;
5281 // (A & B)^(A | B) -> A ^ B
5282 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5283 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5284 if ((A == C && B == D) || (A == D && B == C))
5285 return BinaryOperator::CreateXor(A, B);
5287 // (A | B)^(A & B) -> A ^ B
5288 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5289 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5290 if ((A == C && B == D) || (A == D && B == C))
5291 return BinaryOperator::CreateXor(A, B);
5295 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5296 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5297 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5298 // (X & Y)^(X & Y) -> (Y^Z) & X
5299 Value *X = 0, *Y = 0, *Z = 0;
5301 X = A, Y = B, Z = D;
5303 X = A, Y = B, Z = C;
5305 X = B, Y = A, Z = D;
5307 X = B, Y = A, Z = C;
5310 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
5311 return BinaryOperator::CreateAnd(NewOp, X);
5316 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5317 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5318 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5321 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5322 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5323 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5324 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5325 const Type *SrcTy = Op0C->getOperand(0)->getType();
5326 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5327 // Only do this if the casts both really cause code to be generated.
5328 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5330 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5332 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
5333 Op1C->getOperand(0), I.getName());
5334 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5339 return Changed ? &I : 0;
5342 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5343 LLVMContext *Context) {
5344 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5347 static bool HasAddOverflow(ConstantInt *Result,
5348 ConstantInt *In1, ConstantInt *In2,
5351 if (In2->getValue().isNegative())
5352 return Result->getValue().sgt(In1->getValue());
5354 return Result->getValue().slt(In1->getValue());
5356 return Result->getValue().ult(In1->getValue());
5359 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5360 /// overflowed for this type.
5361 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5362 Constant *In2, LLVMContext *Context,
5363 bool IsSigned = false) {
5364 Result = ConstantExpr::getAdd(In1, In2);
5366 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5367 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5368 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5369 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5370 ExtractElement(In1, Idx, Context),
5371 ExtractElement(In2, Idx, Context),
5378 return HasAddOverflow(cast<ConstantInt>(Result),
5379 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5383 static bool HasSubOverflow(ConstantInt *Result,
5384 ConstantInt *In1, ConstantInt *In2,
5387 if (In2->getValue().isNegative())
5388 return Result->getValue().slt(In1->getValue());
5390 return Result->getValue().sgt(In1->getValue());
5392 return Result->getValue().ugt(In1->getValue());
5395 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5396 /// overflowed for this type.
5397 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5398 Constant *In2, LLVMContext *Context,
5399 bool IsSigned = false) {
5400 Result = ConstantExpr::getSub(In1, In2);
5402 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5403 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5404 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5405 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5406 ExtractElement(In1, Idx, Context),
5407 ExtractElement(In2, Idx, Context),
5414 return HasSubOverflow(cast<ConstantInt>(Result),
5415 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5419 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5420 /// code necessary to compute the offset from the base pointer (without adding
5421 /// in the base pointer). Return the result as a signed integer of intptr size.
5422 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5423 TargetData &TD = *IC.getTargetData();
5424 gep_type_iterator GTI = gep_type_begin(GEP);
5425 const Type *IntPtrTy = TD.getIntPtrType(I.getContext());
5426 Value *Result = Constant::getNullValue(IntPtrTy);
5428 // Build a mask for high order bits.
5429 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5430 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5432 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5435 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5436 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5437 if (OpC->isZero()) continue;
5439 // Handle a struct index, which adds its field offset to the pointer.
5440 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5441 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5443 Result = IC.Builder->CreateAdd(Result,
5444 ConstantInt::get(IntPtrTy, Size),
5445 GEP->getName()+".offs");
5449 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5451 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5452 Scale = ConstantExpr::getMul(OC, Scale);
5453 // Emit an add instruction.
5454 Result = IC.Builder->CreateAdd(Result, Scale, GEP->getName()+".offs");
5457 // Convert to correct type.
5458 if (Op->getType() != IntPtrTy)
5459 Op = IC.Builder->CreateIntCast(Op, IntPtrTy, true, Op->getName()+".c");
5461 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5462 // We'll let instcombine(mul) convert this to a shl if possible.
5463 Op = IC.Builder->CreateMul(Op, Scale, GEP->getName()+".idx");
5466 // Emit an add instruction.
5467 Result = IC.Builder->CreateAdd(Op, Result, GEP->getName()+".offs");
5473 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5474 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5475 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5476 /// be complex, and scales are involved. The above expression would also be
5477 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5478 /// This later form is less amenable to optimization though, and we are allowed
5479 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5481 /// If we can't emit an optimized form for this expression, this returns null.
5483 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5485 TargetData &TD = *IC.getTargetData();
5486 gep_type_iterator GTI = gep_type_begin(GEP);
5488 // Check to see if this gep only has a single variable index. If so, and if
5489 // any constant indices are a multiple of its scale, then we can compute this
5490 // in terms of the scale of the variable index. For example, if the GEP
5491 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5492 // because the expression will cross zero at the same point.
5493 unsigned i, e = GEP->getNumOperands();
5495 for (i = 1; i != e; ++i, ++GTI) {
5496 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5497 // Compute the aggregate offset of constant indices.
5498 if (CI->isZero()) continue;
5500 // Handle a struct index, which adds its field offset to the pointer.
5501 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5502 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5504 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5505 Offset += Size*CI->getSExtValue();
5508 // Found our variable index.
5513 // If there are no variable indices, we must have a constant offset, just
5514 // evaluate it the general way.
5515 if (i == e) return 0;
5517 Value *VariableIdx = GEP->getOperand(i);
5518 // Determine the scale factor of the variable element. For example, this is
5519 // 4 if the variable index is into an array of i32.
5520 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5522 // Verify that there are no other variable indices. If so, emit the hard way.
5523 for (++i, ++GTI; i != e; ++i, ++GTI) {
5524 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5527 // Compute the aggregate offset of constant indices.
5528 if (CI->isZero()) continue;
5530 // Handle a struct index, which adds its field offset to the pointer.
5531 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5532 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5534 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5535 Offset += Size*CI->getSExtValue();
5539 // Okay, we know we have a single variable index, which must be a
5540 // pointer/array/vector index. If there is no offset, life is simple, return
5542 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5544 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5545 // we don't need to bother extending: the extension won't affect where the
5546 // computation crosses zero.
5547 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5548 VariableIdx = new TruncInst(VariableIdx,
5549 TD.getIntPtrType(VariableIdx->getContext()),
5550 VariableIdx->getName(), &I);
5554 // Otherwise, there is an index. The computation we will do will be modulo
5555 // the pointer size, so get it.
5556 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5558 Offset &= PtrSizeMask;
5559 VariableScale &= PtrSizeMask;
5561 // To do this transformation, any constant index must be a multiple of the
5562 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5563 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5564 // multiple of the variable scale.
5565 int64_t NewOffs = Offset / (int64_t)VariableScale;
5566 if (Offset != NewOffs*(int64_t)VariableScale)
5569 // Okay, we can do this evaluation. Start by converting the index to intptr.
5570 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
5571 if (VariableIdx->getType() != IntPtrTy)
5572 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5574 VariableIdx->getName(), &I);
5575 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5576 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5580 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5581 /// else. At this point we know that the GEP is on the LHS of the comparison.
5582 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5583 ICmpInst::Predicate Cond,
5585 // Look through bitcasts.
5586 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5587 RHS = BCI->getOperand(0);
5589 Value *PtrBase = GEPLHS->getOperand(0);
5590 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5591 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5592 // This transformation (ignoring the base and scales) is valid because we
5593 // know pointers can't overflow since the gep is inbounds. See if we can
5594 // output an optimized form.
5595 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5597 // If not, synthesize the offset the hard way.
5599 Offset = EmitGEPOffset(GEPLHS, I, *this);
5600 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5601 Constant::getNullValue(Offset->getType()));
5602 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5603 // If the base pointers are different, but the indices are the same, just
5604 // compare the base pointer.
5605 if (PtrBase != GEPRHS->getOperand(0)) {
5606 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5607 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5608 GEPRHS->getOperand(0)->getType();
5610 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5611 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5612 IndicesTheSame = false;
5616 // If all indices are the same, just compare the base pointers.
5618 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5619 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5621 // Otherwise, the base pointers are different and the indices are
5622 // different, bail out.
5626 // If one of the GEPs has all zero indices, recurse.
5627 bool AllZeros = true;
5628 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5629 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5630 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5635 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5636 ICmpInst::getSwappedPredicate(Cond), I);
5638 // If the other GEP has all zero indices, recurse.
5640 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5641 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5642 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5647 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5649 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5650 // If the GEPs only differ by one index, compare it.
5651 unsigned NumDifferences = 0; // Keep track of # differences.
5652 unsigned DiffOperand = 0; // The operand that differs.
5653 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5654 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5655 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5656 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5657 // Irreconcilable differences.
5661 if (NumDifferences++) break;
5666 if (NumDifferences == 0) // SAME GEP?
5667 return ReplaceInstUsesWith(I, // No comparison is needed here.
5668 ConstantInt::get(Type::getInt1Ty(*Context),
5669 ICmpInst::isTrueWhenEqual(Cond)));
5671 else if (NumDifferences == 1) {
5672 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5673 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5674 // Make sure we do a signed comparison here.
5675 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5679 // Only lower this if the icmp is the only user of the GEP or if we expect
5680 // the result to fold to a constant!
5682 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5683 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5684 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5685 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5686 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5687 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5693 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5695 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5698 if (!isa<ConstantFP>(RHSC)) return 0;
5699 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5701 // Get the width of the mantissa. We don't want to hack on conversions that
5702 // might lose information from the integer, e.g. "i64 -> float"
5703 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5704 if (MantissaWidth == -1) return 0; // Unknown.
5706 // Check to see that the input is converted from an integer type that is small
5707 // enough that preserves all bits. TODO: check here for "known" sign bits.
5708 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5709 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5711 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5712 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5716 // If the conversion would lose info, don't hack on this.
5717 if ((int)InputSize > MantissaWidth)
5720 // Otherwise, we can potentially simplify the comparison. We know that it
5721 // will always come through as an integer value and we know the constant is
5722 // not a NAN (it would have been previously simplified).
5723 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5725 ICmpInst::Predicate Pred;
5726 switch (I.getPredicate()) {
5727 default: llvm_unreachable("Unexpected predicate!");
5728 case FCmpInst::FCMP_UEQ:
5729 case FCmpInst::FCMP_OEQ:
5730 Pred = ICmpInst::ICMP_EQ;
5732 case FCmpInst::FCMP_UGT:
5733 case FCmpInst::FCMP_OGT:
5734 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5736 case FCmpInst::FCMP_UGE:
5737 case FCmpInst::FCMP_OGE:
5738 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5740 case FCmpInst::FCMP_ULT:
5741 case FCmpInst::FCMP_OLT:
5742 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5744 case FCmpInst::FCMP_ULE:
5745 case FCmpInst::FCMP_OLE:
5746 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5748 case FCmpInst::FCMP_UNE:
5749 case FCmpInst::FCMP_ONE:
5750 Pred = ICmpInst::ICMP_NE;
5752 case FCmpInst::FCMP_ORD:
5753 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5754 case FCmpInst::FCMP_UNO:
5755 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5758 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5760 // Now we know that the APFloat is a normal number, zero or inf.
5762 // See if the FP constant is too large for the integer. For example,
5763 // comparing an i8 to 300.0.
5764 unsigned IntWidth = IntTy->getScalarSizeInBits();
5767 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5768 // and large values.
5769 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5770 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5771 APFloat::rmNearestTiesToEven);
5772 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5773 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5774 Pred == ICmpInst::ICMP_SLE)
5775 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5776 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5779 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5780 // +INF and large values.
5781 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5782 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5783 APFloat::rmNearestTiesToEven);
5784 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5785 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5786 Pred == ICmpInst::ICMP_ULE)
5787 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5788 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5793 // See if the RHS value is < SignedMin.
5794 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5795 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5796 APFloat::rmNearestTiesToEven);
5797 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5798 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5799 Pred == ICmpInst::ICMP_SGE)
5800 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5801 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5805 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5806 // [0, UMAX], but it may still be fractional. See if it is fractional by
5807 // casting the FP value to the integer value and back, checking for equality.
5808 // Don't do this for zero, because -0.0 is not fractional.
5809 Constant *RHSInt = LHSUnsigned
5810 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5811 : ConstantExpr::getFPToSI(RHSC, IntTy);
5812 if (!RHS.isZero()) {
5813 bool Equal = LHSUnsigned
5814 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5815 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5817 // If we had a comparison against a fractional value, we have to adjust
5818 // the compare predicate and sometimes the value. RHSC is rounded towards
5819 // zero at this point.
5821 default: llvm_unreachable("Unexpected integer comparison!");
5822 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5823 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5824 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5825 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5826 case ICmpInst::ICMP_ULE:
5827 // (float)int <= 4.4 --> int <= 4
5828 // (float)int <= -4.4 --> false
5829 if (RHS.isNegative())
5830 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5832 case ICmpInst::ICMP_SLE:
5833 // (float)int <= 4.4 --> int <= 4
5834 // (float)int <= -4.4 --> int < -4
5835 if (RHS.isNegative())
5836 Pred = ICmpInst::ICMP_SLT;
5838 case ICmpInst::ICMP_ULT:
5839 // (float)int < -4.4 --> false
5840 // (float)int < 4.4 --> int <= 4
5841 if (RHS.isNegative())
5842 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5843 Pred = ICmpInst::ICMP_ULE;
5845 case ICmpInst::ICMP_SLT:
5846 // (float)int < -4.4 --> int < -4
5847 // (float)int < 4.4 --> int <= 4
5848 if (!RHS.isNegative())
5849 Pred = ICmpInst::ICMP_SLE;
5851 case ICmpInst::ICMP_UGT:
5852 // (float)int > 4.4 --> int > 4
5853 // (float)int > -4.4 --> true
5854 if (RHS.isNegative())
5855 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5857 case ICmpInst::ICMP_SGT:
5858 // (float)int > 4.4 --> int > 4
5859 // (float)int > -4.4 --> int >= -4
5860 if (RHS.isNegative())
5861 Pred = ICmpInst::ICMP_SGE;
5863 case ICmpInst::ICMP_UGE:
5864 // (float)int >= -4.4 --> true
5865 // (float)int >= 4.4 --> int > 4
5866 if (!RHS.isNegative())
5867 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5868 Pred = ICmpInst::ICMP_UGT;
5870 case ICmpInst::ICMP_SGE:
5871 // (float)int >= -4.4 --> int >= -4
5872 // (float)int >= 4.4 --> int > 4
5873 if (!RHS.isNegative())
5874 Pred = ICmpInst::ICMP_SGT;
5880 // Lower this FP comparison into an appropriate integer version of the
5882 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5885 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5886 bool Changed = SimplifyCompare(I);
5887 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5889 // Fold trivial predicates.
5890 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5891 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 0));
5892 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5893 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 1));
5895 // Simplify 'fcmp pred X, X'
5897 switch (I.getPredicate()) {
5898 default: llvm_unreachable("Unknown predicate!");
5899 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5900 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5901 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5902 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 1));
5903 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5904 case FCmpInst::FCMP_OLT: // True if ordered and less than
5905 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5906 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 0));
5908 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5909 case FCmpInst::FCMP_ULT: // True if unordered or less than
5910 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5911 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5912 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5913 I.setPredicate(FCmpInst::FCMP_UNO);
5914 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5917 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5918 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5919 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5920 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5921 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5922 I.setPredicate(FCmpInst::FCMP_ORD);
5923 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5928 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5929 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
5931 // Handle fcmp with constant RHS
5932 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5933 // If the constant is a nan, see if we can fold the comparison based on it.
5934 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5935 if (CFP->getValueAPF().isNaN()) {
5936 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5937 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5938 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5939 "Comparison must be either ordered or unordered!");
5940 // True if unordered.
5941 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5945 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5946 switch (LHSI->getOpcode()) {
5947 case Instruction::PHI:
5948 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5949 // block. If in the same block, we're encouraging jump threading. If
5950 // not, we are just pessimizing the code by making an i1 phi.
5951 if (LHSI->getParent() == I.getParent())
5952 if (Instruction *NV = FoldOpIntoPhi(I, true))
5955 case Instruction::SIToFP:
5956 case Instruction::UIToFP:
5957 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5960 case Instruction::Select:
5961 // If either operand of the select is a constant, we can fold the
5962 // comparison into the select arms, which will cause one to be
5963 // constant folded and the select turned into a bitwise or.
5964 Value *Op1 = 0, *Op2 = 0;
5965 if (LHSI->hasOneUse()) {
5966 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5967 // Fold the known value into the constant operand.
5968 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5969 // Insert a new FCmp of the other select operand.
5970 Op2 = Builder->CreateFCmp(I.getPredicate(),
5971 LHSI->getOperand(2), RHSC, I.getName());
5972 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5973 // Fold the known value into the constant operand.
5974 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5975 // Insert a new FCmp of the other select operand.
5976 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
5982 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5987 return Changed ? &I : 0;
5990 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5991 bool Changed = SimplifyCompare(I);
5992 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5993 const Type *Ty = Op0->getType();
5997 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(),
5998 I.isTrueWhenEqual()));
6000 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
6001 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
6003 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
6004 // addresses never equal each other! We already know that Op0 != Op1.
6005 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
6006 isa<ConstantPointerNull>(Op0)) &&
6007 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
6008 isa<ConstantPointerNull>(Op1)))
6009 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
6010 !I.isTrueWhenEqual()));
6012 // icmp's with boolean values can always be turned into bitwise operations
6013 if (Ty == Type::getInt1Ty(*Context)) {
6014 switch (I.getPredicate()) {
6015 default: llvm_unreachable("Invalid icmp instruction!");
6016 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6017 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
6018 return BinaryOperator::CreateNot(Xor);
6020 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6021 return BinaryOperator::CreateXor(Op0, Op1);
6023 case ICmpInst::ICMP_UGT:
6024 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6026 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6027 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6028 return BinaryOperator::CreateAnd(Not, Op1);
6030 case ICmpInst::ICMP_SGT:
6031 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6033 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6034 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6035 return BinaryOperator::CreateAnd(Not, Op0);
6037 case ICmpInst::ICMP_UGE:
6038 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6040 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6041 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6042 return BinaryOperator::CreateOr(Not, Op1);
6044 case ICmpInst::ICMP_SGE:
6045 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6047 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6048 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6049 return BinaryOperator::CreateOr(Not, Op0);
6054 unsigned BitWidth = 0;
6056 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6057 else if (Ty->isIntOrIntVector())
6058 BitWidth = Ty->getScalarSizeInBits();
6060 bool isSignBit = false;
6062 // See if we are doing a comparison with a constant.
6063 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6064 Value *A = 0, *B = 0;
6066 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6067 if (I.isEquality() && CI->isNullValue() &&
6068 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6069 // (icmp cond A B) if cond is equality
6070 return new ICmpInst(I.getPredicate(), A, B);
6073 // If we have an icmp le or icmp ge instruction, turn it into the
6074 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6075 // them being folded in the code below.
6076 switch (I.getPredicate()) {
6078 case ICmpInst::ICMP_ULE:
6079 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6080 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6081 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
6083 case ICmpInst::ICMP_SLE:
6084 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6085 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6086 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6088 case ICmpInst::ICMP_UGE:
6089 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6090 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6091 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
6093 case ICmpInst::ICMP_SGE:
6094 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6095 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6096 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6100 // If this comparison is a normal comparison, it demands all
6101 // bits, if it is a sign bit comparison, it only demands the sign bit.
6103 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6106 // See if we can fold the comparison based on range information we can get
6107 // by checking whether bits are known to be zero or one in the input.
6108 if (BitWidth != 0) {
6109 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6110 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6112 if (SimplifyDemandedBits(I.getOperandUse(0),
6113 isSignBit ? APInt::getSignBit(BitWidth)
6114 : APInt::getAllOnesValue(BitWidth),
6115 Op0KnownZero, Op0KnownOne, 0))
6117 if (SimplifyDemandedBits(I.getOperandUse(1),
6118 APInt::getAllOnesValue(BitWidth),
6119 Op1KnownZero, Op1KnownOne, 0))
6122 // Given the known and unknown bits, compute a range that the LHS could be
6123 // in. Compute the Min, Max and RHS values based on the known bits. For the
6124 // EQ and NE we use unsigned values.
6125 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6126 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6128 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6130 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6133 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6135 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6139 // If Min and Max are known to be the same, then SimplifyDemandedBits
6140 // figured out that the LHS is a constant. Just constant fold this now so
6141 // that code below can assume that Min != Max.
6142 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6143 return new ICmpInst(I.getPredicate(),
6144 ConstantInt::get(*Context, Op0Min), Op1);
6145 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6146 return new ICmpInst(I.getPredicate(), Op0,
6147 ConstantInt::get(*Context, Op1Min));
6149 // Based on the range information we know about the LHS, see if we can
6150 // simplify this comparison. For example, (x&4) < 8 is always true.
6151 switch (I.getPredicate()) {
6152 default: llvm_unreachable("Unknown icmp opcode!");
6153 case ICmpInst::ICMP_EQ:
6154 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6155 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6157 case ICmpInst::ICMP_NE:
6158 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6159 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6161 case ICmpInst::ICMP_ULT:
6162 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6163 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6164 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6165 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6166 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6167 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6168 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6169 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6170 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6173 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6174 if (CI->isMinValue(true))
6175 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6176 Constant::getAllOnesValue(Op0->getType()));
6179 case ICmpInst::ICMP_UGT:
6180 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6181 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6182 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6183 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6185 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6186 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6187 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6188 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6189 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6192 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6193 if (CI->isMaxValue(true))
6194 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6195 Constant::getNullValue(Op0->getType()));
6198 case ICmpInst::ICMP_SLT:
6199 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6200 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6201 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6202 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6203 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6204 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6205 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6206 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6207 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6211 case ICmpInst::ICMP_SGT:
6212 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6213 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6214 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6215 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6217 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6218 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6219 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6220 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6221 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6225 case ICmpInst::ICMP_SGE:
6226 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6227 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6228 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6229 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6230 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6232 case ICmpInst::ICMP_SLE:
6233 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6234 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6235 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6236 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6237 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6239 case ICmpInst::ICMP_UGE:
6240 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6241 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6242 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6243 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6244 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6246 case ICmpInst::ICMP_ULE:
6247 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6248 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6249 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6250 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6251 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6255 // Turn a signed comparison into an unsigned one if both operands
6256 // are known to have the same sign.
6258 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6259 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6260 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6263 // Test if the ICmpInst instruction is used exclusively by a select as
6264 // part of a minimum or maximum operation. If so, refrain from doing
6265 // any other folding. This helps out other analyses which understand
6266 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6267 // and CodeGen. And in this case, at least one of the comparison
6268 // operands has at least one user besides the compare (the select),
6269 // which would often largely negate the benefit of folding anyway.
6271 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6272 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6273 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6276 // See if we are doing a comparison between a constant and an instruction that
6277 // can be folded into the comparison.
6278 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6279 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6280 // instruction, see if that instruction also has constants so that the
6281 // instruction can be folded into the icmp
6282 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6283 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6287 // Handle icmp with constant (but not simple integer constant) RHS
6288 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6289 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6290 switch (LHSI->getOpcode()) {
6291 case Instruction::GetElementPtr:
6292 if (RHSC->isNullValue()) {
6293 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6294 bool isAllZeros = true;
6295 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6296 if (!isa<Constant>(LHSI->getOperand(i)) ||
6297 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6302 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6303 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6307 case Instruction::PHI:
6308 // Only fold icmp into the PHI if the phi and icmp are in the same
6309 // block. If in the same block, we're encouraging jump threading. If
6310 // not, we are just pessimizing the code by making an i1 phi.
6311 if (LHSI->getParent() == I.getParent())
6312 if (Instruction *NV = FoldOpIntoPhi(I, true))
6315 case Instruction::Select: {
6316 // If either operand of the select is a constant, we can fold the
6317 // comparison into the select arms, which will cause one to be
6318 // constant folded and the select turned into a bitwise or.
6319 Value *Op1 = 0, *Op2 = 0;
6320 if (LHSI->hasOneUse()) {
6321 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6322 // Fold the known value into the constant operand.
6323 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6324 // Insert a new ICmp of the other select operand.
6325 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
6327 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6328 // Fold the known value into the constant operand.
6329 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6330 // Insert a new ICmp of the other select operand.
6331 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
6337 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6340 case Instruction::Call:
6341 // If we have (malloc != null), and if the malloc has a single use, we
6342 // can assume it is successful and remove the malloc.
6343 if (isMalloc(LHSI) && LHSI->hasOneUse() &&
6344 isa<ConstantPointerNull>(RHSC)) {
6345 // Need to explicitly erase malloc call here, instead of adding it to
6346 // Worklist, because it won't get DCE'd from the Worklist since
6347 // isInstructionTriviallyDead() returns false for function calls.
