1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/LLVMContext.h"
40 #include "llvm/Pass.h"
41 #include "llvm/DerivedTypes.h"
42 #include "llvm/GlobalVariable.h"
43 #include "llvm/Operator.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/MallocHelper.h"
46 #include "llvm/Analysis/ValueTracking.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Support/CallSite.h"
51 #include "llvm/Support/ConstantRange.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/ErrorHandling.h"
54 #include "llvm/Support/GetElementPtrTypeIterator.h"
55 #include "llvm/Support/InstVisitor.h"
56 #include "llvm/Support/IRBuilder.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/PatternMatch.h"
59 #include "llvm/Support/TargetFolder.h"
60 #include "llvm/Support/raw_ostream.h"
61 #include "llvm/ADT/DenseMap.h"
62 #include "llvm/ADT/SmallVector.h"
63 #include "llvm/ADT/SmallPtrSet.h"
64 #include "llvm/ADT/Statistic.h"
65 #include "llvm/ADT/STLExtras.h"
69 using namespace llvm::PatternMatch;
71 STATISTIC(NumCombined , "Number of insts combined");
72 STATISTIC(NumConstProp, "Number of constant folds");
73 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
74 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
75 STATISTIC(NumSunkInst , "Number of instructions sunk");
78 /// InstCombineWorklist - This is the worklist management logic for
80 class InstCombineWorklist {
81 SmallVector<Instruction*, 256> Worklist;
82 DenseMap<Instruction*, unsigned> WorklistMap;
84 void operator=(const InstCombineWorklist&RHS); // DO NOT IMPLEMENT
85 InstCombineWorklist(const InstCombineWorklist&); // DO NOT IMPLEMENT
87 InstCombineWorklist() {}
89 bool isEmpty() const { return Worklist.empty(); }
91 /// Add - Add the specified instruction to the worklist if it isn't already
93 void Add(Instruction *I) {
94 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second) {
95 DEBUG(errs() << "IC: ADD: " << *I << '\n');
96 Worklist.push_back(I);
100 void AddValue(Value *V) {
101 if (Instruction *I = dyn_cast<Instruction>(V))
105 /// AddInitialGroup - Add the specified batch of stuff in reverse order.
106 /// which should only be done when the worklist is empty and when the group
107 /// has no duplicates.
108 void AddInitialGroup(Instruction *const *List, unsigned NumEntries) {
109 assert(Worklist.empty() && "Worklist must be empty to add initial group");
110 Worklist.reserve(NumEntries+16);
111 DEBUG(errs() << "IC: ADDING: " << NumEntries << " instrs to worklist\n");
112 for (; NumEntries; --NumEntries) {
113 Instruction *I = List[NumEntries-1];
114 WorklistMap.insert(std::make_pair(I, Worklist.size()));
115 Worklist.push_back(I);
119 // Remove - remove I from the worklist if it exists.
120 void Remove(Instruction *I) {
121 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
122 if (It == WorklistMap.end()) return; // Not in worklist.
124 // Don't bother moving everything down, just null out the slot.
125 Worklist[It->second] = 0;
127 WorklistMap.erase(It);
130 Instruction *RemoveOne() {
131 Instruction *I = Worklist.back();
133 WorklistMap.erase(I);
137 /// AddUsersToWorkList - When an instruction is simplified, add all users of
138 /// the instruction to the work lists because they might get more simplified
141 void AddUsersToWorkList(Instruction &I) {
142 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
144 Add(cast<Instruction>(*UI));
148 /// Zap - check that the worklist is empty and nuke the backing store for
149 /// the map if it is large.
151 assert(WorklistMap.empty() && "Worklist empty, but map not?");
153 // Do an explicit clear, this shrinks the map if needed.
157 } // end anonymous namespace.
161 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works
162 /// just like the normal insertion helper, but also adds any new instructions
163 /// to the instcombine worklist.
164 class InstCombineIRInserter : public IRBuilderDefaultInserter<true> {
165 InstCombineWorklist &Worklist;
167 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
169 void InsertHelper(Instruction *I, const Twine &Name,
170 BasicBlock *BB, BasicBlock::iterator InsertPt) const {
171 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
175 } // end anonymous namespace
179 class InstCombiner : public FunctionPass,
180 public InstVisitor<InstCombiner, Instruction*> {
182 bool MustPreserveLCSSA;
185 /// Worklist - All of the instructions that need to be simplified.
186 InstCombineWorklist Worklist;
188 /// Builder - This is an IRBuilder that automatically inserts new
189 /// instructions into the worklist when they are created.
190 typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
193 static char ID; // Pass identification, replacement for typeid
194 InstCombiner() : FunctionPass(&ID), TD(0), Builder(0) {}
196 LLVMContext *Context;
197 LLVMContext *getContext() const { return Context; }
200 virtual bool runOnFunction(Function &F);
202 bool DoOneIteration(Function &F, unsigned ItNum);
204 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
205 AU.addPreservedID(LCSSAID);
206 AU.setPreservesCFG();
209 TargetData *getTargetData() const { return TD; }
211 // Visitation implementation - Implement instruction combining for different
212 // instruction types. The semantics are as follows:
214 // null - No change was made
215 // I - Change was made, I is still valid, I may be dead though
216 // otherwise - Change was made, replace I with returned instruction
218 Instruction *visitAdd(BinaryOperator &I);
219 Instruction *visitFAdd(BinaryOperator &I);
220 Instruction *visitSub(BinaryOperator &I);
221 Instruction *visitFSub(BinaryOperator &I);
222 Instruction *visitMul(BinaryOperator &I);
223 Instruction *visitFMul(BinaryOperator &I);
224 Instruction *visitURem(BinaryOperator &I);
225 Instruction *visitSRem(BinaryOperator &I);
226 Instruction *visitFRem(BinaryOperator &I);
227 bool SimplifyDivRemOfSelect(BinaryOperator &I);
228 Instruction *commonRemTransforms(BinaryOperator &I);
229 Instruction *commonIRemTransforms(BinaryOperator &I);
230 Instruction *commonDivTransforms(BinaryOperator &I);
231 Instruction *commonIDivTransforms(BinaryOperator &I);
232 Instruction *visitUDiv(BinaryOperator &I);
233 Instruction *visitSDiv(BinaryOperator &I);
234 Instruction *visitFDiv(BinaryOperator &I);
235 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
236 Instruction *FoldAndOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
237 Instruction *visitAnd(BinaryOperator &I);
238 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
239 Instruction *FoldOrOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
240 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
241 Value *A, Value *B, Value *C);
242 Instruction *visitOr (BinaryOperator &I);
243 Instruction *visitXor(BinaryOperator &I);
244 Instruction *visitShl(BinaryOperator &I);
245 Instruction *visitAShr(BinaryOperator &I);
246 Instruction *visitLShr(BinaryOperator &I);
247 Instruction *commonShiftTransforms(BinaryOperator &I);
248 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
250 Instruction *visitFCmpInst(FCmpInst &I);
251 Instruction *visitICmpInst(ICmpInst &I);
252 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
253 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
256 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
257 ConstantInt *DivRHS);
259 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
260 ICmpInst::Predicate Cond, Instruction &I);
261 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
263 Instruction *commonCastTransforms(CastInst &CI);
264 Instruction *commonIntCastTransforms(CastInst &CI);
265 Instruction *commonPointerCastTransforms(CastInst &CI);
266 Instruction *visitTrunc(TruncInst &CI);
267 Instruction *visitZExt(ZExtInst &CI);
268 Instruction *visitSExt(SExtInst &CI);
269 Instruction *visitFPTrunc(FPTruncInst &CI);
270 Instruction *visitFPExt(CastInst &CI);
271 Instruction *visitFPToUI(FPToUIInst &FI);
272 Instruction *visitFPToSI(FPToSIInst &FI);
273 Instruction *visitUIToFP(CastInst &CI);
274 Instruction *visitSIToFP(CastInst &CI);
275 Instruction *visitPtrToInt(PtrToIntInst &CI);
276 Instruction *visitIntToPtr(IntToPtrInst &CI);
277 Instruction *visitBitCast(BitCastInst &CI);
278 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
280 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
281 Instruction *visitSelectInst(SelectInst &SI);
282 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
283 Instruction *visitCallInst(CallInst &CI);
284 Instruction *visitInvokeInst(InvokeInst &II);
285 Instruction *visitPHINode(PHINode &PN);
286 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
287 Instruction *visitAllocationInst(AllocationInst &AI);
288 Instruction *visitFreeInst(FreeInst &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, AllocationInst &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 static inline Value *dyn_castNotVal(Value *V) {
634 if (BinaryOperator::isNot(V))
635 return BinaryOperator::getNotArgument(V);
637 // Constants can be considered to be not'ed values...
638 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
639 return ConstantInt::get(C->getType(), ~C->getValue());
643 // dyn_castFoldableMul - If this value is a multiply that can be folded into
644 // other computations (because it has a constant operand), return the
645 // non-constant operand of the multiply, and set CST to point to the multiplier.
646 // Otherwise, return null.
648 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
649 if (V->hasOneUse() && V->getType()->isInteger())
650 if (Instruction *I = dyn_cast<Instruction>(V)) {
651 if (I->getOpcode() == Instruction::Mul)
652 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
653 return I->getOperand(0);
654 if (I->getOpcode() == Instruction::Shl)
655 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
656 // The multiplier is really 1 << CST.
657 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
658 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
659 CST = ConstantInt::get(V->getType()->getContext(),
660 APInt(BitWidth, 1).shl(CSTVal));
661 return I->getOperand(0);
667 /// AddOne - Add one to a ConstantInt
668 static Constant *AddOne(Constant *C) {
669 return ConstantExpr::getAdd(C,
670 ConstantInt::get(C->getType(), 1));
672 /// SubOne - Subtract one from a ConstantInt
673 static Constant *SubOne(ConstantInt *C) {
674 return ConstantExpr::getSub(C,
675 ConstantInt::get(C->getType(), 1));
677 /// MultiplyOverflows - True if the multiply can not be expressed in an int
679 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
680 uint32_t W = C1->getBitWidth();
681 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
690 APInt MulExt = LHSExt * RHSExt;
693 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
694 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
695 return MulExt.slt(Min) || MulExt.sgt(Max);
697 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
701 /// ShrinkDemandedConstant - Check to see if the specified operand of the
702 /// specified instruction is a constant integer. If so, check to see if there
703 /// are any bits set in the constant that are not demanded. If so, shrink the
704 /// constant and return true.
705 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
707 assert(I && "No instruction?");
708 assert(OpNo < I->getNumOperands() && "Operand index too large");
710 // If the operand is not a constant integer, nothing to do.
711 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
712 if (!OpC) return false;
714 // If there are no bits set that aren't demanded, nothing to do.
715 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
716 if ((~Demanded & OpC->getValue()) == 0)
719 // This instruction is producing bits that are not demanded. Shrink the RHS.
720 Demanded &= OpC->getValue();
721 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
725 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
726 // set of known zero and one bits, compute the maximum and minimum values that
727 // could have the specified known zero and known one bits, returning them in
729 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
730 const APInt& KnownOne,
731 APInt& Min, APInt& Max) {
732 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
733 KnownZero.getBitWidth() == Min.getBitWidth() &&
734 KnownZero.getBitWidth() == Max.getBitWidth() &&
735 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
736 APInt UnknownBits = ~(KnownZero|KnownOne);
738 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
739 // bit if it is unknown.
741 Max = KnownOne|UnknownBits;
743 if (UnknownBits.isNegative()) { // Sign bit is unknown
744 Min.set(Min.getBitWidth()-1);
745 Max.clear(Max.getBitWidth()-1);
749 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
750 // a set of known zero and one bits, compute the maximum and minimum values that
751 // could have the specified known zero and known one bits, returning them in
753 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
754 const APInt &KnownOne,
755 APInt &Min, APInt &Max) {
756 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
757 KnownZero.getBitWidth() == Min.getBitWidth() &&
758 KnownZero.getBitWidth() == Max.getBitWidth() &&
759 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
760 APInt UnknownBits = ~(KnownZero|KnownOne);
762 // The minimum value is when the unknown bits are all zeros.
764 // The maximum value is when the unknown bits are all ones.
765 Max = KnownOne|UnknownBits;
768 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
769 /// SimplifyDemandedBits knows about. See if the instruction has any
770 /// properties that allow us to simplify its operands.
771 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
772 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
773 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
774 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
776 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
777 KnownZero, KnownOne, 0);
778 if (V == 0) return false;
779 if (V == &Inst) return true;
780 ReplaceInstUsesWith(Inst, V);
784 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
785 /// specified instruction operand if possible, updating it in place. It returns
786 /// true if it made any change and false otherwise.
787 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
788 APInt &KnownZero, APInt &KnownOne,
790 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
791 KnownZero, KnownOne, Depth);
792 if (NewVal == 0) return false;
798 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
799 /// value based on the demanded bits. When this function is called, it is known
800 /// that only the bits set in DemandedMask of the result of V are ever used
801 /// downstream. Consequently, depending on the mask and V, it may be possible
802 /// to replace V with a constant or one of its operands. In such cases, this
803 /// function does the replacement and returns true. In all other cases, it
804 /// returns false after analyzing the expression and setting KnownOne and known
805 /// to be one in the expression. KnownZero contains all the bits that are known
806 /// to be zero in the expression. These are provided to potentially allow the
807 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
808 /// the expression. KnownOne and KnownZero always follow the invariant that
809 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
810 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
811 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
812 /// and KnownOne must all be the same.
814 /// This returns null if it did not change anything and it permits no
815 /// simplification. This returns V itself if it did some simplification of V's
816 /// operands based on the information about what bits are demanded. This returns
817 /// some other non-null value if it found out that V is equal to another value
818 /// in the context where the specified bits are demanded, but not for all users.
819 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
820 APInt &KnownZero, APInt &KnownOne,
822 assert(V != 0 && "Null pointer of Value???");
823 assert(Depth <= 6 && "Limit Search Depth");
824 uint32_t BitWidth = DemandedMask.getBitWidth();
825 const Type *VTy = V->getType();
826 assert((TD || !isa<PointerType>(VTy)) &&
827 "SimplifyDemandedBits needs to know bit widths!");
828 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
829 (!VTy->isIntOrIntVector() ||
830 VTy->getScalarSizeInBits() == BitWidth) &&
831 KnownZero.getBitWidth() == BitWidth &&
832 KnownOne.getBitWidth() == BitWidth &&
833 "Value *V, DemandedMask, KnownZero and KnownOne "
834 "must have same BitWidth");
835 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
836 // We know all of the bits for a constant!
837 KnownOne = CI->getValue() & DemandedMask;
838 KnownZero = ~KnownOne & DemandedMask;
841 if (isa<ConstantPointerNull>(V)) {
842 // We know all of the bits for a constant!
844 KnownZero = DemandedMask;
850 if (DemandedMask == 0) { // Not demanding any bits from V.
851 if (isa<UndefValue>(V))
853 return UndefValue::get(VTy);
856 if (Depth == 6) // Limit search depth.
859 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
860 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
862 Instruction *I = dyn_cast<Instruction>(V);
864 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
865 return 0; // Only analyze instructions.
868 // If there are multiple uses of this value and we aren't at the root, then
869 // we can't do any simplifications of the operands, because DemandedMask
870 // only reflects the bits demanded by *one* of the users.
871 if (Depth != 0 && !I->hasOneUse()) {
872 // Despite the fact that we can't simplify this instruction in all User's
873 // context, we can at least compute the knownzero/knownone bits, and we can
874 // do simplifications that apply to *just* the one user if we know that
875 // this instruction has a simpler value in that context.
876 if (I->getOpcode() == Instruction::And) {
877 // If either the LHS or the RHS are Zero, the result is zero.
878 ComputeMaskedBits(I->getOperand(1), DemandedMask,
879 RHSKnownZero, RHSKnownOne, Depth+1);
880 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
881 LHSKnownZero, LHSKnownOne, Depth+1);
883 // If all of the demanded bits are known 1 on one side, return the other.
884 // These bits cannot contribute to the result of the 'and' in this
886 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
887 (DemandedMask & ~LHSKnownZero))
888 return I->getOperand(0);
889 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
890 (DemandedMask & ~RHSKnownZero))
891 return I->getOperand(1);
893 // If all of the demanded bits in the inputs are known zeros, return zero.
894 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
895 return Constant::getNullValue(VTy);
897 } else if (I->getOpcode() == Instruction::Or) {
898 // We can simplify (X|Y) -> X or Y in the user's context if we know that
899 // only bits from X or Y are demanded.
901 // If either the LHS or the RHS are One, the result is One.
902 ComputeMaskedBits(I->getOperand(1), DemandedMask,
903 RHSKnownZero, RHSKnownOne, Depth+1);
904 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
905 LHSKnownZero, LHSKnownOne, Depth+1);
907 // If all of the demanded bits are known zero on one side, return the
908 // other. These bits cannot contribute to the result of the 'or' in this
910 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
911 (DemandedMask & ~LHSKnownOne))
912 return I->getOperand(0);
913 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
914 (DemandedMask & ~RHSKnownOne))
915 return I->getOperand(1);
917 // If all of the potentially set bits on one side are known to be set on
918 // the other side, just use the 'other' side.
919 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
920 (DemandedMask & (~RHSKnownZero)))
921 return I->getOperand(0);
922 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
923 (DemandedMask & (~LHSKnownZero)))
924 return I->getOperand(1);
927 // Compute the KnownZero/KnownOne bits to simplify things downstream.
928 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
932 // If this is the root being simplified, allow it to have multiple uses,
933 // just set the DemandedMask to all bits so that we can try to simplify the
934 // operands. This allows visitTruncInst (for example) to simplify the
935 // operand of a trunc without duplicating all the logic below.
936 if (Depth == 0 && !V->hasOneUse())
937 DemandedMask = APInt::getAllOnesValue(BitWidth);
939 switch (I->getOpcode()) {
941 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
943 case Instruction::And:
944 // If either the LHS or the RHS are Zero, the result is zero.
945 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
946 RHSKnownZero, RHSKnownOne, Depth+1) ||
947 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
948 LHSKnownZero, LHSKnownOne, Depth+1))
950 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
951 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
953 // If all of the demanded bits are known 1 on one side, return the other.
954 // These bits cannot contribute to the result of the 'and'.
955 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
956 (DemandedMask & ~LHSKnownZero))
957 return I->getOperand(0);
958 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
959 (DemandedMask & ~RHSKnownZero))
960 return I->getOperand(1);
962 // If all of the demanded bits in the inputs are known zeros, return zero.
963 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
964 return Constant::getNullValue(VTy);
966 // If the RHS is a constant, see if we can simplify it.
967 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
970 // Output known-1 bits are only known if set in both the LHS & RHS.
971 RHSKnownOne &= LHSKnownOne;
972 // Output known-0 are known to be clear if zero in either the LHS | RHS.
973 RHSKnownZero |= LHSKnownZero;
975 case Instruction::Or:
976 // If either the LHS or the RHS are One, the result is One.
977 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
978 RHSKnownZero, RHSKnownOne, Depth+1) ||
979 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
980 LHSKnownZero, LHSKnownOne, Depth+1))
982 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
983 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
985 // If all of the demanded bits are known zero on one side, return the other.
986 // These bits cannot contribute to the result of the 'or'.
987 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
988 (DemandedMask & ~LHSKnownOne))
989 return I->getOperand(0);
990 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
991 (DemandedMask & ~RHSKnownOne))
992 return I->getOperand(1);
994 // If all of the potentially set bits on one side are known to be set on
995 // the other side, just use the 'other' side.
996 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
997 (DemandedMask & (~RHSKnownZero)))
998 return I->getOperand(0);
999 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1000 (DemandedMask & (~LHSKnownZero)))
1001 return I->getOperand(1);
1003 // If the RHS is a constant, see if we can simplify it.
1004 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1007 // Output known-0 bits are only known if clear in both the LHS & RHS.
1008 RHSKnownZero &= LHSKnownZero;
1009 // Output known-1 are known to be set if set in either the LHS | RHS.
1010 RHSKnownOne |= LHSKnownOne;
1012 case Instruction::Xor: {
1013 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1014 RHSKnownZero, RHSKnownOne, Depth+1) ||
1015 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1016 LHSKnownZero, LHSKnownOne, Depth+1))
1018 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1019 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1021 // If all of the demanded bits are known zero on one side, return the other.
1022 // These bits cannot contribute to the result of the 'xor'.
1023 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1024 return I->getOperand(0);
1025 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1026 return I->getOperand(1);
1028 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1029 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1030 (RHSKnownOne & LHSKnownOne);
1031 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1032 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1033 (RHSKnownOne & LHSKnownZero);
1035 // If all of the demanded bits are known to be zero on one side or the
1036 // other, turn this into an *inclusive* or.
1037 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1038 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1040 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1042 return InsertNewInstBefore(Or, *I);
1045 // If all of the demanded bits on one side are known, and all of the set
1046 // bits on that side are also known to be set on the other side, turn this
1047 // into an AND, as we know the bits will be cleared.
1048 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1049 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1051 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1052 Constant *AndC = Constant::getIntegerValue(VTy,
1053 ~RHSKnownOne & DemandedMask);
1055 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1056 return InsertNewInstBefore(And, *I);
1060 // If the RHS is a constant, see if we can simplify it.
1061 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1062 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1065 // If our LHS is an 'and' and if it has one use, and if any of the bits we
1066 // are flipping are known to be set, then the xor is just resetting those
1067 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
1068 // simplifying both of them.
1069 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0)))
1070 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
1071 isa<ConstantInt>(I->getOperand(1)) &&
1072 isa<ConstantInt>(LHSInst->getOperand(1)) &&
1073 (LHSKnownOne & RHSKnownOne & DemandedMask) != 0) {
1074 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
1075 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
1076 APInt NewMask = ~(LHSKnownOne & RHSKnownOne & DemandedMask);
1079 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
1080 Instruction *NewAnd =
1081 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1082 InsertNewInstBefore(NewAnd, *I);
1085 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
1086 Instruction *NewXor =
1087 BinaryOperator::CreateXor(NewAnd, XorC, "tmp");
1088 return InsertNewInstBefore(NewXor, *I);
1092 RHSKnownZero = KnownZeroOut;
1093 RHSKnownOne = KnownOneOut;
1096 case Instruction::Select:
1097 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1098 RHSKnownZero, RHSKnownOne, Depth+1) ||
1099 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1100 LHSKnownZero, LHSKnownOne, Depth+1))
1102 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1103 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1105 // If the operands are constants, see if we can simplify them.
1106 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1107 ShrinkDemandedConstant(I, 2, DemandedMask))
1110 // Only known if known in both the LHS and RHS.
1111 RHSKnownOne &= LHSKnownOne;
1112 RHSKnownZero &= LHSKnownZero;
1114 case Instruction::Trunc: {
1115 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1116 DemandedMask.zext(truncBf);
1117 RHSKnownZero.zext(truncBf);
1118 RHSKnownOne.zext(truncBf);
1119 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1120 RHSKnownZero, RHSKnownOne, Depth+1))
1122 DemandedMask.trunc(BitWidth);
1123 RHSKnownZero.trunc(BitWidth);
1124 RHSKnownOne.trunc(BitWidth);
1125 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1128 case Instruction::BitCast:
1129 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1130 return false; // vector->int or fp->int?
1132 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1133 if (const VectorType *SrcVTy =
1134 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1135 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1136 // Don't touch a bitcast between vectors of different element counts.
1139 // Don't touch a scalar-to-vector bitcast.
1141 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1142 // Don't touch a vector-to-scalar bitcast.
1145 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1146 RHSKnownZero, RHSKnownOne, Depth+1))
1148 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1150 case Instruction::ZExt: {
1151 // Compute the bits in the result that are not present in the input.
1152 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1154 DemandedMask.trunc(SrcBitWidth);
1155 RHSKnownZero.trunc(SrcBitWidth);
1156 RHSKnownOne.trunc(SrcBitWidth);
1157 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1158 RHSKnownZero, RHSKnownOne, Depth+1))
1160 DemandedMask.zext(BitWidth);
1161 RHSKnownZero.zext(BitWidth);
1162 RHSKnownOne.zext(BitWidth);
1163 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1164 // The top bits are known to be zero.
1165 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1168 case Instruction::SExt: {
1169 // Compute the bits in the result that are not present in the input.
1170 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1172 APInt InputDemandedBits = DemandedMask &
1173 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1175 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1176 // If any of the sign extended bits are demanded, we know that the sign
1178 if ((NewBits & DemandedMask) != 0)
1179 InputDemandedBits.set(SrcBitWidth-1);
1181 InputDemandedBits.trunc(SrcBitWidth);
1182 RHSKnownZero.trunc(SrcBitWidth);
1183 RHSKnownOne.trunc(SrcBitWidth);
1184 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1185 RHSKnownZero, RHSKnownOne, Depth+1))
1187 InputDemandedBits.zext(BitWidth);
1188 RHSKnownZero.zext(BitWidth);
1189 RHSKnownOne.zext(BitWidth);
1190 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1192 // If the sign bit of the input is known set or clear, then we know the
1193 // top bits of the result.
1195 // If the input sign bit is known zero, or if the NewBits are not demanded
1196 // convert this into a zero extension.
1197 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1198 // Convert to ZExt cast
1199 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1200 return InsertNewInstBefore(NewCast, *I);
1201 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1202 RHSKnownOne |= NewBits;
1206 case Instruction::Add: {
1207 // Figure out what the input bits are. If the top bits of the and result
1208 // are not demanded, then the add doesn't demand them from its input
1210 unsigned NLZ = DemandedMask.countLeadingZeros();
1212 // If there is a constant on the RHS, there are a variety of xformations
1214 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1215 // If null, this should be simplified elsewhere. Some of the xforms here
1216 // won't work if the RHS is zero.
1220 // If the top bit of the output is demanded, demand everything from the
1221 // input. Otherwise, we demand all the input bits except NLZ top bits.
1222 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1224 // Find information about known zero/one bits in the input.
1225 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1226 LHSKnownZero, LHSKnownOne, Depth+1))
1229 // If the RHS of the add has bits set that can't affect the input, reduce
1231 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1234 // Avoid excess work.
1235 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1238 // Turn it into OR if input bits are zero.
1239 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1241 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1243 return InsertNewInstBefore(Or, *I);
1246 // We can say something about the output known-zero and known-one bits,
1247 // depending on potential carries from the input constant and the
1248 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1249 // bits set and the RHS constant is 0x01001, then we know we have a known
1250 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1252 // To compute this, we first compute the potential carry bits. These are
1253 // the bits which may be modified. I'm not aware of a better way to do
1255 const APInt &RHSVal = RHS->getValue();
1256 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1258 // Now that we know which bits have carries, compute the known-1/0 sets.
1260 // Bits are known one if they are known zero in one operand and one in the
1261 // other, and there is no input carry.
1262 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1263 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1265 // Bits are known zero if they are known zero in both operands and there
1266 // is no input carry.
1267 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1269 // If the high-bits of this ADD are not demanded, then it does not demand
1270 // the high bits of its LHS or RHS.
1271 if (DemandedMask[BitWidth-1] == 0) {
1272 // Right fill the mask of bits for this ADD to demand the most
1273 // significant bit and all those below it.
1274 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1275 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1276 LHSKnownZero, LHSKnownOne, Depth+1) ||
1277 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1278 LHSKnownZero, LHSKnownOne, Depth+1))
1284 case Instruction::Sub:
1285 // If the high-bits of this SUB are not demanded, then it does not demand
1286 // the high bits of its LHS or RHS.
1287 if (DemandedMask[BitWidth-1] == 0) {
1288 // Right fill the mask of bits for this SUB to demand the most
1289 // significant bit and all those below it.
1290 uint32_t NLZ = DemandedMask.countLeadingZeros();
1291 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1292 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1293 LHSKnownZero, LHSKnownOne, Depth+1) ||
1294 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1295 LHSKnownZero, LHSKnownOne, Depth+1))
1298 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1299 // the known zeros and ones.
1300 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1302 case Instruction::Shl:
1303 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1304 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1305 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1306 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1307 RHSKnownZero, RHSKnownOne, Depth+1))
1309 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1310 RHSKnownZero <<= ShiftAmt;
1311 RHSKnownOne <<= ShiftAmt;
1312 // low bits known zero.
1314 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1317 case Instruction::LShr:
1318 // For a logical shift right
1319 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1320 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1322 // Unsigned shift right.
1323 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1324 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1325 RHSKnownZero, RHSKnownOne, Depth+1))
1327 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1328 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1329 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1331 // Compute the new bits that are at the top now.
1332 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1333 RHSKnownZero |= HighBits; // high bits known zero.
1337 case Instruction::AShr:
1338 // If this is an arithmetic shift right and only the low-bit is set, we can
1339 // always convert this into a logical shr, even if the shift amount is
1340 // variable. The low bit of the shift cannot be an input sign bit unless
1341 // the shift amount is >= the size of the datatype, which is undefined.
1342 if (DemandedMask == 1) {
1343 // Perform the logical shift right.
1344 Instruction *NewVal = BinaryOperator::CreateLShr(
1345 I->getOperand(0), I->getOperand(1), I->getName());
1346 return InsertNewInstBefore(NewVal, *I);
1349 // If the sign bit is the only bit demanded by this ashr, then there is no
1350 // need to do it, the shift doesn't change the high bit.
1351 if (DemandedMask.isSignBit())
1352 return I->getOperand(0);
1354 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1355 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1357 // Signed shift right.
1358 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1359 // If any of the "high bits" are demanded, we should set the sign bit as
1361 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1362 DemandedMaskIn.set(BitWidth-1);
1363 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1364 RHSKnownZero, RHSKnownOne, Depth+1))
1366 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1367 // Compute the new bits that are at the top now.
1368 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1369 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1370 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1372 // Handle the sign bits.
1373 APInt SignBit(APInt::getSignBit(BitWidth));
1374 // Adjust to where it is now in the mask.
1375 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1377 // If the input sign bit is known to be zero, or if none of the top bits
1378 // are demanded, turn this into an unsigned shift right.
1379 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1380 (HighBits & ~DemandedMask) == HighBits) {
1381 // Perform the logical shift right.
1382 Instruction *NewVal = BinaryOperator::CreateLShr(
1383 I->getOperand(0), SA, I->getName());
1384 return InsertNewInstBefore(NewVal, *I);
1385 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1386 RHSKnownOne |= HighBits;
1390 case Instruction::SRem:
1391 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1392 APInt RA = Rem->getValue().abs();
1393 if (RA.isPowerOf2()) {
1394 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1395 return I->getOperand(0);
1397 APInt LowBits = RA - 1;
1398 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1399 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1400 LHSKnownZero, LHSKnownOne, Depth+1))
1403 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1404 LHSKnownZero |= ~LowBits;
1406 KnownZero |= LHSKnownZero & DemandedMask;
1408 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1412 case Instruction::URem: {
1413 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1414 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1415 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1416 KnownZero2, KnownOne2, Depth+1) ||
1417 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1418 KnownZero2, KnownOne2, Depth+1))
1421 unsigned Leaders = KnownZero2.countLeadingOnes();
1422 Leaders = std::max(Leaders,
1423 KnownZero2.countLeadingOnes());
1424 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1427 case Instruction::Call:
1428 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1429 switch (II->getIntrinsicID()) {
1431 case Intrinsic::bswap: {
1432 // If the only bits demanded come from one byte of the bswap result,
1433 // just shift the input byte into position to eliminate the bswap.
1434 unsigned NLZ = DemandedMask.countLeadingZeros();
1435 unsigned NTZ = DemandedMask.countTrailingZeros();
1437 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1438 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1439 // have 14 leading zeros, round to 8.
1442 // If we need exactly one byte, we can do this transformation.
1443 if (BitWidth-NLZ-NTZ == 8) {
1444 unsigned ResultBit = NTZ;
1445 unsigned InputBit = BitWidth-NTZ-8;
1447 // Replace this with either a left or right shift to get the byte into
1449 Instruction *NewVal;
1450 if (InputBit > ResultBit)
1451 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1452 ConstantInt::get(I->getType(), InputBit-ResultBit));
1454 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1455 ConstantInt::get(I->getType(), ResultBit-InputBit));
1456 NewVal->takeName(I);
1457 return InsertNewInstBefore(NewVal, *I);
1460 // TODO: Could compute known zero/one bits based on the input.
1465 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1469 // If the client is only demanding bits that we know, return the known
1471 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1472 return Constant::getIntegerValue(VTy, RHSKnownOne);
1477 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1478 /// any number of elements. DemandedElts contains the set of elements that are
1479 /// actually used by the caller. This method analyzes which elements of the
1480 /// operand are undef and returns that information in UndefElts.
1482 /// If the information about demanded elements can be used to simplify the
1483 /// operation, the operation is simplified, then the resultant value is
1484 /// returned. This returns null if no change was made.
1485 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1488 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1489 APInt EltMask(APInt::getAllOnesValue(VWidth));
1490 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1492 if (isa<UndefValue>(V)) {
1493 // If the entire vector is undefined, just return this info.
1494 UndefElts = EltMask;
1496 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1497 UndefElts = EltMask;
1498 return UndefValue::get(V->getType());
1502 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1503 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1504 Constant *Undef = UndefValue::get(EltTy);
1506 std::vector<Constant*> Elts;
1507 for (unsigned i = 0; i != VWidth; ++i)
1508 if (!DemandedElts[i]) { // If not demanded, set to undef.
1509 Elts.push_back(Undef);
1511 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1512 Elts.push_back(Undef);
1514 } else { // Otherwise, defined.
1515 Elts.push_back(CP->getOperand(i));
1518 // If we changed the constant, return it.
1519 Constant *NewCP = ConstantVector::get(Elts);
1520 return NewCP != CP ? NewCP : 0;
1521 } else if (isa<ConstantAggregateZero>(V)) {
1522 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1525 // Check if this is identity. If so, return 0 since we are not simplifying
1527 if (DemandedElts == ((1ULL << VWidth) -1))
1530 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1531 Constant *Zero = Constant::getNullValue(EltTy);
1532 Constant *Undef = UndefValue::get(EltTy);
1533 std::vector<Constant*> Elts;
1534 for (unsigned i = 0; i != VWidth; ++i) {
1535 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1536 Elts.push_back(Elt);
1538 UndefElts = DemandedElts ^ EltMask;
1539 return ConstantVector::get(Elts);
1542 // Limit search depth.
1546 // If multiple users are using the root value, procede with
1547 // simplification conservatively assuming that all elements
1549 if (!V->hasOneUse()) {
1550 // Quit if we find multiple users of a non-root value though.
1551 // They'll be handled when it's their turn to be visited by
1552 // the main instcombine process.
1554 // TODO: Just compute the UndefElts information recursively.
1557 // Conservatively assume that all elements are needed.
1558 DemandedElts = EltMask;
1561 Instruction *I = dyn_cast<Instruction>(V);
1562 if (!I) return 0; // Only analyze instructions.
1564 bool MadeChange = false;
1565 APInt UndefElts2(VWidth, 0);
1567 switch (I->getOpcode()) {
1570 case Instruction::InsertElement: {
1571 // If this is a variable index, we don't know which element it overwrites.
1572 // demand exactly the same input as we produce.
1573 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1575 // Note that we can't propagate undef elt info, because we don't know
1576 // which elt is getting updated.
1577 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1578 UndefElts2, Depth+1);
1579 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1583 // If this is inserting an element that isn't demanded, remove this
1585 unsigned IdxNo = Idx->getZExtValue();
1586 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1588 return I->getOperand(0);
1591 // Otherwise, the element inserted overwrites whatever was there, so the
1592 // input demanded set is simpler than the output set.
1593 APInt DemandedElts2 = DemandedElts;
1594 DemandedElts2.clear(IdxNo);
1595 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1596 UndefElts, Depth+1);
1597 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1599 // The inserted element is defined.
1600 UndefElts.clear(IdxNo);
1603 case Instruction::ShuffleVector: {
1604 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1605 uint64_t LHSVWidth =
1606 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1607 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1608 for (unsigned i = 0; i < VWidth; i++) {
1609 if (DemandedElts[i]) {
1610 unsigned MaskVal = Shuffle->getMaskValue(i);
1611 if (MaskVal != -1u) {
1612 assert(MaskVal < LHSVWidth * 2 &&
1613 "shufflevector mask index out of range!");
1614 if (MaskVal < LHSVWidth)
1615 LeftDemanded.set(MaskVal);
1617 RightDemanded.set(MaskVal - LHSVWidth);
1622 APInt UndefElts4(LHSVWidth, 0);
1623 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1624 UndefElts4, Depth+1);
1625 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1627 APInt UndefElts3(LHSVWidth, 0);
1628 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1629 UndefElts3, Depth+1);
1630 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1632 bool NewUndefElts = false;
1633 for (unsigned i = 0; i < VWidth; i++) {
1634 unsigned MaskVal = Shuffle->getMaskValue(i);
1635 if (MaskVal == -1u) {
1637 } else if (MaskVal < LHSVWidth) {
1638 if (UndefElts4[MaskVal]) {
1639 NewUndefElts = true;
1643 if (UndefElts3[MaskVal - LHSVWidth]) {
1644 NewUndefElts = true;
1651 // Add additional discovered undefs.
1652 std::vector<Constant*> Elts;
1653 for (unsigned i = 0; i < VWidth; ++i) {
1655 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1657 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1658 Shuffle->getMaskValue(i)));
1660 I->setOperand(2, ConstantVector::get(Elts));
1665 case Instruction::BitCast: {
1666 // Vector->vector casts only.
1667 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1669 unsigned InVWidth = VTy->getNumElements();
1670 APInt InputDemandedElts(InVWidth, 0);
1673 if (VWidth == InVWidth) {
1674 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1675 // elements as are demanded of us.
1677 InputDemandedElts = DemandedElts;
1678 } else if (VWidth > InVWidth) {
1682 // If there are more elements in the result than there are in the source,
1683 // then an input element is live if any of the corresponding output
1684 // elements are live.
1685 Ratio = VWidth/InVWidth;
1686 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1687 if (DemandedElts[OutIdx])
1688 InputDemandedElts.set(OutIdx/Ratio);
1694 // If there are more elements in the source than there are in the result,
1695 // then an input element is live if the corresponding output element is
1697 Ratio = InVWidth/VWidth;
1698 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1699 if (DemandedElts[InIdx/Ratio])
1700 InputDemandedElts.set(InIdx);
1703 // div/rem demand all inputs, because they don't want divide by zero.
1704 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1705 UndefElts2, Depth+1);
1707 I->setOperand(0, TmpV);
1711 UndefElts = UndefElts2;
1712 if (VWidth > InVWidth) {
1713 llvm_unreachable("Unimp");
1714 // If there are more elements in the result than there are in the source,
1715 // then an output element is undef if the corresponding input element is
1717 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1718 if (UndefElts2[OutIdx/Ratio])
1719 UndefElts.set(OutIdx);
1720 } else if (VWidth < InVWidth) {
1721 llvm_unreachable("Unimp");
1722 // If there are more elements in the source than there are in the result,
1723 // then a result element is undef if all of the corresponding input
1724 // elements are undef.
1725 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1726 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1727 if (!UndefElts2[InIdx]) // Not undef?
1728 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1732 case Instruction::And:
1733 case Instruction::Or:
1734 case Instruction::Xor:
1735 case Instruction::Add:
1736 case Instruction::Sub:
1737 case Instruction::Mul:
1738 // div/rem demand all inputs, because they don't want divide by zero.
1739 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1740 UndefElts, Depth+1);
1741 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1742 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1743 UndefElts2, Depth+1);
1744 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1746 // Output elements are undefined if both are undefined. Consider things
1747 // like undef&0. The result is known zero, not undef.
1748 UndefElts &= UndefElts2;
1751 case Instruction::Call: {
1752 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1754 switch (II->getIntrinsicID()) {
1757 // Binary vector operations that work column-wise. A dest element is a
1758 // function of the corresponding input elements from the two inputs.
1759 case Intrinsic::x86_sse_sub_ss:
1760 case Intrinsic::x86_sse_mul_ss:
1761 case Intrinsic::x86_sse_min_ss:
1762 case Intrinsic::x86_sse_max_ss:
1763 case Intrinsic::x86_sse2_sub_sd:
1764 case Intrinsic::x86_sse2_mul_sd:
1765 case Intrinsic::x86_sse2_min_sd:
1766 case Intrinsic::x86_sse2_max_sd:
1767 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1768 UndefElts, Depth+1);
1769 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1770 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1771 UndefElts2, Depth+1);
1772 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1774 // If only the low elt is demanded and this is a scalarizable intrinsic,
1775 // scalarize it now.
1776 if (DemandedElts == 1) {
1777 switch (II->getIntrinsicID()) {
1779 case Intrinsic::x86_sse_sub_ss:
1780 case Intrinsic::x86_sse_mul_ss:
1781 case Intrinsic::x86_sse2_sub_sd:
1782 case Intrinsic::x86_sse2_mul_sd:
1783 // TODO: Lower MIN/MAX/ABS/etc
1784 Value *LHS = II->getOperand(1);
1785 Value *RHS = II->getOperand(2);
1786 // Extract the element as scalars.
1787 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1788 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1789 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1790 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1792 switch (II->getIntrinsicID()) {
1793 default: llvm_unreachable("Case stmts out of sync!");
1794 case Intrinsic::x86_sse_sub_ss:
1795 case Intrinsic::x86_sse2_sub_sd:
1796 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1797 II->getName()), *II);
1799 case Intrinsic::x86_sse_mul_ss:
1800 case Intrinsic::x86_sse2_mul_sd:
1801 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1802 II->getName()), *II);
1807 InsertElementInst::Create(
1808 UndefValue::get(II->getType()), TmpV,
1809 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1810 InsertNewInstBefore(New, *II);
1815 // Output elements are undefined if both are undefined. Consider things
1816 // like undef&0. The result is known zero, not undef.
1817 UndefElts &= UndefElts2;
1823 return MadeChange ? I : 0;
1827 /// AssociativeOpt - Perform an optimization on an associative operator. This
1828 /// function is designed to check a chain of associative operators for a
1829 /// potential to apply a certain optimization. Since the optimization may be
1830 /// applicable if the expression was reassociated, this checks the chain, then
1831 /// reassociates the expression as necessary to expose the optimization
1832 /// opportunity. This makes use of a special Functor, which must define
1833 /// 'shouldApply' and 'apply' methods.
1835 template<typename Functor>
1836 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1837 unsigned Opcode = Root.getOpcode();
1838 Value *LHS = Root.getOperand(0);
1840 // Quick check, see if the immediate LHS matches...
1841 if (F.shouldApply(LHS))
1842 return F.apply(Root);
1844 // Otherwise, if the LHS is not of the same opcode as the root, return.
1845 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1846 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1847 // Should we apply this transform to the RHS?
1848 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1850 // If not to the RHS, check to see if we should apply to the LHS...
1851 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1852 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1856 // If the functor wants to apply the optimization to the RHS of LHSI,
1857 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1859 // Now all of the instructions are in the current basic block, go ahead
1860 // and perform the reassociation.
1861 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1863 // First move the selected RHS to the LHS of the root...
1864 Root.setOperand(0, LHSI->getOperand(1));
1866 // Make what used to be the LHS of the root be the user of the root...
1867 Value *ExtraOperand = TmpLHSI->getOperand(1);
1868 if (&Root == TmpLHSI) {
1869 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1872 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1873 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1874 BasicBlock::iterator ARI = &Root; ++ARI;
1875 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1878 // Now propagate the ExtraOperand down the chain of instructions until we
1880 while (TmpLHSI != LHSI) {
1881 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1882 // Move the instruction to immediately before the chain we are
1883 // constructing to avoid breaking dominance properties.
1884 NextLHSI->moveBefore(ARI);
1887 Value *NextOp = NextLHSI->getOperand(1);
1888 NextLHSI->setOperand(1, ExtraOperand);
1890 ExtraOperand = NextOp;
1893 // Now that the instructions are reassociated, have the functor perform
1894 // the transformation...
1895 return F.apply(Root);
1898 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1905 // AddRHS - Implements: X + X --> X << 1
1908 explicit AddRHS(Value *rhs) : RHS(rhs) {}
1909 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1910 Instruction *apply(BinaryOperator &Add) const {
1911 return BinaryOperator::CreateShl(Add.getOperand(0),
1912 ConstantInt::get(Add.getType(), 1));
1916 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1918 struct AddMaskingAnd {
1920 explicit AddMaskingAnd(Constant *c) : C2(c) {}
1921 bool shouldApply(Value *LHS) const {
1923 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1924 ConstantExpr::getAnd(C1, C2)->isNullValue();
1926 Instruction *apply(BinaryOperator &Add) const {
1927 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1933 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1935 if (CastInst *CI = dyn_cast<CastInst>(&I))
1936 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
1938 // Figure out if the constant is the left or the right argument.
1939 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1940 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1942 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1944 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1945 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1948 Value *Op0 = SO, *Op1 = ConstOperand;
1950 std::swap(Op0, Op1);
1952 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1953 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
1954 SO->getName()+".op");
1955 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
1956 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1957 SO->getName()+".cmp");
1958 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
1959 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1960 SO->getName()+".cmp");
1961 llvm_unreachable("Unknown binary instruction type!");
1964 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1965 // constant as the other operand, try to fold the binary operator into the
1966 // select arguments. This also works for Cast instructions, which obviously do
1967 // not have a second operand.
1968 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1970 // Don't modify shared select instructions
1971 if (!SI->hasOneUse()) return 0;
1972 Value *TV = SI->getOperand(1);
1973 Value *FV = SI->getOperand(2);
1975 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1976 // Bool selects with constant operands can be folded to logical ops.
1977 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
1979 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1980 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1982 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1989 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
1990 /// has a PHI node as operand #0, see if we can fold the instruction into the
1991 /// PHI (which is only possible if all operands to the PHI are constants).
1993 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
1994 /// that would normally be unprofitable because they strongly encourage jump
1996 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
1997 bool AllowAggressive) {
1998 AllowAggressive = false;
1999 PHINode *PN = cast<PHINode>(I.getOperand(0));
2000 unsigned NumPHIValues = PN->getNumIncomingValues();
2001 if (NumPHIValues == 0 ||
2002 // We normally only transform phis with a single use, unless we're trying
2003 // hard to make jump threading happen.
2004 (!PN->hasOneUse() && !AllowAggressive))
2008 // Check to see if all of the operands of the PHI are simple constants
2009 // (constantint/constantfp/undef). If there is one non-constant value,
2010 // remember the BB it is in. If there is more than one or if *it* is a PHI,
2011 // bail out. We don't do arbitrary constant expressions here because moving
2012 // their computation can be expensive without a cost model.
2013 BasicBlock *NonConstBB = 0;
2014 for (unsigned i = 0; i != NumPHIValues; ++i)
2015 if (!isa<Constant>(PN->getIncomingValue(i)) ||
2016 isa<ConstantExpr>(PN->getIncomingValue(i))) {
2017 if (NonConstBB) return 0; // More than one non-const value.
2018 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2019 NonConstBB = PN->getIncomingBlock(i);
2021 // If the incoming non-constant value is in I's block, we have an infinite
2023 if (NonConstBB == I.getParent())
2027 // If there is exactly one non-constant value, we can insert a copy of the
2028 // operation in that block. However, if this is a critical edge, we would be
2029 // inserting the computation one some other paths (e.g. inside a loop). Only
2030 // do this if the pred block is unconditionally branching into the phi block.
2031 if (NonConstBB != 0 && !AllowAggressive) {
2032 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2033 if (!BI || !BI->isUnconditional()) return 0;
2036 // Okay, we can do the transformation: create the new PHI node.
2037 PHINode *NewPN = PHINode::Create(I.getType(), "");
2038 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2039 InsertNewInstBefore(NewPN, *PN);
2040 NewPN->takeName(PN);
2042 // Next, add all of the operands to the PHI.
2043 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
2044 // We only currently try to fold the condition of a select when it is a phi,
2045 // not the true/false values.
2046 Value *TrueV = SI->getTrueValue();
2047 Value *FalseV = SI->getFalseValue();
2048 BasicBlock *PhiTransBB = PN->getParent();
2049 for (unsigned i = 0; i != NumPHIValues; ++i) {
2050 BasicBlock *ThisBB = PN->getIncomingBlock(i);
2051 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
2052 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
2054 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2055 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
2057 assert(PN->getIncomingBlock(i) == NonConstBB);
2058 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
2060 "phitmp", NonConstBB->getTerminator());
2061 Worklist.Add(cast<Instruction>(InV));
2063 NewPN->addIncoming(InV, ThisBB);
2065 } else if (I.getNumOperands() == 2) {
2066 Constant *C = cast<Constant>(I.getOperand(1));
2067 for (unsigned i = 0; i != NumPHIValues; ++i) {
2069 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2070 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2071 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2073 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2075 assert(PN->getIncomingBlock(i) == NonConstBB);
2076 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2077 InV = BinaryOperator::Create(BO->getOpcode(),
2078 PN->getIncomingValue(i), C, "phitmp",
2079 NonConstBB->getTerminator());
2080 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2081 InV = CmpInst::Create(CI->getOpcode(),
2083 PN->getIncomingValue(i), C, "phitmp",
2084 NonConstBB->getTerminator());
2086 llvm_unreachable("Unknown binop!");
2088 Worklist.Add(cast<Instruction>(InV));
2090 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2093 CastInst *CI = cast<CastInst>(&I);
2094 const Type *RetTy = CI->getType();
2095 for (unsigned i = 0; i != NumPHIValues; ++i) {
2097 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2098 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2100 assert(PN->getIncomingBlock(i) == NonConstBB);
2101 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2102 I.getType(), "phitmp",
2103 NonConstBB->getTerminator());
2104 Worklist.Add(cast<Instruction>(InV));
2106 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2109 return ReplaceInstUsesWith(I, NewPN);
2113 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2114 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2115 /// This basically requires proving that the add in the original type would not
2116 /// overflow to change the sign bit or have a carry out.
2117 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2118 // There are different heuristics we can use for this. Here are some simple
2121 // Add has the property that adding any two 2's complement numbers can only
2122 // have one carry bit which can change a sign. As such, if LHS and RHS each
2123 // have at least two sign bits, we know that the addition of the two values will
2124 // sign extend fine.
2125 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2129 // If one of the operands only has one non-zero bit, and if the other operand
2130 // has a known-zero bit in a more significant place than it (not including the
2131 // sign bit) the ripple may go up to and fill the zero, but won't change the
2132 // sign. For example, (X & ~4) + 1.
2140 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2141 bool Changed = SimplifyCommutative(I);
2142 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2144 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2145 // X + undef -> undef
2146 if (isa<UndefValue>(RHS))
2147 return ReplaceInstUsesWith(I, RHS);
2150 if (RHSC->isNullValue())
2151 return ReplaceInstUsesWith(I, LHS);
2153 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2154 // X + (signbit) --> X ^ signbit
2155 const APInt& Val = CI->getValue();
2156 uint32_t BitWidth = Val.getBitWidth();
2157 if (Val == APInt::getSignBit(BitWidth))
2158 return BinaryOperator::CreateXor(LHS, RHS);
2160 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2161 // (X & 254)+1 -> (X&254)|1
2162 if (SimplifyDemandedInstructionBits(I))
2165 // zext(bool) + C -> bool ? C + 1 : C
2166 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2167 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2168 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2171 if (isa<PHINode>(LHS))
2172 if (Instruction *NV = FoldOpIntoPhi(I))
2175 ConstantInt *XorRHS = 0;
2177 if (isa<ConstantInt>(RHSC) &&
2178 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2179 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2180 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2182 uint32_t Size = TySizeBits / 2;
2183 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2184 APInt CFF80Val(-C0080Val);
2186 if (TySizeBits > Size) {
2187 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2188 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2189 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2190 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2191 // This is a sign extend if the top bits are known zero.
2192 if (!MaskedValueIsZero(XorLHS,
2193 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2194 Size = 0; // Not a sign ext, but can't be any others either.
2199 C0080Val = APIntOps::lshr(C0080Val, Size);
2200 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2201 } while (Size >= 1);
2203 // FIXME: This shouldn't be necessary. When the backends can handle types
2204 // with funny bit widths then this switch statement should be removed. It
2205 // is just here to get the size of the "middle" type back up to something
2206 // that the back ends can handle.
2207 const Type *MiddleType = 0;
2210 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2211 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2212 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2215 Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext");
2216 return new SExtInst(NewTrunc, I.getType(), I.getName());
2221 if (I.getType() == Type::getInt1Ty(*Context))
2222 return BinaryOperator::CreateXor(LHS, RHS);
2225 if (I.getType()->isInteger()) {
2226 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2229 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2230 if (RHSI->getOpcode() == Instruction::Sub)
2231 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2232 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2234 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2235 if (LHSI->getOpcode() == Instruction::Sub)
2236 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2237 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2242 // -A + -B --> -(A + B)
2243 if (Value *LHSV = dyn_castNegVal(LHS)) {
2244 if (LHS->getType()->isIntOrIntVector()) {
2245 if (Value *RHSV = dyn_castNegVal(RHS)) {
2246 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
2247 return BinaryOperator::CreateNeg(NewAdd);
2251 return BinaryOperator::CreateSub(RHS, LHSV);
2255 if (!isa<Constant>(RHS))
2256 if (Value *V = dyn_castNegVal(RHS))
2257 return BinaryOperator::CreateSub(LHS, V);
2261 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2262 if (X == RHS) // X*C + X --> X * (C+1)
2263 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2265 // X*C1 + X*C2 --> X * (C1+C2)
2267 if (X == dyn_castFoldableMul(RHS, C1))
2268 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2271 // X + X*C --> X * (C+1)
2272 if (dyn_castFoldableMul(RHS, C2) == LHS)
2273 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2275 // X + ~X --> -1 since ~X = -X-1
2276 if (dyn_castNotVal(LHS) == RHS ||
2277 dyn_castNotVal(RHS) == LHS)
2278 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2281 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2282 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2283 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2286 // A+B --> A|B iff A and B have no bits set in common.
2287 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2288 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2289 APInt LHSKnownOne(IT->getBitWidth(), 0);
2290 APInt LHSKnownZero(IT->getBitWidth(), 0);
2291 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2292 if (LHSKnownZero != 0) {
2293 APInt RHSKnownOne(IT->getBitWidth(), 0);
2294 APInt RHSKnownZero(IT->getBitWidth(), 0);
2295 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2297 // No bits in common -> bitwise or.
2298 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2299 return BinaryOperator::CreateOr(LHS, RHS);
2303 // W*X + Y*Z --> W * (X+Z) iff W == Y
2304 if (I.getType()->isIntOrIntVector()) {
2305 Value *W, *X, *Y, *Z;
2306 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2307 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2311 } else if (Y == X) {
2313 } else if (X == Z) {
2320 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
2321 return BinaryOperator::CreateMul(W, NewAdd);
2326 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2328 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2329 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2331 // (X & FF00) + xx00 -> (X+xx00) & FF00
2332 if (LHS->hasOneUse() &&
2333 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2334 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2335 if (Anded == CRHS) {
2336 // See if all bits from the first bit set in the Add RHS up are included
2337 // in the mask. First, get the rightmost bit.
2338 const APInt& AddRHSV = CRHS->getValue();
2340 // Form a mask of all bits from the lowest bit added through the top.
2341 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2343 // See if the and mask includes all of these bits.
2344 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2346 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2347 // Okay, the xform is safe. Insert the new add pronto.
2348 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
2349 return BinaryOperator::CreateAnd(NewAdd, C2);
2354 // Try to fold constant add into select arguments.
2355 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2356 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2360 // add (select X 0 (sub n A)) A --> select X A n
2362 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2365 SI = dyn_cast<SelectInst>(RHS);
2368 if (SI && SI->hasOneUse()) {
2369 Value *TV = SI->getTrueValue();
2370 Value *FV = SI->getFalseValue();
2373 // Can we fold the add into the argument of the select?
2374 // We check both true and false select arguments for a matching subtract.
2375 if (match(FV, m_Zero()) &&
2376 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2377 // Fold the add into the true select value.
2378 return SelectInst::Create(SI->getCondition(), N, A);
2379 if (match(TV, m_Zero()) &&
2380 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2381 // Fold the add into the false select value.
2382 return SelectInst::Create(SI->getCondition(), A, N);
2386 // Check for (add (sext x), y), see if we can merge this into an
2387 // integer add followed by a sext.
2388 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2389 // (add (sext x), cst) --> (sext (add x, cst'))
2390 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2392 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2393 if (LHSConv->hasOneUse() &&
2394 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2395 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2396 // Insert the new, smaller add.
2397 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2399 return new SExtInst(NewAdd, I.getType());
2403 // (add (sext x), (sext y)) --> (sext (add int x, y))
2404 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2405 // Only do this if x/y have the same type, if at last one of them has a
2406 // single use (so we don't increase the number of sexts), and if the
2407 // integer add will not overflow.
2408 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2409 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2410 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2411 RHSConv->getOperand(0))) {
2412 // Insert the new integer add.
2413 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2414 RHSConv->getOperand(0), "addconv");
2415 return new SExtInst(NewAdd, I.getType());
2420 return Changed ? &I : 0;
2423 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2424 bool Changed = SimplifyCommutative(I);
2425 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2427 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2429 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2430 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2431 (I.getType())->getValueAPF()))
2432 return ReplaceInstUsesWith(I, LHS);
2435 if (isa<PHINode>(LHS))
2436 if (Instruction *NV = FoldOpIntoPhi(I))
2441 // -A + -B --> -(A + B)
2442 if (Value *LHSV = dyn_castFNegVal(LHS))
2443 return BinaryOperator::CreateFSub(RHS, LHSV);
2446 if (!isa<Constant>(RHS))
2447 if (Value *V = dyn_castFNegVal(RHS))
2448 return BinaryOperator::CreateFSub(LHS, V);
2450 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2451 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2452 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2453 return ReplaceInstUsesWith(I, LHS);
2455 // Check for (add double (sitofp x), y), see if we can merge this into an
2456 // integer add followed by a promotion.
2457 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2458 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2459 // ... if the constant fits in the integer value. This is useful for things
2460 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2461 // requires a constant pool load, and generally allows the add to be better
2463 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2465 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2466 if (LHSConv->hasOneUse() &&
2467 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2468 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2469 // Insert the new integer add.
2470 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2472 return new SIToFPInst(NewAdd, I.getType());
2476 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2477 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2478 // Only do this if x/y have the same type, if at last one of them has a
2479 // single use (so we don't increase the number of int->fp conversions),
2480 // and if the integer add will not overflow.
2481 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2482 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2483 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2484 RHSConv->getOperand(0))) {
2485 // Insert the new integer add.
2486 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2487 RHSConv->getOperand(0), "addconv");
2488 return new SIToFPInst(NewAdd, I.getType());
2493 return Changed ? &I : 0;
2496 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2497 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2499 if (Op0 == Op1) // sub X, X -> 0
2500 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2502 // If this is a 'B = x-(-A)', change to B = x+A...
2503 if (Value *V = dyn_castNegVal(Op1))
2504 return BinaryOperator::CreateAdd(Op0, V);
2506 if (isa<UndefValue>(Op0))
2507 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2508 if (isa<UndefValue>(Op1))
2509 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2511 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2512 // Replace (-1 - A) with (~A)...
2513 if (C->isAllOnesValue())
2514 return BinaryOperator::CreateNot(Op1);
2516 // C - ~X == X + (1+C)
2518 if (match(Op1, m_Not(m_Value(X))))
2519 return BinaryOperator::CreateAdd(X, AddOne(C));
2521 // -(X >>u 31) -> (X >>s 31)
2522 // -(X >>s 31) -> (X >>u 31)
2524 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2525 if (SI->getOpcode() == Instruction::LShr) {
2526 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2527 // Check to see if we are shifting out everything but the sign bit.
2528 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2529 SI->getType()->getPrimitiveSizeInBits()-1) {
2530 // Ok, the transformation is safe. Insert AShr.
2531 return BinaryOperator::Create(Instruction::AShr,
2532 SI->getOperand(0), CU, SI->getName());
2536 else if (SI->getOpcode() == Instruction::AShr) {
2537 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2538 // Check to see if we are shifting out everything but the sign bit.
2539 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2540 SI->getType()->getPrimitiveSizeInBits()-1) {
2541 // Ok, the transformation is safe. Insert LShr.
2542 return BinaryOperator::CreateLShr(
2543 SI->getOperand(0), CU, SI->getName());
2550 // Try to fold constant sub into select arguments.
2551 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2552 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2555 // C - zext(bool) -> bool ? C - 1 : C
2556 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2557 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2558 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2561 if (I.getType() == Type::getInt1Ty(*Context))
2562 return BinaryOperator::CreateXor(Op0, Op1);
2564 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2565 if (Op1I->getOpcode() == Instruction::Add) {
2566 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2567 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2569 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2570 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2572 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2573 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2574 // C1-(X+C2) --> (C1-C2)-X
2575 return BinaryOperator::CreateSub(
2576 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2580 if (Op1I->hasOneUse()) {
2581 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2582 // is not used by anyone else...
2584 if (Op1I->getOpcode() == Instruction::Sub) {
2585 // Swap the two operands of the subexpr...
2586 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2587 Op1I->setOperand(0, IIOp1);
2588 Op1I->setOperand(1, IIOp0);
2590 // Create the new top level add instruction...
2591 return BinaryOperator::CreateAdd(Op0, Op1);
2594 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2596 if (Op1I->getOpcode() == Instruction::And &&
2597 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2598 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2600 Value *NewNot = Builder->CreateNot(OtherOp, "B.not");
2601 return BinaryOperator::CreateAnd(Op0, NewNot);
2604 // 0 - (X sdiv C) -> (X sdiv -C)
2605 if (Op1I->getOpcode() == Instruction::SDiv)
2606 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2608 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2609 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2610 ConstantExpr::getNeg(DivRHS));
2612 // X - X*C --> X * (1-C)
2613 ConstantInt *C2 = 0;
2614 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2616 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2618 return BinaryOperator::CreateMul(Op0, CP1);
2623 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2624 if (Op0I->getOpcode() == Instruction::Add) {
2625 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2626 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2627 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2628 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2629 } else if (Op0I->getOpcode() == Instruction::Sub) {
2630 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2631 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2637 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2638 if (X == Op1) // X*C - X --> X * (C-1)
2639 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2641 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2642 if (X == dyn_castFoldableMul(Op1, C2))
2643 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2648 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2649 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2651 // If this is a 'B = x-(-A)', change to B = x+A...
2652 if (Value *V = dyn_castFNegVal(Op1))
2653 return BinaryOperator::CreateFAdd(Op0, V);
2655 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2656 if (Op1I->getOpcode() == Instruction::FAdd) {
2657 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2658 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2660 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2661 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2669 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2670 /// comparison only checks the sign bit. If it only checks the sign bit, set
2671 /// TrueIfSigned if the result of the comparison is true when the input value is
2673 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2674 bool &TrueIfSigned) {
2676 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2677 TrueIfSigned = true;
2678 return RHS->isZero();
2679 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2680 TrueIfSigned = true;
2681 return RHS->isAllOnesValue();
2682 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2683 TrueIfSigned = false;
2684 return RHS->isAllOnesValue();
2685 case ICmpInst::ICMP_UGT:
2686 // True if LHS u> RHS and RHS == high-bit-mask - 1
2687 TrueIfSigned = true;
2688 return RHS->getValue() ==
2689 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2690 case ICmpInst::ICMP_UGE:
2691 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2692 TrueIfSigned = true;
2693 return RHS->getValue().isSignBit();
2699 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2700 bool Changed = SimplifyCommutative(I);
2701 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2703 if (isa<UndefValue>(Op1)) // undef * X -> 0
2704 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2706 // Simplify mul instructions with a constant RHS.
2707 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2708 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1C)) {
2710 // ((X << C1)*C2) == (X * (C2 << C1))
2711 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2712 if (SI->getOpcode() == Instruction::Shl)
2713 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2714 return BinaryOperator::CreateMul(SI->getOperand(0),
2715 ConstantExpr::getShl(CI, ShOp));
2718 return ReplaceInstUsesWith(I, Op1C); // X * 0 == 0
2719 if (CI->equalsInt(1)) // X * 1 == X
2720 return ReplaceInstUsesWith(I, Op0);
2721 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2722 return BinaryOperator::CreateNeg(Op0, I.getName());
2724 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2725 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2726 return BinaryOperator::CreateShl(Op0,
2727 ConstantInt::get(Op0->getType(), Val.logBase2()));
2729 } else if (isa<VectorType>(Op1C->getType())) {
2730 if (Op1C->isNullValue())
2731 return ReplaceInstUsesWith(I, Op1C);
2733 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
2734 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2735 return BinaryOperator::CreateNeg(Op0, I.getName());
2737 // As above, vector X*splat(1.0) -> X in all defined cases.
2738 if (Constant *Splat = Op1V->getSplatValue()) {
2739 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2740 if (CI->equalsInt(1))
2741 return ReplaceInstUsesWith(I, Op0);
2746 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2747 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2748 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1C)) {
2749 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2750 Value *Add = Builder->CreateMul(Op0I->getOperand(0), Op1C, "tmp");
2751 Value *C1C2 = Builder->CreateMul(Op1C, Op0I->getOperand(1));
2752 return BinaryOperator::CreateAdd(Add, C1C2);
2756 // Try to fold constant mul into select arguments.