6348 // It is OK to replace LHSI/MallocCall with Undef because the
6349 // instruction that uses it will be erased via Worklist.
6350 if (extractMallocCall(LHSI)) {
6351 LHSI->replaceAllUsesWith(UndefValue::get(LHSI->getType()));
6352 EraseInstFromFunction(*LHSI);
6353 return ReplaceInstUsesWith(I,
6354 ConstantInt::get(Type::getInt1Ty(*Context),
6355 !I.isTrueWhenEqual()));
6357 if (CallInst* MallocCall = extractMallocCallFromBitCast(LHSI))
6358 if (MallocCall->hasOneUse()) {
6359 MallocCall->replaceAllUsesWith(
6360 UndefValue::get(MallocCall->getType()));
6361 EraseInstFromFunction(*MallocCall);
6362 Worklist.Add(LHSI); // The malloc's bitcast use.
6363 return ReplaceInstUsesWith(I,
6364 ConstantInt::get(Type::getInt1Ty(*Context),
6365 !I.isTrueWhenEqual()));
6372 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6373 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6374 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6376 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6377 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6378 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6381 // Test to see if the operands of the icmp are casted versions of other
6382 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6384 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6385 if (isa<PointerType>(Op0->getType()) &&
6386 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6387 // We keep moving the cast from the left operand over to the right
6388 // operand, where it can often be eliminated completely.
6389 Op0 = CI->getOperand(0);
6391 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6392 // so eliminate it as well.
6393 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6394 Op1 = CI2->getOperand(0);
6396 // If Op1 is a constant, we can fold the cast into the constant.
6397 if (Op0->getType() != Op1->getType()) {
6398 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6399 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6401 // Otherwise, cast the RHS right before the icmp
6402 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
6405 return new ICmpInst(I.getPredicate(), Op0, Op1);
6409 if (isa<CastInst>(Op0)) {
6410 // Handle the special case of: icmp (cast bool to X), <cst>
6411 // This comes up when you have code like
6414 // For generality, we handle any zero-extension of any operand comparison
6415 // with a constant or another cast from the same type.
6416 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6417 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6421 // See if it's the same type of instruction on the left and right.
6422 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6423 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6424 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6425 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6426 switch (Op0I->getOpcode()) {
6428 case Instruction::Add:
6429 case Instruction::Sub:
6430 case Instruction::Xor:
6431 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6432 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6433 Op1I->getOperand(0));
6434 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6435 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6436 if (CI->getValue().isSignBit()) {
6437 ICmpInst::Predicate Pred = I.isSigned()
6438 ? I.getUnsignedPredicate()
6439 : I.getSignedPredicate();
6440 return new ICmpInst(Pred, Op0I->getOperand(0),
6441 Op1I->getOperand(0));
6444 if (CI->getValue().isMaxSignedValue()) {
6445 ICmpInst::Predicate Pred = I.isSigned()
6446 ? I.getUnsignedPredicate()
6447 : I.getSignedPredicate();
6448 Pred = I.getSwappedPredicate(Pred);
6449 return new ICmpInst(Pred, Op0I->getOperand(0),
6450 Op1I->getOperand(0));
6454 case Instruction::Mul:
6455 if (!I.isEquality())
6458 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6459 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6460 // Mask = -1 >> count-trailing-zeros(Cst).
6461 if (!CI->isZero() && !CI->isOne()) {
6462 const APInt &AP = CI->getValue();
6463 ConstantInt *Mask = ConstantInt::get(*Context,
6464 APInt::getLowBitsSet(AP.getBitWidth(),
6466 AP.countTrailingZeros()));
6467 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
6468 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
6469 return new ICmpInst(I.getPredicate(), And1, And2);
6478 // ~x < ~y --> y < x
6480 if (match(Op0, m_Not(m_Value(A))) &&
6481 match(Op1, m_Not(m_Value(B))))
6482 return new ICmpInst(I.getPredicate(), B, A);
6485 if (I.isEquality()) {
6486 Value *A, *B, *C, *D;
6488 // -x == -y --> x == y
6489 if (match(Op0, m_Neg(m_Value(A))) &&
6490 match(Op1, m_Neg(m_Value(B))))
6491 return new ICmpInst(I.getPredicate(), A, B);
6493 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6494 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6495 Value *OtherVal = A == Op1 ? B : A;
6496 return new ICmpInst(I.getPredicate(), OtherVal,
6497 Constant::getNullValue(A->getType()));
6500 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6501 // A^c1 == C^c2 --> A == C^(c1^c2)
6502 ConstantInt *C1, *C2;
6503 if (match(B, m_ConstantInt(C1)) &&
6504 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6506 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6507 Value *Xor = Builder->CreateXor(C, NC, "tmp");
6508 return new ICmpInst(I.getPredicate(), A, Xor);
6511 // A^B == A^D -> B == D
6512 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6513 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6514 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6515 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6519 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6520 (A == Op0 || B == Op0)) {
6521 // A == (A^B) -> B == 0
6522 Value *OtherVal = A == Op0 ? B : A;
6523 return new ICmpInst(I.getPredicate(), OtherVal,
6524 Constant::getNullValue(A->getType()));
6527 // (A-B) == A -> B == 0
6528 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6529 return new ICmpInst(I.getPredicate(), B,
6530 Constant::getNullValue(B->getType()));
6532 // A == (A-B) -> B == 0
6533 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6534 return new ICmpInst(I.getPredicate(), B,
6535 Constant::getNullValue(B->getType()));
6537 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6538 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6539 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6540 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6541 Value *X = 0, *Y = 0, *Z = 0;
6544 X = B; Y = D; Z = A;
6545 } else if (A == D) {
6546 X = B; Y = C; Z = A;
6547 } else if (B == C) {
6548 X = A; Y = D; Z = B;
6549 } else if (B == D) {
6550 X = A; Y = C; Z = B;
6553 if (X) { // Build (X^Y) & Z
6554 Op1 = Builder->CreateXor(X, Y, "tmp");
6555 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
6556 I.setOperand(0, Op1);
6557 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6562 return Changed ? &I : 0;
6566 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6567 /// and CmpRHS are both known to be integer constants.
6568 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6569 ConstantInt *DivRHS) {
6570 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6571 const APInt &CmpRHSV = CmpRHS->getValue();
6573 // FIXME: If the operand types don't match the type of the divide
6574 // then don't attempt this transform. The code below doesn't have the
6575 // logic to deal with a signed divide and an unsigned compare (and
6576 // vice versa). This is because (x /s C1) <s C2 produces different
6577 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6578 // (x /u C1) <u C2. Simply casting the operands and result won't
6579 // work. :( The if statement below tests that condition and bails
6581 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6582 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
6584 if (DivRHS->isZero())
6585 return 0; // The ProdOV computation fails on divide by zero.
6586 if (DivIsSigned && DivRHS->isAllOnesValue())
6587 return 0; // The overflow computation also screws up here
6588 if (DivRHS->isOne())
6589 return 0; // Not worth bothering, and eliminates some funny cases
6592 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6593 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6594 // C2 (CI). By solving for X we can turn this into a range check
6595 // instead of computing a divide.
6596 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6598 // Determine if the product overflows by seeing if the product is
6599 // not equal to the divide. Make sure we do the same kind of divide
6600 // as in the LHS instruction that we're folding.
6601 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6602 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6604 // Get the ICmp opcode
6605 ICmpInst::Predicate Pred = ICI.getPredicate();
6607 // Figure out the interval that is being checked. For example, a comparison
6608 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6609 // Compute this interval based on the constants involved and the signedness of
6610 // the compare/divide. This computes a half-open interval, keeping track of
6611 // whether either value in the interval overflows. After analysis each
6612 // overflow variable is set to 0 if it's corresponding bound variable is valid
6613 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6614 int LoOverflow = 0, HiOverflow = 0;
6615 Constant *LoBound = 0, *HiBound = 0;
6617 if (!DivIsSigned) { // udiv
6618 // e.g. X/5 op 3 --> [15, 20)
6620 HiOverflow = LoOverflow = ProdOV;
6622 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6623 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6624 if (CmpRHSV == 0) { // (X / pos) op 0
6625 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6626 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6628 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6629 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6630 HiOverflow = LoOverflow = ProdOV;
6632 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6633 } else { // (X / pos) op neg
6634 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6635 HiBound = AddOne(Prod);
6636 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6638 ConstantInt* DivNeg =
6639 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6640 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6644 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6645 if (CmpRHSV == 0) { // (X / neg) op 0
6646 // e.g. X/-5 op 0 --> [-4, 5)
6647 LoBound = AddOne(DivRHS);
6648 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6649 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6650 HiOverflow = 1; // [INTMIN+1, overflow)
6651 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6653 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6654 // e.g. X/-5 op 3 --> [-19, -14)
6655 HiBound = AddOne(Prod);
6656 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6658 LoOverflow = AddWithOverflow(LoBound, HiBound,
6659 DivRHS, Context, true) ? -1 : 0;
6660 } else { // (X / neg) op neg
6661 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6662 LoOverflow = HiOverflow = ProdOV;
6664 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6667 // Dividing by a negative swaps the condition. LT <-> GT
6668 Pred = ICmpInst::getSwappedPredicate(Pred);
6671 Value *X = DivI->getOperand(0);
6673 default: llvm_unreachable("Unhandled icmp opcode!");
6674 case ICmpInst::ICMP_EQ:
6675 if (LoOverflow && HiOverflow)
6676 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6677 else if (HiOverflow)
6678 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6679 ICmpInst::ICMP_UGE, X, LoBound);
6680 else if (LoOverflow)
6681 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6682 ICmpInst::ICMP_ULT, X, HiBound);
6684 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6685 case ICmpInst::ICMP_NE:
6686 if (LoOverflow && HiOverflow)
6687 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6688 else if (HiOverflow)
6689 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6690 ICmpInst::ICMP_ULT, X, LoBound);
6691 else if (LoOverflow)
6692 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6693 ICmpInst::ICMP_UGE, X, HiBound);
6695 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6696 case ICmpInst::ICMP_ULT:
6697 case ICmpInst::ICMP_SLT:
6698 if (LoOverflow == +1) // Low bound is greater than input range.
6699 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6700 if (LoOverflow == -1) // Low bound is less than input range.
6701 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6702 return new ICmpInst(Pred, X, LoBound);
6703 case ICmpInst::ICMP_UGT:
6704 case ICmpInst::ICMP_SGT:
6705 if (HiOverflow == +1) // High bound greater than input range.
6706 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6707 else if (HiOverflow == -1) // High bound less than input range.
6708 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6709 if (Pred == ICmpInst::ICMP_UGT)
6710 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6712 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6717 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6719 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6722 const APInt &RHSV = RHS->getValue();
6724 switch (LHSI->getOpcode()) {
6725 case Instruction::Trunc:
6726 if (ICI.isEquality() && LHSI->hasOneUse()) {
6727 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6728 // of the high bits truncated out of x are known.
6729 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6730 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6731 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6732 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6733 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6735 // If all the high bits are known, we can do this xform.
6736 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6737 // Pull in the high bits from known-ones set.
6738 APInt NewRHS(RHS->getValue());
6739 NewRHS.zext(SrcBits);
6741 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6742 ConstantInt::get(*Context, NewRHS));
6747 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6748 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6749 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6751 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6752 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6753 Value *CompareVal = LHSI->getOperand(0);
6755 // If the sign bit of the XorCST is not set, there is no change to
6756 // the operation, just stop using the Xor.
6757 if (!XorCST->getValue().isNegative()) {
6758 ICI.setOperand(0, CompareVal);
6763 // Was the old condition true if the operand is positive?
6764 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6766 // If so, the new one isn't.
6767 isTrueIfPositive ^= true;
6769 if (isTrueIfPositive)
6770 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
6773 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
6777 if (LHSI->hasOneUse()) {
6778 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6779 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6780 const APInt &SignBit = XorCST->getValue();
6781 ICmpInst::Predicate Pred = ICI.isSigned()
6782 ? ICI.getUnsignedPredicate()
6783 : ICI.getSignedPredicate();
6784 return new ICmpInst(Pred, LHSI->getOperand(0),
6785 ConstantInt::get(*Context, RHSV ^ SignBit));
6788 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6789 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6790 const APInt &NotSignBit = XorCST->getValue();
6791 ICmpInst::Predicate Pred = ICI.isSigned()
6792 ? ICI.getUnsignedPredicate()
6793 : ICI.getSignedPredicate();
6794 Pred = ICI.getSwappedPredicate(Pred);
6795 return new ICmpInst(Pred, LHSI->getOperand(0),
6796 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6801 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6802 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6803 LHSI->getOperand(0)->hasOneUse()) {
6804 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6806 // If the LHS is an AND of a truncating cast, we can widen the
6807 // and/compare to be the input width without changing the value
6808 // produced, eliminating a cast.
6809 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6810 // We can do this transformation if either the AND constant does not
6811 // have its sign bit set or if it is an equality comparison.
6812 // Extending a relational comparison when we're checking the sign
6813 // bit would not work.
6814 if (Cast->hasOneUse() &&
6815 (ICI.isEquality() ||
6816 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6818 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6819 APInt NewCST = AndCST->getValue();
6820 NewCST.zext(BitWidth);
6822 NewCI.zext(BitWidth);
6824 Builder->CreateAnd(Cast->getOperand(0),
6825 ConstantInt::get(*Context, NewCST), LHSI->getName());
6826 return new ICmpInst(ICI.getPredicate(), NewAnd,
6827 ConstantInt::get(*Context, NewCI));
6831 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6832 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6833 // happens a LOT in code produced by the C front-end, for bitfield
6835 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6836 if (Shift && !Shift->isShift())
6840 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6841 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6842 const Type *AndTy = AndCST->getType(); // Type of the and.
6844 // We can fold this as long as we can't shift unknown bits
6845 // into the mask. This can only happen with signed shift
6846 // rights, as they sign-extend.
6848 bool CanFold = Shift->isLogicalShift();
6850 // To test for the bad case of the signed shr, see if any
6851 // of the bits shifted in could be tested after the mask.
6852 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6853 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6855 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6856 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6857 AndCST->getValue()) == 0)
6863 if (Shift->getOpcode() == Instruction::Shl)
6864 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6866 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6868 // Check to see if we are shifting out any of the bits being
6870 if (ConstantExpr::get(Shift->getOpcode(),
6871 NewCst, ShAmt) != RHS) {
6872 // If we shifted bits out, the fold is not going to work out.
6873 // As a special case, check to see if this means that the
6874 // result is always true or false now.
6875 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6876 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6877 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6878 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6880 ICI.setOperand(1, NewCst);
6881 Constant *NewAndCST;
6882 if (Shift->getOpcode() == Instruction::Shl)
6883 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6885 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6886 LHSI->setOperand(1, NewAndCST);
6887 LHSI->setOperand(0, Shift->getOperand(0));
6888 Worklist.Add(Shift); // Shift is dead.
6894 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6895 // preferable because it allows the C<<Y expression to be hoisted out
6896 // of a loop if Y is invariant and X is not.
6897 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6898 ICI.isEquality() && !Shift->isArithmeticShift() &&
6899 !isa<Constant>(Shift->getOperand(0))) {
6902 if (Shift->getOpcode() == Instruction::LShr) {
6903 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
6905 // Insert a logical shift.
6906 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
6909 // Compute X & (C << Y).
6911 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6913 ICI.setOperand(0, NewAnd);
6919 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6920 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6923 uint32_t TypeBits = RHSV.getBitWidth();
6925 // Check that the shift amount is in range. If not, don't perform
6926 // undefined shifts. When the shift is visited it will be
6928 if (ShAmt->uge(TypeBits))
6931 if (ICI.isEquality()) {
6932 // If we are comparing against bits always shifted out, the
6933 // comparison cannot succeed.
6935 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6937 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6938 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6939 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6940 return ReplaceInstUsesWith(ICI, Cst);
6943 if (LHSI->hasOneUse()) {
6944 // Otherwise strength reduce the shift into an and.
6945 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6947 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6948 TypeBits-ShAmtVal));
6951 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
6952 return new ICmpInst(ICI.getPredicate(), And,
6953 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6957 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6958 bool TrueIfSigned = false;
6959 if (LHSI->hasOneUse() &&
6960 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6961 // (X << 31) <s 0 --> (X&1) != 0
6962 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6963 (TypeBits-ShAmt->getZExtValue()-1));
6965 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
6966 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6967 And, Constant::getNullValue(And->getType()));
6972 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6973 case Instruction::AShr: {
6974 // Only handle equality comparisons of shift-by-constant.
6975 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6976 if (!ShAmt || !ICI.isEquality()) break;
6978 // Check that the shift amount is in range. If not, don't perform
6979 // undefined shifts. When the shift is visited it will be
6981 uint32_t TypeBits = RHSV.getBitWidth();
6982 if (ShAmt->uge(TypeBits))
6985 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6987 // If we are comparing against bits always shifted out, the
6988 // comparison cannot succeed.
6989 APInt Comp = RHSV << ShAmtVal;
6990 if (LHSI->getOpcode() == Instruction::LShr)
6991 Comp = Comp.lshr(ShAmtVal);
6993 Comp = Comp.ashr(ShAmtVal);
6995 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6996 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6997 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6998 return ReplaceInstUsesWith(ICI, Cst);
7001 // Otherwise, check to see if the bits shifted out are known to be zero.
7002 // If so, we can compare against the unshifted value:
7003 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
7004 if (LHSI->hasOneUse() &&
7005 MaskedValueIsZero(LHSI->getOperand(0),
7006 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
7007 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
7008 ConstantExpr::getShl(RHS, ShAmt));
7011 if (LHSI->hasOneUse()) {
7012 // Otherwise strength reduce the shift into an and.
7013 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7014 Constant *Mask = ConstantInt::get(*Context, Val);
7016 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
7017 Mask, LHSI->getName()+".mask");
7018 return new ICmpInst(ICI.getPredicate(), And,
7019 ConstantExpr::getShl(RHS, ShAmt));
7024 case Instruction::SDiv:
7025 case Instruction::UDiv:
7026 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7027 // Fold this div into the comparison, producing a range check.
7028 // Determine, based on the divide type, what the range is being
7029 // checked. If there is an overflow on the low or high side, remember
7030 // it, otherwise compute the range [low, hi) bounding the new value.
7031 // See: InsertRangeTest above for the kinds of replacements possible.
7032 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7033 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7038 case Instruction::Add:
7039 // Fold: icmp pred (add, X, C1), C2
7041 if (!ICI.isEquality()) {
7042 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7044 const APInt &LHSV = LHSC->getValue();
7046 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7049 if (ICI.isSigned()) {
7050 if (CR.getLower().isSignBit()) {
7051 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7052 ConstantInt::get(*Context, CR.getUpper()));
7053 } else if (CR.getUpper().isSignBit()) {
7054 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7055 ConstantInt::get(*Context, CR.getLower()));
7058 if (CR.getLower().isMinValue()) {
7059 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7060 ConstantInt::get(*Context, CR.getUpper()));
7061 } else if (CR.getUpper().isMinValue()) {
7062 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7063 ConstantInt::get(*Context, CR.getLower()));
7070 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7071 if (ICI.isEquality()) {
7072 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7074 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7075 // the second operand is a constant, simplify a bit.
7076 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7077 switch (BO->getOpcode()) {
7078 case Instruction::SRem:
7079 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7080 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7081 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7082 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7084 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
7086 return new ICmpInst(ICI.getPredicate(), NewRem,
7087 Constant::getNullValue(BO->getType()));
7091 case Instruction::Add:
7092 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7093 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7094 if (BO->hasOneUse())
7095 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7096 ConstantExpr::getSub(RHS, BOp1C));
7097 } else if (RHSV == 0) {
7098 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7099 // efficiently invertible, or if the add has just this one use.
7100 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7102 if (Value *NegVal = dyn_castNegVal(BOp1))
7103 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
7104 else if (Value *NegVal = dyn_castNegVal(BOp0))
7105 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
7106 else if (BO->hasOneUse()) {
7107 Value *Neg = Builder->CreateNeg(BOp1);
7109 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
7113 case Instruction::Xor:
7114 // For the xor case, we can xor two constants together, eliminating
7115 // the explicit xor.
7116 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7117 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7118 ConstantExpr::getXor(RHS, BOC));
7121 case Instruction::Sub:
7122 // Replace (([sub|xor] A, B) != 0) with (A != B)
7124 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7128 case Instruction::Or:
7129 // If bits are being or'd in that are not present in the constant we
7130 // are comparing against, then the comparison could never succeed!
7131 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7132 Constant *NotCI = ConstantExpr::getNot(RHS);
7133 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7134 return ReplaceInstUsesWith(ICI,
7135 ConstantInt::get(Type::getInt1Ty(*Context),
7140 case Instruction::And:
7141 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7142 // If bits are being compared against that are and'd out, then the
7143 // comparison can never succeed!
7144 if ((RHSV & ~BOC->getValue()) != 0)
7145 return ReplaceInstUsesWith(ICI,
7146 ConstantInt::get(Type::getInt1Ty(*Context),
7149 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7150 if (RHS == BOC && RHSV.isPowerOf2())
7151 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7152 ICmpInst::ICMP_NE, LHSI,
7153 Constant::getNullValue(RHS->getType()));
7155 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7156 if (BOC->getValue().isSignBit()) {
7157 Value *X = BO->getOperand(0);
7158 Constant *Zero = Constant::getNullValue(X->getType());
7159 ICmpInst::Predicate pred = isICMP_NE ?
7160 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7161 return new ICmpInst(pred, X, Zero);
7164 // ((X & ~7) == 0) --> X < 8
7165 if (RHSV == 0 && isHighOnes(BOC)) {
7166 Value *X = BO->getOperand(0);
7167 Constant *NegX = ConstantExpr::getNeg(BOC);
7168 ICmpInst::Predicate pred = isICMP_NE ?
7169 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7170 return new ICmpInst(pred, X, NegX);
7175 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7176 // Handle icmp {eq|ne} <intrinsic>, intcst.
7177 if (II->getIntrinsicID() == Intrinsic::bswap) {
7179 ICI.setOperand(0, II->getOperand(1));
7180 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7188 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7189 /// We only handle extending casts so far.
7191 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7192 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7193 Value *LHSCIOp = LHSCI->getOperand(0);
7194 const Type *SrcTy = LHSCIOp->getType();
7195 const Type *DestTy = LHSCI->getType();
7198 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7199 // integer type is the same size as the pointer type.
7200 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7201 TD->getPointerSizeInBits() ==
7202 cast<IntegerType>(DestTy)->getBitWidth()) {
7204 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7205 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7206 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7207 RHSOp = RHSC->getOperand(0);
7208 // If the pointer types don't match, insert a bitcast.
7209 if (LHSCIOp->getType() != RHSOp->getType())
7210 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
7214 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7217 // The code below only handles extension cast instructions, so far.
7219 if (LHSCI->getOpcode() != Instruction::ZExt &&
7220 LHSCI->getOpcode() != Instruction::SExt)
7223 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7224 bool isSignedCmp = ICI.isSigned();
7226 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7227 // Not an extension from the same type?
7228 RHSCIOp = CI->getOperand(0);
7229 if (RHSCIOp->getType() != LHSCIOp->getType())
7232 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7233 // and the other is a zext), then we can't handle this.
7234 if (CI->getOpcode() != LHSCI->getOpcode())
7237 // Deal with equality cases early.
7238 if (ICI.isEquality())
7239 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7241 // A signed comparison of sign extended values simplifies into a
7242 // signed comparison.
7243 if (isSignedCmp && isSignedExt)
7244 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7246 // The other three cases all fold into an unsigned comparison.
7247 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7250 // If we aren't dealing with a constant on the RHS, exit early
7251 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7255 // Compute the constant that would happen if we truncated to SrcTy then
7256 // reextended to DestTy.
7257 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7258 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7261 // If the re-extended constant didn't change...
7263 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7264 // For example, we might have:
7265 // %A = sext i16 %X to i32
7266 // %B = icmp ugt i32 %A, 1330
7267 // It is incorrect to transform this into
7268 // %B = icmp ugt i16 %X, 1330
7269 // because %A may have negative value.
7271 // However, we allow this when the compare is EQ/NE, because they are
7273 if (isSignedExt == isSignedCmp || ICI.isEquality())
7274 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7278 // The re-extended constant changed so the constant cannot be represented
7279 // in the shorter type. Consequently, we cannot emit a simple comparison.
7281 // First, handle some easy cases. We know the result cannot be equal at this
7282 // point so handle the ICI.isEquality() cases
7283 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7284 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7285 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7286 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7288 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7289 // should have been folded away previously and not enter in here.
7292 // We're performing a signed comparison.
7293 if (cast<ConstantInt>(CI)->getValue().isNegative())
7294 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7296 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7298 // We're performing an unsigned comparison.
7300 // We're performing an unsigned comp with a sign extended value.
7301 // This is true if the input is >= 0. [aka >s -1]
7302 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7303 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
7305 // Unsigned extend & unsigned compare -> always true.
7306 Result = ConstantInt::getTrue(*Context);
7310 // Finally, return the value computed.
7311 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7312 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7313 return ReplaceInstUsesWith(ICI, Result);
7315 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7316 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7317 "ICmp should be folded!");
7318 if (Constant *CI = dyn_cast<Constant>(Result))
7319 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7320 return BinaryOperator::CreateNot(Result);
7323 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7324 return commonShiftTransforms(I);
7327 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7328 return commonShiftTransforms(I);
7331 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7332 if (Instruction *R = commonShiftTransforms(I))
7335 Value *Op0 = I.getOperand(0);
7337 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7338 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7339 if (CSI->isAllOnesValue())
7340 return ReplaceInstUsesWith(I, CSI);
7342 // See if we can turn a signed shr into an unsigned shr.
7343 if (MaskedValueIsZero(Op0,
7344 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7345 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7347 // Arithmetic shifting an all-sign-bit value is a no-op.
7348 unsigned NumSignBits = ComputeNumSignBits(Op0);
7349 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7350 return ReplaceInstUsesWith(I, Op0);
7355 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7356 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7357 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7359 // shl X, 0 == X and shr X, 0 == X
7360 // shl 0, X == 0 and shr 0, X == 0
7361 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7362 Op0 == Constant::getNullValue(Op0->getType()))
7363 return ReplaceInstUsesWith(I, Op0);
7365 if (isa<UndefValue>(Op0)) {
7366 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7367 return ReplaceInstUsesWith(I, Op0);
7368 else // undef << X -> 0, undef >>u X -> 0
7369 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7371 if (isa<UndefValue>(Op1)) {
7372 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7373 return ReplaceInstUsesWith(I, Op0);
7374 else // X << undef, X >>u undef -> 0
7375 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7378 // See if we can fold away this shift.
7379 if (SimplifyDemandedInstructionBits(I))
7382 // Try to fold constant and into select arguments.
7383 if (isa<Constant>(Op0))
7384 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7385 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7388 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7389 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7394 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7395 BinaryOperator &I) {
7396 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7398 // See if we can simplify any instructions used by the instruction whose sole
7399 // purpose is to compute bits we don't care about.
7400 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7402 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7405 if (Op1->uge(TypeBits)) {
7406 if (I.getOpcode() != Instruction::AShr)
7407 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7409 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7414 // ((X*C1) << C2) == (X * (C1 << C2))
7415 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7416 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7417 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7418 return BinaryOperator::CreateMul(BO->getOperand(0),
7419 ConstantExpr::getShl(BOOp, Op1));
7421 // Try to fold constant and into select arguments.
7422 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7423 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7425 if (isa<PHINode>(Op0))
7426 if (Instruction *NV = FoldOpIntoPhi(I))
7429 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7430 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7431 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7432 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7433 // place. Don't try to do this transformation in this case. Also, we
7434 // require that the input operand is a shift-by-constant so that we have
7435 // confidence that the shifts will get folded together. We could do this
7436 // xform in more cases, but it is unlikely to be profitable.
7437 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7438 isa<ConstantInt>(TrOp->getOperand(1))) {
7439 // Okay, we'll do this xform. Make the shift of shift.
7440 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7441 // (shift2 (shift1 & 0x00FF), c2)
7442 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
7444 // For logical shifts, the truncation has the effect of making the high
7445 // part of the register be zeros. Emulate this by inserting an AND to
7446 // clear the top bits as needed. This 'and' will usually be zapped by
7447 // other xforms later if dead.
7448 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7449 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7450 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7452 // The mask we constructed says what the trunc would do if occurring
7453 // between the shifts. We want to know the effect *after* the second
7454 // shift. We know that it is a logical shift by a constant, so adjust the
7455 // mask as appropriate.
7456 if (I.getOpcode() == Instruction::Shl)
7457 MaskV <<= Op1->getZExtValue();
7459 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7460 MaskV = MaskV.lshr(Op1->getZExtValue());
7464 Value *And = Builder->CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7467 // Return the value truncated to the interesting size.
7468 return new TruncInst(And, I.getType());
7472 if (Op0->hasOneUse()) {
7473 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7474 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7477 switch (Op0BO->getOpcode()) {
7479 case Instruction::Add:
7480 case Instruction::And:
7481 case Instruction::Or:
7482 case Instruction::Xor: {
7483 // These operators commute.
7484 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7485 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7486 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7487 m_Specific(Op1)))) {
7488 Value *YS = // (Y << C)
7489 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
7491 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
7492 Op0BO->getOperand(1)->getName());
7493 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7494 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7495 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7498 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7499 Value *Op0BOOp1 = Op0BO->getOperand(1);
7500 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7502 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7503 m_ConstantInt(CC))) &&
7504 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7505 Value *YS = // (Y << C)
7506 Builder->CreateShl(Op0BO->getOperand(0), Op1,
7509 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7510 V1->getName()+".mask");
7511 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7516 case Instruction::Sub: {
7517 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7518 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7519 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7520 m_Specific(Op1)))) {
7521 Value *YS = // (Y << C)
7522 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7524 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
7525 Op0BO->getOperand(0)->getName());
7526 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7527 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7528 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7531 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7532 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7533 match(Op0BO->getOperand(0),
7534 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7535 m_ConstantInt(CC))) && V2 == Op1 &&
7536 cast<BinaryOperator>(Op0BO->getOperand(0))
7537 ->getOperand(0)->hasOneUse()) {
7538 Value *YS = // (Y << C)
7539 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7541 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7542 V1->getName()+".mask");
7544 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7552 // If the operand is an bitwise operator with a constant RHS, and the
7553 // shift is the only use, we can pull it out of the shift.
7554 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7555 bool isValid = true; // Valid only for And, Or, Xor
7556 bool highBitSet = false; // Transform if high bit of constant set?
7558 switch (Op0BO->getOpcode()) {
7559 default: isValid = false; break; // Do not perform transform!
7560 case Instruction::Add:
7561 isValid = isLeftShift;
7563 case Instruction::Or:
7564 case Instruction::Xor:
7567 case Instruction::And:
7572 // If this is a signed shift right, and the high bit is modified
7573 // by the logical operation, do not perform the transformation.
7574 // The highBitSet boolean indicates the value of the high bit of
7575 // the constant which would cause it to be modified for this
7578 if (isValid && I.getOpcode() == Instruction::AShr)
7579 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7582 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7585 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
7586 NewShift->takeName(Op0BO);
7588 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7595 // Find out if this is a shift of a shift by a constant.
7596 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7597 if (ShiftOp && !ShiftOp->isShift())
7600 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7601 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7602 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7603 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7604 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7605 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7606 Value *X = ShiftOp->getOperand(0);
7608 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7610 const IntegerType *Ty = cast<IntegerType>(I.getType());
7612 // Check for (X << c1) << c2 and (X >> c1) >> c2
7613 if (I.getOpcode() == ShiftOp->getOpcode()) {
7614 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7616 if (AmtSum >= TypeBits) {
7617 if (I.getOpcode() != Instruction::AShr)
7618 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7619 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7622 return BinaryOperator::Create(I.getOpcode(), X,
7623 ConstantInt::get(Ty, AmtSum));
7626 if (ShiftOp->getOpcode() == Instruction::LShr &&
7627 I.getOpcode() == Instruction::AShr) {
7628 if (AmtSum >= TypeBits)
7629 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7631 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7632 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7635 if (ShiftOp->getOpcode() == Instruction::AShr &&
7636 I.getOpcode() == Instruction::LShr) {
7637 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7638 if (AmtSum >= TypeBits)
7639 AmtSum = TypeBits-1;
7641 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7643 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7644 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7647 // Okay, if we get here, one shift must be left, and the other shift must be
7648 // right. See if the amounts are equal.
7649 if (ShiftAmt1 == ShiftAmt2) {
7650 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7651 if (I.getOpcode() == Instruction::Shl) {
7652 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7653 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7655 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7656 if (I.getOpcode() == Instruction::LShr) {
7657 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7658 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7660 // We can simplify ((X << C) >>s C) into a trunc + sext.
7661 // NOTE: we could do this for any C, but that would make 'unusual' integer
7662 // types. For now, just stick to ones well-supported by the code
7664 const Type *SExtType = 0;
7665 switch (Ty->getBitWidth() - ShiftAmt1) {
7672 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
7677 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
7678 // Otherwise, we can't handle it yet.
7679 } else if (ShiftAmt1 < ShiftAmt2) {
7680 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7682 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7683 if (I.getOpcode() == Instruction::Shl) {
7684 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7685 ShiftOp->getOpcode() == Instruction::AShr);
7686 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7688 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7689 return BinaryOperator::CreateAnd(Shift,
7690 ConstantInt::get(*Context, Mask));
7693 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7694 if (I.getOpcode() == Instruction::LShr) {
7695 assert(ShiftOp->getOpcode() == Instruction::Shl);
7696 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7698 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7699 return BinaryOperator::CreateAnd(Shift,
7700 ConstantInt::get(*Context, Mask));
7703 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7705 assert(ShiftAmt2 < ShiftAmt1);
7706 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7708 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7709 if (I.getOpcode() == Instruction::Shl) {
7710 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7711 ShiftOp->getOpcode() == Instruction::AShr);
7712 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
7713 ConstantInt::get(Ty, ShiftDiff));
7715 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7716 return BinaryOperator::CreateAnd(Shift,
7717 ConstantInt::get(*Context, Mask));
7720 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7721 if (I.getOpcode() == Instruction::LShr) {
7722 assert(ShiftOp->getOpcode() == Instruction::Shl);
7723 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7725 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7726 return BinaryOperator::CreateAnd(Shift,
7727 ConstantInt::get(*Context, Mask));
7730 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7737 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7738 /// expression. If so, decompose it, returning some value X, such that Val is
7741 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7742 int &Offset, LLVMContext *Context) {
7743 assert(Val->getType() == Type::getInt32Ty(*Context) &&
7744 "Unexpected allocation size type!");
7745 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7746 Offset = CI->getZExtValue();
7748 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
7749 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7750 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7751 if (I->getOpcode() == Instruction::Shl) {
7752 // This is a value scaled by '1 << the shift amt'.
7753 Scale = 1U << RHS->getZExtValue();
7755 return I->getOperand(0);
7756 } else if (I->getOpcode() == Instruction::Mul) {
7757 // This value is scaled by 'RHS'.
7758 Scale = RHS->getZExtValue();
7760 return I->getOperand(0);
7761 } else if (I->getOpcode() == Instruction::Add) {
7762 // We have X+C. Check to see if we really have (X*C2)+C1,
7763 // where C1 is divisible by C2.
7766 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7768 Offset += RHS->getZExtValue();
7775 // Otherwise, we can't look past this.
7782 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7783 /// try to eliminate the cast by moving the type information into the alloc.
7784 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7786 const PointerType *PTy = cast<PointerType>(CI.getType());
7788 BuilderTy AllocaBuilder(*Builder);
7789 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
7791 // Remove any uses of AI that are dead.
7792 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7794 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7795 Instruction *User = cast<Instruction>(*UI++);
7796 if (isInstructionTriviallyDead(User)) {
7797 while (UI != E && *UI == User)
7798 ++UI; // If this instruction uses AI more than once, don't break UI.
7801 DEBUG(errs() << "IC: DCE: " << *User << '\n');
7802 EraseInstFromFunction(*User);
7806 // This requires TargetData to get the alloca alignment and size information.
7809 // Get the type really allocated and the type casted to.
7810 const Type *AllocElTy = AI.getAllocatedType();
7811 const Type *CastElTy = PTy->getElementType();
7812 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7814 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7815 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7816 if (CastElTyAlign < AllocElTyAlign) return 0;
7818 // If the allocation has multiple uses, only promote it if we are strictly
7819 // increasing the alignment of the resultant allocation. If we keep it the
7820 // same, we open the door to infinite loops of various kinds. (A reference
7821 // from a dbg.declare doesn't count as a use for this purpose.)
7822 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7823 CastElTyAlign == AllocElTyAlign) return 0;
7825 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7826 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7827 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7829 // See if we can satisfy the modulus by pulling a scale out of the array
7831 unsigned ArraySizeScale;
7833 Value *NumElements = // See if the array size is a decomposable linear expr.
7834 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7835 ArrayOffset, Context);
7837 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7839 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7840 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7842 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7847 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
7848 // Insert before the alloca, not before the cast.
7849 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
7852 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7853 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
7854 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
7857 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
7858 New->setAlignment(AI.getAlignment());
7861 // If the allocation has one real use plus a dbg.declare, just remove the
7863 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7864 EraseInstFromFunction(*DI);
7866 // If the allocation has multiple real uses, insert a cast and change all
7867 // things that used it to use the new cast. This will also hack on CI, but it
7869 else if (!AI.hasOneUse()) {
7870 // New is the allocation instruction, pointer typed. AI is the original
7871 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7872 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
7873 AI.replaceAllUsesWith(NewCast);
7875 return ReplaceInstUsesWith(CI, New);
7878 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7879 /// and return it as type Ty without inserting any new casts and without
7880 /// changing the computed value. This is used by code that tries to decide
7881 /// whether promoting or shrinking integer operations to wider or smaller types
7882 /// will allow us to eliminate a truncate or extend.
7884 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7885 /// extension operation if Ty is larger.
7887 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7888 /// should return true if trunc(V) can be computed by computing V in the smaller
7889 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7890 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7891 /// efficiently truncated.
7893 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7894 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7895 /// the final result.
7896 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7898 int &NumCastsRemoved){
7899 // We can always evaluate constants in another type.
7900 if (isa<Constant>(V))
7903 Instruction *I = dyn_cast<Instruction>(V);
7904 if (!I) return false;
7906 const Type *OrigTy = V->getType();
7908 // If this is an extension or truncate, we can often eliminate it.
7909 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7910 // If this is a cast from the destination type, we can trivially eliminate
7911 // it, and this will remove a cast overall.
7912 if (I->getOperand(0)->getType() == Ty) {
7913 // If the first operand is itself a cast, and is eliminable, do not count
7914 // this as an eliminable cast. We would prefer to eliminate those two
7916 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7922 // We can't extend or shrink something that has multiple uses: doing so would
7923 // require duplicating the instruction in general, which isn't profitable.
7924 if (!I->hasOneUse()) return false;
7926 unsigned Opc = I->getOpcode();
7928 case Instruction::Add:
7929 case Instruction::Sub:
7930 case Instruction::Mul:
7931 case Instruction::And:
7932 case Instruction::Or:
7933 case Instruction::Xor:
7934 // These operators can all arbitrarily be extended or truncated.
7935 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7937 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7940 case Instruction::UDiv:
7941 case Instruction::URem: {
7942 // UDiv and URem can be truncated if all the truncated bits are zero.
7943 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7944 uint32_t BitWidth = Ty->getScalarSizeInBits();
7945 if (BitWidth < OrigBitWidth) {
7946 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7947 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7948 MaskedValueIsZero(I->getOperand(1), Mask)) {
7949 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7951 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7957 case Instruction::Shl:
7958 // If we are truncating the result of this SHL, and if it's a shift of a
7959 // constant amount, we can always perform a SHL in a smaller type.
7960 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7961 uint32_t BitWidth = Ty->getScalarSizeInBits();
7962 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7963 CI->getLimitedValue(BitWidth) < BitWidth)
7964 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7968 case Instruction::LShr:
7969 // If this is a truncate of a logical shr, we can truncate it to a smaller
7970 // lshr iff we know that the bits we would otherwise be shifting in are
7972 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7973 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7974 uint32_t BitWidth = Ty->getScalarSizeInBits();
7975 if (BitWidth < OrigBitWidth &&
7976 MaskedValueIsZero(I->getOperand(0),
7977 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7978 CI->getLimitedValue(BitWidth) < BitWidth) {
7979 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7984 case Instruction::ZExt:
7985 case Instruction::SExt:
7986 case Instruction::Trunc:
7987 // If this is the same kind of case as our original (e.g. zext+zext), we
7988 // can safely replace it. Note that replacing it does not reduce the number
7989 // of casts in the input.
7993 // sext (zext ty1), ty2 -> zext ty2
7994 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7997 case Instruction::Select: {
7998 SelectInst *SI = cast<SelectInst>(I);
7999 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8001 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8004 case Instruction::PHI: {
8005 // We can change a phi if we can change all operands.
8006 PHINode *PN = cast<PHINode>(I);
8007 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8008 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8014 // TODO: Can handle more cases here.
8021 /// EvaluateInDifferentType - Given an expression that
8022 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8023 /// evaluate the expression.
8024 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8026 if (Constant *C = dyn_cast<Constant>(V))
8027 return ConstantExpr::getIntegerCast(C, Ty,
8028 isSigned /*Sext or ZExt*/);
8030 // Otherwise, it must be an instruction.
8031 Instruction *I = cast<Instruction>(V);
8032 Instruction *Res = 0;
8033 unsigned Opc = I->getOpcode();
8035 case Instruction::Add:
8036 case Instruction::Sub:
8037 case Instruction::Mul:
8038 case Instruction::And:
8039 case Instruction::Or:
8040 case Instruction::Xor:
8041 case Instruction::AShr:
8042 case Instruction::LShr:
8043 case Instruction::Shl:
8044 case Instruction::UDiv:
8045 case Instruction::URem: {
8046 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8047 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8048 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8051 case Instruction::Trunc:
8052 case Instruction::ZExt:
8053 case Instruction::SExt:
8054 // If the source type of the cast is the type we're trying for then we can
8055 // just return the source. There's no need to insert it because it is not
8057 if (I->getOperand(0)->getType() == Ty)
8058 return I->getOperand(0);
8060 // Otherwise, must be the same type of cast, so just reinsert a new one.
8061 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8064 case Instruction::Select: {
8065 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8066 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8067 Res = SelectInst::Create(I->getOperand(0), True, False);
8070 case Instruction::PHI: {
8071 PHINode *OPN = cast<PHINode>(I);
8072 PHINode *NPN = PHINode::Create(Ty);
8073 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8074 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8075 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8081 // TODO: Can handle more cases here.
8082 llvm_unreachable("Unreachable!");
8087 return InsertNewInstBefore(Res, *I);
8090 /// @brief Implement the transforms common to all CastInst visitors.
8091 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8092 Value *Src = CI.getOperand(0);
8094 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8095 // eliminate it now.
8096 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8097 if (Instruction::CastOps opc =
8098 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8099 // The first cast (CSrc) is eliminable so we need to fix up or replace
8100 // the second cast (CI). CSrc will then have a good chance of being dead.
8101 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8105 // If we are casting a select then fold the cast into the select
8106 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8107 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8110 // If we are casting a PHI then fold the cast into the PHI
8111 if (isa<PHINode>(Src))
8112 if (Instruction *NV = FoldOpIntoPhi(CI))
8118 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8119 /// or not there is a sequence of GEP indices into the type that will land us at
8120 /// the specified offset. If so, fill them into NewIndices and return the
8121 /// resultant element type, otherwise return null.
8122 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8123 SmallVectorImpl<Value*> &NewIndices,
8124 const TargetData *TD,
8125 LLVMContext *Context) {
8127 if (!Ty->isSized()) return 0;
8129 // Start with the index over the outer type. Note that the type size
8130 // might be zero (even if the offset isn't zero) if the indexed type
8131 // is something like [0 x {int, int}]
8132 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8133 int64_t FirstIdx = 0;
8134 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8135 FirstIdx = Offset/TySize;
8136 Offset -= FirstIdx*TySize;
8138 // Handle hosts where % returns negative instead of values [0..TySize).
8142 assert(Offset >= 0);
8144 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8147 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8149 // Index into the types. If we fail, set OrigBase to null.
8151 // Indexing into tail padding between struct/array elements.
8152 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8155 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8156 const StructLayout *SL = TD->getStructLayout(STy);
8157 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8158 "Offset must stay within the indexed type");
8160 unsigned Elt = SL->getElementContainingOffset(Offset);
8161 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8163 Offset -= SL->getElementOffset(Elt);
8164 Ty = STy->getElementType(Elt);
8165 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8166 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8167 assert(EltSize && "Cannot index into a zero-sized array");
8168 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8170 Ty = AT->getElementType();
8172 // Otherwise, we can't index into the middle of this atomic type, bail.
8180 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8181 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8182 Value *Src = CI.getOperand(0);
8184 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8185 // If casting the result of a getelementptr instruction with no offset, turn
8186 // this into a cast of the original pointer!
8187 if (GEP->hasAllZeroIndices()) {
8188 // Changing the cast operand is usually not a good idea but it is safe
8189 // here because the pointer operand is being replaced with another
8190 // pointer operand so the opcode doesn't need to change.
8192 CI.setOperand(0, GEP->getOperand(0));
8196 // If the GEP has a single use, and the base pointer is a bitcast, and the
8197 // GEP computes a constant offset, see if we can convert these three
8198 // instructions into fewer. This typically happens with unions and other
8199 // non-type-safe code.
8200 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8201 if (GEP->hasAllConstantIndices()) {
8202 // We are guaranteed to get a constant from EmitGEPOffset.
8203 ConstantInt *OffsetV =
8204 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8205 int64_t Offset = OffsetV->getSExtValue();
8207 // Get the base pointer input of the bitcast, and the type it points to.
8208 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8209 const Type *GEPIdxTy =
8210 cast<PointerType>(OrigBase->getType())->getElementType();
8211 SmallVector<Value*, 8> NewIndices;
8212 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8213 // If we were able to index down into an element, create the GEP
8214 // and bitcast the result. This eliminates one bitcast, potentially
8216 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
8217 Builder->CreateInBoundsGEP(OrigBase,
8218 NewIndices.begin(), NewIndices.end()) :
8219 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
8220 NGEP->takeName(GEP);
8222 if (isa<BitCastInst>(CI))
8223 return new BitCastInst(NGEP, CI.getType());
8224 assert(isa<PtrToIntInst>(CI));
8225 return new PtrToIntInst(NGEP, CI.getType());
8231 return commonCastTransforms(CI);
8234 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8235 /// type like i42. We don't want to introduce operations on random non-legal
8236 /// integer types where they don't already exist in the code. In the future,
8237 /// we should consider making this based off target-data, so that 32-bit targets
8238 /// won't get i64 operations etc.
8239 static bool isSafeIntegerType(const Type *Ty) {
8240 switch (Ty->getPrimitiveSizeInBits()) {
8251 /// commonIntCastTransforms - This function implements the common transforms
8252 /// for trunc, zext, and sext.
8253 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8254 if (Instruction *Result = commonCastTransforms(CI))
8257 Value *Src = CI.getOperand(0);
8258 const Type *SrcTy = Src->getType();
8259 const Type *DestTy = CI.getType();
8260 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8261 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8263 // See if we can simplify any instructions used by the LHS whose sole
8264 // purpose is to compute bits we don't care about.
8265 if (SimplifyDemandedInstructionBits(CI))
8268 // If the source isn't an instruction or has more than one use then we
8269 // can't do anything more.
8270 Instruction *SrcI = dyn_cast<Instruction>(Src);
8271 if (!SrcI || !Src->hasOneUse())
8274 // Attempt to propagate the cast into the instruction for int->int casts.
8275 int NumCastsRemoved = 0;
8276 // Only do this if the dest type is a simple type, don't convert the
8277 // expression tree to something weird like i93 unless the source is also
8279 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8280 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8281 CanEvaluateInDifferentType(SrcI, DestTy,
8282 CI.getOpcode(), NumCastsRemoved)) {
8283 // If this cast is a truncate, evaluting in a different type always
8284 // eliminates the cast, so it is always a win. If this is a zero-extension,
8285 // we need to do an AND to maintain the clear top-part of the computation,
8286 // so we require that the input have eliminated at least one cast. If this
8287 // is a sign extension, we insert two new casts (to do the extension) so we
8288 // require that two casts have been eliminated.
8289 bool DoXForm = false;
8290 bool JustReplace = false;
8291 switch (CI.getOpcode()) {
8293 // All the others use floating point so we shouldn't actually
8294 // get here because of the check above.
8295 llvm_unreachable("Unknown cast type");
8296 case Instruction::Trunc:
8299 case Instruction::ZExt: {
8300 DoXForm = NumCastsRemoved >= 1;
8301 if (!DoXForm && 0) {
8302 // If it's unnecessary to issue an AND to clear the high bits, it's
8303 // always profitable to do this xform.