2757 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2758 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2761 if (isa<PHINode>(Op0))
2762 if (Instruction *NV = FoldOpIntoPhi(I))
2766 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2767 if (Value *Op1v = dyn_castNegVal(Op1))
2768 return BinaryOperator::CreateMul(Op0v, Op1v);
2770 // (X / Y) * Y = X - (X % Y)
2771 // (X / Y) * -Y = (X % Y) - X
2774 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2776 (BO->getOpcode() != Instruction::UDiv &&
2777 BO->getOpcode() != Instruction::SDiv)) {
2779 BO = dyn_cast<BinaryOperator>(Op1);
2781 Value *Neg = dyn_castNegVal(Op1C);
2782 if (BO && BO->hasOneUse() &&
2783 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
2784 (BO->getOpcode() == Instruction::UDiv ||
2785 BO->getOpcode() == Instruction::SDiv)) {
2786 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2788 // If the division is exact, X % Y is zero.
2789 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
2790 if (SDiv->isExact()) {
2792 return ReplaceInstUsesWith(I, Op0BO);
2793 return BinaryOperator::CreateNeg(Op0BO);
2797 if (BO->getOpcode() == Instruction::UDiv)
2798 Rem = Builder->CreateURem(Op0BO, Op1BO);
2800 Rem = Builder->CreateSRem(Op0BO, Op1BO);
2804 return BinaryOperator::CreateSub(Op0BO, Rem);
2805 return BinaryOperator::CreateSub(Rem, Op0BO);
2809 /// i1 mul -> i1 and.
2810 if (I.getType() == Type::getInt1Ty(*Context))
2811 return BinaryOperator::CreateAnd(Op0, Op1);
2813 // X*(1 << Y) --> X << Y
2814 // (1 << Y)*X --> X << Y
2817 if (match(Op0, m_Shl(m_One(), m_Value(Y))))
2818 return BinaryOperator::CreateShl(Op1, Y);
2819 if (match(Op1, m_Shl(m_One(), m_Value(Y))))
2820 return BinaryOperator::CreateShl(Op0, Y);
2823 // If one of the operands of the multiply is a cast from a boolean value, then
2824 // we know the bool is either zero or one, so this is a 'masking' multiply.
2825 // X * Y (where Y is 0 or 1) -> X & (0-Y)
2826 if (!isa<VectorType>(I.getType())) {
2827 // -2 is "-1 << 1" so it is all bits set except the low one.
2828 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
2830 Value *BoolCast = 0, *OtherOp = 0;
2831 if (MaskedValueIsZero(Op0, Negative2))
2832 BoolCast = Op0, OtherOp = Op1;
2833 else if (MaskedValueIsZero(Op1, Negative2))
2834 BoolCast = Op1, OtherOp = Op0;
2837 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
2839 return BinaryOperator::CreateAnd(V, OtherOp);
2843 return Changed ? &I : 0;
2846 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2847 bool Changed = SimplifyCommutative(I);
2848 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2850 // Simplify mul instructions with a constant RHS...
2851 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2852 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1C)) {
2853 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2854 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2855 if (Op1F->isExactlyValue(1.0))
2856 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2857 } else if (isa<VectorType>(Op1C->getType())) {
2858 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1C)) {
2859 // As above, vector X*splat(1.0) -> X in all defined cases.
2860 if (Constant *Splat = Op1V->getSplatValue()) {
2861 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2862 if (F->isExactlyValue(1.0))
2863 return ReplaceInstUsesWith(I, Op0);
2868 // Try to fold constant mul into select arguments.
2869 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2870 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2873 if (isa<PHINode>(Op0))
2874 if (Instruction *NV = FoldOpIntoPhi(I))
2878 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
2879 if (Value *Op1v = dyn_castFNegVal(Op1))
2880 return BinaryOperator::CreateFMul(Op0v, Op1v);
2882 return Changed ? &I : 0;
2885 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2887 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2888 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2890 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2891 int NonNullOperand = -1;
2892 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2893 if (ST->isNullValue())
2895 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2896 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2897 if (ST->isNullValue())
2900 if (NonNullOperand == -1)
2903 Value *SelectCond = SI->getOperand(0);
2905 // Change the div/rem to use 'Y' instead of the select.
2906 I.setOperand(1, SI->getOperand(NonNullOperand));
2908 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2909 // problem. However, the select, or the condition of the select may have
2910 // multiple uses. Based on our knowledge that the operand must be non-zero,
2911 // propagate the known value for the select into other uses of it, and
2912 // propagate a known value of the condition into its other users.
2914 // If the select and condition only have a single use, don't bother with this,
2916 if (SI->use_empty() && SelectCond->hasOneUse())
2919 // Scan the current block backward, looking for other uses of SI.
2920 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2922 while (BBI != BBFront) {
2924 // If we found a call to a function, we can't assume it will return, so
2925 // information from below it cannot be propagated above it.
2926 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2929 // Replace uses of the select or its condition with the known values.
2930 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2933 *I = SI->getOperand(NonNullOperand);
2935 } else if (*I == SelectCond) {
2936 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2937 ConstantInt::getFalse(*Context);
2942 // If we past the instruction, quit looking for it.
2945 if (&*BBI == SelectCond)
2948 // If we ran out of things to eliminate, break out of the loop.
2949 if (SelectCond == 0 && SI == 0)
2957 /// This function implements the transforms on div instructions that work
2958 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2959 /// used by the visitors to those instructions.
2960 /// @brief Transforms common to all three div instructions
2961 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2962 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2964 // undef / X -> 0 for integer.
2965 // undef / X -> undef for FP (the undef could be a snan).
2966 if (isa<UndefValue>(Op0)) {
2967 if (Op0->getType()->isFPOrFPVector())
2968 return ReplaceInstUsesWith(I, Op0);
2969 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2972 // X / undef -> undef
2973 if (isa<UndefValue>(Op1))
2974 return ReplaceInstUsesWith(I, Op1);
2979 /// This function implements the transforms common to both integer division
2980 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2981 /// division instructions.
2982 /// @brief Common integer divide transforms
2983 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2984 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2986 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2988 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2989 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
2990 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2991 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2994 Constant *CI = ConstantInt::get(I.getType(), 1);
2995 return ReplaceInstUsesWith(I, CI);
2998 if (Instruction *Common = commonDivTransforms(I))
3001 // Handle cases involving: [su]div X, (select Cond, Y, Z)
3002 // This does not apply for fdiv.
3003 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3006 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3008 if (RHS->equalsInt(1))
3009 return ReplaceInstUsesWith(I, Op0);
3011 // (X / C1) / C2 -> X / (C1*C2)
3012 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3013 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3014 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3015 if (MultiplyOverflows(RHS, LHSRHS,
3016 I.getOpcode()==Instruction::SDiv))
3017 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3019 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3020 ConstantExpr::getMul(RHS, LHSRHS));
3023 if (!RHS->isZero()) { // avoid X udiv 0
3024 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3025 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3027 if (isa<PHINode>(Op0))
3028 if (Instruction *NV = FoldOpIntoPhi(I))
3033 // 0 / X == 0, we don't need to preserve faults!
3034 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3035 if (LHS->equalsInt(0))
3036 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3038 // It can't be division by zero, hence it must be division by one.
3039 if (I.getType() == Type::getInt1Ty(*Context))
3040 return ReplaceInstUsesWith(I, Op0);
3042 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3043 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3046 return ReplaceInstUsesWith(I, Op0);
3052 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3053 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3055 // Handle the integer div common cases
3056 if (Instruction *Common = commonIDivTransforms(I))
3059 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3060 // X udiv C^2 -> X >> C
3061 // Check to see if this is an unsigned division with an exact power of 2,
3062 // if so, convert to a right shift.
3063 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3064 return BinaryOperator::CreateLShr(Op0,
3065 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3067 // X udiv C, where C >= signbit
3068 if (C->getValue().isNegative()) {
3069 Value *IC = Builder->CreateICmpULT( Op0, C);
3070 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3071 ConstantInt::get(I.getType(), 1));
3075 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3076 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3077 if (RHSI->getOpcode() == Instruction::Shl &&
3078 isa<ConstantInt>(RHSI->getOperand(0))) {
3079 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3080 if (C1.isPowerOf2()) {
3081 Value *N = RHSI->getOperand(1);
3082 const Type *NTy = N->getType();
3083 if (uint32_t C2 = C1.logBase2())
3084 N = Builder->CreateAdd(N, ConstantInt::get(NTy, C2), "tmp");
3085 return BinaryOperator::CreateLShr(Op0, N);
3090 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3091 // where C1&C2 are powers of two.
3092 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3093 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3094 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3095 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3096 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3097 // Compute the shift amounts
3098 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3099 // Construct the "on true" case of the select
3100 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3101 Value *TSI = Builder->CreateLShr(Op0, TC, SI->getName()+".t");
3103 // Construct the "on false" case of the select
3104 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3105 Value *FSI = Builder->CreateLShr(Op0, FC, SI->getName()+".f");
3107 // construct the select instruction and return it.
3108 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3114 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3115 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3117 // Handle the integer div common cases
3118 if (Instruction *Common = commonIDivTransforms(I))
3121 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3123 if (RHS->isAllOnesValue())
3124 return BinaryOperator::CreateNeg(Op0);
3126 // sdiv X, C --> ashr X, log2(C)
3127 if (cast<SDivOperator>(&I)->isExact() &&
3128 RHS->getValue().isNonNegative() &&
3129 RHS->getValue().isPowerOf2()) {
3130 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3131 RHS->getValue().exactLogBase2());
3132 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3135 // -X/C --> X/-C provided the negation doesn't overflow.
3136 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3137 if (isa<Constant>(Sub->getOperand(0)) &&
3138 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3139 Sub->hasNoSignedWrap())
3140 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3141 ConstantExpr::getNeg(RHS));
3144 // If the sign bits of both operands are zero (i.e. we can prove they are
3145 // unsigned inputs), turn this into a udiv.
3146 if (I.getType()->isInteger()) {
3147 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3148 if (MaskedValueIsZero(Op0, Mask)) {
3149 if (MaskedValueIsZero(Op1, Mask)) {
3150 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3151 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3153 ConstantInt *ShiftedInt;
3154 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3155 ShiftedInt->getValue().isPowerOf2()) {
3156 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3157 // Safe because the only negative value (1 << Y) can take on is
3158 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3159 // the sign bit set.
3160 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3168 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3169 return commonDivTransforms(I);
3172 /// This function implements the transforms on rem instructions that work
3173 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3174 /// is used by the visitors to those instructions.
3175 /// @brief Transforms common to all three rem instructions
3176 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3177 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3179 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3180 if (I.getType()->isFPOrFPVector())
3181 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3182 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3184 if (isa<UndefValue>(Op1))
3185 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3187 // Handle cases involving: rem X, (select Cond, Y, Z)
3188 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3194 /// This function implements the transforms common to both integer remainder
3195 /// instructions (urem and srem). It is called by the visitors to those integer
3196 /// remainder instructions.
3197 /// @brief Common integer remainder transforms
3198 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3199 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3201 if (Instruction *common = commonRemTransforms(I))
3204 // 0 % X == 0 for integer, we don't need to preserve faults!
3205 if (Constant *LHS = dyn_cast<Constant>(Op0))
3206 if (LHS->isNullValue())
3207 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3209 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3210 // X % 0 == undef, we don't need to preserve faults!
3211 if (RHS->equalsInt(0))
3212 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3214 if (RHS->equalsInt(1)) // X % 1 == 0
3215 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3217 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3218 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3219 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3221 } else if (isa<PHINode>(Op0I)) {
3222 if (Instruction *NV = FoldOpIntoPhi(I))
3226 // See if we can fold away this rem instruction.
3227 if (SimplifyDemandedInstructionBits(I))
3235 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3236 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3238 if (Instruction *common = commonIRemTransforms(I))
3241 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3242 // X urem C^2 -> X and C
3243 // Check to see if this is an unsigned remainder with an exact power of 2,
3244 // if so, convert to a bitwise and.
3245 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3246 if (C->getValue().isPowerOf2())
3247 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3250 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3251 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3252 if (RHSI->getOpcode() == Instruction::Shl &&
3253 isa<ConstantInt>(RHSI->getOperand(0))) {
3254 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3255 Constant *N1 = Constant::getAllOnesValue(I.getType());
3256 Value *Add = Builder->CreateAdd(RHSI, N1, "tmp");
3257 return BinaryOperator::CreateAnd(Op0, Add);
3262 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3263 // where C1&C2 are powers of two.
3264 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3265 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3266 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3267 // STO == 0 and SFO == 0 handled above.
3268 if ((STO->getValue().isPowerOf2()) &&
3269 (SFO->getValue().isPowerOf2())) {
3270 Value *TrueAnd = Builder->CreateAnd(Op0, SubOne(STO),
3271 SI->getName()+".t");
3272 Value *FalseAnd = Builder->CreateAnd(Op0, SubOne(SFO),
3273 SI->getName()+".f");
3274 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3282 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3283 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3285 // Handle the integer rem common cases
3286 if (Instruction *Common = commonIRemTransforms(I))
3289 if (Value *RHSNeg = dyn_castNegVal(Op1))
3290 if (!isa<Constant>(RHSNeg) ||
3291 (isa<ConstantInt>(RHSNeg) &&
3292 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3294 Worklist.AddValue(I.getOperand(1));
3295 I.setOperand(1, RHSNeg);
3299 // If the sign bits of both operands are zero (i.e. we can prove they are
3300 // unsigned inputs), turn this into a urem.
3301 if (I.getType()->isInteger()) {
3302 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3303 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3304 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3305 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3309 // If it's a constant vector, flip any negative values positive.
3310 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3311 unsigned VWidth = RHSV->getNumOperands();
3313 bool hasNegative = false;
3314 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3315 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3316 if (RHS->getValue().isNegative())
3320 std::vector<Constant *> Elts(VWidth);
3321 for (unsigned i = 0; i != VWidth; ++i) {
3322 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3323 if (RHS->getValue().isNegative())
3324 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3330 Constant *NewRHSV = ConstantVector::get(Elts);
3331 if (NewRHSV != RHSV) {
3332 Worklist.AddValue(I.getOperand(1));
3333 I.setOperand(1, NewRHSV);
3342 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3343 return commonRemTransforms(I);
3346 // isOneBitSet - Return true if there is exactly one bit set in the specified
3348 static bool isOneBitSet(const ConstantInt *CI) {
3349 return CI->getValue().isPowerOf2();
3352 // isHighOnes - Return true if the constant is of the form 1+0+.
3353 // This is the same as lowones(~X).
3354 static bool isHighOnes(const ConstantInt *CI) {
3355 return (~CI->getValue() + 1).isPowerOf2();
3358 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3359 /// are carefully arranged to allow folding of expressions such as:
3361 /// (A < B) | (A > B) --> (A != B)
3363 /// Note that this is only valid if the first and second predicates have the
3364 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3366 /// Three bits are used to represent the condition, as follows:
3371 /// <=> Value Definition
3372 /// 000 0 Always false
3379 /// 111 7 Always true
3381 static unsigned getICmpCode(const ICmpInst *ICI) {
3382 switch (ICI->getPredicate()) {
3384 case ICmpInst::ICMP_UGT: return 1; // 001
3385 case ICmpInst::ICMP_SGT: return 1; // 001
3386 case ICmpInst::ICMP_EQ: return 2; // 010
3387 case ICmpInst::ICMP_UGE: return 3; // 011
3388 case ICmpInst::ICMP_SGE: return 3; // 011
3389 case ICmpInst::ICMP_ULT: return 4; // 100
3390 case ICmpInst::ICMP_SLT: return 4; // 100
3391 case ICmpInst::ICMP_NE: return 5; // 101
3392 case ICmpInst::ICMP_ULE: return 6; // 110
3393 case ICmpInst::ICMP_SLE: return 6; // 110
3396 llvm_unreachable("Invalid ICmp predicate!");
3401 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3402 /// predicate into a three bit mask. It also returns whether it is an ordered
3403 /// predicate by reference.
3404 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3407 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3408 case FCmpInst::FCMP_UNO: return 0; // 000
3409 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3410 case FCmpInst::FCMP_UGT: return 1; // 001
3411 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3412 case FCmpInst::FCMP_UEQ: return 2; // 010
3413 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3414 case FCmpInst::FCMP_UGE: return 3; // 011
3415 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3416 case FCmpInst::FCMP_ULT: return 4; // 100
3417 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3418 case FCmpInst::FCMP_UNE: return 5; // 101
3419 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3420 case FCmpInst::FCMP_ULE: return 6; // 110
3423 // Not expecting FCMP_FALSE and FCMP_TRUE;
3424 llvm_unreachable("Unexpected FCmp predicate!");
3429 /// getICmpValue - This is the complement of getICmpCode, which turns an
3430 /// opcode and two operands into either a constant true or false, or a brand
3431 /// new ICmp instruction. The sign is passed in to determine which kind
3432 /// of predicate to use in the new icmp instruction.
3433 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3434 LLVMContext *Context) {
3436 default: llvm_unreachable("Illegal ICmp code!");
3437 case 0: return ConstantInt::getFalse(*Context);
3440 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3442 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3443 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3446 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3448 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3451 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3453 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3454 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3457 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3459 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3460 case 7: return ConstantInt::getTrue(*Context);
3464 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3465 /// opcode and two operands into either a FCmp instruction. isordered is passed
3466 /// in to determine which kind of predicate to use in the new fcmp instruction.
3467 static Value *getFCmpValue(bool isordered, unsigned code,
3468 Value *LHS, Value *RHS, LLVMContext *Context) {
3470 default: llvm_unreachable("Illegal FCmp code!");
3473 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3475 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3478 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3480 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3483 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3485 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3488 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3490 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3493 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3495 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3498 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3500 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3503 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3505 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3506 case 7: return ConstantInt::getTrue(*Context);
3510 /// PredicatesFoldable - Return true if both predicates match sign or if at
3511 /// least one of them is an equality comparison (which is signless).
3512 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3513 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3514 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3515 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3519 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3520 struct FoldICmpLogical {
3523 ICmpInst::Predicate pred;
3524 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3525 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3526 pred(ICI->getPredicate()) {}
3527 bool shouldApply(Value *V) const {
3528 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3529 if (PredicatesFoldable(pred, ICI->getPredicate()))
3530 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3531 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3534 Instruction *apply(Instruction &Log) const {
3535 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3536 if (ICI->getOperand(0) != LHS) {
3537 assert(ICI->getOperand(1) == LHS);
3538 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3541 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3542 unsigned LHSCode = getICmpCode(ICI);
3543 unsigned RHSCode = getICmpCode(RHSICI);
3545 switch (Log.getOpcode()) {
3546 case Instruction::And: Code = LHSCode & RHSCode; break;
3547 case Instruction::Or: Code = LHSCode | RHSCode; break;
3548 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3549 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3552 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3553 ICmpInst::isSignedPredicate(ICI->getPredicate());
3555 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3556 if (Instruction *I = dyn_cast<Instruction>(RV))
3558 // Otherwise, it's a constant boolean value...
3559 return IC.ReplaceInstUsesWith(Log, RV);
3562 } // end anonymous namespace
3564 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3565 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3566 // guaranteed to be a binary operator.
3567 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3569 ConstantInt *AndRHS,
3570 BinaryOperator &TheAnd) {
3571 Value *X = Op->getOperand(0);
3572 Constant *Together = 0;
3574 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3576 switch (Op->getOpcode()) {
3577 case Instruction::Xor:
3578 if (Op->hasOneUse()) {
3579 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3580 Value *And = Builder->CreateAnd(X, AndRHS);
3582 return BinaryOperator::CreateXor(And, Together);
3585 case Instruction::Or:
3586 if (Together == AndRHS) // (X | C) & C --> C
3587 return ReplaceInstUsesWith(TheAnd, AndRHS);
3589 if (Op->hasOneUse() && Together != OpRHS) {
3590 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3591 Value *Or = Builder->CreateOr(X, Together);
3593 return BinaryOperator::CreateAnd(Or, AndRHS);
3596 case Instruction::Add:
3597 if (Op->hasOneUse()) {
3598 // Adding a one to a single bit bit-field should be turned into an XOR
3599 // of the bit. First thing to check is to see if this AND is with a
3600 // single bit constant.
3601 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3603 // If there is only one bit set...
3604 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3605 // Ok, at this point, we know that we are masking the result of the
3606 // ADD down to exactly one bit. If the constant we are adding has
3607 // no bits set below this bit, then we can eliminate the ADD.
3608 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3610 // Check to see if any bits below the one bit set in AndRHSV are set.
3611 if ((AddRHS & (AndRHSV-1)) == 0) {
3612 // If not, the only thing that can effect the output of the AND is
3613 // the bit specified by AndRHSV. If that bit is set, the effect of
3614 // the XOR is to toggle the bit. If it is clear, then the ADD has
3616 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3617 TheAnd.setOperand(0, X);
3620 // Pull the XOR out of the AND.
3621 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
3622 NewAnd->takeName(Op);
3623 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3630 case Instruction::Shl: {
3631 // We know that the AND will not produce any of the bits shifted in, so if
3632 // the anded constant includes them, clear them now!
3634 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3635 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3636 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3637 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3639 if (CI->getValue() == ShlMask) {
3640 // Masking out bits that the shift already masks
3641 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3642 } else if (CI != AndRHS) { // Reducing bits set in and.
3643 TheAnd.setOperand(1, CI);
3648 case Instruction::LShr:
3650 // We know that the AND will not produce any of the bits shifted in, so if
3651 // the anded constant includes them, clear them now! This only applies to
3652 // unsigned shifts, because a signed shr may bring in set bits!
3654 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3655 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3656 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3657 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3659 if (CI->getValue() == ShrMask) {
3660 // Masking out bits that the shift already masks.
3661 return ReplaceInstUsesWith(TheAnd, Op);
3662 } else if (CI != AndRHS) {
3663 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3668 case Instruction::AShr:
3670 // See if this is shifting in some sign extension, then masking it out
3672 if (Op->hasOneUse()) {
3673 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3674 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3675 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3676 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3677 if (C == AndRHS) { // Masking out bits shifted in.
3678 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3679 // Make the argument unsigned.
3680 Value *ShVal = Op->getOperand(0);
3681 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
3682 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3691 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3692 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3693 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3694 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3695 /// insert new instructions.
3696 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3697 bool isSigned, bool Inside,
3699 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3700 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3701 "Lo is not <= Hi in range emission code!");
3704 if (Lo == Hi) // Trivially false.
3705 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3707 // V >= Min && V < Hi --> V < Hi
3708 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3709 ICmpInst::Predicate pred = (isSigned ?
3710 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3711 return new ICmpInst(pred, V, Hi);
3714 // Emit V-Lo <u Hi-Lo
3715 Constant *NegLo = ConstantExpr::getNeg(Lo);
3716 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3717 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3718 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3721 if (Lo == Hi) // Trivially true.
3722 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3724 // V < Min || V >= Hi -> V > Hi-1
3725 Hi = SubOne(cast<ConstantInt>(Hi));
3726 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3727 ICmpInst::Predicate pred = (isSigned ?
3728 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3729 return new ICmpInst(pred, V, Hi);
3732 // Emit V-Lo >u Hi-1-Lo
3733 // Note that Hi has already had one subtracted from it, above.
3734 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3735 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3736 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3737 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3740 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3741 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3742 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3743 // not, since all 1s are not contiguous.
3744 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3745 const APInt& V = Val->getValue();
3746 uint32_t BitWidth = Val->getType()->getBitWidth();
3747 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3749 // look for the first zero bit after the run of ones
3750 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3751 // look for the first non-zero bit
3752 ME = V.getActiveBits();
3756 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3757 /// where isSub determines whether the operator is a sub. If we can fold one of
3758 /// the following xforms:
3760 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3761 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3762 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3764 /// return (A +/- B).
3766 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3767 ConstantInt *Mask, bool isSub,
3769 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3770 if (!LHSI || LHSI->getNumOperands() != 2 ||
3771 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3773 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3775 switch (LHSI->getOpcode()) {
3777 case Instruction::And:
3778 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3779 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3780 if ((Mask->getValue().countLeadingZeros() +
3781 Mask->getValue().countPopulation()) ==
3782 Mask->getValue().getBitWidth())
3785 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3786 // part, we don't need any explicit masks to take them out of A. If that
3787 // is all N is, ignore it.
3788 uint32_t MB = 0, ME = 0;
3789 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3790 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3791 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3792 if (MaskedValueIsZero(RHS, Mask))
3797 case Instruction::Or:
3798 case Instruction::Xor:
3799 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3800 if ((Mask->getValue().countLeadingZeros() +
3801 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3802 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3808 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
3809 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
3812 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3813 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3814 ICmpInst *LHS, ICmpInst *RHS) {
3816 ConstantInt *LHSCst, *RHSCst;
3817 ICmpInst::Predicate LHSCC, RHSCC;
3819 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3820 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3821 m_ConstantInt(LHSCst))) ||
3822 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3823 m_ConstantInt(RHSCst))))
3826 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3827 // where C is a power of 2
3828 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3829 LHSCst->getValue().isPowerOf2()) {
3830 Value *NewOr = Builder->CreateOr(Val, Val2);
3831 return new ICmpInst(LHSCC, NewOr, LHSCst);
3834 // From here on, we only handle:
3835 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3836 if (Val != Val2) return 0;
3838 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3839 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3840 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3841 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3842 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3845 // We can't fold (ugt x, C) & (sgt x, C2).
3846 if (!PredicatesFoldable(LHSCC, RHSCC))
3849 // Ensure that the larger constant is on the RHS.
3851 if (ICmpInst::isSignedPredicate(LHSCC) ||
3852 (ICmpInst::isEquality(LHSCC) &&
3853 ICmpInst::isSignedPredicate(RHSCC)))
3854 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3856 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3859 std::swap(LHS, RHS);
3860 std::swap(LHSCst, RHSCst);
3861 std::swap(LHSCC, RHSCC);
3864 // At this point, we know we have have two icmp instructions
3865 // comparing a value against two constants and and'ing the result
3866 // together. Because of the above check, we know that we only have
3867 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3868 // (from the FoldICmpLogical check above), that the two constants
3869 // are not equal and that the larger constant is on the RHS
3870 assert(LHSCst != RHSCst && "Compares not folded above?");
3873 default: llvm_unreachable("Unknown integer condition code!");
3874 case ICmpInst::ICMP_EQ:
3876 default: llvm_unreachable("Unknown integer condition code!");
3877 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3878 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3879 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3880 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3881 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3882 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3883 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3884 return ReplaceInstUsesWith(I, LHS);
3886 case ICmpInst::ICMP_NE:
3888 default: llvm_unreachable("Unknown integer condition code!");
3889 case ICmpInst::ICMP_ULT:
3890 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3891 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3892 break; // (X != 13 & X u< 15) -> no change
3893 case ICmpInst::ICMP_SLT:
3894 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3895 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3896 break; // (X != 13 & X s< 15) -> no change
3897 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3898 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3899 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3900 return ReplaceInstUsesWith(I, RHS);
3901 case ICmpInst::ICMP_NE:
3902 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3903 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3904 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
3905 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3906 ConstantInt::get(Add->getType(), 1));
3908 break; // (X != 13 & X != 15) -> no change
3911 case ICmpInst::ICMP_ULT:
3913 default: llvm_unreachable("Unknown integer condition code!");
3914 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3915 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3916 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3917 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3919 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3920 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3921 return ReplaceInstUsesWith(I, LHS);
3922 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3926 case ICmpInst::ICMP_SLT:
3928 default: llvm_unreachable("Unknown integer condition code!");
3929 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3930 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3931 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3932 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3934 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3935 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3936 return ReplaceInstUsesWith(I, LHS);
3937 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3941 case ICmpInst::ICMP_UGT:
3943 default: llvm_unreachable("Unknown integer condition code!");
3944 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3945 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3946 return ReplaceInstUsesWith(I, RHS);
3947 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3949 case ICmpInst::ICMP_NE:
3950 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3951 return new ICmpInst(LHSCC, Val, RHSCst);
3952 break; // (X u> 13 & X != 15) -> no change
3953 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3954 return InsertRangeTest(Val, AddOne(LHSCst),
3955 RHSCst, false, true, I);
3956 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3960 case ICmpInst::ICMP_SGT:
3962 default: llvm_unreachable("Unknown integer condition code!");
3963 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3964 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3965 return ReplaceInstUsesWith(I, RHS);
3966 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3968 case ICmpInst::ICMP_NE:
3969 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3970 return new ICmpInst(LHSCC, Val, RHSCst);
3971 break; // (X s> 13 & X != 15) -> no change
3972 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3973 return InsertRangeTest(Val, AddOne(LHSCst),
3974 RHSCst, true, true, I);
3975 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3984 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
3987 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3988 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3989 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3990 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3991 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3992 // If either of the constants are nans, then the whole thing returns
3994 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3995 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3996 return new FCmpInst(FCmpInst::FCMP_ORD,
3997 LHS->getOperand(0), RHS->getOperand(0));
4000 // Handle vector zeros. This occurs because the canonical form of
4001 // "fcmp ord x,x" is "fcmp ord x, 0".
4002 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4003 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4004 return new FCmpInst(FCmpInst::FCMP_ORD,
4005 LHS->getOperand(0), RHS->getOperand(0));
4009 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4010 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4011 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4014 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4015 // Swap RHS operands to match LHS.
4016 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4017 std::swap(Op1LHS, Op1RHS);
4020 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4021 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4023 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4025 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
4026 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4027 if (Op0CC == FCmpInst::FCMP_TRUE)
4028 return ReplaceInstUsesWith(I, RHS);
4029 if (Op1CC == FCmpInst::FCMP_TRUE)
4030 return ReplaceInstUsesWith(I, LHS);
4034 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4035 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4037 std::swap(LHS, RHS);
4038 std::swap(Op0Pred, Op1Pred);
4039 std::swap(Op0Ordered, Op1Ordered);
4042 // uno && ueq -> uno && (uno || eq) -> ueq
4043 // ord && olt -> ord && (ord && lt) -> olt
4044 if (Op0Ordered == Op1Ordered)
4045 return ReplaceInstUsesWith(I, RHS);
4047 // uno && oeq -> uno && (ord && eq) -> false
4048 // uno && ord -> false
4050 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4051 // ord && ueq -> ord && (uno || eq) -> oeq
4052 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4053 Op0LHS, Op0RHS, Context));
4061 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4062 bool Changed = SimplifyCommutative(I);
4063 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4065 if (isa<UndefValue>(Op1)) // X & undef -> 0
4066 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4070 return ReplaceInstUsesWith(I, Op1);
4072 // See if we can simplify any instructions used by the instruction whose sole
4073 // purpose is to compute bits we don't care about.
4074 if (SimplifyDemandedInstructionBits(I))
4076 if (isa<VectorType>(I.getType())) {
4077 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4078 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4079 return ReplaceInstUsesWith(I, I.getOperand(0));
4080 } else if (isa<ConstantAggregateZero>(Op1)) {
4081 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4085 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4086 const APInt &AndRHSMask = AndRHS->getValue();
4087 APInt NotAndRHS(~AndRHSMask);
4089 // Optimize a variety of ((val OP C1) & C2) combinations...
4090 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4091 Value *Op0LHS = Op0I->getOperand(0);
4092 Value *Op0RHS = Op0I->getOperand(1);
4093 switch (Op0I->getOpcode()) {
4095 case Instruction::Xor:
4096 case Instruction::Or:
4097 // If the mask is only needed on one incoming arm, push it up.
4098 if (!Op0I->hasOneUse()) break;
4100 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4101 // Not masking anything out for the LHS, move to RHS.
4102 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
4103 Op0RHS->getName()+".masked");
4104 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
4106 if (!isa<Constant>(Op0RHS) &&
4107 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4108 // Not masking anything out for the RHS, move to LHS.
4109 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
4110 Op0LHS->getName()+".masked");
4111 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
4115 case Instruction::Add:
4116 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4117 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4118 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4119 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4120 return BinaryOperator::CreateAnd(V, AndRHS);
4121 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4122 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4125 case Instruction::Sub:
4126 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4127 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4128 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4129 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4130 return BinaryOperator::CreateAnd(V, AndRHS);
4132 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4133 // has 1's for all bits that the subtraction with A might affect.
4134 if (Op0I->hasOneUse()) {
4135 uint32_t BitWidth = AndRHSMask.getBitWidth();
4136 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4137 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4139 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4140 if (!(A && A->isZero()) && // avoid infinite recursion.
4141 MaskedValueIsZero(Op0LHS, Mask)) {
4142 Value *NewNeg = Builder->CreateNeg(Op0RHS);
4143 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4148 case Instruction::Shl:
4149 case Instruction::LShr:
4150 // (1 << x) & 1 --> zext(x == 0)
4151 // (1 >> x) & 1 --> zext(x == 0)
4152 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4154 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
4155 return new ZExtInst(NewICmp, I.getType());
4160 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4161 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4163 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4164 // If this is an integer truncation or change from signed-to-unsigned, and
4165 // if the source is an and/or with immediate, transform it. This
4166 // frequently occurs for bitfield accesses.