8304 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8305 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8306 if (MaskedValueIsZero(TryRes, Mask))
8307 return ReplaceInstUsesWith(CI, TryRes);
8309 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8310 if (TryI->use_empty())
8311 EraseInstFromFunction(*TryI);
8315 case Instruction::SExt: {
8316 DoXForm = NumCastsRemoved >= 2;
8317 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8318 // If we do not have to emit the truncate + sext pair, then it's always
8319 // profitable to do this xform.
8321 // It's not safe to eliminate the trunc + sext pair if one of the
8322 // eliminated cast is a truncate. e.g.
8323 // t2 = trunc i32 t1 to i16
8324 // t3 = sext i16 t2 to i32
8327 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8328 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8329 if (NumSignBits > (DestBitSize - SrcBitSize))
8330 return ReplaceInstUsesWith(CI, TryRes);
8332 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8333 if (TryI->use_empty())
8334 EraseInstFromFunction(*TryI);
8341 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8342 " to avoid cast: " << CI);
8343 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8344 CI.getOpcode() == Instruction::SExt);
8346 // Just replace this cast with the result.
8347 return ReplaceInstUsesWith(CI, Res);
8349 assert(Res->getType() == DestTy);
8350 switch (CI.getOpcode()) {
8351 default: llvm_unreachable("Unknown cast type!");
8352 case Instruction::Trunc:
8353 // Just replace this cast with the result.
8354 return ReplaceInstUsesWith(CI, Res);
8355 case Instruction::ZExt: {
8356 assert(SrcBitSize < DestBitSize && "Not a zext?");
8358 // If the high bits are already zero, just replace this cast with the
8360 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8361 if (MaskedValueIsZero(Res, Mask))
8362 return ReplaceInstUsesWith(CI, Res);
8364 // We need to emit an AND to clear the high bits.
8365 Constant *C = ConstantInt::get(*Context,
8366 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8367 return BinaryOperator::CreateAnd(Res, C);
8369 case Instruction::SExt: {
8370 // If the high bits are already filled with sign bit, just replace this
8371 // cast with the result.
8372 unsigned NumSignBits = ComputeNumSignBits(Res);
8373 if (NumSignBits > (DestBitSize - SrcBitSize))
8374 return ReplaceInstUsesWith(CI, Res);
8376 // We need to emit a cast to truncate, then a cast to sext.
8377 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
8383 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8384 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8386 switch (SrcI->getOpcode()) {
8387 case Instruction::Add:
8388 case Instruction::Mul:
8389 case Instruction::And:
8390 case Instruction::Or:
8391 case Instruction::Xor:
8392 // If we are discarding information, rewrite.
8393 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8394 // Don't insert two casts unless at least one can be eliminated.
8395 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8396 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8397 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8398 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8399 return BinaryOperator::Create(
8400 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8404 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8405 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8406 SrcI->getOpcode() == Instruction::Xor &&
8407 Op1 == ConstantInt::getTrue(*Context) &&
8408 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8409 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName());
8410 return BinaryOperator::CreateXor(New,
8411 ConstantInt::get(CI.getType(), 1));
8415 case Instruction::Shl: {
8416 // Canonicalize trunc inside shl, if we can.
8417 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8418 if (CI && DestBitSize < SrcBitSize &&
8419 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8420 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8421 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8422 return BinaryOperator::CreateShl(Op0c, Op1c);
8430 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8431 if (Instruction *Result = commonIntCastTransforms(CI))
8434 Value *Src = CI.getOperand(0);
8435 const Type *Ty = CI.getType();
8436 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8437 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8439 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8440 if (DestBitWidth == 1) {
8441 Constant *One = ConstantInt::get(Src->getType(), 1);
8442 Src = Builder->CreateAnd(Src, One, "tmp");
8443 Value *Zero = Constant::getNullValue(Src->getType());
8444 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8447 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8448 ConstantInt *ShAmtV = 0;
8450 if (Src->hasOneUse() &&
8451 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8452 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8454 // Get a mask for the bits shifting in.
8455 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8456 if (MaskedValueIsZero(ShiftOp, Mask)) {
8457 if (ShAmt >= DestBitWidth) // All zeros.
8458 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8460 // Okay, we can shrink this. Truncate the input, then return a new
8462 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName());
8463 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8464 return BinaryOperator::CreateLShr(V1, V2);
8471 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8472 /// in order to eliminate the icmp.
8473 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8475 // If we are just checking for a icmp eq of a single bit and zext'ing it
8476 // to an integer, then shift the bit to the appropriate place and then
8477 // cast to integer to avoid the comparison.
8478 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8479 const APInt &Op1CV = Op1C->getValue();
8481 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8482 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8483 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8484 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8485 if (!DoXform) return ICI;
8487 Value *In = ICI->getOperand(0);
8488 Value *Sh = ConstantInt::get(In->getType(),
8489 In->getType()->getScalarSizeInBits()-1);
8490 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
8491 if (In->getType() != CI.getType())
8492 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
8494 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8495 Constant *One = ConstantInt::get(In->getType(), 1);
8496 In = Builder->CreateXor(In, One, In->getName()+".not");
8499 return ReplaceInstUsesWith(CI, In);
8504 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8505 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8506 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8507 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8508 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8509 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8510 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8511 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8512 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8513 // This only works for EQ and NE
8514 ICI->isEquality()) {
8515 // If Op1C some other power of two, convert:
8516 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8517 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8518 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8519 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8521 APInt KnownZeroMask(~KnownZero);
8522 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8523 if (!DoXform) return ICI;
8525 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8526 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8527 // (X&4) == 2 --> false
8528 // (X&4) != 2 --> true
8529 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8530 Res = ConstantExpr::getZExt(Res, CI.getType());
8531 return ReplaceInstUsesWith(CI, Res);
8534 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8535 Value *In = ICI->getOperand(0);
8537 // Perform a logical shr by shiftamt.
8538 // Insert the shift to put the result in the low bit.
8539 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
8540 In->getName()+".lobit");
8543 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8544 Constant *One = ConstantInt::get(In->getType(), 1);
8545 In = Builder->CreateXor(In, One, "tmp");
8548 if (CI.getType() == In->getType())
8549 return ReplaceInstUsesWith(CI, In);
8551 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8559 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8560 // If one of the common conversion will work ..
8561 if (Instruction *Result = commonIntCastTransforms(CI))
8564 Value *Src = CI.getOperand(0);
8566 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8567 // types and if the sizes are just right we can convert this into a logical
8568 // 'and' which will be much cheaper than the pair of casts.
8569 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8570 // Get the sizes of the types involved. We know that the intermediate type
8571 // will be smaller than A or C, but don't know the relation between A and C.
8572 Value *A = CSrc->getOperand(0);
8573 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8574 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8575 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8576 // If we're actually extending zero bits, then if
8577 // SrcSize < DstSize: zext(a & mask)
8578 // SrcSize == DstSize: a & mask
8579 // SrcSize > DstSize: trunc(a) & mask
8580 if (SrcSize < DstSize) {
8581 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8582 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8583 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
8584 return new ZExtInst(And, CI.getType());
8587 if (SrcSize == DstSize) {
8588 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8589 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8592 if (SrcSize > DstSize) {
8593 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
8594 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8595 return BinaryOperator::CreateAnd(Trunc,
8596 ConstantInt::get(Trunc->getType(),
8601 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8602 return transformZExtICmp(ICI, CI);
8604 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8605 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8606 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8607 // of the (zext icmp) will be transformed.
8608 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8609 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8610 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8611 (transformZExtICmp(LHS, CI, false) ||
8612 transformZExtICmp(RHS, CI, false))) {
8613 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
8614 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
8615 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8619 // zext(trunc(t) & C) -> (t & zext(C)).
8620 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8621 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8622 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8623 Value *TI0 = TI->getOperand(0);
8624 if (TI0->getType() == CI.getType())
8626 BinaryOperator::CreateAnd(TI0,
8627 ConstantExpr::getZExt(C, CI.getType()));
8630 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8631 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8632 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8633 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8634 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8635 And->getOperand(1) == C)
8636 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8637 Value *TI0 = TI->getOperand(0);
8638 if (TI0->getType() == CI.getType()) {
8639 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8640 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
8641 return BinaryOperator::CreateXor(NewAnd, ZC);
8648 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8649 if (Instruction *I = commonIntCastTransforms(CI))
8652 Value *Src = CI.getOperand(0);
8654 // Canonicalize sign-extend from i1 to a select.
8655 if (Src->getType() == Type::getInt1Ty(*Context))
8656 return SelectInst::Create(Src,
8657 Constant::getAllOnesValue(CI.getType()),
8658 Constant::getNullValue(CI.getType()));
8660 // See if the value being truncated is already sign extended. If so, just
8661 // eliminate the trunc/sext pair.
8662 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8663 Value *Op = cast<User>(Src)->getOperand(0);
8664 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8665 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8666 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8667 unsigned NumSignBits = ComputeNumSignBits(Op);
8669 if (OpBits == DestBits) {
8670 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8671 // bits, it is already ready.
8672 if (NumSignBits > DestBits-MidBits)
8673 return ReplaceInstUsesWith(CI, Op);
8674 } else if (OpBits < DestBits) {
8675 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8676 // bits, just sext from i32.
8677 if (NumSignBits > OpBits-MidBits)
8678 return new SExtInst(Op, CI.getType(), "tmp");
8680 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8681 // bits, just truncate to i32.
8682 if (NumSignBits > OpBits-MidBits)
8683 return new TruncInst(Op, CI.getType(), "tmp");
8687 // If the input is a shl/ashr pair of a same constant, then this is a sign
8688 // extension from a smaller value. If we could trust arbitrary bitwidth
8689 // integers, we could turn this into a truncate to the smaller bit and then
8690 // use a sext for the whole extension. Since we don't, look deeper and check
8691 // for a truncate. If the source and dest are the same type, eliminate the
8692 // trunc and extend and just do shifts. For example, turn:
8693 // %a = trunc i32 %i to i8
8694 // %b = shl i8 %a, 6
8695 // %c = ashr i8 %b, 6
8696 // %d = sext i8 %c to i32
8698 // %a = shl i32 %i, 30
8699 // %d = ashr i32 %a, 30
8701 ConstantInt *BA = 0, *CA = 0;
8702 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8703 m_ConstantInt(CA))) &&
8704 BA == CA && isa<TruncInst>(A)) {
8705 Value *I = cast<TruncInst>(A)->getOperand(0);
8706 if (I->getType() == CI.getType()) {
8707 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8708 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8709 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8710 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8711 I = Builder->CreateShl(I, ShAmtV, CI.getName());
8712 return BinaryOperator::CreateAShr(I, ShAmtV);
8719 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8720 /// in the specified FP type without changing its value.
8721 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8722 LLVMContext *Context) {
8724 APFloat F = CFP->getValueAPF();
8725 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8727 return ConstantFP::get(*Context, F);
8731 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8732 /// through it until we get the source value.
8733 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8734 if (Instruction *I = dyn_cast<Instruction>(V))
8735 if (I->getOpcode() == Instruction::FPExt)
8736 return LookThroughFPExtensions(I->getOperand(0), Context);
8738 // If this value is a constant, return the constant in the smallest FP type
8739 // that can accurately represent it. This allows us to turn
8740 // (float)((double)X+2.0) into x+2.0f.
8741 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8742 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
8743 return V; // No constant folding of this.
8744 // See if the value can be truncated to float and then reextended.
8745 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8747 if (CFP->getType() == Type::getDoubleTy(*Context))
8748 return V; // Won't shrink.
8749 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8751 // Don't try to shrink to various long double types.
8757 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8758 if (Instruction *I = commonCastTransforms(CI))
8761 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8762 // smaller than the destination type, we can eliminate the truncate by doing
8763 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8764 // many builtins (sqrt, etc).
8765 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8766 if (OpI && OpI->hasOneUse()) {
8767 switch (OpI->getOpcode()) {
8769 case Instruction::FAdd:
8770 case Instruction::FSub:
8771 case Instruction::FMul:
8772 case Instruction::FDiv:
8773 case Instruction::FRem:
8774 const Type *SrcTy = OpI->getType();
8775 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8776 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8777 if (LHSTrunc->getType() != SrcTy &&
8778 RHSTrunc->getType() != SrcTy) {
8779 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8780 // If the source types were both smaller than the destination type of
8781 // the cast, do this xform.
8782 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8783 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8784 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
8785 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
8786 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8795 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8796 return commonCastTransforms(CI);
8799 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8800 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8802 return commonCastTransforms(FI);
8804 // fptoui(uitofp(X)) --> X
8805 // fptoui(sitofp(X)) --> X
8806 // This is safe if the intermediate type has enough bits in its mantissa to
8807 // accurately represent all values of X. For example, do not do this with
8808 // i64->float->i64. This is also safe for sitofp case, because any negative
8809 // 'X' value would cause an undefined result for the fptoui.
8810 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8811 OpI->getOperand(0)->getType() == FI.getType() &&
8812 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8813 OpI->getType()->getFPMantissaWidth())
8814 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8816 return commonCastTransforms(FI);
8819 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8820 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8822 return commonCastTransforms(FI);
8824 // fptosi(sitofp(X)) --> X
8825 // fptosi(uitofp(X)) --> X
8826 // This is safe if the intermediate type has enough bits in its mantissa to
8827 // accurately represent all values of X. For example, do not do this with
8828 // i64->float->i64. This is also safe for sitofp case, because any negative
8829 // 'X' value would cause an undefined result for the fptoui.
8830 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8831 OpI->getOperand(0)->getType() == FI.getType() &&
8832 (int)FI.getType()->getScalarSizeInBits() <=
8833 OpI->getType()->getFPMantissaWidth())
8834 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8836 return commonCastTransforms(FI);
8839 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8840 return commonCastTransforms(CI);
8843 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8844 return commonCastTransforms(CI);
8847 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8848 // If the destination integer type is smaller than the intptr_t type for
8849 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8850 // trunc to be exposed to other transforms. Don't do this for extending
8851 // ptrtoint's, because we don't know if the target sign or zero extends its
8854 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8855 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
8856 TD->getIntPtrType(CI.getContext()),
8858 return new TruncInst(P, CI.getType());
8861 return commonPointerCastTransforms(CI);
8864 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8865 // If the source integer type is larger than the intptr_t type for
8866 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8867 // allows the trunc to be exposed to other transforms. Don't do this for
8868 // extending inttoptr's, because we don't know if the target sign or zero
8869 // extends to pointers.
8870 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
8871 TD->getPointerSizeInBits()) {
8872 Value *P = Builder->CreateTrunc(CI.getOperand(0),
8873 TD->getIntPtrType(CI.getContext()), "tmp");
8874 return new IntToPtrInst(P, CI.getType());
8877 if (Instruction *I = commonCastTransforms(CI))
8883 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8884 // If the operands are integer typed then apply the integer transforms,
8885 // otherwise just apply the common ones.
8886 Value *Src = CI.getOperand(0);
8887 const Type *SrcTy = Src->getType();
8888 const Type *DestTy = CI.getType();
8890 if (isa<PointerType>(SrcTy)) {
8891 if (Instruction *I = commonPointerCastTransforms(CI))
8894 if (Instruction *Result = commonCastTransforms(CI))
8899 // Get rid of casts from one type to the same type. These are useless and can
8900 // be replaced by the operand.
8901 if (DestTy == Src->getType())
8902 return ReplaceInstUsesWith(CI, Src);
8904 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8905 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8906 const Type *DstElTy = DstPTy->getElementType();
8907 const Type *SrcElTy = SrcPTy->getElementType();
8909 // If the address spaces don't match, don't eliminate the bitcast, which is
8910 // required for changing types.
8911 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8914 // If we are casting a alloca to a pointer to a type of the same
8915 // size, rewrite the allocation instruction to allocate the "right" type.
8916 // There is no need to modify malloc calls because it is their bitcast that
8917 // needs to be cleaned up.
8918 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
8919 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8922 // If the source and destination are pointers, and this cast is equivalent
8923 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8924 // This can enhance SROA and other transforms that want type-safe pointers.
8925 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
8926 unsigned NumZeros = 0;
8927 while (SrcElTy != DstElTy &&
8928 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8929 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8930 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8934 // If we found a path from the src to dest, create the getelementptr now.
8935 if (SrcElTy == DstElTy) {
8936 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8937 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(), "",
8938 ((Instruction*) NULL));
8942 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
8943 if (DestVTy->getNumElements() == 1) {
8944 if (!isa<VectorType>(SrcTy)) {
8945 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
8946 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
8947 Constant::getNullValue(Type::getInt32Ty(*Context)));
8949 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
8953 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
8954 if (SrcVTy->getNumElements() == 1) {
8955 if (!isa<VectorType>(DestTy)) {
8957 Builder->CreateExtractElement(Src,
8958 Constant::getNullValue(Type::getInt32Ty(*Context)));
8959 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
8964 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8965 if (SVI->hasOneUse()) {
8966 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8967 // a bitconvert to a vector with the same # elts.
8968 if (isa<VectorType>(DestTy) &&
8969 cast<VectorType>(DestTy)->getNumElements() ==
8970 SVI->getType()->getNumElements() &&
8971 SVI->getType()->getNumElements() ==
8972 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8974 // If either of the operands is a cast from CI.getType(), then
8975 // evaluating the shuffle in the casted destination's type will allow
8976 // us to eliminate at least one cast.
8977 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8978 Tmp->getOperand(0)->getType() == DestTy) ||
8979 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8980 Tmp->getOperand(0)->getType() == DestTy)) {
8981 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
8982 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
8983 // Return a new shuffle vector. Use the same element ID's, as we
8984 // know the vector types match #elts.
8985 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8993 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8995 /// %D = select %cond, %C, %A
8997 /// %C = select %cond, %B, 0
9000 /// Assuming that the specified instruction is an operand to the select, return
9001 /// a bitmask indicating which operands of this instruction are foldable if they
9002 /// equal the other incoming value of the select.
9004 static unsigned GetSelectFoldableOperands(Instruction *I) {
9005 switch (I->getOpcode()) {
9006 case Instruction::Add:
9007 case Instruction::Mul:
9008 case Instruction::And:
9009 case Instruction::Or:
9010 case Instruction::Xor:
9011 return 3; // Can fold through either operand.
9012 case Instruction::Sub: // Can only fold on the amount subtracted.
9013 case Instruction::Shl: // Can only fold on the shift amount.
9014 case Instruction::LShr:
9015 case Instruction::AShr:
9018 return 0; // Cannot fold
9022 /// GetSelectFoldableConstant - For the same transformation as the previous
9023 /// function, return the identity constant that goes into the select.
9024 static Constant *GetSelectFoldableConstant(Instruction *I,
9025 LLVMContext *Context) {
9026 switch (I->getOpcode()) {
9027 default: llvm_unreachable("This cannot happen!");
9028 case Instruction::Add:
9029 case Instruction::Sub:
9030 case Instruction::Or:
9031 case Instruction::Xor:
9032 case Instruction::Shl:
9033 case Instruction::LShr:
9034 case Instruction::AShr:
9035 return Constant::getNullValue(I->getType());
9036 case Instruction::And:
9037 return Constant::getAllOnesValue(I->getType());
9038 case Instruction::Mul:
9039 return ConstantInt::get(I->getType(), 1);
9043 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9044 /// have the same opcode and only one use each. Try to simplify this.
9045 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9047 if (TI->getNumOperands() == 1) {
9048 // If this is a non-volatile load or a cast from the same type,
9051 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9054 return 0; // unknown unary op.
9057 // Fold this by inserting a select from the input values.
9058 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9059 FI->getOperand(0), SI.getName()+".v");
9060 InsertNewInstBefore(NewSI, SI);
9061 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9065 // Only handle binary operators here.
9066 if (!isa<BinaryOperator>(TI))
9069 // Figure out if the operations have any operands in common.
9070 Value *MatchOp, *OtherOpT, *OtherOpF;
9072 if (TI->getOperand(0) == FI->getOperand(0)) {
9073 MatchOp = TI->getOperand(0);
9074 OtherOpT = TI->getOperand(1);
9075 OtherOpF = FI->getOperand(1);
9076 MatchIsOpZero = true;
9077 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9078 MatchOp = TI->getOperand(1);
9079 OtherOpT = TI->getOperand(0);
9080 OtherOpF = FI->getOperand(0);
9081 MatchIsOpZero = false;
9082 } else if (!TI->isCommutative()) {
9084 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9085 MatchOp = TI->getOperand(0);
9086 OtherOpT = TI->getOperand(1);
9087 OtherOpF = FI->getOperand(0);
9088 MatchIsOpZero = true;
9089 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9090 MatchOp = TI->getOperand(1);
9091 OtherOpT = TI->getOperand(0);
9092 OtherOpF = FI->getOperand(1);
9093 MatchIsOpZero = true;
9098 // If we reach here, they do have operations in common.
9099 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9100 OtherOpF, SI.getName()+".v");
9101 InsertNewInstBefore(NewSI, SI);
9103 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9105 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9107 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9109 llvm_unreachable("Shouldn't get here");
9113 static bool isSelect01(Constant *C1, Constant *C2) {
9114 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9117 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9120 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9123 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9124 /// facilitate further optimization.
9125 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9127 // See the comment above GetSelectFoldableOperands for a description of the
9128 // transformation we are doing here.
9129 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9130 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9131 !isa<Constant>(FalseVal)) {
9132 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9133 unsigned OpToFold = 0;
9134 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9136 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9141 Constant *C = GetSelectFoldableConstant(TVI, Context);
9142 Value *OOp = TVI->getOperand(2-OpToFold);
9143 // Avoid creating select between 2 constants unless it's selecting
9145 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9146 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9147 InsertNewInstBefore(NewSel, SI);
9148 NewSel->takeName(TVI);
9149 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9150 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9151 llvm_unreachable("Unknown instruction!!");
9158 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9159 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9160 !isa<Constant>(TrueVal)) {
9161 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9162 unsigned OpToFold = 0;
9163 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9165 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9170 Constant *C = GetSelectFoldableConstant(FVI, Context);
9171 Value *OOp = FVI->getOperand(2-OpToFold);
9172 // Avoid creating select between 2 constants unless it's selecting
9174 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9175 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9176 InsertNewInstBefore(NewSel, SI);
9177 NewSel->takeName(FVI);
9178 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9179 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9180 llvm_unreachable("Unknown instruction!!");
9190 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9191 /// ICmpInst as its first operand.
9193 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9195 bool Changed = false;
9196 ICmpInst::Predicate Pred = ICI->getPredicate();
9197 Value *CmpLHS = ICI->getOperand(0);
9198 Value *CmpRHS = ICI->getOperand(1);
9199 Value *TrueVal = SI.getTrueValue();
9200 Value *FalseVal = SI.getFalseValue();
9202 // Check cases where the comparison is with a constant that
9203 // can be adjusted to fit the min/max idiom. We may edit ICI in
9204 // place here, so make sure the select is the only user.
9205 if (ICI->hasOneUse())
9206 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9209 case ICmpInst::ICMP_ULT:
9210 case ICmpInst::ICMP_SLT: {
9211 // X < MIN ? T : F --> F
9212 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9213 return ReplaceInstUsesWith(SI, FalseVal);
9214 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9215 Constant *AdjustedRHS = SubOne(CI);
9216 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9217 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9218 Pred = ICmpInst::getSwappedPredicate(Pred);
9219 CmpRHS = AdjustedRHS;
9220 std::swap(FalseVal, TrueVal);
9221 ICI->setPredicate(Pred);
9222 ICI->setOperand(1, CmpRHS);
9223 SI.setOperand(1, TrueVal);
9224 SI.setOperand(2, FalseVal);
9229 case ICmpInst::ICMP_UGT:
9230 case ICmpInst::ICMP_SGT: {
9231 // X > MAX ? T : F --> F
9232 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9233 return ReplaceInstUsesWith(SI, FalseVal);
9234 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9235 Constant *AdjustedRHS = AddOne(CI);
9236 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9237 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9238 Pred = ICmpInst::getSwappedPredicate(Pred);
9239 CmpRHS = AdjustedRHS;
9240 std::swap(FalseVal, TrueVal);
9241 ICI->setPredicate(Pred);
9242 ICI->setOperand(1, CmpRHS);
9243 SI.setOperand(1, TrueVal);
9244 SI.setOperand(2, FalseVal);
9251 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9252 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9253 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9254 if (match(TrueVal, m_ConstantInt<-1>()) &&
9255 match(FalseVal, m_ConstantInt<0>()))
9256 Pred = ICI->getPredicate();
9257 else if (match(TrueVal, m_ConstantInt<0>()) &&
9258 match(FalseVal, m_ConstantInt<-1>()))
9259 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9261 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9262 // If we are just checking for a icmp eq of a single bit and zext'ing it
9263 // to an integer, then shift the bit to the appropriate place and then
9264 // cast to integer to avoid the comparison.