4167 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4168 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4169 CastOp->getNumOperands() == 2)
4170 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4171 if (CastOp->getOpcode() == Instruction::And) {
4172 // Change: and (cast (and X, C1) to T), C2
4173 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4174 // This will fold the two constants together, which may allow
4175 // other simplifications.
4176 Value *NewCast = Builder->CreateTruncOrBitCast(
4177 CastOp->getOperand(0), I.getType(),
4178 CastOp->getName()+".shrunk");
4179 // trunc_or_bitcast(C1)&C2
4180 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4181 C3 = ConstantExpr::getAnd(C3, AndRHS);
4182 return BinaryOperator::CreateAnd(NewCast, C3);
4183 } else if (CastOp->getOpcode() == Instruction::Or) {
4184 // Change: and (cast (or X, C1) to T), C2
4185 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4186 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4187 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4189 return ReplaceInstUsesWith(I, AndRHS);
4195 // Try to fold constant and into select arguments.
4196 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4197 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4199 if (isa<PHINode>(Op0))
4200 if (Instruction *NV = FoldOpIntoPhi(I))
4204 Value *Op0NotVal = dyn_castNotVal(Op0);
4205 Value *Op1NotVal = dyn_castNotVal(Op1);
4207 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4208 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4210 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4211 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4212 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
4213 I.getName()+".demorgan");
4214 return BinaryOperator::CreateNot(Or);
4218 Value *A = 0, *B = 0, *C = 0, *D = 0;
4219 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4220 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4221 return ReplaceInstUsesWith(I, Op1);
4223 // (A|B) & ~(A&B) -> A^B
4224 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4225 if ((A == C && B == D) || (A == D && B == C))
4226 return BinaryOperator::CreateXor(A, B);
4230 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4231 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4232 return ReplaceInstUsesWith(I, Op0);
4234 // ~(A&B) & (A|B) -> A^B
4235 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4236 if ((A == C && B == D) || (A == D && B == C))
4237 return BinaryOperator::CreateXor(A, B);
4241 if (Op0->hasOneUse() &&
4242 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4243 if (A == Op1) { // (A^B)&A -> A&(A^B)
4244 I.swapOperands(); // Simplify below
4245 std::swap(Op0, Op1);
4246 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4247 cast<BinaryOperator>(Op0)->swapOperands();
4248 I.swapOperands(); // Simplify below
4249 std::swap(Op0, Op1);
4253 if (Op1->hasOneUse() &&
4254 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4255 if (B == Op0) { // B&(A^B) -> B&(B^A)
4256 cast<BinaryOperator>(Op1)->swapOperands();
4259 if (A == Op0) // A&(A^B) -> A & ~B
4260 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
4263 // (A&((~A)|B)) -> A&B
4264 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4265 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4266 return BinaryOperator::CreateAnd(A, Op1);
4267 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4268 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4269 return BinaryOperator::CreateAnd(A, Op0);
4272 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4273 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4274 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4277 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4278 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4282 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4283 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4284 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4285 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4286 const Type *SrcTy = Op0C->getOperand(0)->getType();
4287 if (SrcTy == Op1C->getOperand(0)->getType() &&
4288 SrcTy->isIntOrIntVector() &&
4289 // Only do this if the casts both really cause code to be generated.
4290 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4292 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4294 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
4295 Op1C->getOperand(0), I.getName());
4296 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4300 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4301 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4302 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4303 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4304 SI0->getOperand(1) == SI1->getOperand(1) &&
4305 (SI0->hasOneUse() || SI1->hasOneUse())) {
4307 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
4309 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4310 SI1->getOperand(1));
4314 // If and'ing two fcmp, try combine them into one.
4315 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4316 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4317 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4321 return Changed ? &I : 0;
4324 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4325 /// capable of providing pieces of a bswap. The subexpression provides pieces
4326 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4327 /// the expression came from the corresponding "byte swapped" byte in some other
4328 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4329 /// we know that the expression deposits the low byte of %X into the high byte
4330 /// of the bswap result and that all other bytes are zero. This expression is
4331 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4334 /// This function returns true if the match was unsuccessful and false if so.
4335 /// On entry to the function the "OverallLeftShift" is a signed integer value
4336 /// indicating the number of bytes that the subexpression is later shifted. For
4337 /// example, if the expression is later right shifted by 16 bits, the
4338 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4339 /// byte of ByteValues is actually being set.
4341 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4342 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4343 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4344 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4345 /// always in the local (OverallLeftShift) coordinate space.
4347 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4348 SmallVector<Value*, 8> &ByteValues) {
4349 if (Instruction *I = dyn_cast<Instruction>(V)) {
4350 // If this is an or instruction, it may be an inner node of the bswap.
4351 if (I->getOpcode() == Instruction::Or) {
4352 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4354 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4358 // If this is a logical shift by a constant multiple of 8, recurse with
4359 // OverallLeftShift and ByteMask adjusted.
4360 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4362 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4363 // Ensure the shift amount is defined and of a byte value.
4364 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4367 unsigned ByteShift = ShAmt >> 3;
4368 if (I->getOpcode() == Instruction::Shl) {
4369 // X << 2 -> collect(X, +2)
4370 OverallLeftShift += ByteShift;
4371 ByteMask >>= ByteShift;
4373 // X >>u 2 -> collect(X, -2)
4374 OverallLeftShift -= ByteShift;
4375 ByteMask <<= ByteShift;
4376 ByteMask &= (~0U >> (32-ByteValues.size()));
4379 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4380 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4382 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4386 // If this is a logical 'and' with a mask that clears bytes, clear the
4387 // corresponding bytes in ByteMask.
4388 if (I->getOpcode() == Instruction::And &&
4389 isa<ConstantInt>(I->getOperand(1))) {
4390 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4391 unsigned NumBytes = ByteValues.size();
4392 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4393 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4395 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4396 // If this byte is masked out by a later operation, we don't care what
4398 if ((ByteMask & (1 << i)) == 0)
4401 // If the AndMask is all zeros for this byte, clear the bit.
4402 APInt MaskB = AndMask & Byte;
4404 ByteMask &= ~(1U << i);
4408 // If the AndMask is not all ones for this byte, it's not a bytezap.
4412 // Otherwise, this byte is kept.
4415 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4420 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4421 // the input value to the bswap. Some observations: 1) if more than one byte
4422 // is demanded from this input, then it could not be successfully assembled
4423 // into a byteswap. At least one of the two bytes would not be aligned with
4424 // their ultimate destination.
4425 if (!isPowerOf2_32(ByteMask)) return true;
4426 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4428 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4429 // is demanded, it needs to go into byte 0 of the result. This means that the
4430 // byte needs to be shifted until it lands in the right byte bucket. The
4431 // shift amount depends on the position: if the byte is coming from the high
4432 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4433 // low part, it must be shifted left.
4434 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4435 if (InputByteNo < ByteValues.size()/2) {
4436 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4439 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4443 // If the destination byte value is already defined, the values are or'd
4444 // together, which isn't a bswap (unless it's an or of the same bits).
4445 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4447 ByteValues[DestByteNo] = V;
4451 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4452 /// If so, insert the new bswap intrinsic and return it.
4453 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4454 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4455 if (!ITy || ITy->getBitWidth() % 16 ||
4456 // ByteMask only allows up to 32-byte values.
4457 ITy->getBitWidth() > 32*8)
4458 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4460 /// ByteValues - For each byte of the result, we keep track of which value
4461 /// defines each byte.
4462 SmallVector<Value*, 8> ByteValues;
4463 ByteValues.resize(ITy->getBitWidth()/8);
4465 // Try to find all the pieces corresponding to the bswap.
4466 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4467 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4470 // Check to see if all of the bytes come from the same value.
4471 Value *V = ByteValues[0];
4472 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4474 // Check to make sure that all of the bytes come from the same value.
4475 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4476 if (ByteValues[i] != V)
4478 const Type *Tys[] = { ITy };
4479 Module *M = I.getParent()->getParent()->getParent();
4480 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4481 return CallInst::Create(F, V);
4484 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4485 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4486 /// we can simplify this expression to "cond ? C : D or B".
4487 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4489 LLVMContext *Context) {
4490 // If A is not a select of -1/0, this cannot match.
4492 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4495 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4496 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4497 return SelectInst::Create(Cond, C, B);
4498 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4499 return SelectInst::Create(Cond, C, B);
4500 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4501 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4502 return SelectInst::Create(Cond, C, D);
4503 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4504 return SelectInst::Create(Cond, C, D);
4508 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4509 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4510 ICmpInst *LHS, ICmpInst *RHS) {
4512 ConstantInt *LHSCst, *RHSCst;
4513 ICmpInst::Predicate LHSCC, RHSCC;
4515 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4516 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4517 m_ConstantInt(LHSCst))) ||
4518 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4519 m_ConstantInt(RHSCst))))
4522 // From here on, we only handle:
4523 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4524 if (Val != Val2) return 0;
4526 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4527 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4528 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4529 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4530 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4533 // We can't fold (ugt x, C) | (sgt x, C2).
4534 if (!PredicatesFoldable(LHSCC, RHSCC))
4537 // Ensure that the larger constant is on the RHS.
4539 if (ICmpInst::isSignedPredicate(LHSCC) ||
4540 (ICmpInst::isEquality(LHSCC) &&
4541 ICmpInst::isSignedPredicate(RHSCC)))
4542 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4544 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4547 std::swap(LHS, RHS);
4548 std::swap(LHSCst, RHSCst);
4549 std::swap(LHSCC, RHSCC);
4552 // At this point, we know we have have two icmp instructions
4553 // comparing a value against two constants and or'ing the result
4554 // together. Because of the above check, we know that we only have
4555 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4556 // FoldICmpLogical check above), that the two constants are not
4558 assert(LHSCst != RHSCst && "Compares not folded above?");
4561 default: llvm_unreachable("Unknown integer condition code!");
4562 case ICmpInst::ICMP_EQ:
4564 default: llvm_unreachable("Unknown integer condition code!");
4565 case ICmpInst::ICMP_EQ:
4566 if (LHSCst == SubOne(RHSCst)) {
4567 // (X == 13 | X == 14) -> X-13 <u 2
4568 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4569 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4570 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4571 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4573 break; // (X == 13 | X == 15) -> no change
4574 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4575 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4577 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4578 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4579 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4580 return ReplaceInstUsesWith(I, RHS);
4583 case ICmpInst::ICMP_NE:
4585 default: llvm_unreachable("Unknown integer condition code!");
4586 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4587 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4588 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4589 return ReplaceInstUsesWith(I, LHS);
4590 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4591 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4592 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4593 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4596 case ICmpInst::ICMP_ULT:
4598 default: llvm_unreachable("Unknown integer condition code!");
4599 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4601 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4602 // If RHSCst is [us]MAXINT, it is always false. Not handling
4603 // this can cause overflow.
4604 if (RHSCst->isMaxValue(false))
4605 return ReplaceInstUsesWith(I, LHS);
4606 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4608 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4610 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4611 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4612 return ReplaceInstUsesWith(I, RHS);
4613 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4617 case ICmpInst::ICMP_SLT:
4619 default: llvm_unreachable("Unknown integer condition code!");
4620 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4622 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4623 // If RHSCst is [us]MAXINT, it is always false. Not handling
4624 // this can cause overflow.
4625 if (RHSCst->isMaxValue(true))
4626 return ReplaceInstUsesWith(I, LHS);
4627 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4629 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4631 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4632 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4633 return ReplaceInstUsesWith(I, RHS);
4634 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4638 case ICmpInst::ICMP_UGT:
4640 default: llvm_unreachable("Unknown integer condition code!");
4641 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4642 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4643 return ReplaceInstUsesWith(I, LHS);
4644 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4646 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4647 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4648 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4649 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4653 case ICmpInst::ICMP_SGT:
4655 default: llvm_unreachable("Unknown integer condition code!");
4656 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4657 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4658 return ReplaceInstUsesWith(I, LHS);
4659 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4661 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4662 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4663 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4664 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4672 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4674 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4675 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4676 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4677 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4678 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4679 // If either of the constants are nans, then the whole thing returns
4681 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4682 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4684 // Otherwise, no need to compare the two constants, compare the
4686 return new FCmpInst(FCmpInst::FCMP_UNO,
4687 LHS->getOperand(0), RHS->getOperand(0));
4690 // Handle vector zeros. This occurs because the canonical form of
4691 // "fcmp uno x,x" is "fcmp uno x, 0".
4692 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4693 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4694 return new FCmpInst(FCmpInst::FCMP_UNO,
4695 LHS->getOperand(0), RHS->getOperand(0));
4700 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4701 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4702 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4704 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4705 // Swap RHS operands to match LHS.
4706 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4707 std::swap(Op1LHS, Op1RHS);
4709 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4710 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4712 return new FCmpInst((FCmpInst::Predicate)Op0CC,
4714 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4715 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4716 if (Op0CC == FCmpInst::FCMP_FALSE)
4717 return ReplaceInstUsesWith(I, RHS);
4718 if (Op1CC == FCmpInst::FCMP_FALSE)
4719 return ReplaceInstUsesWith(I, LHS);
4722 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4723 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4724 if (Op0Ordered == Op1Ordered) {
4725 // If both are ordered or unordered, return a new fcmp with
4726 // or'ed predicates.
4727 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4728 Op0LHS, Op0RHS, Context);
4729 if (Instruction *I = dyn_cast<Instruction>(RV))
4731 // Otherwise, it's a constant boolean value...
4732 return ReplaceInstUsesWith(I, RV);
4738 /// FoldOrWithConstants - This helper function folds:
4740 /// ((A | B) & C1) | (B & C2)
4746 /// when the XOR of the two constants is "all ones" (-1).
4747 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4748 Value *A, Value *B, Value *C) {
4749 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4753 ConstantInt *CI2 = 0;
4754 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4756 APInt Xor = CI1->getValue() ^ CI2->getValue();
4757 if (!Xor.isAllOnesValue()) return 0;
4759 if (V1 == A || V1 == B) {
4760 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
4761 return BinaryOperator::CreateOr(NewOp, V1);
4767 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4768 bool Changed = SimplifyCommutative(I);
4769 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4771 if (isa<UndefValue>(Op1)) // X | undef -> -1
4772 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4776 return ReplaceInstUsesWith(I, Op0);
4778 // See if we can simplify any instructions used by the instruction whose sole
4779 // purpose is to compute bits we don't care about.
4780 if (SimplifyDemandedInstructionBits(I))
4782 if (isa<VectorType>(I.getType())) {
4783 if (isa<ConstantAggregateZero>(Op1)) {
4784 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4785 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4786 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4787 return ReplaceInstUsesWith(I, I.getOperand(1));
4792 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4793 ConstantInt *C1 = 0; Value *X = 0;
4794 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4795 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
4797 Value *Or = Builder->CreateOr(X, RHS);
4799 return BinaryOperator::CreateAnd(Or,
4800 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4803 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4804 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
4806 Value *Or = Builder->CreateOr(X, RHS);
4808 return BinaryOperator::CreateXor(Or,
4809 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4812 // Try to fold constant and into select arguments.
4813 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4814 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4816 if (isa<PHINode>(Op0))
4817 if (Instruction *NV = FoldOpIntoPhi(I))
4821 Value *A = 0, *B = 0;
4822 ConstantInt *C1 = 0, *C2 = 0;
4824 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4825 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4826 return ReplaceInstUsesWith(I, Op1);
4827 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4828 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4829 return ReplaceInstUsesWith(I, Op0);
4831 // (A | B) | C and A | (B | C) -> bswap if possible.
4832 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4833 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4834 match(Op1, m_Or(m_Value(), m_Value())) ||
4835 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4836 match(Op1, m_Shift(m_Value(), m_Value())))) {
4837 if (Instruction *BSwap = MatchBSwap(I))
4841 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4842 if (Op0->hasOneUse() &&
4843 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4844 MaskedValueIsZero(Op1, C1->getValue())) {
4845 Value *NOr = Builder->CreateOr(A, Op1);
4847 return BinaryOperator::CreateXor(NOr, C1);
4850 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4851 if (Op1->hasOneUse() &&
4852 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4853 MaskedValueIsZero(Op0, C1->getValue())) {
4854 Value *NOr = Builder->CreateOr(A, Op0);
4856 return BinaryOperator::CreateXor(NOr, C1);
4860 Value *C = 0, *D = 0;
4861 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4862 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4863 Value *V1 = 0, *V2 = 0, *V3 = 0;
4864 C1 = dyn_cast<ConstantInt>(C);
4865 C2 = dyn_cast<ConstantInt>(D);
4866 if (C1 && C2) { // (A & C1)|(B & C2)
4867 // If we have: ((V + N) & C1) | (V & C2)
4868 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4869 // replace with V+N.
4870 if (C1->getValue() == ~C2->getValue()) {
4871 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4872 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4873 // Add commutes, try both ways.
4874 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4875 return ReplaceInstUsesWith(I, A);
4876 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4877 return ReplaceInstUsesWith(I, A);
4879 // Or commutes, try both ways.
4880 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4881 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4882 // Add commutes, try both ways.
4883 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4884 return ReplaceInstUsesWith(I, B);
4885 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4886 return ReplaceInstUsesWith(I, B);
4889 V1 = 0; V2 = 0; V3 = 0;
4892 // Check to see if we have any common things being and'ed. If so, find the
4893 // terms for V1 & (V2|V3).
4894 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4895 if (A == B) // (A & C)|(A & D) == A & (C|D)
4896 V1 = A, V2 = C, V3 = D;
4897 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4898 V1 = A, V2 = B, V3 = C;
4899 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4900 V1 = C, V2 = A, V3 = D;
4901 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4902 V1 = C, V2 = A, V3 = B;
4905 Value *Or = Builder->CreateOr(V2, V3, "tmp");
4906 return BinaryOperator::CreateAnd(V1, Or);
4910 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4911 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4913 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4915 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4917 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4920 // ((A&~B)|(~A&B)) -> A^B
4921 if ((match(C, m_Not(m_Specific(D))) &&
4922 match(B, m_Not(m_Specific(A)))))
4923 return BinaryOperator::CreateXor(A, D);
4924 // ((~B&A)|(~A&B)) -> A^B
4925 if ((match(A, m_Not(m_Specific(D))) &&
4926 match(B, m_Not(m_Specific(C)))))
4927 return BinaryOperator::CreateXor(C, D);
4928 // ((A&~B)|(B&~A)) -> A^B
4929 if ((match(C, m_Not(m_Specific(B))) &&
4930 match(D, m_Not(m_Specific(A)))))
4931 return BinaryOperator::CreateXor(A, B);
4932 // ((~B&A)|(B&~A)) -> A^B
4933 if ((match(A, m_Not(m_Specific(B))) &&
4934 match(D, m_Not(m_Specific(C)))))
4935 return BinaryOperator::CreateXor(C, B);
4938 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4939 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4940 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4941 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4942 SI0->getOperand(1) == SI1->getOperand(1) &&
4943 (SI0->hasOneUse() || SI1->hasOneUse())) {
4944 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
4946 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4947 SI1->getOperand(1));
4951 // ((A|B)&1)|(B&-2) -> (A&1) | B
4952 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4953 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4954 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4955 if (Ret) return Ret;
4957 // (B&-2)|((A|B)&1) -> (A&1) | B
4958 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4959 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4960 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4961 if (Ret) return Ret;
4964 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4965 if (A == Op1) // ~A | A == -1
4966 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4970 // Note, A is still live here!
4971 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4973 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4975 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4976 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4977 Value *And = Builder->CreateAnd(A, B, I.getName()+".demorgan");
4978 return BinaryOperator::CreateNot(And);
4982 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4983 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4984 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4987 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4988 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4992 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4993 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4994 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4995 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4996 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4997 !isa<ICmpInst>(Op1C->getOperand(0))) {
4998 const Type *SrcTy = Op0C->getOperand(0)->getType();
4999 if (SrcTy == Op1C->getOperand(0)->getType() &&
5000 SrcTy->isIntOrIntVector() &&
5001 // Only do this if the casts both really cause code to be
5003 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5005 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5007 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
5008 Op1C->getOperand(0), I.getName());
5009 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5016 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
5017 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
5018 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
5019 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5023 return Changed ? &I : 0;
5028 // XorSelf - Implements: X ^ X --> 0
5031 XorSelf(Value *rhs) : RHS(rhs) {}
5032 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5033 Instruction *apply(BinaryOperator &Xor) const {
5040 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5041 bool Changed = SimplifyCommutative(I);
5042 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5044 if (isa<UndefValue>(Op1)) {
5045 if (isa<UndefValue>(Op0))
5046 // Handle undef ^ undef -> 0 special case. This is a common
5048 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5049 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5052 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5053 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5054 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5055 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5058 // See if we can simplify any instructions used by the instruction whose sole
5059 // purpose is to compute bits we don't care about.
5060 if (SimplifyDemandedInstructionBits(I))
5062 if (isa<VectorType>(I.getType()))
5063 if (isa<ConstantAggregateZero>(Op1))
5064 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5066 // Is this a ~ operation?
5067 if (Value *NotOp = dyn_castNotVal(&I)) {
5068 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5069 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5070 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5071 if (Op0I->getOpcode() == Instruction::And ||
5072 Op0I->getOpcode() == Instruction::Or) {
5073 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
5074 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5076 Builder->CreateNot(Op0I->getOperand(1),
5077 Op0I->getOperand(1)->getName()+".not");
5078 if (Op0I->getOpcode() == Instruction::And)
5079 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5080 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5087 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5088 if (RHS->isOne() && Op0->hasOneUse()) {
5089 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5090 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5091 return new ICmpInst(ICI->getInversePredicate(),
5092 ICI->getOperand(0), ICI->getOperand(1));
5094 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5095 return new FCmpInst(FCI->getInversePredicate(),
5096 FCI->getOperand(0), FCI->getOperand(1));
5099 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5100 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5101 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5102 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5103 Instruction::CastOps Opcode = Op0C->getOpcode();
5104 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
5105 (RHS == ConstantExpr::getCast(Opcode,
5106 ConstantInt::getTrue(*Context),
5107 Op0C->getDestTy()))) {
5108 CI->setPredicate(CI->getInversePredicate());
5109 return CastInst::Create(Opcode, CI, Op0C->getType());
5115 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5116 // ~(c-X) == X-c-1 == X+(-c-1)
5117 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5118 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5119 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5120 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5121 ConstantInt::get(I.getType(), 1));
5122 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5125 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5126 if (Op0I->getOpcode() == Instruction::Add) {
5127 // ~(X-c) --> (-c-1)-X
5128 if (RHS->isAllOnesValue()) {
5129 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5130 return BinaryOperator::CreateSub(
5131 ConstantExpr::getSub(NegOp0CI,
5132 ConstantInt::get(I.getType(), 1)),
5133 Op0I->getOperand(0));
5134 } else if (RHS->getValue().isSignBit()) {
5135 // (X + C) ^ signbit -> (X + C + signbit)
5136 Constant *C = ConstantInt::get(*Context,
5137 RHS->getValue() + Op0CI->getValue());
5138 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5141 } else if (Op0I->getOpcode() == Instruction::Or) {
5142 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5143 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5144 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5145 // Anything in both C1 and C2 is known to be zero, remove it from
5147 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5148 NewRHS = ConstantExpr::getAnd(NewRHS,
5149 ConstantExpr::getNot(CommonBits));
5151 I.setOperand(0, Op0I->getOperand(0));
5152 I.setOperand(1, NewRHS);
5159 // Try to fold constant and into select arguments.
5160 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5161 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5163 if (isa<PHINode>(Op0))
5164 if (Instruction *NV = FoldOpIntoPhi(I))
5168 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5170 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5172 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5174 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5177 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5180 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5181 if (A == Op0) { // B^(B|A) == (A|B)^B
5182 Op1I->swapOperands();
5184 std::swap(Op0, Op1);
5185 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5186 I.swapOperands(); // Simplified below.
5187 std::swap(Op0, Op1);
5189 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5190 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5191 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5192 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5193 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5195 if (A == Op0) { // A^(A&B) -> A^(B&A)
5196 Op1I->swapOperands();
5199 if (B == Op0) { // A^(B&A) -> (B&A)^A
5200 I.swapOperands(); // Simplified below.
5201 std::swap(Op0, Op1);
5206 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5209 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5210 Op0I->hasOneUse()) {
5211 if (A == Op1) // (B|A)^B == (A|B)^B
5213 if (B == Op1) // (A|B)^B == A & ~B
5214 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
5215 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5216 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5217 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5218 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5219 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5221 if (A == Op1) // (A&B)^A -> (B&A)^A
5223 if (B == Op1 && // (B&A)^A == ~B & A
5224 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5225 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
5230 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5231 if (Op0I && Op1I && Op0I->isShift() &&
5232 Op0I->getOpcode() == Op1I->getOpcode() &&
5233 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5234 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5236 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
5238 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5239 Op1I->getOperand(1));
5243 Value *A, *B, *C, *D;
5244 // (A & B)^(A | B) -> A ^ B
5245 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5246 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5247 if ((A == C && B == D) || (A == D && B == C))
5248 return BinaryOperator::CreateXor(A, B);
5250 // (A | B)^(A & B) -> A ^ B
5251 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5252 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5253 if ((A == C && B == D) || (A == D && B == C))
5254 return BinaryOperator::CreateXor(A, B);
5258 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5259 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5260 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5261 // (X & Y)^(X & Y) -> (Y^Z) & X
5262 Value *X = 0, *Y = 0, *Z = 0;
5264 X = A, Y = B, Z = D;
5266 X = A, Y = B, Z = C;
5268 X = B, Y = A, Z = D;
5270 X = B, Y = A, Z = C;
5273 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
5274 return BinaryOperator::CreateAnd(NewOp, X);
5279 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5280 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5281 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5284 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5285 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5286 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5287 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5288 const Type *SrcTy = Op0C->getOperand(0)->getType();
5289 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5290 // Only do this if the casts both really cause code to be generated.
5291 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5293 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5295 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
5296 Op1C->getOperand(0), I.getName());
5297 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5302 return Changed ? &I : 0;
5305 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5306 LLVMContext *Context) {
5307 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5310 static bool HasAddOverflow(ConstantInt *Result,
5311 ConstantInt *In1, ConstantInt *In2,
5314 if (In2->getValue().isNegative())
5315 return Result->getValue().sgt(In1->getValue());
5317 return Result->getValue().slt(In1->getValue());
5319 return Result->getValue().ult(In1->getValue());
5322 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5323 /// overflowed for this type.
5324 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5325 Constant *In2, LLVMContext *Context,
5326 bool IsSigned = false) {
5327 Result = ConstantExpr::getAdd(In1, In2);
5329 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5330 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5331 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5332 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5333 ExtractElement(In1, Idx, Context),
5334 ExtractElement(In2, Idx, Context),
5341 return HasAddOverflow(cast<ConstantInt>(Result),
5342 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5346 static bool HasSubOverflow(ConstantInt *Result,
5347 ConstantInt *In1, ConstantInt *In2,
5350 if (In2->getValue().isNegative())
5351 return Result->getValue().slt(In1->getValue());
5353 return Result->getValue().sgt(In1->getValue());
5355 return Result->getValue().ugt(In1->getValue());
5358 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5359 /// overflowed for this type.
5360 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5361 Constant *In2, LLVMContext *Context,
5362 bool IsSigned = false) {
5363 Result = ConstantExpr::getSub(In1, In2);
5365 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5366 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5367 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5368 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5369 ExtractElement(In1, Idx, Context),
5370 ExtractElement(In2, Idx, Context),
5377 return HasSubOverflow(cast<ConstantInt>(Result),
5378 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5382 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5383 /// code necessary to compute the offset from the base pointer (without adding
5384 /// in the base pointer). Return the result as a signed integer of intptr size.
5385 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5386 TargetData &TD = *IC.getTargetData();
5387 gep_type_iterator GTI = gep_type_begin(GEP);
5388 const Type *IntPtrTy = TD.getIntPtrType(I.getContext());
5389 Value *Result = Constant::getNullValue(IntPtrTy);
5391 // Build a mask for high order bits.
5392 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5393 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5395 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5398 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5399 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5400 if (OpC->isZero()) continue;
5402 // Handle a struct index, which adds its field offset to the pointer.
5403 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5404 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5406 Result = IC.Builder->CreateAdd(Result,
5407 ConstantInt::get(IntPtrTy, Size),
5408 GEP->getName()+".offs");
5412 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5414 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5415 Scale = ConstantExpr::getMul(OC, Scale);
5416 // Emit an add instruction.
5417 Result = IC.Builder->CreateAdd(Result, Scale, GEP->getName()+".offs");
5420 // Convert to correct type.
5421 if (Op->getType() != IntPtrTy)
5422 Op = IC.Builder->CreateIntCast(Op, IntPtrTy, true, Op->getName()+".c");
5424 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5425 // We'll let instcombine(mul) convert this to a shl if possible.
5426 Op = IC.Builder->CreateMul(Op, Scale, GEP->getName()+".idx");
5429 // Emit an add instruction.
5430 Result = IC.Builder->CreateAdd(Op, Result, GEP->getName()+".offs");
5436 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5437 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5438 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5439 /// be complex, and scales are involved. The above expression would also be
5440 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5441 /// This later form is less amenable to optimization though, and we are allowed
5442 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5444 /// If we can't emit an optimized form for this expression, this returns null.
5446 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5448 TargetData &TD = *IC.getTargetData();
5449 gep_type_iterator GTI = gep_type_begin(GEP);
5451 // Check to see if this gep only has a single variable index. If so, and if
5452 // any constant indices are a multiple of its scale, then we can compute this
5453 // in terms of the scale of the variable index. For example, if the GEP
5454 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5455 // because the expression will cross zero at the same point.
5456 unsigned i, e = GEP->getNumOperands();
5458 for (i = 1; i != e; ++i, ++GTI) {
5459 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5460 // Compute the aggregate offset of constant indices.
5461 if (CI->isZero()) continue;
5463 // Handle a struct index, which adds its field offset to the pointer.
5464 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5465 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5467 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5468 Offset += Size*CI->getSExtValue();
5471 // Found our variable index.
5476 // If there are no variable indices, we must have a constant offset, just
5477 // evaluate it the general way.
5478 if (i == e) return 0;
5480 Value *VariableIdx = GEP->getOperand(i);
5481 // Determine the scale factor of the variable element. For example, this is
5482 // 4 if the variable index is into an array of i32.
5483 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5485 // Verify that there are no other variable indices. If so, emit the hard way.
5486 for (++i, ++GTI; i != e; ++i, ++GTI) {
5487 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5490 // Compute the aggregate offset of constant indices.
5491 if (CI->isZero()) continue;
5493 // Handle a struct index, which adds its field offset to the pointer.
5494 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5495 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5497 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5498 Offset += Size*CI->getSExtValue();
5502 // Okay, we know we have a single variable index, which must be a
5503 // pointer/array/vector index. If there is no offset, life is simple, return
5505 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5507 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5508 // we don't need to bother extending: the extension won't affect where the
5509 // computation crosses zero.
5510 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5511 VariableIdx = new TruncInst(VariableIdx,
5512 TD.getIntPtrType(VariableIdx->getContext()),
5513 VariableIdx->getName(), &I);
5517 // Otherwise, there is an index. The computation we will do will be modulo
5518 // the pointer size, so get it.
5519 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5521 Offset &= PtrSizeMask;
5522 VariableScale &= PtrSizeMask;
5524 // To do this transformation, any constant index must be a multiple of the
5525 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5526 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5527 // multiple of the variable scale.
5528 int64_t NewOffs = Offset / (int64_t)VariableScale;
5529 if (Offset != NewOffs*(int64_t)VariableScale)
5532 // Okay, we can do this evaluation. Start by converting the index to intptr.
5533 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
5534 if (VariableIdx->getType() != IntPtrTy)
5535 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5537 VariableIdx->getName(), &I);
5538 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5539 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5543 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5544 /// else. At this point we know that the GEP is on the LHS of the comparison.
5545 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5546 ICmpInst::Predicate Cond,
5548 // Look through bitcasts.
5549 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5550 RHS = BCI->getOperand(0);
5552 Value *PtrBase = GEPLHS->getOperand(0);
5553 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5554 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5555 // This transformation (ignoring the base and scales) is valid because we
5556 // know pointers can't overflow since the gep is inbounds. See if we can
5557 // output an optimized form.
5558 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5560 // If not, synthesize the offset the hard way.
5562 Offset = EmitGEPOffset(GEPLHS, I, *this);
5563 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5564 Constant::getNullValue(Offset->getType()));
5565 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5566 // If the base pointers are different, but the indices are the same, just
5567 // compare the base pointer.
5568 if (PtrBase != GEPRHS->getOperand(0)) {
5569 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5570 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5571 GEPRHS->getOperand(0)->getType();
5573 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5574 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5575 IndicesTheSame = false;
5579 // If all indices are the same, just compare the base pointers.
5581 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5582 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5584 // Otherwise, the base pointers are different and the indices are
5585 // different, bail out.
5589 // If one of the GEPs has all zero indices, recurse.
5590 bool AllZeros = true;
5591 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5592 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5593 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5598 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5599 ICmpInst::getSwappedPredicate(Cond), I);
5601 // If the other GEP has all zero indices, recurse.
5603 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5604 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5605 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5610 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5612 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5613 // If the GEPs only differ by one index, compare it.
5614 unsigned NumDifferences = 0; // Keep track of # differences.
5615 unsigned DiffOperand = 0; // The operand that differs.
5616 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5617 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5618 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5619 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5620 // Irreconcilable differences.
5624 if (NumDifferences++) break;
5629 if (NumDifferences == 0) // SAME GEP?
5630 return ReplaceInstUsesWith(I, // No comparison is needed here.
5631 ConstantInt::get(Type::getInt1Ty(*Context),
5632 ICmpInst::isTrueWhenEqual(Cond)));
5634 else if (NumDifferences == 1) {
5635 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5636 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5637 // Make sure we do a signed comparison here.
5638 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5642 // Only lower this if the icmp is the only user of the GEP or if we expect
5643 // the result to fold to a constant!
5645 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5646 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5647 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5648 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5649 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5650 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5656 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5658 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5661 if (!isa<ConstantFP>(RHSC)) return 0;
5662 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5664 // Get the width of the mantissa. We don't want to hack on conversions that
5665 // might lose information from the integer, e.g. "i64 -> float"
5666 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5667 if (MantissaWidth == -1) return 0; // Unknown.
5669 // Check to see that the input is converted from an integer type that is small
5670 // enough that preserves all bits. TODO: check here for "known" sign bits.
5671 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5672 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5674 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5675 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5679 // If the conversion would lose info, don't hack on this.
5680 if ((int)InputSize > MantissaWidth)
5683 // Otherwise, we can potentially simplify the comparison. We know that it
5684 // will always come through as an integer value and we know the constant is
5685 // not a NAN (it would have been previously simplified).
5686 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5688 ICmpInst::Predicate Pred;
5689 switch (I.getPredicate()) {
5690 default: llvm_unreachable("Unexpected predicate!");
5691 case FCmpInst::FCMP_UEQ:
5692 case FCmpInst::FCMP_OEQ:
5693 Pred = ICmpInst::ICMP_EQ;
5695 case FCmpInst::FCMP_UGT:
5696 case FCmpInst::FCMP_OGT:
5697 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5699 case FCmpInst::FCMP_UGE:
5700 case FCmpInst::FCMP_OGE:
5701 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5703 case FCmpInst::FCMP_ULT:
5704 case FCmpInst::FCMP_OLT:
5705 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5707 case FCmpInst::FCMP_ULE:
5708 case FCmpInst::FCMP_OLE:
5709 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5711 case FCmpInst::FCMP_UNE:
5712 case FCmpInst::FCMP_ONE:
5713 Pred = ICmpInst::ICMP_NE;
5715 case FCmpInst::FCMP_ORD:
5716 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5717 case FCmpInst::FCMP_UNO:
5718 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5721 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5723 // Now we know that the APFloat is a normal number, zero or inf.
5725 // See if the FP constant is too large for the integer. For example,
5726 // comparing an i8 to 300.0.
5727 unsigned IntWidth = IntTy->getScalarSizeInBits();
5730 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5731 // and large values.
5732 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5733 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5734 APFloat::rmNearestTiesToEven);
5735 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5736 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5737 Pred == ICmpInst::ICMP_SLE)
5738 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5739 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5742 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5743 // +INF and large values.
5744 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5745 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5746 APFloat::rmNearestTiesToEven);
5747 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5748 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5749 Pred == ICmpInst::ICMP_ULE)
5750 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5751 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5756 // See if the RHS value is < SignedMin.
5757 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5758 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5759 APFloat::rmNearestTiesToEven);
5760 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5761 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5762 Pred == ICmpInst::ICMP_SGE)
5763 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5764 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5768 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5769 // [0, UMAX], but it may still be fractional. See if it is fractional by
5770 // casting the FP value to the integer value and back, checking for equality.
5771 // Don't do this for zero, because -0.0 is not fractional.
5772 Constant *RHSInt = LHSUnsigned
5773 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5774 : ConstantExpr::getFPToSI(RHSC, IntTy);
5775 if (!RHS.isZero()) {
5776 bool Equal = LHSUnsigned
5777 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5778 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5780 // If we had a comparison against a fractional value, we have to adjust
5781 // the compare predicate and sometimes the value. RHSC is rounded towards
5782 // zero at this point.
5784 default: llvm_unreachable("Unexpected integer comparison!");
5785 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5786 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5787 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5788 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5789 case ICmpInst::ICMP_ULE:
5790 // (float)int <= 4.4 --> int <= 4
5791 // (float)int <= -4.4 --> false
5792 if (RHS.isNegative())
5793 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5795 case ICmpInst::ICMP_SLE:
5796 // (float)int <= 4.4 --> int <= 4
5797 // (float)int <= -4.4 --> int < -4
5798 if (RHS.isNegative())
5799 Pred = ICmpInst::ICMP_SLT;
5801 case ICmpInst::ICMP_ULT:
5802 // (float)int < -4.4 --> false
5803 // (float)int < 4.4 --> int <= 4
5804 if (RHS.isNegative())
5805 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5806 Pred = ICmpInst::ICMP_ULE;
5808 case ICmpInst::ICMP_SLT:
5809 // (float)int < -4.4 --> int < -4
5810 // (float)int < 4.4 --> int <= 4
5811 if (!RHS.isNegative())
5812 Pred = ICmpInst::ICMP_SLE;
5814 case ICmpInst::ICMP_UGT:
5815 // (float)int > 4.4 --> int > 4
5816 // (float)int > -4.4 --> true
5817 if (RHS.isNegative())
5818 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5820 case ICmpInst::ICMP_SGT:
5821 // (float)int > 4.4 --> int > 4
5822 // (float)int > -4.4 --> int >= -4
5823 if (RHS.isNegative())
5824 Pred = ICmpInst::ICMP_SGE;
5826 case ICmpInst::ICMP_UGE:
5827 // (float)int >= -4.4 --> true
5828 // (float)int >= 4.4 --> int > 4
5829 if (!RHS.isNegative())
5830 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5831 Pred = ICmpInst::ICMP_UGT;
5833 case ICmpInst::ICMP_SGE:
5834 // (float)int >= -4.4 --> int >= -4
5835 // (float)int >= 4.4 --> int > 4
5836 if (!RHS.isNegative())
5837 Pred = ICmpInst::ICMP_SGT;
5843 // Lower this FP comparison into an appropriate integer version of the
5845 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5848 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5849 bool Changed = SimplifyCompare(I);
5850 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5852 // Fold trivial predicates.
5853 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5854 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 0));
5855 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5856 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 1));
5858 // Simplify 'fcmp pred X, X'
5860 switch (I.getPredicate()) {
5861 default: llvm_unreachable("Unknown predicate!");
5862 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5863 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5864 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5865 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 1));
5866 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5867 case FCmpInst::FCMP_OLT: // True if ordered and less than
5868 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5869 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(), 0));
5871 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5872 case FCmpInst::FCMP_ULT: // True if unordered or less than
5873 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5874 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5875 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5876 I.setPredicate(FCmpInst::FCMP_UNO);
5877 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5880 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5881 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5882 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5883 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5884 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5885 I.setPredicate(FCmpInst::FCMP_ORD);
5886 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5891 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5892 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
5894 // Handle fcmp with constant RHS
5895 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5896 // If the constant is a nan, see if we can fold the comparison based on it.
5897 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5898 if (CFP->getValueAPF().isNaN()) {
5899 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5900 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5901 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5902 "Comparison must be either ordered or unordered!");
5903 // True if unordered.
5904 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5908 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5909 switch (LHSI->getOpcode()) {
5910 case Instruction::PHI:
5911 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5912 // block. If in the same block, we're encouraging jump threading. If
5913 // not, we are just pessimizing the code by making an i1 phi.
5914 if (LHSI->getParent() == I.getParent())
5915 if (Instruction *NV = FoldOpIntoPhi(I, true))
5918 case Instruction::SIToFP:
5919 case Instruction::UIToFP:
5920 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5923 case Instruction::Select:
5924 // If either operand of the select is a constant, we can fold the
5925 // comparison into the select arms, which will cause one to be
5926 // constant folded and the select turned into a bitwise or.
5927 Value *Op1 = 0, *Op2 = 0;
5928 if (LHSI->hasOneUse()) {
5929 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5930 // Fold the known value into the constant operand.
5931 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5932 // Insert a new FCmp of the other select operand.
5933 Op2 = Builder->CreateFCmp(I.getPredicate(),
5934 LHSI->getOperand(2), RHSC, I.getName());
5935 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5936 // Fold the known value into the constant operand.
5937 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5938 // Insert a new FCmp of the other select operand.
5939 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
5945 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5950 return Changed ? &I : 0;
5953 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5954 bool Changed = SimplifyCompare(I);
5955 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5956 const Type *Ty = Op0->getType();
5960 return ReplaceInstUsesWith(I, ConstantInt::get(I.getType(),
5961 I.isTrueWhenEqual()));
5963 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5964 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
5966 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5967 // addresses never equal each other! We already know that Op0 != Op1.
5968 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5969 isa<ConstantPointerNull>(Op0)) &&
5970 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5971 isa<ConstantPointerNull>(Op1)))
5972 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
5973 !I.isTrueWhenEqual()));
5975 // icmp's with boolean values can always be turned into bitwise operations
5976 if (Ty == Type::getInt1Ty(*Context)) {
5977 switch (I.getPredicate()) {
5978 default: llvm_unreachable("Invalid icmp instruction!");
5979 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5980 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
5981 return BinaryOperator::CreateNot(Xor);
5983 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5984 return BinaryOperator::CreateXor(Op0, Op1);
5986 case ICmpInst::ICMP_UGT:
5987 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5989 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5990 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
5991 return BinaryOperator::CreateAnd(Not, Op1);
5993 case ICmpInst::ICMP_SGT:
5994 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5996 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5997 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
5998 return BinaryOperator::CreateAnd(Not, Op0);
6000 case ICmpInst::ICMP_UGE:
6001 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6003 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6004 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
6005 return BinaryOperator::CreateOr(Not, Op1);
6007 case ICmpInst::ICMP_SGE:
6008 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6010 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6011 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
6012 return BinaryOperator::CreateOr(Not, Op0);
6017 unsigned BitWidth = 0;
6019 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6020 else if (Ty->isIntOrIntVector())
6021 BitWidth = Ty->getScalarSizeInBits();
6023 bool isSignBit = false;
6025 // See if we are doing a comparison with a constant.
6026 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6027 Value *A = 0, *B = 0;
6029 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6030 if (I.isEquality() && CI->isNullValue() &&
6031 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6032 // (icmp cond A B) if cond is equality
6033 return new ICmpInst(I.getPredicate(), A, B);
6036 // If we have an icmp le or icmp ge instruction, turn it into the
6037 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6038 // them being folded in the code below.
6039 switch (I.getPredicate()) {
6041 case ICmpInst::ICMP_ULE:
6042 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6043 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6044 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
6046 case ICmpInst::ICMP_SLE:
6047 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6048 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6049 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6051 case ICmpInst::ICMP_UGE:
6052 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6053 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6054 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
6056 case ICmpInst::ICMP_SGE:
6057 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6058 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6059 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6063 // If this comparison is a normal comparison, it demands all
6064 // bits, if it is a sign bit comparison, it only demands the sign bit.
6066 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6069 // See if we can fold the comparison based on range information we can get
6070 // by checking whether bits are known to be zero or one in the input.
6071 if (BitWidth != 0) {
6072 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6073 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6075 if (SimplifyDemandedBits(I.getOperandUse(0),
6076 isSignBit ? APInt::getSignBit(BitWidth)
6077 : APInt::getAllOnesValue(BitWidth),
6078 Op0KnownZero, Op0KnownOne, 0))
6080 if (SimplifyDemandedBits(I.getOperandUse(1),
6081 APInt::getAllOnesValue(BitWidth),
6082 Op1KnownZero, Op1KnownOne, 0))
6085 // Given the known and unknown bits, compute a range that the LHS could be
6086 // in. Compute the Min, Max and RHS values based on the known bits. For the
6087 // EQ and NE we use unsigned values.
6088 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6089 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6090 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6091 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6093 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6096 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6098 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6102 // If Min and Max are known to be the same, then SimplifyDemandedBits
6103 // figured out that the LHS is a constant. Just constant fold this now so
6104 // that code below can assume that Min != Max.
6105 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6106 return new ICmpInst(I.getPredicate(),
6107 ConstantInt::get(*Context, Op0Min), Op1);
6108 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6109 return new ICmpInst(I.getPredicate(), Op0,
6110 ConstantInt::get(*Context, Op1Min));
6112 // Based on the range information we know about the LHS, see if we can
6113 // simplify this comparison. For example, (x&4) < 8 is always true.
6114 switch (I.getPredicate()) {
6115 default: llvm_unreachable("Unknown icmp opcode!");
6116 case ICmpInst::ICMP_EQ:
6117 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6118 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6120 case ICmpInst::ICMP_NE:
6121 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6122 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6124 case ICmpInst::ICMP_ULT:
6125 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6126 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6127 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6128 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6129 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6130 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6131 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6132 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6133 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6136 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6137 if (CI->isMinValue(true))
6138 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6139 Constant::getAllOnesValue(Op0->getType()));
6142 case ICmpInst::ICMP_UGT:
6143 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6144 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6145 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6146 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6148 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6149 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6150 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6151 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6152 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6155 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6156 if (CI->isMaxValue(true))
6157 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6158 Constant::getNullValue(Op0->getType()));
6161 case ICmpInst::ICMP_SLT:
6162 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6163 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6164 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6165 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6166 if (Op1Min == Op0Max) // A <s 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 <s C -> A == C-1 if min(A)+1 == C
6170 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6174 case ICmpInst::ICMP_SGT:
6175 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6176 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6177 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6178 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6180 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6181 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6182 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6183 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6184 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6188 case ICmpInst::ICMP_SGE:
6189 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6190 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6191 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6192 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6193 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6195 case ICmpInst::ICMP_SLE:
6196 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6197 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6198 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6199 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6200 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6202 case ICmpInst::ICMP_UGE:
6203 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6204 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6205 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6206 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6207 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6209 case ICmpInst::ICMP_ULE:
6210 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6211 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6212 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6213 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6214 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6218 // Turn a signed comparison into an unsigned one if both operands
6219 // are known to have the same sign.
6220 if (I.isSignedPredicate() &&
6221 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6222 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6223 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6226 // Test if the ICmpInst instruction is used exclusively by a select as
6227 // part of a minimum or maximum operation. If so, refrain from doing
6228 // any other folding. This helps out other analyses which understand
6229 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6230 // and CodeGen. And in this case, at least one of the comparison
6231 // operands has at least one user besides the compare (the select),
6232 // which would often largely negate the benefit of folding anyway.
6234 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6235 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6236 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6239 // See if we are doing a comparison between a constant and an instruction that
6240 // can be folded into the comparison.
6241 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6242 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6243 // instruction, see if that instruction also has constants so that the
6244 // instruction can be folded into the icmp
6245 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6246 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6250 // Handle icmp with constant (but not simple integer constant) RHS
6251 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6252 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6253 switch (LHSI->getOpcode()) {
6254 case Instruction::GetElementPtr:
6255 if (RHSC->isNullValue()) {
6256 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6257 bool isAllZeros = true;
6258 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6259 if (!isa<Constant>(LHSI->getOperand(i)) ||
6260 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6265 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6266 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6270 case Instruction::PHI:
6271 // Only fold icmp into the PHI if the phi and icmp are in the same
6272 // block. If in the same block, we're encouraging jump threading. If
6273 // not, we are just pessimizing the code by making an i1 phi.
6274 if (LHSI->getParent() == I.getParent())
6275 if (Instruction *NV = FoldOpIntoPhi(I, true))
6278 case Instruction::Select: {
6279 // If either operand of the select is a constant, we can fold the
6280 // comparison into the select arms, which will cause one to be
6281 // constant folded and the select turned into a bitwise or.
6282 Value *Op1 = 0, *Op2 = 0;
6283 if (LHSI->hasOneUse()) {
6284 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6285 // Fold the known value into the constant operand.
6286 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6287 // Insert a new ICmp of the other select operand.
6288 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
6290 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6291 // Fold the known value into the constant operand.
6292 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6293 // Insert a new ICmp of the other select operand.
6294 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
6300 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6303 case Instruction::Call:
6304 // If we have (malloc != null), and if the malloc has a single use, we
6305 // can assume it is successful and remove the malloc.
6306 if (isMalloc(LHSI) && LHSI->hasOneUse() &&
6307 isa<ConstantPointerNull>(RHSC)) {
6309 return ReplaceInstUsesWith(I,
6310 ConstantInt::get(Type::getInt1Ty(*Context),
6311 !I.isTrueWhenEqual()));
6317 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6318 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6319 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6321 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6322 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6323 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6326 // Test to see if the operands of the icmp are casted versions of other
6327 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6329 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6330 if (isa<PointerType>(Op0->getType()) &&
6331 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6332 // We keep moving the cast from the left operand over to the right
6333 // operand, where it can often be eliminated completely.
6334 Op0 = CI->getOperand(0);
6336 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6337 // so eliminate it as well.
6338 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6339 Op1 = CI2->getOperand(0);
6341 // If Op1 is a constant, we can fold the cast into the constant.
6342 if (Op0->getType() != Op1->getType()) {
6343 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6344 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6346 // Otherwise, cast the RHS right before the icmp
6347 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
6350 return new ICmpInst(I.getPredicate(), Op0, Op1);
6354 if (isa<CastInst>(Op0)) {
6355 // Handle the special case of: icmp (cast bool to X), <cst>
6356 // This comes up when you have code like
6359 // For generality, we handle any zero-extension of any operand comparison
6360 // with a constant or another cast from the same type.
6361 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6362 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6366 // See if it's the same type of instruction on the left and right.
6367 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6368 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6369 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6370 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6371 switch (Op0I->getOpcode()) {
6373 case Instruction::Add:
6374 case Instruction::Sub:
6375 case Instruction::Xor:
6376 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6377 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6378 Op1I->getOperand(0));
6379 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6380 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6381 if (CI->getValue().isSignBit()) {
6382 ICmpInst::Predicate Pred = I.isSignedPredicate()
6383 ? I.getUnsignedPredicate()
6384 : I.getSignedPredicate();
6385 return new ICmpInst(Pred, Op0I->getOperand(0),
6386 Op1I->getOperand(0));
6389 if (CI->getValue().isMaxSignedValue()) {
6390 ICmpInst::Predicate Pred = I.isSignedPredicate()
6391 ? I.getUnsignedPredicate()
6392 : I.getSignedPredicate();
6393 Pred = I.getSwappedPredicate(Pred);
6394 return new ICmpInst(Pred, Op0I->getOperand(0),
6395 Op1I->getOperand(0));
6399 case Instruction::Mul:
6400 if (!I.isEquality())
6403 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6404 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6405 // Mask = -1 >> count-trailing-zeros(Cst).
6406 if (!CI->isZero() && !CI->isOne()) {
6407 const APInt &AP = CI->getValue();
6408 ConstantInt *Mask = ConstantInt::get(*Context,
6409 APInt::getLowBitsSet(AP.getBitWidth(),
6411 AP.countTrailingZeros()));
6412 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
6413 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
6414 return new ICmpInst(I.getPredicate(), And1, And2);
6423 // ~x < ~y --> y < x
6425 if (match(Op0, m_Not(m_Value(A))) &&
6426 match(Op1, m_Not(m_Value(B))))
6427 return new ICmpInst(I.getPredicate(), B, A);
6430 if (I.isEquality()) {
6431 Value *A, *B, *C, *D;
6433 // -x == -y --> x == y
6434 if (match(Op0, m_Neg(m_Value(A))) &&
6435 match(Op1, m_Neg(m_Value(B))))
6436 return new ICmpInst(I.getPredicate(), A, B);
6438 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6439 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6440 Value *OtherVal = A == Op1 ? B : A;
6441 return new ICmpInst(I.getPredicate(), OtherVal,
6442 Constant::getNullValue(A->getType()));
6445 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6446 // A^c1 == C^c2 --> A == C^(c1^c2)
6447 ConstantInt *C1, *C2;
6448 if (match(B, m_ConstantInt(C1)) &&
6449 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6451 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6452 Value *Xor = Builder->CreateXor(C, NC, "tmp");
6453 return new ICmpInst(I.getPredicate(), A, Xor);
6456 // A^B == A^D -> B == D
6457 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6458 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6459 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6460 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6464 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6465 (A == Op0 || B == Op0)) {
6466 // A == (A^B) -> B == 0
6467 Value *OtherVal = A == Op0 ? B : A;
6468 return new ICmpInst(I.getPredicate(), OtherVal,
6469 Constant::getNullValue(A->getType()));
6472 // (A-B) == A -> B == 0
6473 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6474 return new ICmpInst(I.getPredicate(), B,
6475 Constant::getNullValue(B->getType()));
6477 // A == (A-B) -> B == 0
6478 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6479 return new ICmpInst(I.getPredicate(), B,
6480 Constant::getNullValue(B->getType()));
6482 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6483 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6484 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6485 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6486 Value *X = 0, *Y = 0, *Z = 0;
6489 X = B; Y = D; Z = A;
6490 } else if (A == D) {
6491 X = B; Y = C; Z = A;
6492 } else if (B == C) {
6493 X = A; Y = D; Z = B;
6494 } else if (B == D) {
6495 X = A; Y = C; Z = B;
6498 if (X) { // Build (X^Y) & Z
6499 Op1 = Builder->CreateXor(X, Y, "tmp");
6500 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
6501 I.setOperand(0, Op1);
6502 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6507 return Changed ? &I : 0;
6511 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6512 /// and CmpRHS are both known to be integer constants.
6513 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6514 ConstantInt *DivRHS) {
6515 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6516 const APInt &CmpRHSV = CmpRHS->getValue();
6518 // FIXME: If the operand types don't match the type of the divide
6519 // then don't attempt this transform. The code below doesn't have the
6520 // logic to deal with a signed divide and an unsigned compare (and
6521 // vice versa). This is because (x /s C1) <s C2 produces different
6522 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6523 // (x /u C1) <u C2. Simply casting the operands and result won't
6524 // work. :( The if statement below tests that condition and bails
6526 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6527 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6529 if (DivRHS->isZero())
6530 return 0; // The ProdOV computation fails on divide by zero.
6531 if (DivIsSigned && DivRHS->isAllOnesValue())
6532 return 0; // The overflow computation also screws up here
6533 if (DivRHS->isOne())
6534 return 0; // Not worth bothering, and eliminates some funny cases
6537 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6538 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6539 // C2 (CI). By solving for X we can turn this into a range check
6540 // instead of computing a divide.
6541 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6543 // Determine if the product overflows by seeing if the product is
6544 // not equal to the divide. Make sure we do the same kind of divide
6545 // as in the LHS instruction that we're folding.
6546 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6547 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6549 // Get the ICmp opcode
6550 ICmpInst::Predicate Pred = ICI.getPredicate();
6552 // Figure out the interval that is being checked. For example, a comparison
6553 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6554 // Compute this interval based on the constants involved and the signedness of
6555 // the compare/divide. This computes a half-open interval, keeping track of
6556 // whether either value in the interval overflows. After analysis each
6557 // overflow variable is set to 0 if it's corresponding bound variable is valid
6558 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6559 int LoOverflow = 0, HiOverflow = 0;
6560 Constant *LoBound = 0, *HiBound = 0;
6562 if (!DivIsSigned) { // udiv
6563 // e.g. X/5 op 3 --> [15, 20)
6565 HiOverflow = LoOverflow = ProdOV;
6567 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6568 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6569 if (CmpRHSV == 0) { // (X / pos) op 0
6570 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6571 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6573 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6574 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6575 HiOverflow = LoOverflow = ProdOV;
6577 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6578 } else { // (X / pos) op neg
6579 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6580 HiBound = AddOne(Prod);
6581 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6583 ConstantInt* DivNeg =
6584 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6585 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6589 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6590 if (CmpRHSV == 0) { // (X / neg) op 0
6591 // e.g. X/-5 op 0 --> [-4, 5)
6592 LoBound = AddOne(DivRHS);
6593 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6594 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6595 HiOverflow = 1; // [INTMIN+1, overflow)
6596 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6598 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6599 // e.g. X/-5 op 3 --> [-19, -14)
6600 HiBound = AddOne(Prod);
6601 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6603 LoOverflow = AddWithOverflow(LoBound, HiBound,
6604 DivRHS, Context, true) ? -1 : 0;
6605 } else { // (X / neg) op neg
6606 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6607 LoOverflow = HiOverflow = ProdOV;
6609 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6612 // Dividing by a negative swaps the condition. LT <-> GT
6613 Pred = ICmpInst::getSwappedPredicate(Pred);
6616 Value *X = DivI->getOperand(0);
6618 default: llvm_unreachable("Unhandled icmp opcode!");
6619 case ICmpInst::ICMP_EQ:
6620 if (LoOverflow && HiOverflow)
6621 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6622 else if (HiOverflow)
6623 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6624 ICmpInst::ICMP_UGE, X, LoBound);
6625 else if (LoOverflow)
6626 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6627 ICmpInst::ICMP_ULT, X, HiBound);
6629 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6630 case ICmpInst::ICMP_NE:
6631 if (LoOverflow && HiOverflow)
6632 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6633 else if (HiOverflow)
6634 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6635 ICmpInst::ICMP_ULT, X, LoBound);
6636 else if (LoOverflow)
6637 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6638 ICmpInst::ICMP_UGE, X, HiBound);
6640 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6641 case ICmpInst::ICMP_ULT:
6642 case ICmpInst::ICMP_SLT:
6643 if (LoOverflow == +1) // Low bound is greater than input range.
6644 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6645 if (LoOverflow == -1) // Low bound is less than input range.
6646 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6647 return new ICmpInst(Pred, X, LoBound);
6648 case ICmpInst::ICMP_UGT:
6649 case ICmpInst::ICMP_SGT:
6650 if (HiOverflow == +1) // High bound greater than input range.
6651 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6652 else if (HiOverflow == -1) // High bound less than input range.
6653 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6654 if (Pred == ICmpInst::ICMP_UGT)
6655 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6657 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6662 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6664 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6667 const APInt &RHSV = RHS->getValue();
6669 switch (LHSI->getOpcode()) {
6670 case Instruction::Trunc:
6671 if (ICI.isEquality() && LHSI->hasOneUse()) {
6672 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6673 // of the high bits truncated out of x are known.
6674 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6675 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6676 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6677 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6678 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6680 // If all the high bits are known, we can do this xform.
6681 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6682 // Pull in the high bits from known-ones set.
6683 APInt NewRHS(RHS->getValue());
6684 NewRHS.zext(SrcBits);
6686 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6687 ConstantInt::get(*Context, NewRHS));
6692 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6693 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6694 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6696 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6697 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6698 Value *CompareVal = LHSI->getOperand(0);
6700 // If the sign bit of the XorCST is not set, there is no change to
6701 // the operation, just stop using the Xor.
6702 if (!XorCST->getValue().isNegative()) {
6703 ICI.setOperand(0, CompareVal);
6708 // Was the old condition true if the operand is positive?
6709 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6711 // If so, the new one isn't.
6712 isTrueIfPositive ^= true;
6714 if (isTrueIfPositive)
6715 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
6718 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
6722 if (LHSI->hasOneUse()) {
6723 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6724 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6725 const APInt &SignBit = XorCST->getValue();
6726 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6727 ? ICI.getUnsignedPredicate()
6728 : ICI.getSignedPredicate();
6729 return new ICmpInst(Pred, LHSI->getOperand(0),
6730 ConstantInt::get(*Context, RHSV ^ SignBit));
6733 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6734 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6735 const APInt &NotSignBit = XorCST->getValue();
6736 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6737 ? ICI.getUnsignedPredicate()
6738 : ICI.getSignedPredicate();
6739 Pred = ICI.getSwappedPredicate(Pred);
6740 return new ICmpInst(Pred, LHSI->getOperand(0),
6741 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6746 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6747 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6748 LHSI->getOperand(0)->hasOneUse()) {
6749 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6751 // If the LHS is an AND of a truncating cast, we can widen the
6752 // and/compare to be the input width without changing the value
6753 // produced, eliminating a cast.
6754 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6755 // We can do this transformation if either the AND constant does not
6756 // have its sign bit set or if it is an equality comparison.
6757 // Extending a relational comparison when we're checking the sign
6758 // bit would not work.
6759 if (Cast->hasOneUse() &&
6760 (ICI.isEquality() ||
6761 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6763 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6764 APInt NewCST = AndCST->getValue();
6765 NewCST.zext(BitWidth);
6767 NewCI.zext(BitWidth);
6769 Builder->CreateAnd(Cast->getOperand(0),
6770 ConstantInt::get(*Context, NewCST), LHSI->getName());
6771 return new ICmpInst(ICI.getPredicate(), NewAnd,
6772 ConstantInt::get(*Context, NewCI));
6776 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6777 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6778 // happens a LOT in code produced by the C front-end, for bitfield
6780 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6781 if (Shift && !Shift->isShift())
6785 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6786 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6787 const Type *AndTy = AndCST->getType(); // Type of the and.
6789 // We can fold this as long as we can't shift unknown bits
6790 // into the mask. This can only happen with signed shift
6791 // rights, as they sign-extend.
6793 bool CanFold = Shift->isLogicalShift();
6795 // To test for the bad case of the signed shr, see if any
6796 // of the bits shifted in could be tested after the mask.
6797 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6798 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6800 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6801 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6802 AndCST->getValue()) == 0)
6808 if (Shift->getOpcode() == Instruction::Shl)
6809 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6811 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6813 // Check to see if we are shifting out any of the bits being
6815 if (ConstantExpr::get(Shift->getOpcode(),
6816 NewCst, ShAmt) != RHS) {
6817 // If we shifted bits out, the fold is not going to work out.
6818 // As a special case, check to see if this means that the
6819 // result is always true or false now.
6820 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6821 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6822 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6823 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6825 ICI.setOperand(1, NewCst);
6826 Constant *NewAndCST;
6827 if (Shift->getOpcode() == Instruction::Shl)
6828 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6830 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6831 LHSI->setOperand(1, NewAndCST);
6832 LHSI->setOperand(0, Shift->getOperand(0));
6833 Worklist.Add(Shift); // Shift is dead.
6839 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6840 // preferable because it allows the C<<Y expression to be hoisted out
6841 // of a loop if Y is invariant and X is not.
6842 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6843 ICI.isEquality() && !Shift->isArithmeticShift() &&
6844 !isa<Constant>(Shift->getOperand(0))) {
6847 if (Shift->getOpcode() == Instruction::LShr) {
6848 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
6850 // Insert a logical shift.
6851 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
6854 // Compute X & (C << Y).
6856 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6858 ICI.setOperand(0, NewAnd);
6864 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6865 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6868 uint32_t TypeBits = RHSV.getBitWidth();
6870 // Check that the shift amount is in range. If not, don't perform
6871 // undefined shifts. When the shift is visited it will be
6873 if (ShAmt->uge(TypeBits))
6876 if (ICI.isEquality()) {
6877 // If we are comparing against bits always shifted out, the
6878 // comparison cannot succeed.
6880 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6882 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6883 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6884 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6885 return ReplaceInstUsesWith(ICI, Cst);
6888 if (LHSI->hasOneUse()) {
6889 // Otherwise strength reduce the shift into an and.
6890 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6892 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6893 TypeBits-ShAmtVal));
6896 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
6897 return new ICmpInst(ICI.getPredicate(), And,
6898 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6902 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6903 bool TrueIfSigned = false;
6904 if (LHSI->hasOneUse() &&
6905 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6906 // (X << 31) <s 0 --> (X&1) != 0
6907 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6908 (TypeBits-ShAmt->getZExtValue()-1));
6910 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
6911 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6912 And, Constant::getNullValue(And->getType()));
6917 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6918 case Instruction::AShr: {
6919 // Only handle equality comparisons of shift-by-constant.
6920 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6921 if (!ShAmt || !ICI.isEquality()) break;
6923 // Check that the shift amount is in range. If not, don't perform
6924 // undefined shifts. When the shift is visited it will be
6926 uint32_t TypeBits = RHSV.getBitWidth();
6927 if (ShAmt->uge(TypeBits))
6930 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6932 // If we are comparing against bits always shifted out, the
6933 // comparison cannot succeed.
6934 APInt Comp = RHSV << ShAmtVal;
6935 if (LHSI->getOpcode() == Instruction::LShr)
6936 Comp = Comp.lshr(ShAmtVal);
6938 Comp = Comp.ashr(ShAmtVal);
6940 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6941 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6942 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6943 return ReplaceInstUsesWith(ICI, Cst);
6946 // Otherwise, check to see if the bits shifted out are known to be zero.