9265 const APInt &Op1CV = CI->getValue();
9267 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9268 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9269 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9270 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9271 Value *In = ICI->getOperand(0);
9272 Value *Sh = ConstantInt::get(In->getType(),
9273 In->getType()->getScalarSizeInBits()-1);
9274 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9275 In->getName()+".lobit"),
9277 if (In->getType() != SI.getType())
9278 In = CastInst::CreateIntegerCast(In, SI.getType(),
9279 true/*SExt*/, "tmp", ICI);
9281 if (Pred == ICmpInst::ICMP_SGT)
9282 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9283 In->getName()+".not"), *ICI);
9285 return ReplaceInstUsesWith(SI, In);
9290 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9291 // Transform (X == Y) ? X : Y -> Y
9292 if (Pred == ICmpInst::ICMP_EQ)
9293 return ReplaceInstUsesWith(SI, FalseVal);
9294 // Transform (X != Y) ? X : Y -> X
9295 if (Pred == ICmpInst::ICMP_NE)
9296 return ReplaceInstUsesWith(SI, TrueVal);
9297 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9299 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9300 // Transform (X == Y) ? Y : X -> X
9301 if (Pred == ICmpInst::ICMP_EQ)
9302 return ReplaceInstUsesWith(SI, FalseVal);
9303 // Transform (X != Y) ? Y : X -> Y
9304 if (Pred == ICmpInst::ICMP_NE)
9305 return ReplaceInstUsesWith(SI, TrueVal);
9306 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9309 /// NOTE: if we wanted to, this is where to detect integer ABS
9311 return Changed ? &SI : 0;
9315 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a
9316 /// PHI node (but the two may be in different blocks). See if the true/false
9317 /// values (V) are live in all of the predecessor blocks of the PHI. For
9318 /// example, cases like this cannot be mapped:
9320 /// X = phi [ C1, BB1], [C2, BB2]
9322 /// Z = select X, Y, 0
9324 /// because Y is not live in BB1/BB2.
9326 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V,
9327 const SelectInst &SI) {
9328 // If the value is a non-instruction value like a constant or argument, it
9329 // can always be mapped.
9330 const Instruction *I = dyn_cast<Instruction>(V);
9331 if (I == 0) return true;
9333 // If V is a PHI node defined in the same block as the condition PHI, we can
9334 // map the arguments.
9335 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
9337 if (const PHINode *VP = dyn_cast<PHINode>(I))
9338 if (VP->getParent() == CondPHI->getParent())
9341 // Otherwise, if the PHI and select are defined in the same block and if V is
9342 // defined in a different block, then we can transform it.
9343 if (SI.getParent() == CondPHI->getParent() &&
9344 I->getParent() != CondPHI->getParent())
9347 // Otherwise we have a 'hard' case and we can't tell without doing more
9348 // detailed dominator based analysis, punt.
9352 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9353 Value *CondVal = SI.getCondition();
9354 Value *TrueVal = SI.getTrueValue();
9355 Value *FalseVal = SI.getFalseValue();
9357 // select true, X, Y -> X
9358 // select false, X, Y -> Y
9359 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9360 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9362 // select C, X, X -> X
9363 if (TrueVal == FalseVal)
9364 return ReplaceInstUsesWith(SI, TrueVal);
9366 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9367 return ReplaceInstUsesWith(SI, FalseVal);
9368 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9369 return ReplaceInstUsesWith(SI, TrueVal);
9370 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9371 if (isa<Constant>(TrueVal))
9372 return ReplaceInstUsesWith(SI, TrueVal);
9374 return ReplaceInstUsesWith(SI, FalseVal);
9377 if (SI.getType() == Type::getInt1Ty(*Context)) {
9378 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9379 if (C->getZExtValue()) {
9380 // Change: A = select B, true, C --> A = or B, C
9381 return BinaryOperator::CreateOr(CondVal, FalseVal);
9383 // Change: A = select B, false, C --> A = and !B, C
9385 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9386 "not."+CondVal->getName()), SI);
9387 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9389 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9390 if (C->getZExtValue() == false) {
9391 // Change: A = select B, C, false --> A = and B, C
9392 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9394 // Change: A = select B, C, true --> A = or !B, C
9396 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9397 "not."+CondVal->getName()), SI);
9398 return BinaryOperator::CreateOr(NotCond, TrueVal);
9402 // select a, b, a -> a&b
9403 // select a, a, b -> a|b
9404 if (CondVal == TrueVal)
9405 return BinaryOperator::CreateOr(CondVal, FalseVal);
9406 else if (CondVal == FalseVal)
9407 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9410 // Selecting between two integer constants?
9411 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9412 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9413 // select C, 1, 0 -> zext C to int
9414 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9415 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9416 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9417 // select C, 0, 1 -> zext !C to int
9419 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9420 "not."+CondVal->getName()), SI);
9421 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9424 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9425 // If one of the constants is zero (we know they can't both be) and we
9426 // have an icmp instruction with zero, and we have an 'and' with the
9427 // non-constant value, eliminate this whole mess. This corresponds to
9428 // cases like this: ((X & 27) ? 27 : 0)
9429 if (TrueValC->isZero() || FalseValC->isZero())
9430 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9431 cast<Constant>(IC->getOperand(1))->isNullValue())
9432 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9433 if (ICA->getOpcode() == Instruction::And &&
9434 isa<ConstantInt>(ICA->getOperand(1)) &&
9435 (ICA->getOperand(1) == TrueValC ||
9436 ICA->getOperand(1) == FalseValC) &&
9437 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9438 // Okay, now we know that everything is set up, we just don't
9439 // know whether we have a icmp_ne or icmp_eq and whether the
9440 // true or false val is the zero.
9441 bool ShouldNotVal = !TrueValC->isZero();
9442 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9445 V = InsertNewInstBefore(BinaryOperator::Create(
9446 Instruction::Xor, V, ICA->getOperand(1)), SI);
9447 return ReplaceInstUsesWith(SI, V);
9452 // See if we are selecting two values based on a comparison of the two values.
9453 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9454 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9455 // Transform (X == Y) ? X : Y -> Y
9456 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9457 // This is not safe in general for floating point:
9458 // consider X== -0, Y== +0.
9459 // It becomes safe if either operand is a nonzero constant.
9460 ConstantFP *CFPt, *CFPf;
9461 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9462 !CFPt->getValueAPF().isZero()) ||
9463 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9464 !CFPf->getValueAPF().isZero()))
9465 return ReplaceInstUsesWith(SI, FalseVal);
9467 // Transform (X != Y) ? X : Y -> X
9468 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9469 return ReplaceInstUsesWith(SI, TrueVal);
9470 // NOTE: if we wanted to, this is where to detect MIN/MAX
9472 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9473 // Transform (X == Y) ? Y : X -> X
9474 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9475 // This is not safe in general for floating point:
9476 // consider X== -0, Y== +0.
9477 // It becomes safe if either operand is a nonzero constant.
9478 ConstantFP *CFPt, *CFPf;
9479 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9480 !CFPt->getValueAPF().isZero()) ||
9481 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9482 !CFPf->getValueAPF().isZero()))
9483 return ReplaceInstUsesWith(SI, FalseVal);
9485 // Transform (X != Y) ? Y : X -> Y
9486 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9487 return ReplaceInstUsesWith(SI, TrueVal);
9488 // NOTE: if we wanted to, this is where to detect MIN/MAX
9490 // NOTE: if we wanted to, this is where to detect ABS
9493 // See if we are selecting two values based on a comparison of the two values.
9494 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9495 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9498 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9499 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9500 if (TI->hasOneUse() && FI->hasOneUse()) {
9501 Instruction *AddOp = 0, *SubOp = 0;
9503 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9504 if (TI->getOpcode() == FI->getOpcode())
9505 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9508 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9509 // even legal for FP.
9510 if ((TI->getOpcode() == Instruction::Sub &&
9511 FI->getOpcode() == Instruction::Add) ||
9512 (TI->getOpcode() == Instruction::FSub &&
9513 FI->getOpcode() == Instruction::FAdd)) {
9514 AddOp = FI; SubOp = TI;
9515 } else if ((FI->getOpcode() == Instruction::Sub &&
9516 TI->getOpcode() == Instruction::Add) ||
9517 (FI->getOpcode() == Instruction::FSub &&
9518 TI->getOpcode() == Instruction::FAdd)) {
9519 AddOp = TI; SubOp = FI;
9523 Value *OtherAddOp = 0;
9524 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9525 OtherAddOp = AddOp->getOperand(1);
9526 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9527 OtherAddOp = AddOp->getOperand(0);
9531 // So at this point we know we have (Y -> OtherAddOp):
9532 // select C, (add X, Y), (sub X, Z)
9533 Value *NegVal; // Compute -Z
9534 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9535 NegVal = ConstantExpr::getNeg(C);
9537 NegVal = InsertNewInstBefore(
9538 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9542 Value *NewTrueOp = OtherAddOp;
9543 Value *NewFalseOp = NegVal;
9545 std::swap(NewTrueOp, NewFalseOp);
9546 Instruction *NewSel =
9547 SelectInst::Create(CondVal, NewTrueOp,
9548 NewFalseOp, SI.getName() + ".p");
9550 NewSel = InsertNewInstBefore(NewSel, SI);
9551 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9556 // See if we can fold the select into one of our operands.
9557 if (SI.getType()->isInteger()) {
9558 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9563 // See if we can fold the select into a phi node if the condition is a select.
9564 if (isa<PHINode>(SI.getCondition()))
9565 // The true/false values have to be live in the PHI predecessor's blocks.
9566 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
9567 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
9568 if (Instruction *NV = FoldOpIntoPhi(SI))
9571 if (BinaryOperator::isNot(CondVal)) {
9572 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9573 SI.setOperand(1, FalseVal);
9574 SI.setOperand(2, TrueVal);
9581 /// EnforceKnownAlignment - If the specified pointer points to an object that
9582 /// we control, modify the object's alignment to PrefAlign. This isn't
9583 /// often possible though. If alignment is important, a more reliable approach
9584 /// is to simply align all global variables and allocation instructions to
9585 /// their preferred alignment from the beginning.
9587 static unsigned EnforceKnownAlignment(Value *V,
9588 unsigned Align, unsigned PrefAlign) {
9590 User *U = dyn_cast<User>(V);
9591 if (!U) return Align;
9593 switch (Operator::getOpcode(U)) {
9595 case Instruction::BitCast:
9596 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9597 case Instruction::GetElementPtr: {
9598 // If all indexes are zero, it is just the alignment of the base pointer.
9599 bool AllZeroOperands = true;
9600 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9601 if (!isa<Constant>(*i) ||
9602 !cast<Constant>(*i)->isNullValue()) {
9603 AllZeroOperands = false;
9607 if (AllZeroOperands) {
9608 // Treat this like a bitcast.
9609 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9615 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9616 // If there is a large requested alignment and we can, bump up the alignment
9618 if (!GV->isDeclaration()) {
9619 if (GV->getAlignment() >= PrefAlign)
9620 Align = GV->getAlignment();
9622 GV->setAlignment(PrefAlign);
9626 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
9627 // If there is a requested alignment and if this is an alloca, round up.
9628 if (AI->getAlignment() >= PrefAlign)
9629 Align = AI->getAlignment();
9631 AI->setAlignment(PrefAlign);
9639 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9640 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9641 /// and it is more than the alignment of the ultimate object, see if we can
9642 /// increase the alignment of the ultimate object, making this check succeed.
9643 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9644 unsigned PrefAlign) {
9645 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9646 sizeof(PrefAlign) * CHAR_BIT;
9647 APInt Mask = APInt::getAllOnesValue(BitWidth);
9648 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9649 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9650 unsigned TrailZ = KnownZero.countTrailingOnes();
9651 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9653 if (PrefAlign > Align)
9654 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9656 // We don't need to make any adjustment.
9660 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9661 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9662 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9663 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9664 unsigned CopyAlign = MI->getAlignment();
9666 if (CopyAlign < MinAlign) {
9667 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9672 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9674 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9675 if (MemOpLength == 0) return 0;
9677 // Source and destination pointer types are always "i8*" for intrinsic. See
9678 // if the size is something we can handle with a single primitive load/store.
9679 // A single load+store correctly handles overlapping memory in the memmove
9681 unsigned Size = MemOpLength->getZExtValue();
9682 if (Size == 0) return MI; // Delete this mem transfer.
9684 if (Size > 8 || (Size&(Size-1)))
9685 return 0; // If not 1/2/4/8 bytes, exit.
9687 // Use an integer load+store unless we can find something better.
9689 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
9691 // Memcpy forces the use of i8* for the source and destination. That means
9692 // that if you're using memcpy to move one double around, you'll get a cast
9693 // from double* to i8*. We'd much rather use a double load+store rather than
9694 // an i64 load+store, here because this improves the odds that the source or
9695 // dest address will be promotable. See if we can find a better type than the
9696 // integer datatype.
9697 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9698 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9699 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9700 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9701 // down through these levels if so.
9702 while (!SrcETy->isSingleValueType()) {
9703 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9704 if (STy->getNumElements() == 1)
9705 SrcETy = STy->getElementType(0);
9708 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9709 if (ATy->getNumElements() == 1)
9710 SrcETy = ATy->getElementType();
9717 if (SrcETy->isSingleValueType())
9718 NewPtrTy = PointerType::getUnqual(SrcETy);
9723 // If the memcpy/memmove provides better alignment info than we can
9725 SrcAlign = std::max(SrcAlign, CopyAlign);
9726 DstAlign = std::max(DstAlign, CopyAlign);
9728 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy);
9729 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy);
9730 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9731 InsertNewInstBefore(L, *MI);
9732 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9734 // Set the size of the copy to 0, it will be deleted on the next iteration.
9735 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9739 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9740 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9741 if (MI->getAlignment() < Alignment) {
9742 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9747 // Extract the length and alignment and fill if they are constant.
9748 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9749 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9750 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
9752 uint64_t Len = LenC->getZExtValue();
9753 Alignment = MI->getAlignment();
9755 // If the length is zero, this is a no-op
9756 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9758 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9759 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9760 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
9762 Value *Dest = MI->getDest();
9763 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
9765 // Alignment 0 is identity for alignment 1 for memset, but not store.
9766 if (Alignment == 0) Alignment = 1;
9768 // Extract the fill value and store.
9769 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9770 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9771 Dest, false, Alignment), *MI);
9773 // Set the size of the copy to 0, it will be deleted on the next iteration.
9774 MI->setLength(Constant::getNullValue(LenC->getType()));
9782 /// visitCallInst - CallInst simplification. This mostly only handles folding
9783 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9784 /// the heavy lifting.
9786 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9787 if (isFreeCall(&CI))
9788 return visitFree(CI);
9790 // If the caller function is nounwind, mark the call as nounwind, even if the
9792 if (CI.getParent()->getParent()->doesNotThrow() &&
9793 !CI.doesNotThrow()) {
9794 CI.setDoesNotThrow();
9798 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9799 if (!II) return visitCallSite(&CI);
9801 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9803 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9804 bool Changed = false;
9806 // memmove/cpy/set of zero bytes is a noop.
9807 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9808 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9810 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9811 if (CI->getZExtValue() == 1) {
9812 // Replace the instruction with just byte operations. We would
9813 // transform other cases to loads/stores, but we don't know if
9814 // alignment is sufficient.
9818 // If we have a memmove and the source operation is a constant global,
9819 // then the source and dest pointers can't alias, so we can change this
9820 // into a call to memcpy.
9821 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9822 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9823 if (GVSrc->isConstant()) {
9824 Module *M = CI.getParent()->getParent()->getParent();
9825 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9827 Tys[0] = CI.getOperand(3)->getType();
9829 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9833 // memmove(x,x,size) -> noop.
9834 if (MMI->getSource() == MMI->getDest())
9835 return EraseInstFromFunction(CI);
9838 // If we can determine a pointer alignment that is bigger than currently
9839 // set, update the alignment.
9840 if (isa<MemTransferInst>(MI)) {
9841 if (Instruction *I = SimplifyMemTransfer(MI))
9843 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9844 if (Instruction *I = SimplifyMemSet(MSI))
9848 if (Changed) return II;
9851 switch (II->getIntrinsicID()) {
9853 case Intrinsic::bswap:
9854 // bswap(bswap(x)) -> x
9855 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9856 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9857 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9859 case Intrinsic::ppc_altivec_lvx:
9860 case Intrinsic::ppc_altivec_lvxl:
9861 case Intrinsic::x86_sse_loadu_ps:
9862 case Intrinsic::x86_sse2_loadu_pd:
9863 case Intrinsic::x86_sse2_loadu_dq:
9864 // Turn PPC lvx -> load if the pointer is known aligned.
9865 // Turn X86 loadups -> load if the pointer is known aligned.
9866 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9867 Value *Ptr = Builder->CreateBitCast(II->getOperand(1),
9868 PointerType::getUnqual(II->getType()));
9869 return new LoadInst(Ptr);
9872 case Intrinsic::ppc_altivec_stvx:
9873 case Intrinsic::ppc_altivec_stvxl:
9874 // Turn stvx -> store if the pointer is known aligned.
9875 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9876 const Type *OpPtrTy =
9877 PointerType::getUnqual(II->getOperand(1)->getType());
9878 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy);
9879 return new StoreInst(II->getOperand(1), Ptr);
9882 case Intrinsic::x86_sse_storeu_ps:
9883 case Intrinsic::x86_sse2_storeu_pd:
9884 case Intrinsic::x86_sse2_storeu_dq:
9885 // Turn X86 storeu -> store if the pointer is known aligned.
9886 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9887 const Type *OpPtrTy =
9888 PointerType::getUnqual(II->getOperand(2)->getType());
9889 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
9890 return new StoreInst(II->getOperand(2), Ptr);
9894 case Intrinsic::x86_sse_cvttss2si: {
9895 // These intrinsics only demands the 0th element of its input vector. If
9896 // we can simplify the input based on that, do so now.
9898 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9899 APInt DemandedElts(VWidth, 1);
9900 APInt UndefElts(VWidth, 0);
9901 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9903 II->setOperand(1, V);
9909 case Intrinsic::ppc_altivec_vperm:
9910 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9911 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9912 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9914 // Check that all of the elements are integer constants or undefs.
9915 bool AllEltsOk = true;
9916 for (unsigned i = 0; i != 16; ++i) {
9917 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9918 !isa<UndefValue>(Mask->getOperand(i))) {
9925 // Cast the input vectors to byte vectors.
9926 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
9927 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType());
9928 Value *Result = UndefValue::get(Op0->getType());
9930 // Only extract each element once.
9931 Value *ExtractedElts[32];
9932 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9934 for (unsigned i = 0; i != 16; ++i) {
9935 if (isa<UndefValue>(Mask->getOperand(i)))
9937 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9938 Idx &= 31; // Match the hardware behavior.
9940 if (ExtractedElts[Idx] == 0) {
9941 ExtractedElts[Idx] =
9942 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
9943 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false),
9947 // Insert this value into the result vector.
9948 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
9949 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
9952 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9957 case Intrinsic::stackrestore: {
9958 // If the save is right next to the restore, remove the restore. This can
9959 // happen when variable allocas are DCE'd.
9960 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9961 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9962 BasicBlock::iterator BI = SS;
9964 return EraseInstFromFunction(CI);
9968 // Scan down this block to see if there is another stack restore in the
9969 // same block without an intervening call/alloca.
9970 BasicBlock::iterator BI = II;
9971 TerminatorInst *TI = II->getParent()->getTerminator();
9972 bool CannotRemove = false;
9973 for (++BI; &*BI != TI; ++BI) {
9974 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
9975 CannotRemove = true;
9978 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9979 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9980 // If there is a stackrestore below this one, remove this one.
9981 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9982 return EraseInstFromFunction(CI);
9983 // Otherwise, ignore the intrinsic.
9985 // If we found a non-intrinsic call, we can't remove the stack
9987 CannotRemove = true;
9993 // If the stack restore is in a return/unwind block and if there are no
9994 // allocas or calls between the restore and the return, nuke the restore.
9995 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9996 return EraseInstFromFunction(CI);
10001 return visitCallSite(II);
10004 // InvokeInst simplification
10006 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10007 return visitCallSite(&II);
10010 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10011 /// passed through the varargs area, we can eliminate the use of the cast.
10012 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10013 const CastInst * const CI,
10014 const TargetData * const TD,
10016 if (!CI->isLosslessCast())
10019 // The size of ByVal arguments is derived from the type, so we
10020 // can't change to a type with a different size. If the size were
10021 // passed explicitly we could avoid this check.
10022 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10025 const Type* SrcTy =
10026 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10027 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10028 if (!SrcTy->isSized() || !DstTy->isSized())
10030 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10035 // visitCallSite - Improvements for call and invoke instructions.
10037 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10038 bool Changed = false;
10040 // If the callee is a constexpr cast of a function, attempt to move the cast
10041 // to the arguments of the call/invoke.
10042 if (transformConstExprCastCall(CS)) return 0;
10044 Value *Callee = CS.getCalledValue();
10046 if (Function *CalleeF = dyn_cast<Function>(Callee))
10047 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10048 Instruction *OldCall = CS.getInstruction();
10049 // If the call and callee calling conventions don't match, this call must
10050 // be unreachable, as the call is undefined.
10051 new StoreInst(ConstantInt::getTrue(*Context),
10052 UndefValue::get(Type::getInt1PtrTy(*Context)),
10054 // If OldCall dues not return void then replaceAllUsesWith undef.
10055 // This allows ValueHandlers and custom metadata to adjust itself.
10056 if (!OldCall->getType()->isVoidTy())
10057 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10058 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10059 return EraseInstFromFunction(*OldCall);
10063 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10064 // This instruction is not reachable, just remove it. We insert a store to
10065 // undef so that we know that this code is not reachable, despite the fact
10066 // that we can't modify the CFG here.
10067 new StoreInst(ConstantInt::getTrue(*Context),
10068 UndefValue::get(Type::getInt1PtrTy(*Context)),
10069 CS.getInstruction());
10071 // If CS dues not return void then replaceAllUsesWith undef.
10072 // This allows ValueHandlers and custom metadata to adjust itself.
10073 if (!CS.getInstruction()->getType()->isVoidTy())
10074 CS.getInstruction()->
10075 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10077 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10078 // Don't break the CFG, insert a dummy cond branch.
10079 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10080 ConstantInt::getTrue(*Context), II);
10082 return EraseInstFromFunction(*CS.getInstruction());
10085 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10086 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10087 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10088 return transformCallThroughTrampoline(CS);
10090 const PointerType *PTy = cast<PointerType>(Callee->getType());
10091 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10092 if (FTy->isVarArg()) {
10093 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10094 // See if we can optimize any arguments passed through the varargs area of
10096 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10097 E = CS.arg_end(); I != E; ++I, ++ix) {
10098 CastInst *CI = dyn_cast<CastInst>(*I);
10099 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10100 *I = CI->getOperand(0);
10106 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10107 // Inline asm calls cannot throw - mark them 'nounwind'.
10108 CS.setDoesNotThrow();
10112 return Changed ? CS.getInstruction() : 0;
10115 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10116 // attempt to move the cast to the arguments of the call/invoke.
10118 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10119 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10120 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10121 if (CE->getOpcode() != Instruction::BitCast ||
10122 !isa<Function>(CE->getOperand(0)))
10124 Function *Callee = cast<Function>(CE->getOperand(0));
10125 Instruction *Caller = CS.getInstruction();
10126 const AttrListPtr &CallerPAL = CS.getAttributes();
10128 // Okay, this is a cast from a function to a different type. Unless doing so
10129 // would cause a type conversion of one of our arguments, change this call to
10130 // be a direct call with arguments casted to the appropriate types.
10132 const FunctionType *FT = Callee->getFunctionType();
10133 const Type *OldRetTy = Caller->getType();
10134 const Type *NewRetTy = FT->getReturnType();
10136 if (isa<StructType>(NewRetTy))
10137 return false; // TODO: Handle multiple return values.
10139 // Check to see if we are changing the return type...
10140 if (OldRetTy != NewRetTy) {
10141 if (Callee->isDeclaration() &&
10142 // Conversion is ok if changing from one pointer type to another or from
10143 // a pointer to an integer of the same size.
10144 !((isa<PointerType>(OldRetTy) || !TD ||
10145 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
10146 (isa<PointerType>(NewRetTy) || !TD ||
10147 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
10148 return false; // Cannot transform this return value.
10150 if (!Caller->use_empty() &&
10151 // void -> non-void is handled specially
10152 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
10153 return false; // Cannot transform this return value.
10155 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10156 Attributes RAttrs = CallerPAL.getRetAttributes();
10157 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10158 return false; // Attribute not compatible with transformed value.
10161 // If the callsite is an invoke instruction, and the return value is used by
10162 // a PHI node in a successor, we cannot change the return type of the call
10163 // because there is no place to put the cast instruction (without breaking
10164 // the critical edge). Bail out in this case.
10165 if (!Caller->use_empty())
10166 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10167 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10169 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10170 if (PN->getParent() == II->getNormalDest() ||
10171 PN->getParent() == II->getUnwindDest())
10175 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10176 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10178 CallSite::arg_iterator AI = CS.arg_begin();
10179 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10180 const Type *ParamTy = FT->getParamType(i);
10181 const Type *ActTy = (*AI)->getType();
10183 if (!CastInst::isCastable(ActTy, ParamTy))
10184 return false; // Cannot transform this parameter value.
10186 if (CallerPAL.getParamAttributes(i + 1)
10187 & Attribute::typeIncompatible(ParamTy))
10188 return false; // Attribute not compatible with transformed value.
10190 // Converting from one pointer type to another or between a pointer and an
10191 // integer of the same size is safe even if we do not have a body.
10192 bool isConvertible = ActTy == ParamTy ||
10193 (TD && ((isa<PointerType>(ParamTy) ||
10194 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10195 (isa<PointerType>(ActTy) ||
10196 ActTy == TD->getIntPtrType(Caller->getContext()))));
10197 if (Callee->isDeclaration() && !isConvertible) return false;
10200 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10201 Callee->isDeclaration())
10202 return false; // Do not delete arguments unless we have a function body.
10204 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10205 !CallerPAL.isEmpty())
10206 // In this case we have more arguments than the new function type, but we
10207 // won't be dropping them. Check that these extra arguments have attributes
10208 // that are compatible with being a vararg call argument.
10209 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10210 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10212 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10213 if (PAttrs & Attribute::VarArgsIncompatible)
10217 // Okay, we decided that this is a safe thing to do: go ahead and start
10218 // inserting cast instructions as necessary...
10219 std::vector<Value*> Args;
10220 Args.reserve(NumActualArgs);
10221 SmallVector<AttributeWithIndex, 8> attrVec;
10222 attrVec.reserve(NumCommonArgs);
10224 // Get any return attributes.
10225 Attributes RAttrs = CallerPAL.getRetAttributes();
10227 // If the return value is not being used, the type may not be compatible
10228 // with the existing attributes. Wipe out any problematic attributes.
10229 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10231 // Add the new return attributes.
10233 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10235 AI = CS.arg_begin();
10236 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10237 const Type *ParamTy = FT->getParamType(i);
10238 if ((*AI)->getType() == ParamTy) {
10239 Args.push_back(*AI);
10241 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10242 false, ParamTy, false);
10243 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
10246 // Add any parameter attributes.
10247 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10248 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10251 // If the function takes more arguments than the call was taking, add them
10253 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10254 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10256 // If we are removing arguments to the function, emit an obnoxious warning.
10257 if (FT->getNumParams() < NumActualArgs) {
10258 if (!FT->isVarArg()) {
10259 errs() << "WARNING: While resolving call to function '"
10260 << Callee->getName() << "' arguments were dropped!\n";
10262 // Add all of the arguments in their promoted form to the arg list.
10263 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10264 const Type *PTy = getPromotedType((*AI)->getType());
10265 if (PTy != (*AI)->getType()) {
10266 // Must promote to pass through va_arg area!
10267 Instruction::CastOps opcode =
10268 CastInst::getCastOpcode(*AI, false, PTy, false);
10269 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
10271 Args.push_back(*AI);
10274 // Add any parameter attributes.
10275 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10276 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10281 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10282 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10284 if (NewRetTy->isVoidTy())
10285 Caller->setName(""); // Void type should not have a name.
10287 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10291 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10292 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10293 Args.begin(), Args.end(),
10294 Caller->getName(), Caller);
10295 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10296 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10298 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10299 Caller->getName(), Caller);
10300 CallInst *CI = cast<CallInst>(Caller);
10301 if (CI->isTailCall())
10302 cast<CallInst>(NC)->setTailCall();
10303 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10304 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10307 // Insert a cast of the return type as necessary.
10309 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10310 if (!NV->getType()->isVoidTy()) {
10311 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10313 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10315 // If this is an invoke instruction, we should insert it after the first
10316 // non-phi, instruction in the normal successor block.
10317 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10318 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10319 InsertNewInstBefore(NC, *I);
10321 // Otherwise, it's a call, just insert cast right after the call instr
10322 InsertNewInstBefore(NC, *Caller);
10324 Worklist.AddUsersToWorkList(*Caller);
10326 NV = UndefValue::get(Caller->getType());
10331 if (!Caller->use_empty())
10332 Caller->replaceAllUsesWith(NV);
10334 EraseInstFromFunction(*Caller);
10338 // transformCallThroughTrampoline - Turn a call to a function created by the
10339 // init_trampoline intrinsic into a direct call to the underlying function.
10341 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10342 Value *Callee = CS.getCalledValue();
10343 const PointerType *PTy = cast<PointerType>(Callee->getType());
10344 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10345 const AttrListPtr &Attrs = CS.getAttributes();
10347 // If the call already has the 'nest' attribute somewhere then give up -
10348 // otherwise 'nest' would occur twice after splicing in the chain.
10349 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10352 IntrinsicInst *Tramp =
10353 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10355 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10356 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10357 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10359 const AttrListPtr &NestAttrs = NestF->getAttributes();
10360 if (!NestAttrs.isEmpty()) {
10361 unsigned NestIdx = 1;
10362 const Type *NestTy = 0;
10363 Attributes NestAttr = Attribute::None;
10365 // Look for a parameter marked with the 'nest' attribute.
10366 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10367 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10368 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10369 // Record the parameter type and any other attributes.
10371 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10376 Instruction *Caller = CS.getInstruction();
10377 std::vector<Value*> NewArgs;
10378 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10380 SmallVector<AttributeWithIndex, 8> NewAttrs;
10381 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10383 // Insert the nest argument into the call argument list, which may
10384 // mean appending it. Likewise for attributes.
10386 // Add any result attributes.
10387 if (Attributes Attr = Attrs.getRetAttributes())
10388 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10392 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10394 if (Idx == NestIdx) {
10395 // Add the chain argument and attributes.
10396 Value *NestVal = Tramp->getOperand(3);
10397 if (NestVal->getType() != NestTy)
10398 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10399 NewArgs.push_back(NestVal);
10400 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10406 // Add the original argument and attributes.
10407 NewArgs.push_back(*I);
10408 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10410 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10416 // Add any function attributes.
10417 if (Attributes Attr = Attrs.getFnAttributes())
10418 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10420 // The trampoline may have been bitcast to a bogus type (FTy).
10421 // Handle this by synthesizing a new function type, equal to FTy
10422 // with the chain parameter inserted.
10424 std::vector<const Type*> NewTypes;
10425 NewTypes.reserve(FTy->getNumParams()+1);
10427 // Insert the chain's type into the list of parameter types, which may
10428 // mean appending it.
10431 FunctionType::param_iterator I = FTy->param_begin(),
10432 E = FTy->param_end();
10435 if (Idx == NestIdx)
10436 // Add the chain's type.
10437 NewTypes.push_back(NestTy);
10442 // Add the original type.
10443 NewTypes.push_back(*I);
10449 // Replace the trampoline call with a direct call. Let the generic
10450 // code sort out any function type mismatches.
10451 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10453 Constant *NewCallee =
10454 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10455 NestF : ConstantExpr::getBitCast(NestF,
10456 PointerType::getUnqual(NewFTy));
10457 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10460 Instruction *NewCaller;
10461 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10462 NewCaller = InvokeInst::Create(NewCallee,
10463 II->getNormalDest(), II->getUnwindDest(),
10464 NewArgs.begin(), NewArgs.end(),
10465 Caller->getName(), Caller);
10466 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10467 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10469 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10470 Caller->getName(), Caller);
10471 if (cast<CallInst>(Caller)->isTailCall())
10472 cast<CallInst>(NewCaller)->setTailCall();
10473 cast<CallInst>(NewCaller)->
10474 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10475 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10477 if (!Caller->getType()->isVoidTy())
10478 Caller->replaceAllUsesWith(NewCaller);
10479 Caller->eraseFromParent();
10480 Worklist.Remove(Caller);
10485 // Replace the trampoline call with a direct call. Since there is no 'nest'
10486 // parameter, there is no need to adjust the argument list. Let the generic
10487 // code sort out any function type mismatches.
10488 Constant *NewCallee =
10489 NestF->getType() == PTy ? NestF :
10490 ConstantExpr::getBitCast(NestF, PTy);
10491 CS.setCalledFunction(NewCallee);
10492 return CS.getInstruction();
10495 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
10496 /// and if a/b/c and the add's all have a single use, turn this into a phi
10497 /// and a single binop.
10498 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10499 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10500 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10501 unsigned Opc = FirstInst->getOpcode();
10502 Value *LHSVal = FirstInst->getOperand(0);
10503 Value *RHSVal = FirstInst->getOperand(1);
10505 const Type *LHSType = LHSVal->getType();
10506 const Type *RHSType = RHSVal->getType();
10508 // Scan to see if all operands are the same opcode, and all have one use.
10509 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10510 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10511 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10512 // Verify type of the LHS matches so we don't fold cmp's of different
10513 // types or GEP's with different index types.
10514 I->getOperand(0)->getType() != LHSType ||
10515 I->getOperand(1)->getType() != RHSType)
10518 // If they are CmpInst instructions, check their predicates
10519 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10520 if (cast<CmpInst>(I)->getPredicate() !=
10521 cast<CmpInst>(FirstInst)->getPredicate())
10524 // Keep track of which operand needs a phi node.
10525 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10526 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10529 // If both LHS and RHS would need a PHI, don't do this transformation,
10530 // because it would increase the number of PHIs entering the block,
10531 // which leads to higher register pressure. This is especially
10532 // bad when the PHIs are in the header of a loop.
10533 if (!LHSVal && !RHSVal)
10536 // Otherwise, this is safe to transform!
10538 Value *InLHS = FirstInst->getOperand(0);
10539 Value *InRHS = FirstInst->getOperand(1);
10540 PHINode *NewLHS = 0, *NewRHS = 0;
10542 NewLHS = PHINode::Create(LHSType,
10543 FirstInst->getOperand(0)->getName() + ".pn");
10544 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10545 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10546 InsertNewInstBefore(NewLHS, PN);
10551 NewRHS = PHINode::Create(RHSType,
10552 FirstInst->getOperand(1)->getName() + ".pn");
10553 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10554 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10555 InsertNewInstBefore(NewRHS, PN);
10559 // Add all operands to the new PHIs.
10560 if (NewLHS || NewRHS) {
10561 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10562 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10564 Value *NewInLHS = InInst->getOperand(0);
10565 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10568 Value *NewInRHS = InInst->getOperand(1);
10569 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10574 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10575 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10576 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10577 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10581 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10582 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10584 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10585 FirstInst->op_end());
10586 // This is true if all GEP bases are allocas and if all indices into them are
10588 bool AllBasePointersAreAllocas = true;
10590 // We don't want to replace this phi if the replacement would require
10591 // more than one phi, which leads to higher register pressure. This is
10592 // especially bad when the PHIs are in the header of a loop.
10593 bool NeededPhi = false;
10595 // Scan to see if all operands are the same opcode, and all have one use.
10596 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10597 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10598 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10599 GEP->getNumOperands() != FirstInst->getNumOperands())
10602 // Keep track of whether or not all GEPs are of alloca pointers.
10603 if (AllBasePointersAreAllocas &&
10604 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10605 !GEP->hasAllConstantIndices()))
10606 AllBasePointersAreAllocas = false;
10608 // Compare the operand lists.
10609 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10610 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10613 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10614 // if one of the PHIs has a constant for the index. The index may be
10615 // substantially cheaper to compute for the constants, so making it a
10616 // variable index could pessimize the path. This also handles the case
10617 // for struct indices, which must always be constant.
10618 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10619 isa<ConstantInt>(GEP->getOperand(op)))
10622 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10625 // If we already needed a PHI for an earlier operand, and another operand
10626 // also requires a PHI, we'd be introducing more PHIs than we're
10627 // eliminating, which increases register pressure on entry to the PHI's
10632 FixedOperands[op] = 0; // Needs a PHI.
10637 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10638 // bother doing this transformation. At best, this will just save a bit of
10639 // offset calculation, but all the predecessors will have to materialize the
10640 // stack address into a register anyway. We'd actually rather *clone* the
10641 // load up into the predecessors so that we have a load of a gep of an alloca,
10642 // which can usually all be folded into the load.
10643 if (AllBasePointersAreAllocas)
10646 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10647 // that is variable.
10648 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10650 bool HasAnyPHIs = false;
10651 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10652 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10653 Value *FirstOp = FirstInst->getOperand(i);
10654 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10655 FirstOp->getName()+".pn");
10656 InsertNewInstBefore(NewPN, PN);
10658 NewPN->reserveOperandSpace(e);
10659 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10660 OperandPhis[i] = NewPN;
10661 FixedOperands[i] = NewPN;
10666 // Add all operands to the new PHIs.
10668 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10669 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10670 BasicBlock *InBB = PN.getIncomingBlock(i);
10672 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10673 if (PHINode *OpPhi = OperandPhis[op])
10674 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10678 Value *Base = FixedOperands[0];
10679 return cast<GEPOperator>(FirstInst)->isInBounds() ?
10680 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
10681 FixedOperands.end()) :
10682 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10683 FixedOperands.end());
10687 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10688 /// sink the load out of the block that defines it. This means that it must be
10689 /// obvious the value of the load is not changed from the point of the load to
10690 /// the end of the block it is in.
10692 /// Finally, it is safe, but not profitable, to sink a load targetting a
10693 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10695 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10696 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10698 for (++BBI; BBI != E; ++BBI)
10699 if (BBI->mayWriteToMemory())
10702 // Check for non-address taken alloca. If not address-taken already, it isn't
10703 // profitable to do this xform.
10704 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10705 bool isAddressTaken = false;
10706 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10708 if (isa<LoadInst>(UI)) continue;
10709 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10710 // If storing TO the alloca, then the address isn't taken.
10711 if (SI->getOperand(1) == AI) continue;
10713 isAddressTaken = true;
10717 if (!isAddressTaken && AI->isStaticAlloca())
10721 // If this load is a load from a GEP with a constant offset from an alloca,
10722 // then we don't want to sink it. In its present form, it will be
10723 // load [constant stack offset]. Sinking it will cause us to have to
10724 // materialize the stack addresses in each predecessor in a register only to
10725 // do a shared load from register in the successor.
10726 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10727 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10728 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10735 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10736 // operator and they all are only used by the PHI, PHI together their
10737 // inputs, and do the operation once, to the result of the PHI.
10738 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10739 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10741 // Scan the instruction, looking for input operations that can be folded away.
10742 // If all input operands to the phi are the same instruction (e.g. a cast from
10743 // the same type or "+42") we can pull the operation through the PHI, reducing
10744 // code size and simplifying code.
10745 Constant *ConstantOp = 0;
10746 const Type *CastSrcTy = 0;
10747 bool isVolatile = false;
10748 if (isa<CastInst>(FirstInst)) {
10749 CastSrcTy = FirstInst->getOperand(0)->getType();
10750 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10751 // Can fold binop, compare or shift here if the RHS is a constant,
10752 // otherwise call FoldPHIArgBinOpIntoPHI.
10753 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10754 if (ConstantOp == 0)
10755 return FoldPHIArgBinOpIntoPHI(PN);
10756 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10757 isVolatile = LI->isVolatile();
10758 // We can't sink the load if the loaded value could be modified between the
10759 // load and the PHI.
10760 if (LI->getParent() != PN.getIncomingBlock(0) ||
10761 !isSafeAndProfitableToSinkLoad(LI))
10764 // If the PHI is of volatile loads and the load block has multiple
10765 // successors, sinking it would remove a load of the volatile value from
10766 // the path through the other successor.
10768 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10771 } else if (isa<GetElementPtrInst>(FirstInst)) {
10772 return FoldPHIArgGEPIntoPHI(PN);
10774 return 0; // Cannot fold this operation.
10777 // Check to see if all arguments are the same operation.
10778 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10779 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10780 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10781 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10784 if (I->getOperand(0)->getType() != CastSrcTy)
10785 return 0; // Cast operation must match.
10786 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10787 // We can't sink the load if the loaded value could be modified between
10788 // the load and the PHI.
10789 if (LI->isVolatile() != isVolatile ||
10790 LI->getParent() != PN.getIncomingBlock(i) ||
10791 !isSafeAndProfitableToSinkLoad(LI))
10794 // If the PHI is of volatile loads and the load block has multiple
10795 // successors, sinking it would remove a load of the volatile value from
10796 // the path through the other successor.
10798 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10801 } else if (I->getOperand(1) != ConstantOp) {
10806 // Okay, they are all the same operation. Create a new PHI node of the
10807 // correct type, and PHI together all of the LHS's of the instructions.
10808 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10809 PN.getName()+".in");
10810 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10812 Value *InVal = FirstInst->getOperand(0);
10813 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10815 // Add all operands to the new PHI.
10816 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10817 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10818 if (NewInVal != InVal)
10820 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10825 // The new PHI unions all of the same values together. This is really
10826 // common, so we handle it intelligently here for compile-time speed.
10830 InsertNewInstBefore(NewPN, PN);
10834 // Insert and return the new operation.
10835 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10836 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10837 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10838 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10839 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10840 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10841 PhiVal, ConstantOp);
10842 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10844 // If this was a volatile load that we are merging, make sure to loop through
10845 // and mark all the input loads as non-volatile. If we don't do this, we will
10846 // insert a new volatile load and the old ones will not be deletable.
10848 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10849 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10851 return new LoadInst(PhiVal, "", isVolatile);
10854 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10856 static bool DeadPHICycle(PHINode *PN,
10857 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10858 if (PN->use_empty()) return true;
10859 if (!PN->hasOneUse()) return false;
10861 // Remember this node, and if we find the cycle, return.
10862 if (!PotentiallyDeadPHIs.insert(PN))
10865 // Don't scan crazily complex things.
10866 if (PotentiallyDeadPHIs.size() == 16)
10869 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10870 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10875 /// PHIsEqualValue - Return true if this phi node is always equal to
10876 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10877 /// z = some value; x = phi (y, z); y = phi (x, z)
10878 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10879 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10880 // See if we already saw this PHI node.
10881 if (!ValueEqualPHIs.insert(PN))
10884 // Don't scan crazily complex things.
10885 if (ValueEqualPHIs.size() == 16)
10888 // Scan the operands to see if they are either phi nodes or are equal to
10890 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10891 Value *Op = PN->getIncomingValue(i);
10892 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10893 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10895 } else if (Op != NonPhiInVal)
10903 // PHINode simplification
10905 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10906 // If LCSSA is around, don't mess with Phi nodes
10907 if (MustPreserveLCSSA) return 0;
10909 if (Value *V = PN.hasConstantValue())
10910 return ReplaceInstUsesWith(PN, V);
10912 // If all PHI operands are the same operation, pull them through the PHI,
10913 // reducing code size.
10914 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10915 isa<Instruction>(PN.getIncomingValue(1)) &&
10916 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10917 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10918 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10919 // than themselves more than once.
10920 PN.getIncomingValue(0)->hasOneUse())
10921 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10924 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10925 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10926 // PHI)... break the cycle.
10927 if (PN.hasOneUse()) {
10928 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10929 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10930 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10931 PotentiallyDeadPHIs.insert(&PN);
10932 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10933 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10936 // If this phi has a single use, and if that use just computes a value for
10937 // the next iteration of a loop, delete the phi. This occurs with unused
10938 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10939 // common case here is good because the only other things that catch this
10940 // are induction variable analysis (sometimes) and ADCE, which is only run
10942 if (PHIUser->hasOneUse() &&
10943 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10944 PHIUser->use_back() == &PN) {
10945 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10949 // We sometimes end up with phi cycles that non-obviously end up being the
10950 // same value, for example:
10951 // z = some value; x = phi (y, z); y = phi (x, z)
10952 // where the phi nodes don't necessarily need to be in the same block. Do a
10953 // quick check to see if the PHI node only contains a single non-phi value, if
10954 // so, scan to see if the phi cycle is actually equal to that value.
10956 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10957 // Scan for the first non-phi operand.
10958 while (InValNo != NumOperandVals &&
10959 isa<PHINode>(PN.getIncomingValue(InValNo)))
10962 if (InValNo != NumOperandVals) {
10963 Value *NonPhiInVal = PN.getOperand(InValNo);
10965 // Scan the rest of the operands to see if there are any conflicts, if so
10966 // there is no need to recursively scan other phis.
10967 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10968 Value *OpVal = PN.getIncomingValue(InValNo);
10969 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10973 // If we scanned over all operands, then we have one unique value plus
10974 // phi values. Scan PHI nodes to see if they all merge in each other or
10976 if (InValNo == NumOperandVals) {
10977 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10978 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10979 return ReplaceInstUsesWith(PN, NonPhiInVal);
10984 // Sort the PHI node operands to match the pred iterator order. This will
10985 // help identical PHIs be eliminated by other passes. Other passes shouldn't
10986 // depend on this for correctness however.
10988 for (pred_iterator PI = pred_begin(PN.getParent()),
10989 PE = pred_end(PN.getParent()); PI != PE; ++PI, ++i)
10990 if (PN.getIncomingBlock(i) != *PI) {
10991 unsigned j = PN.getBasicBlockIndex(*PI);
10992 Value *VA = PN.getIncomingValue(i);
10993 BasicBlock *BBA = PN.getIncomingBlock(i);
10994 Value *VB = PN.getIncomingValue(j);
10995 BasicBlock *BBB = PN.getIncomingBlock(j);
10996 PN.setIncomingBlock(i, BBB);
10997 PN.setIncomingValue(i, VB);
10998 PN.setIncomingBlock(j, BBA);
10999 PN.setIncomingValue(j, VA);
11005 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11006 Value *PtrOp = GEP.getOperand(0);
11007 // Eliminate 'getelementptr %P, i32 0' and 'getelementptr %P', they are noops.
11008 if (GEP.getNumOperands() == 1)
11009 return ReplaceInstUsesWith(GEP, PtrOp);
11011 if (isa<UndefValue>(GEP.getOperand(0)))
11012 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
11014 bool HasZeroPointerIndex = false;
11015 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11016 HasZeroPointerIndex = C->isNullValue();
11018 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11019 return ReplaceInstUsesWith(GEP, PtrOp);
11021 // Eliminate unneeded casts for indices.
11023 bool MadeChange = false;
11024 unsigned PtrSize = TD->getPointerSizeInBits();
11026 gep_type_iterator GTI = gep_type_begin(GEP);
11027 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
11028 I != E; ++I, ++GTI) {
11029 if (!isa<SequentialType>(*GTI)) continue;
11031 // If we are using a wider index than needed for this platform, shrink it
11032 // to what we need. If narrower, sign-extend it to what we need. This
11033 // explicit cast can make subsequent optimizations more obvious.
11034 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
11035 if (OpBits == PtrSize)
11038 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
11041 if (MadeChange) return &GEP;
11044 // Combine Indices - If the source pointer to this getelementptr instruction
11045 // is a getelementptr instruction, combine the indices of the two
11046 // getelementptr instructions into a single instruction.
11048 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
11049 // Note that if our source is a gep chain itself that we wait for that
11050 // chain to be resolved before we perform this transformation. This
11051 // avoids us creating a TON of code in some cases.
11053 if (GetElementPtrInst *SrcGEP =
11054 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
11055 if (SrcGEP->getNumOperands() == 2)
11056 return 0; // Wait until our source is folded to completion.
11058 SmallVector<Value*, 8> Indices;
11060 // Find out whether the last index in the source GEP is a sequential idx.
11061 bool EndsWithSequential = false;
11062 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
11064 EndsWithSequential = !isa<StructType>(*I);
11066 // Can we combine the two pointer arithmetics offsets?
11067 if (EndsWithSequential) {
11068 // Replace: gep (gep %P, long B), long A, ...
11069 // With: T = long A+B; gep %P, T, ...
11072 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
11073 Value *GO1 = GEP.getOperand(1);
11074 if (SO1 == Constant::getNullValue(SO1->getType())) {
11076 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
11079 // If they aren't the same type, then the input hasn't been processed
11080 // by the loop above yet (which canonicalizes sequential index types to
11081 // intptr_t). Just avoid transforming this until the input has been
11083 if (SO1->getType() != GO1->getType())
11085 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11088 // Update the GEP in place if possible.