6947 // If so, we can compare against the unshifted value:
6948 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6949 if (LHSI->hasOneUse() &&
6950 MaskedValueIsZero(LHSI->getOperand(0),
6951 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6952 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6953 ConstantExpr::getShl(RHS, ShAmt));
6956 if (LHSI->hasOneUse()) {
6957 // Otherwise strength reduce the shift into an and.
6958 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6959 Constant *Mask = ConstantInt::get(*Context, Val);
6961 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
6962 Mask, LHSI->getName()+".mask");
6963 return new ICmpInst(ICI.getPredicate(), And,
6964 ConstantExpr::getShl(RHS, ShAmt));
6969 case Instruction::SDiv:
6970 case Instruction::UDiv:
6971 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6972 // Fold this div into the comparison, producing a range check.
6973 // Determine, based on the divide type, what the range is being
6974 // checked. If there is an overflow on the low or high side, remember
6975 // it, otherwise compute the range [low, hi) bounding the new value.
6976 // See: InsertRangeTest above for the kinds of replacements possible.
6977 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6978 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6983 case Instruction::Add:
6984 // Fold: icmp pred (add, X, C1), C2
6986 if (!ICI.isEquality()) {
6987 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6989 const APInt &LHSV = LHSC->getValue();
6991 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6994 if (ICI.isSignedPredicate()) {
6995 if (CR.getLower().isSignBit()) {
6996 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6997 ConstantInt::get(*Context, CR.getUpper()));
6998 } else if (CR.getUpper().isSignBit()) {
6999 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7000 ConstantInt::get(*Context, CR.getLower()));
7003 if (CR.getLower().isMinValue()) {
7004 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7005 ConstantInt::get(*Context, CR.getUpper()));
7006 } else if (CR.getUpper().isMinValue()) {
7007 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7008 ConstantInt::get(*Context, CR.getLower()));
7015 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7016 if (ICI.isEquality()) {
7017 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7019 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7020 // the second operand is a constant, simplify a bit.
7021 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7022 switch (BO->getOpcode()) {
7023 case Instruction::SRem:
7024 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7025 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7026 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7027 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7029 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
7031 return new ICmpInst(ICI.getPredicate(), NewRem,
7032 Constant::getNullValue(BO->getType()));
7036 case Instruction::Add:
7037 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7038 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7039 if (BO->hasOneUse())
7040 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7041 ConstantExpr::getSub(RHS, BOp1C));
7042 } else if (RHSV == 0) {
7043 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7044 // efficiently invertible, or if the add has just this one use.
7045 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7047 if (Value *NegVal = dyn_castNegVal(BOp1))
7048 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
7049 else if (Value *NegVal = dyn_castNegVal(BOp0))
7050 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
7051 else if (BO->hasOneUse()) {
7052 Value *Neg = Builder->CreateNeg(BOp1);
7054 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
7058 case Instruction::Xor:
7059 // For the xor case, we can xor two constants together, eliminating
7060 // the explicit xor.
7061 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7062 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7063 ConstantExpr::getXor(RHS, BOC));
7066 case Instruction::Sub:
7067 // Replace (([sub|xor] A, B) != 0) with (A != B)
7069 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7073 case Instruction::Or:
7074 // If bits are being or'd in that are not present in the constant we
7075 // are comparing against, then the comparison could never succeed!
7076 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7077 Constant *NotCI = ConstantExpr::getNot(RHS);
7078 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7079 return ReplaceInstUsesWith(ICI,
7080 ConstantInt::get(Type::getInt1Ty(*Context),
7085 case Instruction::And:
7086 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7087 // If bits are being compared against that are and'd out, then the
7088 // comparison can never succeed!
7089 if ((RHSV & ~BOC->getValue()) != 0)
7090 return ReplaceInstUsesWith(ICI,
7091 ConstantInt::get(Type::getInt1Ty(*Context),
7094 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7095 if (RHS == BOC && RHSV.isPowerOf2())
7096 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7097 ICmpInst::ICMP_NE, LHSI,
7098 Constant::getNullValue(RHS->getType()));
7100 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7101 if (BOC->getValue().isSignBit()) {
7102 Value *X = BO->getOperand(0);
7103 Constant *Zero = Constant::getNullValue(X->getType());
7104 ICmpInst::Predicate pred = isICMP_NE ?
7105 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7106 return new ICmpInst(pred, X, Zero);
7109 // ((X & ~7) == 0) --> X < 8
7110 if (RHSV == 0 && isHighOnes(BOC)) {
7111 Value *X = BO->getOperand(0);
7112 Constant *NegX = ConstantExpr::getNeg(BOC);
7113 ICmpInst::Predicate pred = isICMP_NE ?
7114 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7115 return new ICmpInst(pred, X, NegX);
7120 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7121 // Handle icmp {eq|ne} <intrinsic>, intcst.
7122 if (II->getIntrinsicID() == Intrinsic::bswap) {
7124 ICI.setOperand(0, II->getOperand(1));
7125 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7133 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7134 /// We only handle extending casts so far.
7136 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7137 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7138 Value *LHSCIOp = LHSCI->getOperand(0);
7139 const Type *SrcTy = LHSCIOp->getType();
7140 const Type *DestTy = LHSCI->getType();
7143 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7144 // integer type is the same size as the pointer type.
7145 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7146 TD->getPointerSizeInBits() ==
7147 cast<IntegerType>(DestTy)->getBitWidth()) {
7149 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7150 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7151 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7152 RHSOp = RHSC->getOperand(0);
7153 // If the pointer types don't match, insert a bitcast.
7154 if (LHSCIOp->getType() != RHSOp->getType())
7155 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
7159 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7162 // The code below only handles extension cast instructions, so far.
7164 if (LHSCI->getOpcode() != Instruction::ZExt &&
7165 LHSCI->getOpcode() != Instruction::SExt)
7168 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7169 bool isSignedCmp = ICI.isSignedPredicate();
7171 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7172 // Not an extension from the same type?
7173 RHSCIOp = CI->getOperand(0);
7174 if (RHSCIOp->getType() != LHSCIOp->getType())
7177 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7178 // and the other is a zext), then we can't handle this.
7179 if (CI->getOpcode() != LHSCI->getOpcode())
7182 // Deal with equality cases early.
7183 if (ICI.isEquality())
7184 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7186 // A signed comparison of sign extended values simplifies into a
7187 // signed comparison.
7188 if (isSignedCmp && isSignedExt)
7189 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7191 // The other three cases all fold into an unsigned comparison.
7192 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7195 // If we aren't dealing with a constant on the RHS, exit early
7196 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7200 // Compute the constant that would happen if we truncated to SrcTy then
7201 // reextended to DestTy.
7202 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7203 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7206 // If the re-extended constant didn't change...
7208 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7209 // For example, we might have:
7210 // %A = sext i16 %X to i32
7211 // %B = icmp ugt i32 %A, 1330
7212 // It is incorrect to transform this into
7213 // %B = icmp ugt i16 %X, 1330
7214 // because %A may have negative value.
7216 // However, we allow this when the compare is EQ/NE, because they are
7218 if (isSignedExt == isSignedCmp || ICI.isEquality())
7219 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7223 // The re-extended constant changed so the constant cannot be represented
7224 // in the shorter type. Consequently, we cannot emit a simple comparison.
7226 // First, handle some easy cases. We know the result cannot be equal at this
7227 // point so handle the ICI.isEquality() cases
7228 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7229 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7230 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7231 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7233 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7234 // should have been folded away previously and not enter in here.
7237 // We're performing a signed comparison.
7238 if (cast<ConstantInt>(CI)->getValue().isNegative())
7239 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7241 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7243 // We're performing an unsigned comparison.
7245 // We're performing an unsigned comp with a sign extended value.
7246 // This is true if the input is >= 0. [aka >s -1]
7247 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7248 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
7250 // Unsigned extend & unsigned compare -> always true.
7251 Result = ConstantInt::getTrue(*Context);
7255 // Finally, return the value computed.
7256 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7257 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7258 return ReplaceInstUsesWith(ICI, Result);
7260 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7261 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7262 "ICmp should be folded!");
7263 if (Constant *CI = dyn_cast<Constant>(Result))
7264 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7265 return BinaryOperator::CreateNot(Result);
7268 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7269 return commonShiftTransforms(I);
7272 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7273 return commonShiftTransforms(I);
7276 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7277 if (Instruction *R = commonShiftTransforms(I))
7280 Value *Op0 = I.getOperand(0);
7282 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7283 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7284 if (CSI->isAllOnesValue())
7285 return ReplaceInstUsesWith(I, CSI);
7287 // See if we can turn a signed shr into an unsigned shr.
7288 if (MaskedValueIsZero(Op0,
7289 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7290 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7292 // Arithmetic shifting an all-sign-bit value is a no-op.
7293 unsigned NumSignBits = ComputeNumSignBits(Op0);
7294 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7295 return ReplaceInstUsesWith(I, Op0);
7300 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7301 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7302 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7304 // shl X, 0 == X and shr X, 0 == X
7305 // shl 0, X == 0 and shr 0, X == 0
7306 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7307 Op0 == Constant::getNullValue(Op0->getType()))
7308 return ReplaceInstUsesWith(I, Op0);
7310 if (isa<UndefValue>(Op0)) {
7311 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7312 return ReplaceInstUsesWith(I, Op0);
7313 else // undef << X -> 0, undef >>u X -> 0
7314 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7316 if (isa<UndefValue>(Op1)) {
7317 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7318 return ReplaceInstUsesWith(I, Op0);
7319 else // X << undef, X >>u undef -> 0
7320 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7323 // See if we can fold away this shift.
7324 if (SimplifyDemandedInstructionBits(I))
7327 // Try to fold constant and into select arguments.
7328 if (isa<Constant>(Op0))
7329 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7330 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7333 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7334 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7339 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7340 BinaryOperator &I) {
7341 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7343 // See if we can simplify any instructions used by the instruction whose sole
7344 // purpose is to compute bits we don't care about.
7345 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7347 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7350 if (Op1->uge(TypeBits)) {
7351 if (I.getOpcode() != Instruction::AShr)
7352 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7354 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7359 // ((X*C1) << C2) == (X * (C1 << C2))
7360 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7361 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7362 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7363 return BinaryOperator::CreateMul(BO->getOperand(0),
7364 ConstantExpr::getShl(BOOp, Op1));
7366 // Try to fold constant and into select arguments.
7367 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7368 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7370 if (isa<PHINode>(Op0))
7371 if (Instruction *NV = FoldOpIntoPhi(I))
7374 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7375 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7376 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7377 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7378 // place. Don't try to do this transformation in this case. Also, we
7379 // require that the input operand is a shift-by-constant so that we have
7380 // confidence that the shifts will get folded together. We could do this
7381 // xform in more cases, but it is unlikely to be profitable.
7382 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7383 isa<ConstantInt>(TrOp->getOperand(1))) {
7384 // Okay, we'll do this xform. Make the shift of shift.
7385 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7386 // (shift2 (shift1 & 0x00FF), c2)
7387 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
7389 // For logical shifts, the truncation has the effect of making the high
7390 // part of the register be zeros. Emulate this by inserting an AND to
7391 // clear the top bits as needed. This 'and' will usually be zapped by
7392 // other xforms later if dead.
7393 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7394 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7395 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7397 // The mask we constructed says what the trunc would do if occurring
7398 // between the shifts. We want to know the effect *after* the second
7399 // shift. We know that it is a logical shift by a constant, so adjust the
7400 // mask as appropriate.
7401 if (I.getOpcode() == Instruction::Shl)
7402 MaskV <<= Op1->getZExtValue();
7404 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7405 MaskV = MaskV.lshr(Op1->getZExtValue());
7409 Value *And = Builder->CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7412 // Return the value truncated to the interesting size.
7413 return new TruncInst(And, I.getType());
7417 if (Op0->hasOneUse()) {
7418 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7419 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7422 switch (Op0BO->getOpcode()) {
7424 case Instruction::Add:
7425 case Instruction::And:
7426 case Instruction::Or:
7427 case Instruction::Xor: {
7428 // These operators commute.
7429 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7430 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7431 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7432 m_Specific(Op1)))) {
7433 Value *YS = // (Y << C)
7434 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
7436 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
7437 Op0BO->getOperand(1)->getName());
7438 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7439 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7440 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7443 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7444 Value *Op0BOOp1 = Op0BO->getOperand(1);
7445 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7447 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7448 m_ConstantInt(CC))) &&
7449 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7450 Value *YS = // (Y << C)
7451 Builder->CreateShl(Op0BO->getOperand(0), Op1,
7454 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7455 V1->getName()+".mask");
7456 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7461 case Instruction::Sub: {
7462 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7463 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7464 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7465 m_Specific(Op1)))) {
7466 Value *YS = // (Y << C)
7467 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7469 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
7470 Op0BO->getOperand(0)->getName());
7471 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7472 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7473 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7476 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7477 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7478 match(Op0BO->getOperand(0),
7479 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7480 m_ConstantInt(CC))) && V2 == Op1 &&
7481 cast<BinaryOperator>(Op0BO->getOperand(0))
7482 ->getOperand(0)->hasOneUse()) {
7483 Value *YS = // (Y << C)
7484 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7486 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7487 V1->getName()+".mask");
7489 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7497 // If the operand is an bitwise operator with a constant RHS, and the
7498 // shift is the only use, we can pull it out of the shift.
7499 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7500 bool isValid = true; // Valid only for And, Or, Xor
7501 bool highBitSet = false; // Transform if high bit of constant set?
7503 switch (Op0BO->getOpcode()) {
7504 default: isValid = false; break; // Do not perform transform!
7505 case Instruction::Add:
7506 isValid = isLeftShift;
7508 case Instruction::Or:
7509 case Instruction::Xor:
7512 case Instruction::And:
7517 // If this is a signed shift right, and the high bit is modified
7518 // by the logical operation, do not perform the transformation.
7519 // The highBitSet boolean indicates the value of the high bit of
7520 // the constant which would cause it to be modified for this
7523 if (isValid && I.getOpcode() == Instruction::AShr)
7524 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7527 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7530 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
7531 NewShift->takeName(Op0BO);
7533 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7540 // Find out if this is a shift of a shift by a constant.
7541 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7542 if (ShiftOp && !ShiftOp->isShift())
7545 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7546 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7547 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7548 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7549 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7550 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7551 Value *X = ShiftOp->getOperand(0);
7553 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7555 const IntegerType *Ty = cast<IntegerType>(I.getType());
7557 // Check for (X << c1) << c2 and (X >> c1) >> c2
7558 if (I.getOpcode() == ShiftOp->getOpcode()) {
7559 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7561 if (AmtSum >= TypeBits) {
7562 if (I.getOpcode() != Instruction::AShr)
7563 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7564 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7567 return BinaryOperator::Create(I.getOpcode(), X,
7568 ConstantInt::get(Ty, AmtSum));
7571 if (ShiftOp->getOpcode() == Instruction::LShr &&
7572 I.getOpcode() == Instruction::AShr) {
7573 if (AmtSum >= TypeBits)
7574 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7576 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7577 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7580 if (ShiftOp->getOpcode() == Instruction::AShr &&
7581 I.getOpcode() == Instruction::LShr) {
7582 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7583 if (AmtSum >= TypeBits)
7584 AmtSum = TypeBits-1;
7586 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7588 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7589 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7592 // Okay, if we get here, one shift must be left, and the other shift must be
7593 // right. See if the amounts are equal.
7594 if (ShiftAmt1 == ShiftAmt2) {
7595 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7596 if (I.getOpcode() == Instruction::Shl) {
7597 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7598 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7600 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7601 if (I.getOpcode() == Instruction::LShr) {
7602 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7603 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7605 // We can simplify ((X << C) >>s C) into a trunc + sext.
7606 // NOTE: we could do this for any C, but that would make 'unusual' integer
7607 // types. For now, just stick to ones well-supported by the code
7609 const Type *SExtType = 0;
7610 switch (Ty->getBitWidth() - ShiftAmt1) {
7617 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
7622 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
7623 // Otherwise, we can't handle it yet.
7624 } else if (ShiftAmt1 < ShiftAmt2) {
7625 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7627 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7628 if (I.getOpcode() == Instruction::Shl) {
7629 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7630 ShiftOp->getOpcode() == Instruction::AShr);
7631 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7633 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7634 return BinaryOperator::CreateAnd(Shift,
7635 ConstantInt::get(*Context, Mask));
7638 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7639 if (I.getOpcode() == Instruction::LShr) {
7640 assert(ShiftOp->getOpcode() == Instruction::Shl);
7641 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7643 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7644 return BinaryOperator::CreateAnd(Shift,
7645 ConstantInt::get(*Context, Mask));
7648 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7650 assert(ShiftAmt2 < ShiftAmt1);
7651 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7653 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7654 if (I.getOpcode() == Instruction::Shl) {
7655 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7656 ShiftOp->getOpcode() == Instruction::AShr);
7657 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
7658 ConstantInt::get(Ty, ShiftDiff));
7660 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7661 return BinaryOperator::CreateAnd(Shift,
7662 ConstantInt::get(*Context, Mask));
7665 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7666 if (I.getOpcode() == Instruction::LShr) {
7667 assert(ShiftOp->getOpcode() == Instruction::Shl);
7668 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7670 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7671 return BinaryOperator::CreateAnd(Shift,
7672 ConstantInt::get(*Context, Mask));
7675 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7682 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7683 /// expression. If so, decompose it, returning some value X, such that Val is
7686 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7687 int &Offset, LLVMContext *Context) {
7688 assert(Val->getType() == Type::getInt32Ty(*Context) &&
7689 "Unexpected allocation size type!");
7690 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7691 Offset = CI->getZExtValue();
7693 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
7694 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7695 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7696 if (I->getOpcode() == Instruction::Shl) {
7697 // This is a value scaled by '1 << the shift amt'.
7698 Scale = 1U << RHS->getZExtValue();
7700 return I->getOperand(0);
7701 } else if (I->getOpcode() == Instruction::Mul) {
7702 // This value is scaled by 'RHS'.
7703 Scale = RHS->getZExtValue();
7705 return I->getOperand(0);
7706 } else if (I->getOpcode() == Instruction::Add) {
7707 // We have X+C. Check to see if we really have (X*C2)+C1,
7708 // where C1 is divisible by C2.
7711 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7713 Offset += RHS->getZExtValue();
7720 // Otherwise, we can't look past this.
7727 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7728 /// try to eliminate the cast by moving the type information into the alloc.
7729 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7730 AllocationInst &AI) {
7731 const PointerType *PTy = cast<PointerType>(CI.getType());
7733 BuilderTy AllocaBuilder(*Builder);
7734 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
7736 // Remove any uses of AI that are dead.
7737 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7739 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7740 Instruction *User = cast<Instruction>(*UI++);
7741 if (isInstructionTriviallyDead(User)) {
7742 while (UI != E && *UI == User)
7743 ++UI; // If this instruction uses AI more than once, don't break UI.
7746 DEBUG(errs() << "IC: DCE: " << *User << '\n');
7747 EraseInstFromFunction(*User);
7751 // This requires TargetData to get the alloca alignment and size information.
7754 // Get the type really allocated and the type casted to.
7755 const Type *AllocElTy = AI.getAllocatedType();
7756 const Type *CastElTy = PTy->getElementType();
7757 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7759 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7760 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7761 if (CastElTyAlign < AllocElTyAlign) return 0;
7763 // If the allocation has multiple uses, only promote it if we are strictly
7764 // increasing the alignment of the resultant allocation. If we keep it the
7765 // same, we open the door to infinite loops of various kinds. (A reference
7766 // from a dbg.declare doesn't count as a use for this purpose.)
7767 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7768 CastElTyAlign == AllocElTyAlign) return 0;
7770 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7771 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7772 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7774 // See if we can satisfy the modulus by pulling a scale out of the array
7776 unsigned ArraySizeScale;
7778 Value *NumElements = // See if the array size is a decomposable linear expr.
7779 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7780 ArrayOffset, Context);
7782 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7784 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7785 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7787 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7792 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
7793 // Insert before the alloca, not before the cast.
7794 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
7797 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7798 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
7799 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
7802 AllocationInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
7803 New->setAlignment(AI.getAlignment());
7806 // If the allocation has one real use plus a dbg.declare, just remove the
7808 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7809 EraseInstFromFunction(*DI);
7811 // If the allocation has multiple real uses, insert a cast and change all
7812 // things that used it to use the new cast. This will also hack on CI, but it
7814 else if (!AI.hasOneUse()) {
7815 // New is the allocation instruction, pointer typed. AI is the original
7816 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7817 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
7818 AI.replaceAllUsesWith(NewCast);
7820 return ReplaceInstUsesWith(CI, New);
7823 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7824 /// and return it as type Ty without inserting any new casts and without
7825 /// changing the computed value. This is used by code that tries to decide
7826 /// whether promoting or shrinking integer operations to wider or smaller types
7827 /// will allow us to eliminate a truncate or extend.
7829 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7830 /// extension operation if Ty is larger.
7832 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7833 /// should return true if trunc(V) can be computed by computing V in the smaller
7834 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7835 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7836 /// efficiently truncated.
7838 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7839 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7840 /// the final result.
7841 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7843 int &NumCastsRemoved){
7844 // We can always evaluate constants in another type.
7845 if (isa<Constant>(V))
7848 Instruction *I = dyn_cast<Instruction>(V);
7849 if (!I) return false;
7851 const Type *OrigTy = V->getType();
7853 // If this is an extension or truncate, we can often eliminate it.
7854 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7855 // If this is a cast from the destination type, we can trivially eliminate
7856 // it, and this will remove a cast overall.
7857 if (I->getOperand(0)->getType() == Ty) {
7858 // If the first operand is itself a cast, and is eliminable, do not count
7859 // this as an eliminable cast. We would prefer to eliminate those two
7861 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7867 // We can't extend or shrink something that has multiple uses: doing so would
7868 // require duplicating the instruction in general, which isn't profitable.
7869 if (!I->hasOneUse()) return false;
7871 unsigned Opc = I->getOpcode();
7873 case Instruction::Add:
7874 case Instruction::Sub:
7875 case Instruction::Mul:
7876 case Instruction::And:
7877 case Instruction::Or:
7878 case Instruction::Xor:
7879 // These operators can all arbitrarily be extended or truncated.
7880 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7882 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7885 case Instruction::UDiv:
7886 case Instruction::URem: {
7887 // UDiv and URem can be truncated if all the truncated bits are zero.
7888 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7889 uint32_t BitWidth = Ty->getScalarSizeInBits();
7890 if (BitWidth < OrigBitWidth) {
7891 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7892 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7893 MaskedValueIsZero(I->getOperand(1), Mask)) {
7894 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7896 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7902 case Instruction::Shl:
7903 // If we are truncating the result of this SHL, and if it's a shift of a
7904 // constant amount, we can always perform a SHL in a smaller type.
7905 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7906 uint32_t BitWidth = Ty->getScalarSizeInBits();
7907 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7908 CI->getLimitedValue(BitWidth) < BitWidth)
7909 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7913 case Instruction::LShr:
7914 // If this is a truncate of a logical shr, we can truncate it to a smaller
7915 // lshr iff we know that the bits we would otherwise be shifting in are
7917 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7918 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7919 uint32_t BitWidth = Ty->getScalarSizeInBits();
7920 if (BitWidth < OrigBitWidth &&
7921 MaskedValueIsZero(I->getOperand(0),
7922 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7923 CI->getLimitedValue(BitWidth) < BitWidth) {
7924 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7929 case Instruction::ZExt:
7930 case Instruction::SExt:
7931 case Instruction::Trunc:
7932 // If this is the same kind of case as our original (e.g. zext+zext), we
7933 // can safely replace it. Note that replacing it does not reduce the number
7934 // of casts in the input.
7938 // sext (zext ty1), ty2 -> zext ty2
7939 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7942 case Instruction::Select: {
7943 SelectInst *SI = cast<SelectInst>(I);
7944 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7946 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7949 case Instruction::PHI: {
7950 // We can change a phi if we can change all operands.
7951 PHINode *PN = cast<PHINode>(I);
7952 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7953 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7959 // TODO: Can handle more cases here.
7966 /// EvaluateInDifferentType - Given an expression that
7967 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7968 /// evaluate the expression.
7969 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7971 if (Constant *C = dyn_cast<Constant>(V))
7972 return ConstantExpr::getIntegerCast(C, Ty,
7973 isSigned /*Sext or ZExt*/);
7975 // Otherwise, it must be an instruction.
7976 Instruction *I = cast<Instruction>(V);
7977 Instruction *Res = 0;
7978 unsigned Opc = I->getOpcode();
7980 case Instruction::Add:
7981 case Instruction::Sub:
7982 case Instruction::Mul:
7983 case Instruction::And:
7984 case Instruction::Or:
7985 case Instruction::Xor:
7986 case Instruction::AShr:
7987 case Instruction::LShr:
7988 case Instruction::Shl:
7989 case Instruction::UDiv:
7990 case Instruction::URem: {
7991 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7992 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7993 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
7996 case Instruction::Trunc:
7997 case Instruction::ZExt:
7998 case Instruction::SExt:
7999 // If the source type of the cast is the type we're trying for then we can
8000 // just return the source. There's no need to insert it because it is not
8002 if (I->getOperand(0)->getType() == Ty)
8003 return I->getOperand(0);
8005 // Otherwise, must be the same type of cast, so just reinsert a new one.
8006 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8009 case Instruction::Select: {
8010 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8011 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8012 Res = SelectInst::Create(I->getOperand(0), True, False);
8015 case Instruction::PHI: {
8016 PHINode *OPN = cast<PHINode>(I);
8017 PHINode *NPN = PHINode::Create(Ty);
8018 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8019 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8020 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8026 // TODO: Can handle more cases here.
8027 llvm_unreachable("Unreachable!");
8032 return InsertNewInstBefore(Res, *I);
8035 /// @brief Implement the transforms common to all CastInst visitors.
8036 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8037 Value *Src = CI.getOperand(0);
8039 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8040 // eliminate it now.
8041 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8042 if (Instruction::CastOps opc =
8043 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8044 // The first cast (CSrc) is eliminable so we need to fix up or replace
8045 // the second cast (CI). CSrc will then have a good chance of being dead.
8046 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8050 // If we are casting a select then fold the cast into the select
8051 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8052 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8055 // If we are casting a PHI then fold the cast into the PHI
8056 if (isa<PHINode>(Src))
8057 if (Instruction *NV = FoldOpIntoPhi(CI))
8063 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8064 /// or not there is a sequence of GEP indices into the type that will land us at
8065 /// the specified offset. If so, fill them into NewIndices and return the
8066 /// resultant element type, otherwise return null.
8067 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8068 SmallVectorImpl<Value*> &NewIndices,
8069 const TargetData *TD,
8070 LLVMContext *Context) {
8072 if (!Ty->isSized()) return 0;
8074 // Start with the index over the outer type. Note that the type size
8075 // might be zero (even if the offset isn't zero) if the indexed type
8076 // is something like [0 x {int, int}]
8077 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8078 int64_t FirstIdx = 0;
8079 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8080 FirstIdx = Offset/TySize;
8081 Offset -= FirstIdx*TySize;
8083 // Handle hosts where % returns negative instead of values [0..TySize).
8087 assert(Offset >= 0);
8089 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8092 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8094 // Index into the types. If we fail, set OrigBase to null.
8096 // Indexing into tail padding between struct/array elements.
8097 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8100 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8101 const StructLayout *SL = TD->getStructLayout(STy);
8102 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8103 "Offset must stay within the indexed type");
8105 unsigned Elt = SL->getElementContainingOffset(Offset);
8106 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8108 Offset -= SL->getElementOffset(Elt);
8109 Ty = STy->getElementType(Elt);
8110 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8111 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8112 assert(EltSize && "Cannot index into a zero-sized array");
8113 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8115 Ty = AT->getElementType();
8117 // Otherwise, we can't index into the middle of this atomic type, bail.
8125 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8126 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8127 Value *Src = CI.getOperand(0);
8129 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8130 // If casting the result of a getelementptr instruction with no offset, turn
8131 // this into a cast of the original pointer!
8132 if (GEP->hasAllZeroIndices()) {
8133 // Changing the cast operand is usually not a good idea but it is safe
8134 // here because the pointer operand is being replaced with another
8135 // pointer operand so the opcode doesn't need to change.
8137 CI.setOperand(0, GEP->getOperand(0));
8141 // If the GEP has a single use, and the base pointer is a bitcast, and the
8142 // GEP computes a constant offset, see if we can convert these three
8143 // instructions into fewer. This typically happens with unions and other
8144 // non-type-safe code.
8145 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8146 if (GEP->hasAllConstantIndices()) {
8147 // We are guaranteed to get a constant from EmitGEPOffset.
8148 ConstantInt *OffsetV =
8149 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8150 int64_t Offset = OffsetV->getSExtValue();
8152 // Get the base pointer input of the bitcast, and the type it points to.
8153 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8154 const Type *GEPIdxTy =
8155 cast<PointerType>(OrigBase->getType())->getElementType();
8156 SmallVector<Value*, 8> NewIndices;
8157 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8158 // If we were able to index down into an element, create the GEP
8159 // and bitcast the result. This eliminates one bitcast, potentially
8161 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
8162 Builder->CreateInBoundsGEP(OrigBase,
8163 NewIndices.begin(), NewIndices.end()) :
8164 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
8165 NGEP->takeName(GEP);
8167 if (isa<BitCastInst>(CI))
8168 return new BitCastInst(NGEP, CI.getType());
8169 assert(isa<PtrToIntInst>(CI));
8170 return new PtrToIntInst(NGEP, CI.getType());
8176 return commonCastTransforms(CI);
8179 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8180 /// type like i42. We don't want to introduce operations on random non-legal
8181 /// integer types where they don't already exist in the code. In the future,
8182 /// we should consider making this based off target-data, so that 32-bit targets
8183 /// won't get i64 operations etc.
8184 static bool isSafeIntegerType(const Type *Ty) {
8185 switch (Ty->getPrimitiveSizeInBits()) {
8196 /// commonIntCastTransforms - This function implements the common transforms
8197 /// for trunc, zext, and sext.
8198 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8199 if (Instruction *Result = commonCastTransforms(CI))
8202 Value *Src = CI.getOperand(0);
8203 const Type *SrcTy = Src->getType();
8204 const Type *DestTy = CI.getType();
8205 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8206 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8208 // See if we can simplify any instructions used by the LHS whose sole
8209 // purpose is to compute bits we don't care about.
8210 if (SimplifyDemandedInstructionBits(CI))
8213 // If the source isn't an instruction or has more than one use then we
8214 // can't do anything more.
8215 Instruction *SrcI = dyn_cast<Instruction>(Src);
8216 if (!SrcI || !Src->hasOneUse())
8219 // Attempt to propagate the cast into the instruction for int->int casts.
8220 int NumCastsRemoved = 0;
8221 // Only do this if the dest type is a simple type, don't convert the
8222 // expression tree to something weird like i93 unless the source is also
8224 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8225 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8226 CanEvaluateInDifferentType(SrcI, DestTy,
8227 CI.getOpcode(), NumCastsRemoved)) {
8228 // If this cast is a truncate, evaluting in a different type always
8229 // eliminates the cast, so it is always a win. If this is a zero-extension,
8230 // we need to do an AND to maintain the clear top-part of the computation,
8231 // so we require that the input have eliminated at least one cast. If this
8232 // is a sign extension, we insert two new casts (to do the extension) so we
8233 // require that two casts have been eliminated.
8234 bool DoXForm = false;
8235 bool JustReplace = false;
8236 switch (CI.getOpcode()) {
8238 // All the others use floating point so we shouldn't actually
8239 // get here because of the check above.
8240 llvm_unreachable("Unknown cast type");
8241 case Instruction::Trunc:
8244 case Instruction::ZExt: {
8245 DoXForm = NumCastsRemoved >= 1;
8246 if (!DoXForm && 0) {
8247 // If it's unnecessary to issue an AND to clear the high bits, it's
8248 // always profitable to do this xform.
8249 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8250 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8251 if (MaskedValueIsZero(TryRes, Mask))
8252 return ReplaceInstUsesWith(CI, TryRes);
8254 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8255 if (TryI->use_empty())
8256 EraseInstFromFunction(*TryI);
8260 case Instruction::SExt: {
8261 DoXForm = NumCastsRemoved >= 2;
8262 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8263 // If we do not have to emit the truncate + sext pair, then it's always
8264 // profitable to do this xform.
8266 // It's not safe to eliminate the trunc + sext pair if one of the
8267 // eliminated cast is a truncate. e.g.
8268 // t2 = trunc i32 t1 to i16
8269 // t3 = sext i16 t2 to i32
8272 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8273 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8274 if (NumSignBits > (DestBitSize - SrcBitSize))
8275 return ReplaceInstUsesWith(CI, TryRes);
8277 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8278 if (TryI->use_empty())
8279 EraseInstFromFunction(*TryI);
8286 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8287 " to avoid cast: " << CI);
8288 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8289 CI.getOpcode() == Instruction::SExt);
8291 // Just replace this cast with the result.
8292 return ReplaceInstUsesWith(CI, Res);
8294 assert(Res->getType() == DestTy);
8295 switch (CI.getOpcode()) {
8296 default: llvm_unreachable("Unknown cast type!");
8297 case Instruction::Trunc:
8298 // Just replace this cast with the result.
8299 return ReplaceInstUsesWith(CI, Res);
8300 case Instruction::ZExt: {
8301 assert(SrcBitSize < DestBitSize && "Not a zext?");
8303 // If the high bits are already zero, just replace this cast with the
8305 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8306 if (MaskedValueIsZero(Res, Mask))
8307 return ReplaceInstUsesWith(CI, Res);
8309 // We need to emit an AND to clear the high bits.
8310 Constant *C = ConstantInt::get(*Context,
8311 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8312 return BinaryOperator::CreateAnd(Res, C);
8314 case Instruction::SExt: {
8315 // If the high bits are already filled with sign bit, just replace this
8316 // cast with the result.
8317 unsigned NumSignBits = ComputeNumSignBits(Res);
8318 if (NumSignBits > (DestBitSize - SrcBitSize))
8319 return ReplaceInstUsesWith(CI, Res);
8321 // We need to emit a cast to truncate, then a cast to sext.
8322 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
8328 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8329 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8331 switch (SrcI->getOpcode()) {
8332 case Instruction::Add:
8333 case Instruction::Mul:
8334 case Instruction::And:
8335 case Instruction::Or:
8336 case Instruction::Xor:
8337 // If we are discarding information, rewrite.
8338 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8339 // Don't insert two casts unless at least one can be eliminated.
8340 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8341 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8342 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8343 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8344 return BinaryOperator::Create(
8345 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8349 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8350 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8351 SrcI->getOpcode() == Instruction::Xor &&
8352 Op1 == ConstantInt::getTrue(*Context) &&
8353 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8354 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName());
8355 return BinaryOperator::CreateXor(New,
8356 ConstantInt::get(CI.getType(), 1));
8360 case Instruction::Shl: {
8361 // Canonicalize trunc inside shl, if we can.
8362 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8363 if (CI && DestBitSize < SrcBitSize &&
8364 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8365 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
8366 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
8367 return BinaryOperator::CreateShl(Op0c, Op1c);
8375 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8376 if (Instruction *Result = commonIntCastTransforms(CI))
8379 Value *Src = CI.getOperand(0);
8380 const Type *Ty = CI.getType();
8381 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8382 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8384 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8385 if (DestBitWidth == 1) {
8386 Constant *One = ConstantInt::get(Src->getType(), 1);
8387 Src = Builder->CreateAnd(Src, One, "tmp");
8388 Value *Zero = Constant::getNullValue(Src->getType());
8389 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8392 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8393 ConstantInt *ShAmtV = 0;
8395 if (Src->hasOneUse() &&
8396 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8397 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8399 // Get a mask for the bits shifting in.
8400 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8401 if (MaskedValueIsZero(ShiftOp, Mask)) {
8402 if (ShAmt >= DestBitWidth) // All zeros.
8403 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8405 // Okay, we can shrink this. Truncate the input, then return a new
8407 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName());
8408 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8409 return BinaryOperator::CreateLShr(V1, V2);
8416 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8417 /// in order to eliminate the icmp.
8418 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8420 // If we are just checking for a icmp eq of a single bit and zext'ing it
8421 // to an integer, then shift the bit to the appropriate place and then
8422 // cast to integer to avoid the comparison.
8423 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8424 const APInt &Op1CV = Op1C->getValue();
8426 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8427 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8428 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8429 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8430 if (!DoXform) return ICI;
8432 Value *In = ICI->getOperand(0);
8433 Value *Sh = ConstantInt::get(In->getType(),
8434 In->getType()->getScalarSizeInBits()-1);
8435 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
8436 if (In->getType() != CI.getType())
8437 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
8439 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8440 Constant *One = ConstantInt::get(In->getType(), 1);
8441 In = Builder->CreateXor(In, One, In->getName()+".not");
8444 return ReplaceInstUsesWith(CI, In);
8449 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8450 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8451 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8452 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8453 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8454 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8455 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8456 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8457 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8458 // This only works for EQ and NE
8459 ICI->isEquality()) {
8460 // If Op1C some other power of two, convert:
8461 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8462 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8463 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8464 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8466 APInt KnownZeroMask(~KnownZero);
8467 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8468 if (!DoXform) return ICI;
8470 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8471 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8472 // (X&4) == 2 --> false
8473 // (X&4) != 2 --> true
8474 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8475 Res = ConstantExpr::getZExt(Res, CI.getType());
8476 return ReplaceInstUsesWith(CI, Res);
8479 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8480 Value *In = ICI->getOperand(0);
8482 // Perform a logical shr by shiftamt.
8483 // Insert the shift to put the result in the low bit.
8484 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
8485 In->getName()+".lobit");
8488 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8489 Constant *One = ConstantInt::get(In->getType(), 1);
8490 In = Builder->CreateXor(In, One, "tmp");
8493 if (CI.getType() == In->getType())
8494 return ReplaceInstUsesWith(CI, In);
8496 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8504 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8505 // If one of the common conversion will work ..
8506 if (Instruction *Result = commonIntCastTransforms(CI))
8509 Value *Src = CI.getOperand(0);
8511 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8512 // types and if the sizes are just right we can convert this into a logical
8513 // 'and' which will be much cheaper than the pair of casts.
8514 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8515 // Get the sizes of the types involved. We know that the intermediate type
8516 // will be smaller than A or C, but don't know the relation between A and C.
8517 Value *A = CSrc->getOperand(0);
8518 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8519 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8520 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8521 // If we're actually extending zero bits, then if
8522 // SrcSize < DstSize: zext(a & mask)
8523 // SrcSize == DstSize: a & mask
8524 // SrcSize > DstSize: trunc(a) & mask
8525 if (SrcSize < DstSize) {
8526 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8527 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8528 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
8529 return new ZExtInst(And, CI.getType());
8532 if (SrcSize == DstSize) {
8533 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8534 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8537 if (SrcSize > DstSize) {
8538 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
8539 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8540 return BinaryOperator::CreateAnd(Trunc,
8541 ConstantInt::get(Trunc->getType(),
8546 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8547 return transformZExtICmp(ICI, CI);
8549 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8550 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8551 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8552 // of the (zext icmp) will be transformed.
8553 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8554 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8555 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8556 (transformZExtICmp(LHS, CI, false) ||
8557 transformZExtICmp(RHS, CI, false))) {
8558 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
8559 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
8560 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8564 // zext(trunc(t) & C) -> (t & zext(C)).
8565 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8566 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8567 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8568 Value *TI0 = TI->getOperand(0);
8569 if (TI0->getType() == CI.getType())
8571 BinaryOperator::CreateAnd(TI0,
8572 ConstantExpr::getZExt(C, CI.getType()));
8575 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8576 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8577 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8578 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8579 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8580 And->getOperand(1) == C)
8581 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8582 Value *TI0 = TI->getOperand(0);
8583 if (TI0->getType() == CI.getType()) {
8584 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8585 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
8586 return BinaryOperator::CreateXor(NewAnd, ZC);
8593 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8594 if (Instruction *I = commonIntCastTransforms(CI))
8597 Value *Src = CI.getOperand(0);
8599 // Canonicalize sign-extend from i1 to a select.
8600 if (Src->getType() == Type::getInt1Ty(*Context))
8601 return SelectInst::Create(Src,
8602 Constant::getAllOnesValue(CI.getType()),
8603 Constant::getNullValue(CI.getType()));
8605 // See if the value being truncated is already sign extended. If so, just
8606 // eliminate the trunc/sext pair.
8607 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8608 Value *Op = cast<User>(Src)->getOperand(0);
8609 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8610 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8611 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8612 unsigned NumSignBits = ComputeNumSignBits(Op);
8614 if (OpBits == DestBits) {
8615 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8616 // bits, it is already ready.
8617 if (NumSignBits > DestBits-MidBits)
8618 return ReplaceInstUsesWith(CI, Op);
8619 } else if (OpBits < DestBits) {
8620 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8621 // bits, just sext from i32.
8622 if (NumSignBits > OpBits-MidBits)
8623 return new SExtInst(Op, CI.getType(), "tmp");
8625 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8626 // bits, just truncate to i32.
8627 if (NumSignBits > OpBits-MidBits)
8628 return new TruncInst(Op, CI.getType(), "tmp");
8632 // If the input is a shl/ashr pair of a same constant, then this is a sign
8633 // extension from a smaller value. If we could trust arbitrary bitwidth
8634 // integers, we could turn this into a truncate to the smaller bit and then
8635 // use a sext for the whole extension. Since we don't, look deeper and check
8636 // for a truncate. If the source and dest are the same type, eliminate the
8637 // trunc and extend and just do shifts. For example, turn:
8638 // %a = trunc i32 %i to i8
8639 // %b = shl i8 %a, 6
8640 // %c = ashr i8 %b, 6
8641 // %d = sext i8 %c to i32
8643 // %a = shl i32 %i, 30
8644 // %d = ashr i32 %a, 30
8646 ConstantInt *BA = 0, *CA = 0;
8647 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8648 m_ConstantInt(CA))) &&
8649 BA == CA && isa<TruncInst>(A)) {
8650 Value *I = cast<TruncInst>(A)->getOperand(0);
8651 if (I->getType() == CI.getType()) {
8652 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8653 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8654 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8655 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8656 I = Builder->CreateShl(I, ShAmtV, CI.getName());
8657 return BinaryOperator::CreateAShr(I, ShAmtV);
8664 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8665 /// in the specified FP type without changing its value.
8666 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8667 LLVMContext *Context) {
8669 APFloat F = CFP->getValueAPF();
8670 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8672 return ConstantFP::get(*Context, F);
8676 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8677 /// through it until we get the source value.
8678 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8679 if (Instruction *I = dyn_cast<Instruction>(V))
8680 if (I->getOpcode() == Instruction::FPExt)
8681 return LookThroughFPExtensions(I->getOperand(0), Context);
8683 // If this value is a constant, return the constant in the smallest FP type
8684 // that can accurately represent it. This allows us to turn
8685 // (float)((double)X+2.0) into x+2.0f.
8686 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8687 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
8688 return V; // No constant folding of this.
8689 // See if the value can be truncated to float and then reextended.
8690 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8692 if (CFP->getType() == Type::getDoubleTy(*Context))
8693 return V; // Won't shrink.
8694 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8696 // Don't try to shrink to various long double types.
8702 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8703 if (Instruction *I = commonCastTransforms(CI))
8706 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8707 // smaller than the destination type, we can eliminate the truncate by doing
8708 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8709 // many builtins (sqrt, etc).
8710 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8711 if (OpI && OpI->hasOneUse()) {
8712 switch (OpI->getOpcode()) {
8714 case Instruction::FAdd:
8715 case Instruction::FSub:
8716 case Instruction::FMul:
8717 case Instruction::FDiv:
8718 case Instruction::FRem:
8719 const Type *SrcTy = OpI->getType();
8720 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8721 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8722 if (LHSTrunc->getType() != SrcTy &&
8723 RHSTrunc->getType() != SrcTy) {
8724 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8725 // If the source types were both smaller than the destination type of
8726 // the cast, do this xform.
8727 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8728 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8729 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
8730 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
8731 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8740 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8741 return commonCastTransforms(CI);
8744 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8745 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8747 return commonCastTransforms(FI);
8749 // fptoui(uitofp(X)) --> X
8750 // fptoui(sitofp(X)) --> X
8751 // This is safe if the intermediate type has enough bits in its mantissa to
8752 // accurately represent all values of X. For example, do not do this with
8753 // i64->float->i64. This is also safe for sitofp case, because any negative
8754 // 'X' value would cause an undefined result for the fptoui.
8755 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8756 OpI->getOperand(0)->getType() == FI.getType() &&
8757 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8758 OpI->getType()->getFPMantissaWidth())
8759 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8761 return commonCastTransforms(FI);
8764 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8765 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8767 return commonCastTransforms(FI);
8769 // fptosi(sitofp(X)) --> X
8770 // fptosi(uitofp(X)) --> X
8771 // This is safe if the intermediate type has enough bits in its mantissa to
8772 // accurately represent all values of X. For example, do not do this with
8773 // i64->float->i64. This is also safe for sitofp case, because any negative
8774 // 'X' value would cause an undefined result for the fptoui.
8775 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8776 OpI->getOperand(0)->getType() == FI.getType() &&
8777 (int)FI.getType()->getScalarSizeInBits() <=
8778 OpI->getType()->getFPMantissaWidth())
8779 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8781 return commonCastTransforms(FI);
8784 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8785 return commonCastTransforms(CI);
8788 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8789 return commonCastTransforms(CI);
8792 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8793 // If the destination integer type is smaller than the intptr_t type for
8794 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8795 // trunc to be exposed to other transforms. Don't do this for extending
8796 // ptrtoint's, because we don't know if the target sign or zero extends its
8799 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8800 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
8801 TD->getIntPtrType(CI.getContext()),
8803 return new TruncInst(P, CI.getType());
8806 return commonPointerCastTransforms(CI);
8809 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8810 // If the source integer type is larger than the intptr_t type for
8811 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8812 // allows the trunc to be exposed to other transforms. Don't do this for
8813 // extending inttoptr's, because we don't know if the target sign or zero
8814 // extends to pointers.
8815 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
8816 TD->getPointerSizeInBits()) {
8817 Value *P = Builder->CreateTrunc(CI.getOperand(0),
8818 TD->getIntPtrType(CI.getContext()), "tmp");
8819 return new IntToPtrInst(P, CI.getType());
8822 if (Instruction *I = commonCastTransforms(CI))
8828 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8829 // If the operands are integer typed then apply the integer transforms,
8830 // otherwise just apply the common ones.
8831 Value *Src = CI.getOperand(0);
8832 const Type *SrcTy = Src->getType();
8833 const Type *DestTy = CI.getType();
8835 if (isa<PointerType>(SrcTy)) {
8836 if (Instruction *I = commonPointerCastTransforms(CI))
8839 if (Instruction *Result = commonCastTransforms(CI))
8844 // Get rid of casts from one type to the same type. These are useless and can
8845 // be replaced by the operand.
8846 if (DestTy == Src->getType())
8847 return ReplaceInstUsesWith(CI, Src);
8849 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8850 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8851 const Type *DstElTy = DstPTy->getElementType();
8852 const Type *SrcElTy = SrcPTy->getElementType();
8854 // If the address spaces don't match, don't eliminate the bitcast, which is
8855 // required for changing types.
8856 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8859 // If we are casting a alloca to a pointer to a type of the same
8860 // size, rewrite the allocation instruction to allocate the "right" type.
8861 // There is no need to modify malloc calls because it is their bitcast that
8862 // needs to be cleaned up.
8863 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8864 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8867 // If the source and destination are pointers, and this cast is equivalent
8868 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8869 // This can enhance SROA and other transforms that want type-safe pointers.
8870 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
8871 unsigned NumZeros = 0;
8872 while (SrcElTy != DstElTy &&
8873 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8874 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8875 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8879 // If we found a path from the src to dest, create the getelementptr now.
8880 if (SrcElTy == DstElTy) {
8881 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8882 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(), "",
8883 ((Instruction*) NULL));
8887 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
8888 if (DestVTy->getNumElements() == 1) {
8889 if (!isa<VectorType>(SrcTy)) {
8890 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
8891 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
8892 Constant::getNullValue(Type::getInt32Ty(*Context)));
8894 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
8898 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
8899 if (SrcVTy->getNumElements() == 1) {
8900 if (!isa<VectorType>(DestTy)) {
8902 Builder->CreateExtractElement(Src,
8903 Constant::getNullValue(Type::getInt32Ty(*Context)));
8904 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
8909 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8910 if (SVI->hasOneUse()) {
8911 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8912 // a bitconvert to a vector with the same # elts.
8913 if (isa<VectorType>(DestTy) &&
8914 cast<VectorType>(DestTy)->getNumElements() ==
8915 SVI->getType()->getNumElements() &&
8916 SVI->getType()->getNumElements() ==
8917 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8919 // If either of the operands is a cast from CI.getType(), then
8920 // evaluating the shuffle in the casted destination's type will allow
8921 // us to eliminate at least one cast.
8922 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8923 Tmp->getOperand(0)->getType() == DestTy) ||
8924 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8925 Tmp->getOperand(0)->getType() == DestTy)) {
8926 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
8927 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
8928 // Return a new shuffle vector. Use the same element ID's, as we
8929 // know the vector types match #elts.
8930 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8938 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8940 /// %D = select %cond, %C, %A
8942 /// %C = select %cond, %B, 0
8945 /// Assuming that the specified instruction is an operand to the select, return
8946 /// a bitmask indicating which operands of this instruction are foldable if they
8947 /// equal the other incoming value of the select.
8949 static unsigned GetSelectFoldableOperands(Instruction *I) {
8950 switch (I->getOpcode()) {
8951 case Instruction::Add:
8952 case Instruction::Mul:
8953 case Instruction::And:
8954 case Instruction::Or:
8955 case Instruction::Xor:
8956 return 3; // Can fold through either operand.
8957 case Instruction::Sub: // Can only fold on the amount subtracted.
8958 case Instruction::Shl: // Can only fold on the shift amount.
8959 case Instruction::LShr:
8960 case Instruction::AShr:
8963 return 0; // Cannot fold
8967 /// GetSelectFoldableConstant - For the same transformation as the previous
8968 /// function, return the identity constant that goes into the select.
8969 static Constant *GetSelectFoldableConstant(Instruction *I,
8970 LLVMContext *Context) {
8971 switch (I->getOpcode()) {
8972 default: llvm_unreachable("This cannot happen!");
8973 case Instruction::Add:
8974 case Instruction::Sub:
8975 case Instruction::Or:
8976 case Instruction::Xor:
8977 case Instruction::Shl:
8978 case Instruction::LShr:
8979 case Instruction::AShr:
8980 return Constant::getNullValue(I->getType());
8981 case Instruction::And:
8982 return Constant::getAllOnesValue(I->getType());
8983 case Instruction::Mul:
8984 return ConstantInt::get(I->getType(), 1);
8988 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8989 /// have the same opcode and only one use each. Try to simplify this.
8990 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8992 if (TI->getNumOperands() == 1) {
8993 // If this is a non-volatile load or a cast from the same type,
8996 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8999 return 0; // unknown unary op.
9002 // Fold this by inserting a select from the input values.
9003 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9004 FI->getOperand(0), SI.getName()+".v");
9005 InsertNewInstBefore(NewSI, SI);
9006 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9010 // Only handle binary operators here.
9011 if (!isa<BinaryOperator>(TI))
9014 // Figure out if the operations have any operands in common.
9015 Value *MatchOp, *OtherOpT, *OtherOpF;
9017 if (TI->getOperand(0) == FI->getOperand(0)) {
9018 MatchOp = TI->getOperand(0);
9019 OtherOpT = TI->getOperand(1);
9020 OtherOpF = FI->getOperand(1);
9021 MatchIsOpZero = true;
9022 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9023 MatchOp = TI->getOperand(1);
9024 OtherOpT = TI->getOperand(0);
9025 OtherOpF = FI->getOperand(0);
9026 MatchIsOpZero = false;
9027 } else if (!TI->isCommutative()) {
9029 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9030 MatchOp = TI->getOperand(0);
9031 OtherOpT = TI->getOperand(1);
9032 OtherOpF = FI->getOperand(0);
9033 MatchIsOpZero = true;
9034 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9035 MatchOp = TI->getOperand(1);
9036 OtherOpT = TI->getOperand(0);
9037 OtherOpF = FI->getOperand(1);
9038 MatchIsOpZero = true;
9043 // If we reach here, they do have operations in common.
9044 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9045 OtherOpF, SI.getName()+".v");
9046 InsertNewInstBefore(NewSI, SI);
9048 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9050 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9052 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9054 llvm_unreachable("Shouldn't get here");
9058 static bool isSelect01(Constant *C1, Constant *C2) {
9059 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9062 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9065 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9068 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9069 /// facilitate further optimization.
9070 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9072 // See the comment above GetSelectFoldableOperands for a description of the
9073 // transformation we are doing here.
9074 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9075 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9076 !isa<Constant>(FalseVal)) {
9077 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9078 unsigned OpToFold = 0;
9079 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9081 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9086 Constant *C = GetSelectFoldableConstant(TVI, Context);
9087 Value *OOp = TVI->getOperand(2-OpToFold);
9088 // Avoid creating select between 2 constants unless it's selecting
9090 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9091 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9092 InsertNewInstBefore(NewSel, SI);
9093 NewSel->takeName(TVI);
9094 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9095 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9096 llvm_unreachable("Unknown instruction!!");
9103 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9104 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9105 !isa<Constant>(TrueVal)) {
9106 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9107 unsigned OpToFold = 0;
9108 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9110 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9115 Constant *C = GetSelectFoldableConstant(FVI, Context);
9116 Value *OOp = FVI->getOperand(2-OpToFold);
9117 // Avoid creating select between 2 constants unless it's selecting
9119 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9120 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9121 InsertNewInstBefore(NewSel, SI);
9122 NewSel->takeName(FVI);
9123 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9124 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9125 llvm_unreachable("Unknown instruction!!");
9135 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9136 /// ICmpInst as its first operand.
9138 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9140 bool Changed = false;
9141 ICmpInst::Predicate Pred = ICI->getPredicate();
9142 Value *CmpLHS = ICI->getOperand(0);
9143 Value *CmpRHS = ICI->getOperand(1);
9144 Value *TrueVal = SI.getTrueValue();
9145 Value *FalseVal = SI.getFalseValue();
9147 // Check cases where the comparison is with a constant that
9148 // can be adjusted to fit the min/max idiom. We may edit ICI in
9149 // place here, so make sure the select is the only user.
9150 if (ICI->hasOneUse())
9151 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9154 case ICmpInst::ICMP_ULT:
9155 case ICmpInst::ICMP_SLT: {
9156 // X < MIN ? T : F --> F
9157 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9158 return ReplaceInstUsesWith(SI, FalseVal);
9159 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9160 Constant *AdjustedRHS = SubOne(CI);
9161 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9162 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9163 Pred = ICmpInst::getSwappedPredicate(Pred);
9164 CmpRHS = AdjustedRHS;
9165 std::swap(FalseVal, TrueVal);
9166 ICI->setPredicate(Pred);
9167 ICI->setOperand(1, CmpRHS);
9168 SI.setOperand(1, TrueVal);
9169 SI.setOperand(2, FalseVal);
9174 case ICmpInst::ICMP_UGT:
9175 case ICmpInst::ICMP_SGT: {
9176 // X > MAX ? T : F --> F
9177 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9178 return ReplaceInstUsesWith(SI, FalseVal);
9179 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9180 Constant *AdjustedRHS = AddOne(CI);
9181 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9182 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9183 Pred = ICmpInst::getSwappedPredicate(Pred);
9184 CmpRHS = AdjustedRHS;
9185 std::swap(FalseVal, TrueVal);
9186 ICI->setPredicate(Pred);
9187 ICI->setOperand(1, CmpRHS);
9188 SI.setOperand(1, TrueVal);
9189 SI.setOperand(2, FalseVal);
9196 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9197 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9198 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9199 if (match(TrueVal, m_ConstantInt<-1>()) &&
9200 match(FalseVal, m_ConstantInt<0>()))
9201 Pred = ICI->getPredicate();
9202 else if (match(TrueVal, m_ConstantInt<0>()) &&
9203 match(FalseVal, m_ConstantInt<-1>()))
9204 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9206 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9207 // If we are just checking for a icmp eq of a single bit and zext'ing it
9208 // to an integer, then shift the bit to the appropriate place and then
9209 // cast to integer to avoid the comparison.
9210 const APInt &Op1CV = CI->getValue();
9212 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9213 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9214 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9215 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9216 Value *In = ICI->getOperand(0);
9217 Value *Sh = ConstantInt::get(In->getType(),
9218 In->getType()->getScalarSizeInBits()-1);
9219 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9220 In->getName()+".lobit"),
9222 if (In->getType() != SI.getType())
9223 In = CastInst::CreateIntegerCast(In, SI.getType(),
9224 true/*SExt*/, "tmp", ICI);
9226 if (Pred == ICmpInst::ICMP_SGT)
9227 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9228 In->getName()+".not"), *ICI);
9230 return ReplaceInstUsesWith(SI, In);
9235 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9236 // Transform (X == Y) ? X : Y -> Y
9237 if (Pred == ICmpInst::ICMP_EQ)
9238 return ReplaceInstUsesWith(SI, FalseVal);
9239 // Transform (X != Y) ? X : Y -> X
9240 if (Pred == ICmpInst::ICMP_NE)
9241 return ReplaceInstUsesWith(SI, TrueVal);
9242 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9244 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9245 // Transform (X == Y) ? Y : X -> X
9246 if (Pred == ICmpInst::ICMP_EQ)
9247 return ReplaceInstUsesWith(SI, FalseVal);
9248 // Transform (X != Y) ? Y : X -> Y
9249 if (Pred == ICmpInst::ICMP_NE)
9250 return ReplaceInstUsesWith(SI, TrueVal);
9251 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9254 /// NOTE: if we wanted to, this is where to detect integer ABS
9256 return Changed ? &SI : 0;
9259 /// isDefinedInBB - Return true if the value is an instruction defined in the
9260 /// specified basicblock.
9261 static bool isDefinedInBB(const Value *V, const BasicBlock *BB) {
9262 const Instruction *I = dyn_cast<Instruction>(V);
9263 return I != 0 && I->getParent() == BB;
9267 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9268 Value *CondVal = SI.getCondition();
9269 Value *TrueVal = SI.getTrueValue();
9270 Value *FalseVal = SI.getFalseValue();
9272 // select true, X, Y -> X
9273 // select false, X, Y -> Y
9274 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9275 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9277 // select C, X, X -> X
9278 if (TrueVal == FalseVal)
9279 return ReplaceInstUsesWith(SI, TrueVal);
9281 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9282 return ReplaceInstUsesWith(SI, FalseVal);
9283 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9284 return ReplaceInstUsesWith(SI, TrueVal);
9285 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9286 if (isa<Constant>(TrueVal))
9287 return ReplaceInstUsesWith(SI, TrueVal);
9289 return ReplaceInstUsesWith(SI, FalseVal);
9292 if (SI.getType() == Type::getInt1Ty(*Context)) {
9293 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9294 if (C->getZExtValue()) {
9295 // Change: A = select B, true, C --> A = or B, C
9296 return BinaryOperator::CreateOr(CondVal, FalseVal);
9298 // Change: A = select B, false, C --> A = and !B, C
9300 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9301 "not."+CondVal->getName()), SI);
9302 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9304 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9305 if (C->getZExtValue() == false) {
9306 // Change: A = select B, C, false --> A = and B, C
9307 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9309 // Change: A = select B, C, true --> A = or !B, C
9311 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9312 "not."+CondVal->getName()), SI);
9313 return BinaryOperator::CreateOr(NotCond, TrueVal);
9317 // select a, b, a -> a&b
9318 // select a, a, b -> a|b
9319 if (CondVal == TrueVal)
9320 return BinaryOperator::CreateOr(CondVal, FalseVal);
9321 else if (CondVal == FalseVal)
9322 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9325 // Selecting between two integer constants?
9326 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9327 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9328 // select C, 1, 0 -> zext C to int
9329 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9330 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9331 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9332 // select C, 0, 1 -> zext !C to int
9334 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9335 "not."+CondVal->getName()), SI);
9336 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9339 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9340 // If one of the constants is zero (we know they can't both be) and we
9341 // have an icmp instruction with zero, and we have an 'and' with the
9342 // non-constant value, eliminate this whole mess. This corresponds to
9343 // cases like this: ((X & 27) ? 27 : 0)
9344 if (TrueValC->isZero() || FalseValC->isZero())
9345 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9346 cast<Constant>(IC->getOperand(1))->isNullValue())
9347 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9348 if (ICA->getOpcode() == Instruction::And &&
9349 isa<ConstantInt>(ICA->getOperand(1)) &&
9350 (ICA->getOperand(1) == TrueValC ||
9351 ICA->getOperand(1) == FalseValC) &&
9352 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9353 // Okay, now we know that everything is set up, we just don't
9354 // know whether we have a icmp_ne or icmp_eq and whether the
9355 // true or false val is the zero.
9356 bool ShouldNotVal = !TrueValC->isZero();
9357 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9360 V = InsertNewInstBefore(BinaryOperator::Create(
9361 Instruction::Xor, V, ICA->getOperand(1)), SI);
9362 return ReplaceInstUsesWith(SI, V);
9367 // See if we are selecting two values based on a comparison of the two values.
9368 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9369 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9370 // Transform (X == Y) ? X : Y -> Y
9371 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9372 // This is not safe in general for floating point:
9373 // consider X== -0, Y== +0.
9374 // It becomes safe if either operand is a nonzero constant.
9375 ConstantFP *CFPt, *CFPf;
9376 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9377 !CFPt->getValueAPF().isZero()) ||
9378 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9379 !CFPf->getValueAPF().isZero()))
9380 return ReplaceInstUsesWith(SI, FalseVal);
9382 // Transform (X != Y) ? X : Y -> X
9383 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9384 return ReplaceInstUsesWith(SI, TrueVal);
9385 // NOTE: if we wanted to, this is where to detect MIN/MAX
9387 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9388 // Transform (X == Y) ? Y : X -> X
9389 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9390 // This is not safe in general for floating point:
9391 // consider X== -0, Y== +0.
9392 // It becomes safe if either operand is a nonzero constant.
9393 ConstantFP *CFPt, *CFPf;
9394 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9395 !CFPt->getValueAPF().isZero()) ||
9396 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9397 !CFPf->getValueAPF().isZero()))
9398 return ReplaceInstUsesWith(SI, FalseVal);
9400 // Transform (X != Y) ? Y : X -> Y
9401 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9402 return ReplaceInstUsesWith(SI, TrueVal);
9403 // NOTE: if we wanted to, this is where to detect MIN/MAX
9405 // NOTE: if we wanted to, this is where to detect ABS
9408 // See if we are selecting two values based on a comparison of the two values.
9409 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9410 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9413 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9414 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9415 if (TI->hasOneUse() && FI->hasOneUse()) {
9416 Instruction *AddOp = 0, *SubOp = 0;
9418 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9419 if (TI->getOpcode() == FI->getOpcode())
9420 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9423 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9424 // even legal for FP.
9425 if ((TI->getOpcode() == Instruction::Sub &&
9426 FI->getOpcode() == Instruction::Add) ||
9427 (TI->getOpcode() == Instruction::FSub &&
9428 FI->getOpcode() == Instruction::FAdd)) {
9429 AddOp = FI; SubOp = TI;
9430 } else if ((FI->getOpcode() == Instruction::Sub &&
9431 TI->getOpcode() == Instruction::Add) ||
9432 (FI->getOpcode() == Instruction::FSub &&
9433 TI->getOpcode() == Instruction::FAdd)) {
9434 AddOp = TI; SubOp = FI;
9438 Value *OtherAddOp = 0;
9439 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9440 OtherAddOp = AddOp->getOperand(1);
9441 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9442 OtherAddOp = AddOp->getOperand(0);
9446 // So at this point we know we have (Y -> OtherAddOp):
9447 // select C, (add X, Y), (sub X, Z)
9448 Value *NegVal; // Compute -Z
9449 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9450 NegVal = ConstantExpr::getNeg(C);
9452 NegVal = InsertNewInstBefore(
9453 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9457 Value *NewTrueOp = OtherAddOp;
9458 Value *NewFalseOp = NegVal;
9460 std::swap(NewTrueOp, NewFalseOp);
9461 Instruction *NewSel =
9462 SelectInst::Create(CondVal, NewTrueOp,
9463 NewFalseOp, SI.getName() + ".p");
9465 NewSel = InsertNewInstBefore(NewSel, SI);
9466 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9471 // See if we can fold the select into one of our operands.
9472 if (SI.getType()->isInteger()) {
9473 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9478 // See if we can fold the select into a phi node. The true/false values have
9479 // to be live in the predecessor blocks. If they are instructions in SI's
9480 // block, we can't map to the predecessor.