11089 if (Src->getNumOperands() == 2) {
11090 GEP.setOperand(0, Src->getOperand(0));
11091 GEP.setOperand(1, Sum);
11094 Indices.append(Src->op_begin()+1, Src->op_end()-1);
11095 Indices.push_back(Sum);
11096 Indices.append(GEP.op_begin()+2, GEP.op_end());
11097 } else if (isa<Constant>(*GEP.idx_begin()) &&
11098 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11099 Src->getNumOperands() != 1) {
11100 // Otherwise we can do the fold if the first index of the GEP is a zero
11101 Indices.append(Src->op_begin()+1, Src->op_end());
11102 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
11105 if (!Indices.empty())
11106 return (cast<GEPOperator>(&GEP)->isInBounds() &&
11107 Src->isInBounds()) ?
11108 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
11109 Indices.end(), GEP.getName()) :
11110 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
11111 Indices.end(), GEP.getName());
11114 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
11115 if (Value *X = getBitCastOperand(PtrOp)) {
11116 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
11118 // If the input bitcast is actually "bitcast(bitcast(x))", then we don't
11119 // want to change the gep until the bitcasts are eliminated.
11120 if (getBitCastOperand(X)) {
11121 Worklist.AddValue(PtrOp);
11125 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11126 // into : GEP [10 x i8]* X, i32 0, ...
11128 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11129 // into : GEP i8* X, ...
11131 // This occurs when the program declares an array extern like "int X[];"
11132 if (HasZeroPointerIndex) {
11133 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11134 const PointerType *XTy = cast<PointerType>(X->getType());
11135 if (const ArrayType *CATy =
11136 dyn_cast<ArrayType>(CPTy->getElementType())) {
11137 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11138 if (CATy->getElementType() == XTy->getElementType()) {
11139 // -> GEP i8* X, ...
11140 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11141 return cast<GEPOperator>(&GEP)->isInBounds() ?
11142 GetElementPtrInst::CreateInBounds(X, Indices.begin(), Indices.end(),
11144 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11148 if (const ArrayType *XATy = dyn_cast<ArrayType>(XTy->getElementType())){
11149 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11150 if (CATy->getElementType() == XATy->getElementType()) {
11151 // -> GEP [10 x i8]* X, i32 0, ...
11152 // At this point, we know that the cast source type is a pointer
11153 // to an array of the same type as the destination pointer
11154 // array. Because the array type is never stepped over (there
11155 // is a leading zero) we can fold the cast into this GEP.
11156 GEP.setOperand(0, X);
11161 } else if (GEP.getNumOperands() == 2) {
11162 // Transform things like:
11163 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11164 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11165 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11166 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11167 if (TD && isa<ArrayType>(SrcElTy) &&
11168 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11169 TD->getTypeAllocSize(ResElTy)) {
11171 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11172 Idx[1] = GEP.getOperand(1);
11173 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11174 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11175 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11176 // V and GEP are both pointer types --> BitCast
11177 return new BitCastInst(NewGEP, GEP.getType());
11180 // Transform things like:
11181 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11182 // (where tmp = 8*tmp2) into:
11183 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11185 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
11186 uint64_t ArrayEltSize =
11187 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11189 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11190 // allow either a mul, shift, or constant here.
11192 ConstantInt *Scale = 0;
11193 if (ArrayEltSize == 1) {
11194 NewIdx = GEP.getOperand(1);
11195 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11196 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11197 NewIdx = ConstantInt::get(CI->getType(), 1);
11199 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11200 if (Inst->getOpcode() == Instruction::Shl &&
11201 isa<ConstantInt>(Inst->getOperand(1))) {
11202 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11203 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11204 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11206 NewIdx = Inst->getOperand(0);
11207 } else if (Inst->getOpcode() == Instruction::Mul &&
11208 isa<ConstantInt>(Inst->getOperand(1))) {
11209 Scale = cast<ConstantInt>(Inst->getOperand(1));
11210 NewIdx = Inst->getOperand(0);
11214 // If the index will be to exactly the right offset with the scale taken
11215 // out, perform the transformation. Note, we don't know whether Scale is
11216 // signed or not. We'll use unsigned version of division/modulo
11217 // operation after making sure Scale doesn't have the sign bit set.
11218 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11219 Scale->getZExtValue() % ArrayEltSize == 0) {
11220 Scale = ConstantInt::get(Scale->getType(),
11221 Scale->getZExtValue() / ArrayEltSize);
11222 if (Scale->getZExtValue() != 1) {
11223 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11225 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
11228 // Insert the new GEP instruction.
11230 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11232 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11233 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11234 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11235 // The NewGEP must be pointer typed, so must the old one -> BitCast
11236 return new BitCastInst(NewGEP, GEP.getType());
11242 /// See if we can simplify:
11243 /// X = bitcast A* to B*
11244 /// Y = gep X, <...constant indices...>
11245 /// into a gep of the original struct. This is important for SROA and alias
11246 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11247 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11249 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11250 // Determine how much the GEP moves the pointer. We are guaranteed to get
11251 // a constant back from EmitGEPOffset.
11252 ConstantInt *OffsetV =
11253 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11254 int64_t Offset = OffsetV->getSExtValue();
11256 // If this GEP instruction doesn't move the pointer, just replace the GEP
11257 // with a bitcast of the real input to the dest type.
11259 // If the bitcast is of an allocation, and the allocation will be
11260 // converted to match the type of the cast, don't touch this.
11261 if (isa<AllocaInst>(BCI->getOperand(0)) ||
11262 isMalloc(BCI->getOperand(0))) {
11263 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11264 if (Instruction *I = visitBitCast(*BCI)) {
11267 BCI->getParent()->getInstList().insert(BCI, I);
11268 ReplaceInstUsesWith(*BCI, I);
11273 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11276 // Otherwise, if the offset is non-zero, we need to find out if there is a
11277 // field at Offset in 'A's type. If so, we can pull the cast through the
11279 SmallVector<Value*, 8> NewIndices;
11281 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11282 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11283 Value *NGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11284 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
11285 NewIndices.end()) :
11286 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
11289 if (NGEP->getType() == GEP.getType())
11290 return ReplaceInstUsesWith(GEP, NGEP);
11291 NGEP->takeName(&GEP);
11292 return new BitCastInst(NGEP, GEP.getType());
11300 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
11301 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11302 if (AI.isArrayAllocation()) { // Check C != 1
11303 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11304 const Type *NewTy =
11305 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11306 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11307 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
11308 New->setAlignment(AI.getAlignment());
11310 // Scan to the end of the allocation instructions, to skip over a block of
11311 // allocas if possible...also skip interleaved debug info
11313 BasicBlock::iterator It = New;
11314 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11316 // Now that I is pointing to the first non-allocation-inst in the block,
11317 // insert our getelementptr instruction...
11319 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
11323 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
11324 New->getName()+".sub", It);
11326 // Now make everything use the getelementptr instead of the original
11328 return ReplaceInstUsesWith(AI, V);
11329 } else if (isa<UndefValue>(AI.getArraySize())) {
11330 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11334 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11335 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11336 // Note that we only do this for alloca's, because malloc should allocate
11337 // and return a unique pointer, even for a zero byte allocation.
11338 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11339 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11341 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11342 if (AI.getAlignment() == 0)
11343 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11349 Instruction *InstCombiner::visitFree(Instruction &FI) {
11350 Value *Op = FI.getOperand(1);
11352 // free undef -> unreachable.
11353 if (isa<UndefValue>(Op)) {
11354 // Insert a new store to null because we cannot modify the CFG here.
11355 new StoreInst(ConstantInt::getTrue(*Context),
11356 UndefValue::get(Type::getInt1PtrTy(*Context)), &FI);
11357 return EraseInstFromFunction(FI);
11360 // If we have 'free null' delete the instruction. This can happen in stl code
11361 // when lots of inlining happens.
11362 if (isa<ConstantPointerNull>(Op))
11363 return EraseInstFromFunction(FI);
11365 // If we have a malloc call whose only use is a free call, delete both.
11366 if (isMalloc(Op)) {
11367 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
11368 if (Op->hasOneUse() && CI->hasOneUse()) {
11369 EraseInstFromFunction(FI);
11370 EraseInstFromFunction(*CI);
11371 return EraseInstFromFunction(*cast<Instruction>(Op));
11374 // Op is a call to malloc
11375 if (Op->hasOneUse()) {
11376 EraseInstFromFunction(FI);
11377 return EraseInstFromFunction(*cast<Instruction>(Op));
11385 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11386 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11387 const TargetData *TD) {
11388 User *CI = cast<User>(LI.getOperand(0));
11389 Value *CastOp = CI->getOperand(0);
11390 LLVMContext *Context = IC.getContext();
11392 const PointerType *DestTy = cast<PointerType>(CI->getType());
11393 const Type *DestPTy = DestTy->getElementType();
11394 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11396 // If the address spaces don't match, don't eliminate the cast.
11397 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11400 const Type *SrcPTy = SrcTy->getElementType();
11402 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11403 isa<VectorType>(DestPTy)) {
11404 // If the source is an array, the code below will not succeed. Check to
11405 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11407 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11408 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11409 if (ASrcTy->getNumElements() != 0) {
11411 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11413 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11414 SrcTy = cast<PointerType>(CastOp->getType());
11415 SrcPTy = SrcTy->getElementType();
11418 if (IC.getTargetData() &&
11419 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11420 isa<VectorType>(SrcPTy)) &&
11421 // Do not allow turning this into a load of an integer, which is then
11422 // casted to a pointer, this pessimizes pointer analysis a lot.
11423 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11424 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11425 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11427 // Okay, we are casting from one integer or pointer type to another of
11428 // the same size. Instead of casting the pointer before the load, cast
11429 // the result of the loaded value.
11431 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
11432 // Now cast the result of the load.
11433 return new BitCastInst(NewLoad, LI.getType());
11440 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11441 Value *Op = LI.getOperand(0);
11443 // Attempt to improve the alignment.
11445 unsigned KnownAlign =
11446 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11448 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11449 LI.getAlignment()))
11450 LI.setAlignment(KnownAlign);
11453 // load (cast X) --> cast (load X) iff safe.
11454 if (isa<CastInst>(Op))
11455 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11458 // None of the following transforms are legal for volatile loads.
11459 if (LI.isVolatile()) return 0;
11461 // Do really simple store-to-load forwarding and load CSE, to catch cases
11462 // where there are several consequtive memory accesses to the same location,
11463 // separated by a few arithmetic operations.
11464 BasicBlock::iterator BBI = &LI;
11465 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11466 return ReplaceInstUsesWith(LI, AvailableVal);
11468 // load(gep null, ...) -> unreachable
11469 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11470 const Value *GEPI0 = GEPI->getOperand(0);
11471 // TODO: Consider a target hook for valid address spaces for this xform.
11472 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
11473 // Insert a new store to null instruction before the load to indicate
11474 // that this code is not reachable. We do this instead of inserting
11475 // an unreachable instruction directly because we cannot modify the
11477 new StoreInst(UndefValue::get(LI.getType()),
11478 Constant::getNullValue(Op->getType()), &LI);
11479 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11483 // load null/undef -> unreachable
11484 // TODO: Consider a target hook for valid address spaces for this xform.
11485 if (isa<UndefValue>(Op) ||
11486 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
11487 // Insert a new store to null instruction before the load to indicate that
11488 // this code is not reachable. We do this instead of inserting an
11489 // unreachable instruction directly because we cannot modify the CFG.
11490 new StoreInst(UndefValue::get(LI.getType()),
11491 Constant::getNullValue(Op->getType()), &LI);
11492 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11495 // Instcombine load (constantexpr_cast global) -> cast (load global)
11496 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
11498 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11501 if (Op->hasOneUse()) {
11502 // Change select and PHI nodes to select values instead of addresses: this
11503 // helps alias analysis out a lot, allows many others simplifications, and
11504 // exposes redundancy in the code.
11506 // Note that we cannot do the transformation unless we know that the
11507 // introduced loads cannot trap! Something like this is valid as long as
11508 // the condition is always false: load (select bool %C, int* null, int* %G),
11509 // but it would not be valid if we transformed it to load from null
11510 // unconditionally.
11512 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11513 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11514 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11515 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11516 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
11517 SI->getOperand(1)->getName()+".val");
11518 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
11519 SI->getOperand(2)->getName()+".val");
11520 return SelectInst::Create(SI->getCondition(), V1, V2);
11523 // load (select (cond, null, P)) -> load P
11524 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11525 if (C->isNullValue()) {
11526 LI.setOperand(0, SI->getOperand(2));
11530 // load (select (cond, P, null)) -> load P
11531 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11532 if (C->isNullValue()) {
11533 LI.setOperand(0, SI->getOperand(1));
11541 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11542 /// when possible. This makes it generally easy to do alias analysis and/or
11543 /// SROA/mem2reg of the memory object.
11544 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11545 User *CI = cast<User>(SI.getOperand(1));
11546 Value *CastOp = CI->getOperand(0);
11548 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11549 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11550 if (SrcTy == 0) return 0;
11552 const Type *SrcPTy = SrcTy->getElementType();
11554 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11557 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11558 /// to its first element. This allows us to handle things like:
11559 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11560 /// on 32-bit hosts.
11561 SmallVector<Value*, 4> NewGEPIndices;
11563 // If the source is an array, the code below will not succeed. Check to
11564 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11566 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11567 // Index through pointer.
11568 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
11569 NewGEPIndices.push_back(Zero);
11572 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11573 if (!STy->getNumElements()) /* Struct can be empty {} */
11575 NewGEPIndices.push_back(Zero);
11576 SrcPTy = STy->getElementType(0);
11577 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11578 NewGEPIndices.push_back(Zero);
11579 SrcPTy = ATy->getElementType();
11585 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11588 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11591 // If the pointers point into different address spaces or if they point to
11592 // values with different sizes, we can't do the transformation.
11593 if (!IC.getTargetData() ||
11594 SrcTy->getAddressSpace() !=
11595 cast<PointerType>(CI->getType())->getAddressSpace() ||
11596 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11597 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11600 // Okay, we are casting from one integer or pointer type to another of
11601 // the same size. Instead of casting the pointer before
11602 // the store, cast the value to be stored.
11604 Value *SIOp0 = SI.getOperand(0);
11605 Instruction::CastOps opcode = Instruction::BitCast;
11606 const Type* CastSrcTy = SIOp0->getType();
11607 const Type* CastDstTy = SrcPTy;
11608 if (isa<PointerType>(CastDstTy)) {
11609 if (CastSrcTy->isInteger())
11610 opcode = Instruction::IntToPtr;
11611 } else if (isa<IntegerType>(CastDstTy)) {
11612 if (isa<PointerType>(SIOp0->getType()))
11613 opcode = Instruction::PtrToInt;
11616 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11617 // emit a GEP to index into its first field.
11618 if (!NewGEPIndices.empty())
11619 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
11620 NewGEPIndices.end());
11622 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
11623 SIOp0->getName()+".c");
11624 return new StoreInst(NewCast, CastOp);
11627 /// equivalentAddressValues - Test if A and B will obviously have the same
11628 /// value. This includes recognizing that %t0 and %t1 will have the same
11629 /// value in code like this:
11630 /// %t0 = getelementptr \@a, 0, 3
11631 /// store i32 0, i32* %t0
11632 /// %t1 = getelementptr \@a, 0, 3
11633 /// %t2 = load i32* %t1
11635 static bool equivalentAddressValues(Value *A, Value *B) {
11636 // Test if the values are trivially equivalent.
11637 if (A == B) return true;
11639 // Test if the values come form identical arithmetic instructions.
11640 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
11641 // its only used to compare two uses within the same basic block, which
11642 // means that they'll always either have the same value or one of them
11643 // will have an undefined value.
11644 if (isa<BinaryOperator>(A) ||
11645 isa<CastInst>(A) ||
11647 isa<GetElementPtrInst>(A))
11648 if (Instruction *BI = dyn_cast<Instruction>(B))
11649 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
11652 // Otherwise they may not be equivalent.
11656 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11657 // return the llvm.dbg.declare.
11658 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11659 if (!V->hasNUses(2))
11661 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11663 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11665 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11666 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11673 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11674 Value *Val = SI.getOperand(0);
11675 Value *Ptr = SI.getOperand(1);
11677 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11678 EraseInstFromFunction(SI);
11683 // If the RHS is an alloca with a single use, zapify the store, making the
11685 // If the RHS is an alloca with a two uses, the other one being a
11686 // llvm.dbg.declare, zapify the store and the declare, making the
11687 // alloca dead. We must do this to prevent declare's from affecting
11689 if (!SI.isVolatile()) {
11690 if (Ptr->hasOneUse()) {
11691 if (isa<AllocaInst>(Ptr)) {
11692 EraseInstFromFunction(SI);
11696 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11697 if (isa<AllocaInst>(GEP->getOperand(0))) {
11698 if (GEP->getOperand(0)->hasOneUse()) {
11699 EraseInstFromFunction(SI);
11703 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11704 EraseInstFromFunction(*DI);
11705 EraseInstFromFunction(SI);
11712 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11713 EraseInstFromFunction(*DI);
11714 EraseInstFromFunction(SI);
11720 // Attempt to improve the alignment.
11722 unsigned KnownAlign =
11723 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11725 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11726 SI.getAlignment()))
11727 SI.setAlignment(KnownAlign);
11730 // Do really simple DSE, to catch cases where there are several consecutive
11731 // stores to the same location, separated by a few arithmetic operations. This
11732 // situation often occurs with bitfield accesses.
11733 BasicBlock::iterator BBI = &SI;
11734 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11737 // Don't count debug info directives, lest they affect codegen,
11738 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11739 // It is necessary for correctness to skip those that feed into a
11740 // llvm.dbg.declare, as these are not present when debugging is off.
11741 if (isa<DbgInfoIntrinsic>(BBI) ||
11742 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11747 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11748 // Prev store isn't volatile, and stores to the same location?
11749 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11750 SI.getOperand(1))) {
11753 EraseInstFromFunction(*PrevSI);
11759 // If this is a load, we have to stop. However, if the loaded value is from
11760 // the pointer we're loading and is producing the pointer we're storing,
11761 // then *this* store is dead (X = load P; store X -> P).
11762 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11763 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11764 !SI.isVolatile()) {
11765 EraseInstFromFunction(SI);
11769 // Otherwise, this is a load from some other location. Stores before it
11770 // may not be dead.
11774 // Don't skip over loads or things that can modify memory.
11775 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11780 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11782 // store X, null -> turns into 'unreachable' in SimplifyCFG
11783 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
11784 if (!isa<UndefValue>(Val)) {
11785 SI.setOperand(0, UndefValue::get(Val->getType()));
11786 if (Instruction *U = dyn_cast<Instruction>(Val))
11787 Worklist.Add(U); // Dropped a use.
11790 return 0; // Do not modify these!
11793 // store undef, Ptr -> noop
11794 if (isa<UndefValue>(Val)) {
11795 EraseInstFromFunction(SI);
11800 // If the pointer destination is a cast, see if we can fold the cast into the
11802 if (isa<CastInst>(Ptr))
11803 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11805 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11807 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11811 // If this store is the last instruction in the basic block (possibly
11812 // excepting debug info instructions and the pointer bitcasts that feed
11813 // into them), and if the block ends with an unconditional branch, try
11814 // to move it to the successor block.
11818 } while (isa<DbgInfoIntrinsic>(BBI) ||
11819 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11820 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11821 if (BI->isUnconditional())
11822 if (SimplifyStoreAtEndOfBlock(SI))
11823 return 0; // xform done!
11828 /// SimplifyStoreAtEndOfBlock - Turn things like:
11829 /// if () { *P = v1; } else { *P = v2 }
11830 /// into a phi node with a store in the successor.
11832 /// Simplify things like:
11833 /// *P = v1; if () { *P = v2; }
11834 /// into a phi node with a store in the successor.
11836 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11837 BasicBlock *StoreBB = SI.getParent();
11839 // Check to see if the successor block has exactly two incoming edges. If
11840 // so, see if the other predecessor contains a store to the same location.
11841 // if so, insert a PHI node (if needed) and move the stores down.
11842 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11844 // Determine whether Dest has exactly two predecessors and, if so, compute
11845 // the other predecessor.
11846 pred_iterator PI = pred_begin(DestBB);
11847 BasicBlock *OtherBB = 0;
11848 if (*PI != StoreBB)
11851 if (PI == pred_end(DestBB))
11854 if (*PI != StoreBB) {
11859 if (++PI != pred_end(DestBB))
11862 // Bail out if all the relevant blocks aren't distinct (this can happen,
11863 // for example, if SI is in an infinite loop)
11864 if (StoreBB == DestBB || OtherBB == DestBB)
11867 // Verify that the other block ends in a branch and is not otherwise empty.
11868 BasicBlock::iterator BBI = OtherBB->getTerminator();
11869 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11870 if (!OtherBr || BBI == OtherBB->begin())
11873 // If the other block ends in an unconditional branch, check for the 'if then
11874 // else' case. there is an instruction before the branch.
11875 StoreInst *OtherStore = 0;
11876 if (OtherBr->isUnconditional()) {
11878 // Skip over debugging info.
11879 while (isa<DbgInfoIntrinsic>(BBI) ||
11880 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11881 if (BBI==OtherBB->begin())
11885 // If this isn't a store, or isn't a store to the same location, bail out.
11886 OtherStore = dyn_cast<StoreInst>(BBI);
11887 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11890 // Otherwise, the other block ended with a conditional branch. If one of the
11891 // destinations is StoreBB, then we have the if/then case.
11892 if (OtherBr->getSuccessor(0) != StoreBB &&
11893 OtherBr->getSuccessor(1) != StoreBB)
11896 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11897 // if/then triangle. See if there is a store to the same ptr as SI that
11898 // lives in OtherBB.
11900 // Check to see if we find the matching store.
11901 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11902 if (OtherStore->getOperand(1) != SI.getOperand(1))
11906 // If we find something that may be using or overwriting the stored
11907 // value, or if we run out of instructions, we can't do the xform.
11908 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11909 BBI == OtherBB->begin())
11913 // In order to eliminate the store in OtherBr, we have to
11914 // make sure nothing reads or overwrites the stored value in
11916 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11917 // FIXME: This should really be AA driven.
11918 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11923 // Insert a PHI node now if we need it.
11924 Value *MergedVal = OtherStore->getOperand(0);
11925 if (MergedVal != SI.getOperand(0)) {
11926 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11927 PN->reserveOperandSpace(2);
11928 PN->addIncoming(SI.getOperand(0), SI.getParent());
11929 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11930 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11933 // Advance to a place where it is safe to insert the new store and
11935 BBI = DestBB->getFirstNonPHI();
11936 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11937 OtherStore->isVolatile()), *BBI);
11939 // Nuke the old stores.
11940 EraseInstFromFunction(SI);
11941 EraseInstFromFunction(*OtherStore);
11947 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11948 // Change br (not X), label True, label False to: br X, label False, True
11950 BasicBlock *TrueDest;
11951 BasicBlock *FalseDest;
11952 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11953 !isa<Constant>(X)) {
11954 // Swap Destinations and condition...
11955 BI.setCondition(X);
11956 BI.setSuccessor(0, FalseDest);
11957 BI.setSuccessor(1, TrueDest);
11961 // Cannonicalize fcmp_one -> fcmp_oeq
11962 FCmpInst::Predicate FPred; Value *Y;
11963 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11964 TrueDest, FalseDest)) &&
11965 BI.getCondition()->hasOneUse())
11966 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11967 FPred == FCmpInst::FCMP_OGE) {
11968 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
11969 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
11971 // Swap Destinations and condition.
11972 BI.setSuccessor(0, FalseDest);
11973 BI.setSuccessor(1, TrueDest);
11974 Worklist.Add(Cond);
11978 // Cannonicalize icmp_ne -> icmp_eq
11979 ICmpInst::Predicate IPred;
11980 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11981 TrueDest, FalseDest)) &&
11982 BI.getCondition()->hasOneUse())
11983 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11984 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11985 IPred == ICmpInst::ICMP_SGE) {
11986 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
11987 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
11988 // Swap Destinations and condition.
11989 BI.setSuccessor(0, FalseDest);
11990 BI.setSuccessor(1, TrueDest);
11991 Worklist.Add(Cond);
11998 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11999 Value *Cond = SI.getCondition();
12000 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12001 if (I->getOpcode() == Instruction::Add)
12002 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12003 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12004 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12006 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12008 SI.setOperand(0, I->getOperand(0));
12016 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12017 Value *Agg = EV.getAggregateOperand();
12019 if (!EV.hasIndices())
12020 return ReplaceInstUsesWith(EV, Agg);
12022 if (Constant *C = dyn_cast<Constant>(Agg)) {
12023 if (isa<UndefValue>(C))
12024 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12026 if (isa<ConstantAggregateZero>(C))
12027 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12029 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12030 // Extract the element indexed by the first index out of the constant
12031 Value *V = C->getOperand(*EV.idx_begin());
12032 if (EV.getNumIndices() > 1)
12033 // Extract the remaining indices out of the constant indexed by the
12035 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12037 return ReplaceInstUsesWith(EV, V);
12039 return 0; // Can't handle other constants
12041 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12042 // We're extracting from an insertvalue instruction, compare the indices
12043 const unsigned *exti, *exte, *insi, *inse;
12044 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12045 exte = EV.idx_end(), inse = IV->idx_end();
12046 exti != exte && insi != inse;
12048 if (*insi != *exti)
12049 // The insert and extract both reference distinctly different elements.
12050 // This means the extract is not influenced by the insert, and we can
12051 // replace the aggregate operand of the extract with the aggregate
12052 // operand of the insert. i.e., replace
12053 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12054 // %E = extractvalue { i32, { i32 } } %I, 0
12056 // %E = extractvalue { i32, { i32 } } %A, 0
12057 return ExtractValueInst::Create(IV->getAggregateOperand(),
12058 EV.idx_begin(), EV.idx_end());
12060 if (exti == exte && insi == inse)
12061 // Both iterators are at the end: Index lists are identical. Replace
12062 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12063 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12065 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12066 if (exti == exte) {
12067 // The extract list is a prefix of the insert list. i.e. replace
12068 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12069 // %E = extractvalue { i32, { i32 } } %I, 1
12071 // %X = extractvalue { i32, { i32 } } %A, 1
12072 // %E = insertvalue { i32 } %X, i32 42, 0
12073 // by switching the order of the insert and extract (though the
12074 // insertvalue should be left in, since it may have other uses).
12075 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
12076 EV.idx_begin(), EV.idx_end());
12077 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12081 // The insert list is a prefix of the extract list
12082 // We can simply remove the common indices from the extract and make it
12083 // operate on the inserted value instead of the insertvalue result.
12085 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12086 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12088 // %E extractvalue { i32 } { i32 42 }, 0
12089 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12092 // Can't simplify extracts from other values. Note that nested extracts are
12093 // already simplified implicitely by the above (extract ( extract (insert) )
12094 // will be translated into extract ( insert ( extract ) ) first and then just
12095 // the value inserted, if appropriate).
12099 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12100 /// is to leave as a vector operation.
12101 static bool CheapToScalarize(Value *V, bool isConstant) {
12102 if (isa<ConstantAggregateZero>(V))
12104 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12105 if (isConstant) return true;
12106 // If all elts are the same, we can extract.
12107 Constant *Op0 = C->getOperand(0);
12108 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12109 if (C->getOperand(i) != Op0)
12113 Instruction *I = dyn_cast<Instruction>(V);
12114 if (!I) return false;
12116 // Insert element gets simplified to the inserted element or is deleted if
12117 // this is constant idx extract element and its a constant idx insertelt.
12118 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12119 isa<ConstantInt>(I->getOperand(2)))
12121 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12123 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12124 if (BO->hasOneUse() &&
12125 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12126 CheapToScalarize(BO->getOperand(1), isConstant)))
12128 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12129 if (CI->hasOneUse() &&
12130 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12131 CheapToScalarize(CI->getOperand(1), isConstant)))
12137 /// Read and decode a shufflevector mask.
12139 /// It turns undef elements into values that are larger than the number of
12140 /// elements in the input.
12141 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12142 unsigned NElts = SVI->getType()->getNumElements();
12143 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12144 return std::vector<unsigned>(NElts, 0);
12145 if (isa<UndefValue>(SVI->getOperand(2)))
12146 return std::vector<unsigned>(NElts, 2*NElts);
12148 std::vector<unsigned> Result;
12149 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12150 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12151 if (isa<UndefValue>(*i))
12152 Result.push_back(NElts*2); // undef -> 8
12154 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12158 /// FindScalarElement - Given a vector and an element number, see if the scalar
12159 /// value is already around as a register, for example if it were inserted then
12160 /// extracted from the vector.
12161 static Value *FindScalarElement(Value *V, unsigned EltNo,
12162 LLVMContext *Context) {
12163 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12164 const VectorType *PTy = cast<VectorType>(V->getType());
12165 unsigned Width = PTy->getNumElements();
12166 if (EltNo >= Width) // Out of range access.
12167 return UndefValue::get(PTy->getElementType());
12169 if (isa<UndefValue>(V))
12170 return UndefValue::get(PTy->getElementType());
12171 else if (isa<ConstantAggregateZero>(V))
12172 return Constant::getNullValue(PTy->getElementType());
12173 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12174 return CP->getOperand(EltNo);
12175 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12176 // If this is an insert to a variable element, we don't know what it is.
12177 if (!isa<ConstantInt>(III->getOperand(2)))
12179 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12181 // If this is an insert to the element we are looking for, return the
12183 if (EltNo == IIElt)
12184 return III->getOperand(1);
12186 // Otherwise, the insertelement doesn't modify the value, recurse on its
12188 return FindScalarElement(III->getOperand(0), EltNo, Context);
12189 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12190 unsigned LHSWidth =
12191 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12192 unsigned InEl = getShuffleMask(SVI)[EltNo];
12193 if (InEl < LHSWidth)
12194 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12195 else if (InEl < LHSWidth*2)
12196 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12198 return UndefValue::get(PTy->getElementType());
12201 // Otherwise, we don't know.
12205 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12206 // If vector val is undef, replace extract with scalar undef.
12207 if (isa<UndefValue>(EI.getOperand(0)))
12208 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12210 // If vector val is constant 0, replace extract with scalar 0.
12211 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12212 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12214 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12215 // If vector val is constant with all elements the same, replace EI with
12216 // that element. When the elements are not identical, we cannot replace yet
12217 // (we do that below, but only when the index is constant).
12218 Constant *op0 = C->getOperand(0);
12219 for (unsigned i = 1; i != C->getNumOperands(); ++i)
12220 if (C->getOperand(i) != op0) {
12225 return ReplaceInstUsesWith(EI, op0);
12228 // If extracting a specified index from the vector, see if we can recursively
12229 // find a previously computed scalar that was inserted into the vector.
12230 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12231 unsigned IndexVal = IdxC->getZExtValue();
12232 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
12234 // If this is extracting an invalid index, turn this into undef, to avoid
12235 // crashing the code below.
12236 if (IndexVal >= VectorWidth)
12237 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12239 // This instruction only demands the single element from the input vector.
12240 // If the input vector has a single use, simplify it based on this use
12242 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12243 APInt UndefElts(VectorWidth, 0);
12244 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12245 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12246 DemandedMask, UndefElts)) {
12247 EI.setOperand(0, V);
12252 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12253 return ReplaceInstUsesWith(EI, Elt);
12255 // If the this extractelement is directly using a bitcast from a vector of
12256 // the same number of elements, see if we can find the source element from
12257 // it. In this case, we will end up needing to bitcast the scalars.
12258 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12259 if (const VectorType *VT =
12260 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12261 if (VT->getNumElements() == VectorWidth)
12262 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12263 IndexVal, Context))
12264 return new BitCastInst(Elt, EI.getType());
12268 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12269 // Push extractelement into predecessor operation if legal and
12270 // profitable to do so
12271 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12272 if (I->hasOneUse() &&
12273 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
12275 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
12276 EI.getName()+".lhs");
12278 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
12279 EI.getName()+".rhs");
12280 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12282 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12283 // Extracting the inserted element?
12284 if (IE->getOperand(2) == EI.getOperand(1))
12285 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12286 // If the inserted and extracted elements are constants, they must not
12287 // be the same value, extract from the pre-inserted value instead.
12288 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
12289 Worklist.AddValue(EI.getOperand(0));
12290 EI.setOperand(0, IE->getOperand(0));
12293 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12294 // If this is extracting an element from a shufflevector, figure out where
12295 // it came from and extract from the appropriate input element instead.
12296 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12297 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12299 unsigned LHSWidth =
12300 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12302 if (SrcIdx < LHSWidth)
12303 Src = SVI->getOperand(0);
12304 else if (SrcIdx < LHSWidth*2) {
12305 SrcIdx -= LHSWidth;
12306 Src = SVI->getOperand(1);
12308 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12310 return ExtractElementInst::Create(Src,
12311 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx,
12315 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12320 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12321 /// elements from either LHS or RHS, return the shuffle mask and true.
12322 /// Otherwise, return false.
12323 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12324 std::vector<Constant*> &Mask,
12325 LLVMContext *Context) {
12326 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12327 "Invalid CollectSingleShuffleElements");
12328 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12330 if (isa<UndefValue>(V)) {
12331 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12333 } else if (V == LHS) {
12334 for (unsigned i = 0; i != NumElts; ++i)
12335 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12337 } else if (V == RHS) {
12338 for (unsigned i = 0; i != NumElts; ++i)
12339 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
12341 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12342 // If this is an insert of an extract from some other vector, include it.
12343 Value *VecOp = IEI->getOperand(0);
12344 Value *ScalarOp = IEI->getOperand(1);
12345 Value *IdxOp = IEI->getOperand(2);
12347 if (!isa<ConstantInt>(IdxOp))
12349 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12351 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12352 // Okay, we can handle this if the vector we are insertinting into is
12353 // transitively ok.
12354 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12355 // If so, update the mask to reflect the inserted undef.
12356 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
12359 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12360 if (isa<ConstantInt>(EI->getOperand(1)) &&
12361 EI->getOperand(0)->getType() == V->getType()) {
12362 unsigned ExtractedIdx =
12363 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12365 // This must be extracting from either LHS or RHS.
12366 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12367 // Okay, we can handle this if the vector we are insertinting into is
12368 // transitively ok.
12369 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12370 // If so, update the mask to reflect the inserted value.
12371 if (EI->getOperand(0) == LHS) {
12372 Mask[InsertedIdx % NumElts] =
12373 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12375 assert(EI->getOperand(0) == RHS);
12376 Mask[InsertedIdx % NumElts] =
12377 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
12386 // TODO: Handle shufflevector here!
12391 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12392 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12393 /// that computes V and the LHS value of the shuffle.
12394 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12395 Value *&RHS, LLVMContext *Context) {
12396 assert(isa<VectorType>(V->getType()) &&
12397 (RHS == 0 || V->getType() == RHS->getType()) &&
12398 "Invalid shuffle!");
12399 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12401 if (isa<UndefValue>(V)) {
12402 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12404 } else if (isa<ConstantAggregateZero>(V)) {
12405 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
12407 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12408 // If this is an insert of an extract from some other vector, include it.
12409 Value *VecOp = IEI->getOperand(0);
12410 Value *ScalarOp = IEI->getOperand(1);
12411 Value *IdxOp = IEI->getOperand(2);
12413 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12414 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12415 EI->getOperand(0)->getType() == V->getType()) {
12416 unsigned ExtractedIdx =
12417 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12418 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12420 // Either the extracted from or inserted into vector must be RHSVec,
12421 // otherwise we'd end up with a shuffle of three inputs.
12422 if (EI->getOperand(0) == RHS || RHS == 0) {
12423 RHS = EI->getOperand(0);
12424 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12425 Mask[InsertedIdx % NumElts] =
12426 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
12430 if (VecOp == RHS) {
12431 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12433 // Everything but the extracted element is replaced with the RHS.
12434 for (unsigned i = 0; i != NumElts; ++i) {
12435 if (i != InsertedIdx)
12436 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
12441 // If this insertelement is a chain that comes from exactly these two
12442 // vectors, return the vector and the effective shuffle.
12443 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12445 return EI->getOperand(0);
12450 // TODO: Handle shufflevector here!
12452 // Otherwise, can't do anything fancy. Return an identity vector.
12453 for (unsigned i = 0; i != NumElts; ++i)
12454 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12458 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12459 Value *VecOp = IE.getOperand(0);
12460 Value *ScalarOp = IE.getOperand(1);
12461 Value *IdxOp = IE.getOperand(2);
12463 // Inserting an undef or into an undefined place, remove this.
12464 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12465 ReplaceInstUsesWith(IE, VecOp);
12467 // If the inserted element was extracted from some other vector, and if the
12468 // indexes are constant, try to turn this into a shufflevector operation.
12469 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12470 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12471 EI->getOperand(0)->getType() == IE.getType()) {
12472 unsigned NumVectorElts = IE.getType()->getNumElements();
12473 unsigned ExtractedIdx =
12474 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12475 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12477 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12478 return ReplaceInstUsesWith(IE, VecOp);
12480 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12481 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12483 // If we are extracting a value from a vector, then inserting it right
12484 // back into the same place, just use the input vector.
12485 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12486 return ReplaceInstUsesWith(IE, VecOp);
12488 // If this insertelement isn't used by some other insertelement, turn it
12489 // (and any insertelements it points to), into one big shuffle.
12490 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12491 std::vector<Constant*> Mask;
12493 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12494 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12495 // We now have a shuffle of LHS, RHS, Mask.
12496 return new ShuffleVectorInst(LHS, RHS,
12497 ConstantVector::get(Mask));
12502 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12503 APInt UndefElts(VWidth, 0);
12504 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12505 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12512 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12513 Value *LHS = SVI.getOperand(0);
12514 Value *RHS = SVI.getOperand(1);
12515 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12517 bool MadeChange = false;
12519 // Undefined shuffle mask -> undefined value.
12520 if (isa<UndefValue>(SVI.getOperand(2)))
12521 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12523 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12525 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12528 APInt UndefElts(VWidth, 0);
12529 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12530 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12531 LHS = SVI.getOperand(0);
12532 RHS = SVI.getOperand(1);
12536 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12537 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12538 if (LHS == RHS || isa<UndefValue>(LHS)) {
12539 if (isa<UndefValue>(LHS) && LHS == RHS) {
12540 // shuffle(undef,undef,mask) -> undef.
12541 return ReplaceInstUsesWith(SVI, LHS);
12544 // Remap any references to RHS to use LHS.
12545 std::vector<Constant*> Elts;
12546 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12547 if (Mask[i] >= 2*e)
12548 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12550 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12551 (Mask[i] < e && isa<UndefValue>(LHS))) {
12552 Mask[i] = 2*e; // Turn into undef.
12553 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12555 Mask[i] = Mask[i] % e; // Force to LHS.
12556 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
12560 SVI.setOperand(0, SVI.getOperand(1));
12561 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12562 SVI.setOperand(2, ConstantVector::get(Elts));
12563 LHS = SVI.getOperand(0);
12564 RHS = SVI.getOperand(1);
12568 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12569 bool isLHSID = true, isRHSID = true;
12571 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12572 if (Mask[i] >= e*2) continue; // Ignore undef values.
12573 // Is this an identity shuffle of the LHS value?
12574 isLHSID &= (Mask[i] == i);
12576 // Is this an identity shuffle of the RHS value?
12577 isRHSID &= (Mask[i]-e == i);
12580 // Eliminate identity shuffles.
12581 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12582 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12584 // If the LHS is a shufflevector itself, see if we can combine it with this
12585 // one without producing an unusual shuffle. Here we are really conservative:
12586 // we are absolutely afraid of producing a shuffle mask not in the input
12587 // program, because the code gen may not be smart enough to turn a merged
12588 // shuffle into two specific shuffles: it may produce worse code. As such,
12589 // we only merge two shuffles if the result is one of the two input shuffle
12590 // masks. In this case, merging the shuffles just removes one instruction,
12591 // which we know is safe. This is good for things like turning:
12592 // (splat(splat)) -> splat.
12593 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12594 if (isa<UndefValue>(RHS)) {
12595 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12597 std::vector<unsigned> NewMask;
12598 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12599 if (Mask[i] >= 2*e)
12600 NewMask.push_back(2*e);
12602 NewMask.push_back(LHSMask[Mask[i]]);
12604 // If the result mask is equal to the src shuffle or this shuffle mask, do
12605 // the replacement.
12606 if (NewMask == LHSMask || NewMask == Mask) {
12607 unsigned LHSInNElts =
12608 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12609 std::vector<Constant*> Elts;
12610 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12611 if (NewMask[i] >= LHSInNElts*2) {
12612 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12614 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), NewMask[i]));
12617 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12618 LHSSVI->getOperand(1),
12619 ConstantVector::get(Elts));
12624 return MadeChange ? &SVI : 0;
12630 /// TryToSinkInstruction - Try to move the specified instruction from its
12631 /// current block into the beginning of DestBlock, which can only happen if it's
12632 /// safe to move the instruction past all of the instructions between it and the
12633 /// end of its block.
12634 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12635 assert(I->hasOneUse() && "Invariants didn't hold!");
12637 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12638 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12641 // Do not sink alloca instructions out of the entry block.
12642 if (isa<AllocaInst>(I) && I->getParent() ==
12643 &DestBlock->getParent()->getEntryBlock())
12646 // We can only sink load instructions if there is nothing between the load and
12647 // the end of block that could change the value.
12648 if (I->mayReadFromMemory()) {
12649 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12651 if (Scan->mayWriteToMemory())
12655 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12657 CopyPrecedingStopPoint(I, InsertPos);
12658 I->moveBefore(InsertPos);
12664 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12665 /// all reachable code to the worklist.
12667 /// This has a couple of tricks to make the code faster and more powerful. In
12668 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12669 /// them to the worklist (this significantly speeds up instcombine on code where
12670 /// many instructions are dead or constant). Additionally, if we find a branch
12671 /// whose condition is a known constant, we only visit the reachable successors.
12673 static bool AddReachableCodeToWorklist(BasicBlock *BB,
12674 SmallPtrSet<BasicBlock*, 64> &Visited,
12676 const TargetData *TD) {
12677 bool MadeIRChange = false;
12678 SmallVector<BasicBlock*, 256> Worklist;
12679 Worklist.push_back(BB);
12681 std::vector<Instruction*> InstrsForInstCombineWorklist;
12682 InstrsForInstCombineWorklist.reserve(128);
12684 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
12686 while (!Worklist.empty()) {
12687 BB = Worklist.back();
12688 Worklist.pop_back();
12690 // We have now visited this block! If we've already been here, ignore it.
12691 if (!Visited.insert(BB)) continue;
12693 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12694 Instruction *Inst = BBI++;
12696 // DCE instruction if trivially dead.
12697 if (isInstructionTriviallyDead(Inst)) {
12699 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
12700 Inst->eraseFromParent();
12704 // ConstantProp instruction if trivially constant.
12705 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
12706 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12707 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
12709 Inst->replaceAllUsesWith(C);
12711 Inst->eraseFromParent();
12718 // See if we can constant fold its operands.
12719 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
12721 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
12722 if (CE == 0) continue;
12724 // If we already folded this constant, don't try again.
12725 if (!FoldedConstants.insert(CE))
12729 ConstantFoldConstantExpression(CE, BB->getContext(), TD);
12730 if (NewC && NewC != CE) {
12732 MadeIRChange = true;
12738 InstrsForInstCombineWorklist.push_back(Inst);
12741 // Recursively visit successors. If this is a branch or switch on a
12742 // constant, only visit the reachable successor.
12743 TerminatorInst *TI = BB->getTerminator();
12744 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12745 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12746 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12747 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12748 Worklist.push_back(ReachableBB);
12751 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12752 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12753 // See if this is an explicit destination.
12754 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12755 if (SI->getCaseValue(i) == Cond) {
12756 BasicBlock *ReachableBB = SI->getSuccessor(i);
12757 Worklist.push_back(ReachableBB);
12761 // Otherwise it is the default destination.
12762 Worklist.push_back(SI->getSuccessor(0));
12767 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12768 Worklist.push_back(TI->getSuccessor(i));
12771 // Once we've found all of the instructions to add to instcombine's worklist,
12772 // add them in reverse order. This way instcombine will visit from the top
12773 // of the function down. This jives well with the way that it adds all uses
12774 // of instructions to the worklist after doing a transformation, thus avoiding
12775 // some N^2 behavior in pathological cases.
12776 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
12777 InstrsForInstCombineWorklist.size());
12779 return MadeIRChange;
12782 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12783 MadeIRChange = false;
12785 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12786 << F.getNameStr() << "\n");
12789 // Do a depth-first traversal of the function, populate the worklist with
12790 // the reachable instructions. Ignore blocks that are not reachable. Keep
12791 // track of which blocks we visit.
12792 SmallPtrSet<BasicBlock*, 64> Visited;
12793 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12795 // Do a quick scan over the function. If we find any blocks that are
12796 // unreachable, remove any instructions inside of them. This prevents
12797 // the instcombine code from having to deal with some bad special cases.
12798 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12799 if (!Visited.count(BB)) {
12800 Instruction *Term = BB->getTerminator();
12801 while (Term != BB->begin()) { // Remove instrs bottom-up
12802 BasicBlock::iterator I = Term; --I;
12804 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12805 // A debug intrinsic shouldn't force another iteration if we weren't
12806 // going to do one without it.
12807 if (!isa<DbgInfoIntrinsic>(I)) {
12809 MadeIRChange = true;
12812 // If I is not void type then replaceAllUsesWith undef.
12813 // This allows ValueHandlers and custom metadata to adjust itself.
12814 if (!I->getType()->isVoidTy())
12815 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12816 I->eraseFromParent();
12821 while (!Worklist.isEmpty()) {
12822 Instruction *I = Worklist.RemoveOne();
12823 if (I == 0) continue; // skip null values.
12825 // Check to see if we can DCE the instruction.
12826 if (isInstructionTriviallyDead(I)) {
12827 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12828 EraseInstFromFunction(*I);
12830 MadeIRChange = true;
12834 // Instruction isn't dead, see if we can constant propagate it.
12835 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
12836 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
12837 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
12839 // Add operands to the worklist.
12840 ReplaceInstUsesWith(*I, C);
12842 EraseInstFromFunction(*I);
12843 MadeIRChange = true;
12847 // See if we can trivially sink this instruction to a successor basic block.
12848 if (I->hasOneUse()) {
12849 BasicBlock *BB = I->getParent();
12850 Instruction *UserInst = cast<Instruction>(I->use_back());
12851 BasicBlock *UserParent;
12853 // Get the block the use occurs in.
12854 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
12855 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
12857 UserParent = UserInst->getParent();
12859 if (UserParent != BB) {
12860 bool UserIsSuccessor = false;
12861 // See if the user is one of our successors.
12862 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12863 if (*SI == UserParent) {
12864 UserIsSuccessor = true;
12868 // If the user is one of our immediate successors, and if that successor
12869 // only has us as a predecessors (we'd have to split the critical edge
12870 // otherwise), we can keep going.
12871 if (UserIsSuccessor && UserParent->getSinglePredecessor())
12872 // Okay, the CFG is simple enough, try to sink this instruction.
12873 MadeIRChange |= TryToSinkInstruction(I, UserParent);
12877 // Now that we have an instruction, try combining it to simplify it.
12878 Builder->SetInsertPoint(I->getParent(), I);
12883 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
12884 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
12886 if (Instruction *Result = visit(*I)) {
12888 // Should we replace the old instruction with a new one?
12890 DEBUG(errs() << "IC: Old = " << *I << '\n'
12891 << " New = " << *Result << '\n');
12893 // Everything uses the new instruction now.
12894 I->replaceAllUsesWith(Result);
12896 // Push the new instruction and any users onto the worklist.
12897 Worklist.Add(Result);
12898 Worklist.AddUsersToWorkList(*Result);
12900 // Move the name to the new instruction first.
12901 Result->takeName(I);
12903 // Insert the new instruction into the basic block...
12904 BasicBlock *InstParent = I->getParent();
12905 BasicBlock::iterator InsertPos = I;
12907 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12908 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12911 InstParent->getInstList().insert(InsertPos, Result);
12913 EraseInstFromFunction(*I);
12916 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
12917 << " New = " << *I << '\n');
12920 // If the instruction was modified, it's possible that it is now dead.
12921 // if so, remove it.
12922 if (isInstructionTriviallyDead(I)) {
12923 EraseInstFromFunction(*I);
12926 Worklist.AddUsersToWorkList(*I);
12929 MadeIRChange = true;
12934 return MadeIRChange;
12938 bool InstCombiner::runOnFunction(Function &F) {
12939 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12940 Context = &F.getContext();
12941 TD = getAnalysisIfAvailable<TargetData>();
12944 /// Builder - This is an IRBuilder that automatically inserts new
12945 /// instructions into the worklist when they are created.
12946 IRBuilder<true, TargetFolder, InstCombineIRInserter>
12947 TheBuilder(F.getContext(), TargetFolder(TD, F.getContext()),
12948 InstCombineIRInserter(Worklist));
12949 Builder = &TheBuilder;
12951 bool EverMadeChange = false;
12953 // Iterate while there is work to do.
12954 unsigned Iteration = 0;
12955 while (DoOneIteration(F, Iteration++))
12956 EverMadeChange = true;
12959 return EverMadeChange;
12962 FunctionPass *llvm::createInstructionCombiningPass() {
12963 return new InstCombiner();