9481 if (isa<PHINode>(SI.getCondition()) &&
9482 (!isDefinedInBB(SI.getTrueValue(), SI.getParent()) ||
9483 isa<PHINode>(SI.getTrueValue())) &&
9484 (!isDefinedInBB(SI.getFalseValue(), SI.getParent()) ||
9485 isa<PHINode>(SI.getFalseValue())))
9486 if (Instruction *NV = FoldOpIntoPhi(SI))
9489 if (BinaryOperator::isNot(CondVal)) {
9490 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9491 SI.setOperand(1, FalseVal);
9492 SI.setOperand(2, TrueVal);
9499 /// EnforceKnownAlignment - If the specified pointer points to an object that
9500 /// we control, modify the object's alignment to PrefAlign. This isn't
9501 /// often possible though. If alignment is important, a more reliable approach
9502 /// is to simply align all global variables and allocation instructions to
9503 /// their preferred alignment from the beginning.
9505 static unsigned EnforceKnownAlignment(Value *V,
9506 unsigned Align, unsigned PrefAlign) {
9508 User *U = dyn_cast<User>(V);
9509 if (!U) return Align;
9511 switch (Operator::getOpcode(U)) {
9513 case Instruction::BitCast:
9514 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9515 case Instruction::GetElementPtr: {
9516 // If all indexes are zero, it is just the alignment of the base pointer.
9517 bool AllZeroOperands = true;
9518 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9519 if (!isa<Constant>(*i) ||
9520 !cast<Constant>(*i)->isNullValue()) {
9521 AllZeroOperands = false;
9525 if (AllZeroOperands) {
9526 // Treat this like a bitcast.
9527 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9533 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9534 // If there is a large requested alignment and we can, bump up the alignment
9536 if (!GV->isDeclaration()) {
9537 if (GV->getAlignment() >= PrefAlign)
9538 Align = GV->getAlignment();
9540 GV->setAlignment(PrefAlign);
9544 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
9545 // If there is a requested alignment and if this is an alloca, round up.
9546 if (AI->getAlignment() >= PrefAlign)
9547 Align = AI->getAlignment();
9549 AI->setAlignment(PrefAlign);
9557 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9558 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9559 /// and it is more than the alignment of the ultimate object, see if we can
9560 /// increase the alignment of the ultimate object, making this check succeed.
9561 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9562 unsigned PrefAlign) {
9563 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9564 sizeof(PrefAlign) * CHAR_BIT;
9565 APInt Mask = APInt::getAllOnesValue(BitWidth);
9566 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9567 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9568 unsigned TrailZ = KnownZero.countTrailingOnes();
9569 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9571 if (PrefAlign > Align)
9572 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9574 // We don't need to make any adjustment.
9578 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9579 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9580 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9581 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9582 unsigned CopyAlign = MI->getAlignment();
9584 if (CopyAlign < MinAlign) {
9585 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9590 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9592 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9593 if (MemOpLength == 0) return 0;
9595 // Source and destination pointer types are always "i8*" for intrinsic. See
9596 // if the size is something we can handle with a single primitive load/store.
9597 // A single load+store correctly handles overlapping memory in the memmove
9599 unsigned Size = MemOpLength->getZExtValue();
9600 if (Size == 0) return MI; // Delete this mem transfer.
9602 if (Size > 8 || (Size&(Size-1)))
9603 return 0; // If not 1/2/4/8 bytes, exit.
9605 // Use an integer load+store unless we can find something better.
9607 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
9609 // Memcpy forces the use of i8* for the source and destination. That means
9610 // that if you're using memcpy to move one double around, you'll get a cast
9611 // from double* to i8*. We'd much rather use a double load+store rather than
9612 // an i64 load+store, here because this improves the odds that the source or
9613 // dest address will be promotable. See if we can find a better type than the
9614 // integer datatype.
9615 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9616 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9617 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9618 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9619 // down through these levels if so.
9620 while (!SrcETy->isSingleValueType()) {
9621 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9622 if (STy->getNumElements() == 1)
9623 SrcETy = STy->getElementType(0);
9626 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9627 if (ATy->getNumElements() == 1)
9628 SrcETy = ATy->getElementType();
9635 if (SrcETy->isSingleValueType())
9636 NewPtrTy = PointerType::getUnqual(SrcETy);
9641 // If the memcpy/memmove provides better alignment info than we can
9643 SrcAlign = std::max(SrcAlign, CopyAlign);
9644 DstAlign = std::max(DstAlign, CopyAlign);
9646 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy);
9647 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy);
9648 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9649 InsertNewInstBefore(L, *MI);
9650 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9652 // Set the size of the copy to 0, it will be deleted on the next iteration.
9653 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9657 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9658 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9659 if (MI->getAlignment() < Alignment) {
9660 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9665 // Extract the length and alignment and fill if they are constant.
9666 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9667 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9668 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
9670 uint64_t Len = LenC->getZExtValue();
9671 Alignment = MI->getAlignment();
9673 // If the length is zero, this is a no-op
9674 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9676 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9677 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9678 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
9680 Value *Dest = MI->getDest();
9681 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
9683 // Alignment 0 is identity for alignment 1 for memset, but not store.
9684 if (Alignment == 0) Alignment = 1;
9686 // Extract the fill value and store.
9687 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9688 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9689 Dest, false, Alignment), *MI);
9691 // Set the size of the copy to 0, it will be deleted on the next iteration.
9692 MI->setLength(Constant::getNullValue(LenC->getType()));
9700 /// visitCallInst - CallInst simplification. This mostly only handles folding
9701 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9702 /// the heavy lifting.
9704 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9705 // If the caller function is nounwind, mark the call as nounwind, even if the
9707 if (CI.getParent()->getParent()->doesNotThrow() &&
9708 !CI.doesNotThrow()) {
9709 CI.setDoesNotThrow();
9713 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9714 if (!II) return visitCallSite(&CI);
9716 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9718 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9719 bool Changed = false;
9721 // memmove/cpy/set of zero bytes is a noop.
9722 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9723 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9725 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9726 if (CI->getZExtValue() == 1) {
9727 // Replace the instruction with just byte operations. We would
9728 // transform other cases to loads/stores, but we don't know if
9729 // alignment is sufficient.
9733 // If we have a memmove and the source operation is a constant global,
9734 // then the source and dest pointers can't alias, so we can change this
9735 // into a call to memcpy.
9736 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9737 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9738 if (GVSrc->isConstant()) {
9739 Module *M = CI.getParent()->getParent()->getParent();
9740 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9742 Tys[0] = CI.getOperand(3)->getType();
9744 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9748 // memmove(x,x,size) -> noop.
9749 if (MMI->getSource() == MMI->getDest())
9750 return EraseInstFromFunction(CI);
9753 // If we can determine a pointer alignment that is bigger than currently
9754 // set, update the alignment.
9755 if (isa<MemTransferInst>(MI)) {
9756 if (Instruction *I = SimplifyMemTransfer(MI))
9758 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9759 if (Instruction *I = SimplifyMemSet(MSI))
9763 if (Changed) return II;
9766 switch (II->getIntrinsicID()) {
9768 case Intrinsic::bswap:
9769 // bswap(bswap(x)) -> x
9770 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9771 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9772 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9774 case Intrinsic::ppc_altivec_lvx:
9775 case Intrinsic::ppc_altivec_lvxl:
9776 case Intrinsic::x86_sse_loadu_ps:
9777 case Intrinsic::x86_sse2_loadu_pd:
9778 case Intrinsic::x86_sse2_loadu_dq:
9779 // Turn PPC lvx -> load if the pointer is known aligned.
9780 // Turn X86 loadups -> load if the pointer is known aligned.
9781 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9782 Value *Ptr = Builder->CreateBitCast(II->getOperand(1),
9783 PointerType::getUnqual(II->getType()));
9784 return new LoadInst(Ptr);
9787 case Intrinsic::ppc_altivec_stvx:
9788 case Intrinsic::ppc_altivec_stvxl:
9789 // Turn stvx -> store if the pointer is known aligned.
9790 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9791 const Type *OpPtrTy =
9792 PointerType::getUnqual(II->getOperand(1)->getType());
9793 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy);
9794 return new StoreInst(II->getOperand(1), Ptr);
9797 case Intrinsic::x86_sse_storeu_ps:
9798 case Intrinsic::x86_sse2_storeu_pd:
9799 case Intrinsic::x86_sse2_storeu_dq:
9800 // Turn X86 storeu -> store if the pointer is known aligned.
9801 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9802 const Type *OpPtrTy =
9803 PointerType::getUnqual(II->getOperand(2)->getType());
9804 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
9805 return new StoreInst(II->getOperand(2), Ptr);
9809 case Intrinsic::x86_sse_cvttss2si: {
9810 // These intrinsics only demands the 0th element of its input vector. If
9811 // we can simplify the input based on that, do so now.
9813 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9814 APInt DemandedElts(VWidth, 1);
9815 APInt UndefElts(VWidth, 0);
9816 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9818 II->setOperand(1, V);
9824 case Intrinsic::ppc_altivec_vperm:
9825 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9826 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9827 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9829 // Check that all of the elements are integer constants or undefs.
9830 bool AllEltsOk = true;
9831 for (unsigned i = 0; i != 16; ++i) {
9832 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9833 !isa<UndefValue>(Mask->getOperand(i))) {
9840 // Cast the input vectors to byte vectors.
9841 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
9842 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType());
9843 Value *Result = UndefValue::get(Op0->getType());
9845 // Only extract each element once.
9846 Value *ExtractedElts[32];
9847 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9849 for (unsigned i = 0; i != 16; ++i) {
9850 if (isa<UndefValue>(Mask->getOperand(i)))
9852 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9853 Idx &= 31; // Match the hardware behavior.
9855 if (ExtractedElts[Idx] == 0) {
9856 ExtractedElts[Idx] =
9857 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
9858 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false),
9862 // Insert this value into the result vector.
9863 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
9864 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
9867 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9872 case Intrinsic::stackrestore: {
9873 // If the save is right next to the restore, remove the restore. This can
9874 // happen when variable allocas are DCE'd.
9875 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9876 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9877 BasicBlock::iterator BI = SS;
9879 return EraseInstFromFunction(CI);
9883 // Scan down this block to see if there is another stack restore in the
9884 // same block without an intervening call/alloca.
9885 BasicBlock::iterator BI = II;
9886 TerminatorInst *TI = II->getParent()->getTerminator();
9887 bool CannotRemove = false;
9888 for (++BI; &*BI != TI; ++BI) {
9889 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
9890 CannotRemove = true;
9893 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9894 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9895 // If there is a stackrestore below this one, remove this one.
9896 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9897 return EraseInstFromFunction(CI);
9898 // Otherwise, ignore the intrinsic.
9900 // If we found a non-intrinsic call, we can't remove the stack
9902 CannotRemove = true;
9908 // If the stack restore is in a return/unwind block and if there are no
9909 // allocas or calls between the restore and the return, nuke the restore.
9910 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9911 return EraseInstFromFunction(CI);
9916 return visitCallSite(II);
9919 // InvokeInst simplification
9921 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9922 return visitCallSite(&II);
9925 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9926 /// passed through the varargs area, we can eliminate the use of the cast.
9927 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9928 const CastInst * const CI,
9929 const TargetData * const TD,
9931 if (!CI->isLosslessCast())
9934 // The size of ByVal arguments is derived from the type, so we
9935 // can't change to a type with a different size. If the size were
9936 // passed explicitly we could avoid this check.
9937 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9941 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9942 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9943 if (!SrcTy->isSized() || !DstTy->isSized())
9945 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
9950 // visitCallSite - Improvements for call and invoke instructions.
9952 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9953 bool Changed = false;
9955 // If the callee is a constexpr cast of a function, attempt to move the cast
9956 // to the arguments of the call/invoke.
9957 if (transformConstExprCastCall(CS)) return 0;
9959 Value *Callee = CS.getCalledValue();
9961 if (Function *CalleeF = dyn_cast<Function>(Callee))
9962 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9963 Instruction *OldCall = CS.getInstruction();
9964 // If the call and callee calling conventions don't match, this call must
9965 // be unreachable, as the call is undefined.
9966 new StoreInst(ConstantInt::getTrue(*Context),
9967 UndefValue::get(Type::getInt1PtrTy(*Context)),
9969 // If OldCall dues not return void then replaceAllUsesWith undef.
9970 // This allows ValueHandlers and custom metadata to adjust itself.
9971 if (!OldCall->getType()->isVoidTy())
9972 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9973 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9974 return EraseInstFromFunction(*OldCall);
9978 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9979 // This instruction is not reachable, just remove it. We insert a store to
9980 // undef so that we know that this code is not reachable, despite the fact
9981 // that we can't modify the CFG here.
9982 new StoreInst(ConstantInt::getTrue(*Context),
9983 UndefValue::get(Type::getInt1PtrTy(*Context)),
9984 CS.getInstruction());
9986 // If CS dues not return void then replaceAllUsesWith undef.
9987 // This allows ValueHandlers and custom metadata to adjust itself.
9988 if (!CS.getInstruction()->getType()->isVoidTy())
9989 CS.getInstruction()->
9990 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9992 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9993 // Don't break the CFG, insert a dummy cond branch.
9994 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9995 ConstantInt::getTrue(*Context), II);
9997 return EraseInstFromFunction(*CS.getInstruction());
10000 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10001 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10002 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10003 return transformCallThroughTrampoline(CS);
10005 const PointerType *PTy = cast<PointerType>(Callee->getType());
10006 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10007 if (FTy->isVarArg()) {
10008 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10009 // See if we can optimize any arguments passed through the varargs area of
10011 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10012 E = CS.arg_end(); I != E; ++I, ++ix) {
10013 CastInst *CI = dyn_cast<CastInst>(*I);
10014 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10015 *I = CI->getOperand(0);
10021 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10022 // Inline asm calls cannot throw - mark them 'nounwind'.
10023 CS.setDoesNotThrow();
10027 return Changed ? CS.getInstruction() : 0;
10030 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10031 // attempt to move the cast to the arguments of the call/invoke.
10033 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10034 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10035 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10036 if (CE->getOpcode() != Instruction::BitCast ||
10037 !isa<Function>(CE->getOperand(0)))
10039 Function *Callee = cast<Function>(CE->getOperand(0));
10040 Instruction *Caller = CS.getInstruction();
10041 const AttrListPtr &CallerPAL = CS.getAttributes();
10043 // Okay, this is a cast from a function to a different type. Unless doing so
10044 // would cause a type conversion of one of our arguments, change this call to
10045 // be a direct call with arguments casted to the appropriate types.
10047 const FunctionType *FT = Callee->getFunctionType();
10048 const Type *OldRetTy = Caller->getType();
10049 const Type *NewRetTy = FT->getReturnType();
10051 if (isa<StructType>(NewRetTy))
10052 return false; // TODO: Handle multiple return values.
10054 // Check to see if we are changing the return type...
10055 if (OldRetTy != NewRetTy) {
10056 if (Callee->isDeclaration() &&
10057 // Conversion is ok if changing from one pointer type to another or from
10058 // a pointer to an integer of the same size.
10059 !((isa<PointerType>(OldRetTy) || !TD ||
10060 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
10061 (isa<PointerType>(NewRetTy) || !TD ||
10062 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
10063 return false; // Cannot transform this return value.
10065 if (!Caller->use_empty() &&
10066 // void -> non-void is handled specially
10067 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
10068 return false; // Cannot transform this return value.
10070 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10071 Attributes RAttrs = CallerPAL.getRetAttributes();
10072 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10073 return false; // Attribute not compatible with transformed value.
10076 // If the callsite is an invoke instruction, and the return value is used by
10077 // a PHI node in a successor, we cannot change the return type of the call
10078 // because there is no place to put the cast instruction (without breaking
10079 // the critical edge). Bail out in this case.
10080 if (!Caller->use_empty())
10081 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10082 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10084 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10085 if (PN->getParent() == II->getNormalDest() ||
10086 PN->getParent() == II->getUnwindDest())
10090 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10091 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10093 CallSite::arg_iterator AI = CS.arg_begin();
10094 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10095 const Type *ParamTy = FT->getParamType(i);
10096 const Type *ActTy = (*AI)->getType();
10098 if (!CastInst::isCastable(ActTy, ParamTy))
10099 return false; // Cannot transform this parameter value.
10101 if (CallerPAL.getParamAttributes(i + 1)
10102 & Attribute::typeIncompatible(ParamTy))
10103 return false; // Attribute not compatible with transformed value.
10105 // Converting from one pointer type to another or between a pointer and an
10106 // integer of the same size is safe even if we do not have a body.
10107 bool isConvertible = ActTy == ParamTy ||
10108 (TD && ((isa<PointerType>(ParamTy) ||
10109 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10110 (isa<PointerType>(ActTy) ||
10111 ActTy == TD->getIntPtrType(Caller->getContext()))));
10112 if (Callee->isDeclaration() && !isConvertible) return false;
10115 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10116 Callee->isDeclaration())
10117 return false; // Do not delete arguments unless we have a function body.
10119 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10120 !CallerPAL.isEmpty())
10121 // In this case we have more arguments than the new function type, but we
10122 // won't be dropping them. Check that these extra arguments have attributes
10123 // that are compatible with being a vararg call argument.
10124 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10125 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10127 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10128 if (PAttrs & Attribute::VarArgsIncompatible)
10132 // Okay, we decided that this is a safe thing to do: go ahead and start
10133 // inserting cast instructions as necessary...
10134 std::vector<Value*> Args;
10135 Args.reserve(NumActualArgs);
10136 SmallVector<AttributeWithIndex, 8> attrVec;
10137 attrVec.reserve(NumCommonArgs);
10139 // Get any return attributes.
10140 Attributes RAttrs = CallerPAL.getRetAttributes();
10142 // If the return value is not being used, the type may not be compatible
10143 // with the existing attributes. Wipe out any problematic attributes.
10144 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10146 // Add the new return attributes.
10148 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10150 AI = CS.arg_begin();
10151 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10152 const Type *ParamTy = FT->getParamType(i);
10153 if ((*AI)->getType() == ParamTy) {
10154 Args.push_back(*AI);
10156 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10157 false, ParamTy, false);
10158 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
10161 // Add any parameter attributes.
10162 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10163 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10166 // If the function takes more arguments than the call was taking, add them
10168 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10169 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10171 // If we are removing arguments to the function, emit an obnoxious warning.
10172 if (FT->getNumParams() < NumActualArgs) {
10173 if (!FT->isVarArg()) {
10174 errs() << "WARNING: While resolving call to function '"
10175 << Callee->getName() << "' arguments were dropped!\n";
10177 // Add all of the arguments in their promoted form to the arg list.
10178 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10179 const Type *PTy = getPromotedType((*AI)->getType());
10180 if (PTy != (*AI)->getType()) {
10181 // Must promote to pass through va_arg area!
10182 Instruction::CastOps opcode =
10183 CastInst::getCastOpcode(*AI, false, PTy, false);
10184 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
10186 Args.push_back(*AI);
10189 // Add any parameter attributes.
10190 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10191 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10196 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10197 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10199 if (NewRetTy->isVoidTy())
10200 Caller->setName(""); // Void type should not have a name.
10202 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10206 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10207 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10208 Args.begin(), Args.end(),
10209 Caller->getName(), Caller);
10210 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10211 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10213 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10214 Caller->getName(), Caller);
10215 CallInst *CI = cast<CallInst>(Caller);
10216 if (CI->isTailCall())
10217 cast<CallInst>(NC)->setTailCall();
10218 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10219 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10222 // Insert a cast of the return type as necessary.
10224 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10225 if (!NV->getType()->isVoidTy()) {
10226 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10228 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10230 // If this is an invoke instruction, we should insert it after the first
10231 // non-phi, instruction in the normal successor block.
10232 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10233 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10234 InsertNewInstBefore(NC, *I);
10236 // Otherwise, it's a call, just insert cast right after the call instr
10237 InsertNewInstBefore(NC, *Caller);
10239 Worklist.AddUsersToWorkList(*Caller);
10241 NV = UndefValue::get(Caller->getType());
10246 if (!Caller->use_empty())
10247 Caller->replaceAllUsesWith(NV);
10249 EraseInstFromFunction(*Caller);
10253 // transformCallThroughTrampoline - Turn a call to a function created by the
10254 // init_trampoline intrinsic into a direct call to the underlying function.
10256 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10257 Value *Callee = CS.getCalledValue();
10258 const PointerType *PTy = cast<PointerType>(Callee->getType());
10259 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10260 const AttrListPtr &Attrs = CS.getAttributes();
10262 // If the call already has the 'nest' attribute somewhere then give up -
10263 // otherwise 'nest' would occur twice after splicing in the chain.
10264 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10267 IntrinsicInst *Tramp =
10268 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10270 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10271 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10272 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10274 const AttrListPtr &NestAttrs = NestF->getAttributes();
10275 if (!NestAttrs.isEmpty()) {
10276 unsigned NestIdx = 1;
10277 const Type *NestTy = 0;
10278 Attributes NestAttr = Attribute::None;
10280 // Look for a parameter marked with the 'nest' attribute.
10281 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10282 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10283 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10284 // Record the parameter type and any other attributes.
10286 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10291 Instruction *Caller = CS.getInstruction();
10292 std::vector<Value*> NewArgs;
10293 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10295 SmallVector<AttributeWithIndex, 8> NewAttrs;
10296 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10298 // Insert the nest argument into the call argument list, which may
10299 // mean appending it. Likewise for attributes.
10301 // Add any result attributes.
10302 if (Attributes Attr = Attrs.getRetAttributes())
10303 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10307 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10309 if (Idx == NestIdx) {
10310 // Add the chain argument and attributes.
10311 Value *NestVal = Tramp->getOperand(3);
10312 if (NestVal->getType() != NestTy)
10313 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10314 NewArgs.push_back(NestVal);
10315 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10321 // Add the original argument and attributes.
10322 NewArgs.push_back(*I);
10323 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10325 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10331 // Add any function attributes.
10332 if (Attributes Attr = Attrs.getFnAttributes())
10333 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10335 // The trampoline may have been bitcast to a bogus type (FTy).
10336 // Handle this by synthesizing a new function type, equal to FTy
10337 // with the chain parameter inserted.
10339 std::vector<const Type*> NewTypes;
10340 NewTypes.reserve(FTy->getNumParams()+1);
10342 // Insert the chain's type into the list of parameter types, which may
10343 // mean appending it.
10346 FunctionType::param_iterator I = FTy->param_begin(),
10347 E = FTy->param_end();
10350 if (Idx == NestIdx)
10351 // Add the chain's type.
10352 NewTypes.push_back(NestTy);
10357 // Add the original type.
10358 NewTypes.push_back(*I);
10364 // Replace the trampoline call with a direct call. Let the generic
10365 // code sort out any function type mismatches.
10366 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10368 Constant *NewCallee =
10369 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10370 NestF : ConstantExpr::getBitCast(NestF,
10371 PointerType::getUnqual(NewFTy));
10372 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10375 Instruction *NewCaller;
10376 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10377 NewCaller = InvokeInst::Create(NewCallee,
10378 II->getNormalDest(), II->getUnwindDest(),
10379 NewArgs.begin(), NewArgs.end(),
10380 Caller->getName(), Caller);
10381 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10382 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10384 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10385 Caller->getName(), Caller);
10386 if (cast<CallInst>(Caller)->isTailCall())
10387 cast<CallInst>(NewCaller)->setTailCall();
10388 cast<CallInst>(NewCaller)->
10389 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10390 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10392 if (!Caller->getType()->isVoidTy())
10393 Caller->replaceAllUsesWith(NewCaller);
10394 Caller->eraseFromParent();
10395 Worklist.Remove(Caller);
10400 // Replace the trampoline call with a direct call. Since there is no 'nest'
10401 // parameter, there is no need to adjust the argument list. Let the generic
10402 // code sort out any function type mismatches.
10403 Constant *NewCallee =
10404 NestF->getType() == PTy ? NestF :
10405 ConstantExpr::getBitCast(NestF, PTy);
10406 CS.setCalledFunction(NewCallee);
10407 return CS.getInstruction();
10410 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
10411 /// and if a/b/c and the add's all have a single use, turn this into a phi
10412 /// and a single binop.
10413 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10414 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10415 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10416 unsigned Opc = FirstInst->getOpcode();
10417 Value *LHSVal = FirstInst->getOperand(0);
10418 Value *RHSVal = FirstInst->getOperand(1);
10420 const Type *LHSType = LHSVal->getType();
10421 const Type *RHSType = RHSVal->getType();
10423 // Scan to see if all operands are the same opcode, and all have one use.
10424 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10425 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10426 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10427 // Verify type of the LHS matches so we don't fold cmp's of different
10428 // types or GEP's with different index types.
10429 I->getOperand(0)->getType() != LHSType ||
10430 I->getOperand(1)->getType() != RHSType)
10433 // If they are CmpInst instructions, check their predicates
10434 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10435 if (cast<CmpInst>(I)->getPredicate() !=
10436 cast<CmpInst>(FirstInst)->getPredicate())
10439 // Keep track of which operand needs a phi node.
10440 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10441 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10444 // If both LHS and RHS would need a PHI, don't do this transformation,
10445 // because it would increase the number of PHIs entering the block,
10446 // which leads to higher register pressure. This is especially
10447 // bad when the PHIs are in the header of a loop.
10448 if (!LHSVal && !RHSVal)
10451 // Otherwise, this is safe to transform!
10453 Value *InLHS = FirstInst->getOperand(0);
10454 Value *InRHS = FirstInst->getOperand(1);
10455 PHINode *NewLHS = 0, *NewRHS = 0;
10457 NewLHS = PHINode::Create(LHSType,
10458 FirstInst->getOperand(0)->getName() + ".pn");
10459 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10460 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10461 InsertNewInstBefore(NewLHS, PN);
10466 NewRHS = PHINode::Create(RHSType,
10467 FirstInst->getOperand(1)->getName() + ".pn");
10468 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10469 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10470 InsertNewInstBefore(NewRHS, PN);
10474 // Add all operands to the new PHIs.
10475 if (NewLHS || NewRHS) {
10476 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10477 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10479 Value *NewInLHS = InInst->getOperand(0);
10480 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10483 Value *NewInRHS = InInst->getOperand(1);
10484 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10489 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10490 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10491 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10492 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10496 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10497 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10499 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10500 FirstInst->op_end());
10501 // This is true if all GEP bases are allocas and if all indices into them are
10503 bool AllBasePointersAreAllocas = true;
10505 // We don't want to replace this phi if the replacement would require
10506 // more than one phi, which leads to higher register pressure. This is
10507 // especially bad when the PHIs are in the header of a loop.
10508 bool NeededPhi = false;
10510 // Scan to see if all operands are the same opcode, and all have one use.
10511 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10512 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10513 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10514 GEP->getNumOperands() != FirstInst->getNumOperands())
10517 // Keep track of whether or not all GEPs are of alloca pointers.
10518 if (AllBasePointersAreAllocas &&
10519 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10520 !GEP->hasAllConstantIndices()))
10521 AllBasePointersAreAllocas = false;
10523 // Compare the operand lists.
10524 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10525 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10528 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10529 // if one of the PHIs has a constant for the index. The index may be
10530 // substantially cheaper to compute for the constants, so making it a
10531 // variable index could pessimize the path. This also handles the case
10532 // for struct indices, which must always be constant.
10533 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10534 isa<ConstantInt>(GEP->getOperand(op)))
10537 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10540 // If we already needed a PHI for an earlier operand, and another operand
10541 // also requires a PHI, we'd be introducing more PHIs than we're
10542 // eliminating, which increases register pressure on entry to the PHI's
10547 FixedOperands[op] = 0; // Needs a PHI.
10552 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10553 // bother doing this transformation. At best, this will just save a bit of
10554 // offset calculation, but all the predecessors will have to materialize the
10555 // stack address into a register anyway. We'd actually rather *clone* the
10556 // load up into the predecessors so that we have a load of a gep of an alloca,
10557 // which can usually all be folded into the load.
10558 if (AllBasePointersAreAllocas)
10561 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10562 // that is variable.
10563 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10565 bool HasAnyPHIs = false;
10566 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10567 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10568 Value *FirstOp = FirstInst->getOperand(i);
10569 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10570 FirstOp->getName()+".pn");
10571 InsertNewInstBefore(NewPN, PN);
10573 NewPN->reserveOperandSpace(e);
10574 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10575 OperandPhis[i] = NewPN;
10576 FixedOperands[i] = NewPN;
10581 // Add all operands to the new PHIs.
10583 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10584 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10585 BasicBlock *InBB = PN.getIncomingBlock(i);
10587 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10588 if (PHINode *OpPhi = OperandPhis[op])
10589 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10593 Value *Base = FixedOperands[0];
10594 return cast<GEPOperator>(FirstInst)->isInBounds() ?
10595 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
10596 FixedOperands.end()) :
10597 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10598 FixedOperands.end());
10602 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10603 /// sink the load out of the block that defines it. This means that it must be
10604 /// obvious the value of the load is not changed from the point of the load to
10605 /// the end of the block it is in.
10607 /// Finally, it is safe, but not profitable, to sink a load targetting a
10608 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10610 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10611 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10613 for (++BBI; BBI != E; ++BBI)
10614 if (BBI->mayWriteToMemory())
10617 // Check for non-address taken alloca. If not address-taken already, it isn't
10618 // profitable to do this xform.
10619 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10620 bool isAddressTaken = false;
10621 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10623 if (isa<LoadInst>(UI)) continue;
10624 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10625 // If storing TO the alloca, then the address isn't taken.
10626 if (SI->getOperand(1) == AI) continue;
10628 isAddressTaken = true;
10632 if (!isAddressTaken && AI->isStaticAlloca())
10636 // If this load is a load from a GEP with a constant offset from an alloca,
10637 // then we don't want to sink it. In its present form, it will be
10638 // load [constant stack offset]. Sinking it will cause us to have to
10639 // materialize the stack addresses in each predecessor in a register only to
10640 // do a shared load from register in the successor.
10641 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10642 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10643 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10650 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10651 // operator and they all are only used by the PHI, PHI together their
10652 // inputs, and do the operation once, to the result of the PHI.
10653 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10654 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10656 // Scan the instruction, looking for input operations that can be folded away.
10657 // If all input operands to the phi are the same instruction (e.g. a cast from
10658 // the same type or "+42") we can pull the operation through the PHI, reducing
10659 // code size and simplifying code.
10660 Constant *ConstantOp = 0;
10661 const Type *CastSrcTy = 0;
10662 bool isVolatile = false;
10663 if (isa<CastInst>(FirstInst)) {
10664 CastSrcTy = FirstInst->getOperand(0)->getType();
10665 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10666 // Can fold binop, compare or shift here if the RHS is a constant,
10667 // otherwise call FoldPHIArgBinOpIntoPHI.
10668 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10669 if (ConstantOp == 0)
10670 return FoldPHIArgBinOpIntoPHI(PN);
10671 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10672 isVolatile = LI->isVolatile();
10673 // We can't sink the load if the loaded value could be modified between the
10674 // load and the PHI.
10675 if (LI->getParent() != PN.getIncomingBlock(0) ||
10676 !isSafeAndProfitableToSinkLoad(LI))
10679 // If the PHI is of volatile loads and the load block has multiple
10680 // successors, sinking it would remove a load of the volatile value from
10681 // the path through the other successor.
10683 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10686 } else if (isa<GetElementPtrInst>(FirstInst)) {
10687 return FoldPHIArgGEPIntoPHI(PN);
10689 return 0; // Cannot fold this operation.
10692 // Check to see if all arguments are the same operation.
10693 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10694 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10695 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10696 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10699 if (I->getOperand(0)->getType() != CastSrcTy)
10700 return 0; // Cast operation must match.
10701 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10702 // We can't sink the load if the loaded value could be modified between
10703 // the load and the PHI.
10704 if (LI->isVolatile() != isVolatile ||
10705 LI->getParent() != PN.getIncomingBlock(i) ||
10706 !isSafeAndProfitableToSinkLoad(LI))
10709 // If the PHI is of volatile loads and the load block has multiple
10710 // successors, sinking it would remove a load of the volatile value from
10711 // the path through the other successor.
10713 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10716 } else if (I->getOperand(1) != ConstantOp) {
10721 // Okay, they are all the same operation. Create a new PHI node of the
10722 // correct type, and PHI together all of the LHS's of the instructions.
10723 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10724 PN.getName()+".in");
10725 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10727 Value *InVal = FirstInst->getOperand(0);
10728 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10730 // Add all operands to the new PHI.
10731 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10732 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10733 if (NewInVal != InVal)
10735 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10740 // The new PHI unions all of the same values together. This is really
10741 // common, so we handle it intelligently here for compile-time speed.
10745 InsertNewInstBefore(NewPN, PN);
10749 // Insert and return the new operation.
10750 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10751 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10752 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10753 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10754 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10755 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10756 PhiVal, ConstantOp);
10757 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10759 // If this was a volatile load that we are merging, make sure to loop through
10760 // and mark all the input loads as non-volatile. If we don't do this, we will
10761 // insert a new volatile load and the old ones will not be deletable.
10763 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10764 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10766 return new LoadInst(PhiVal, "", isVolatile);
10769 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10771 static bool DeadPHICycle(PHINode *PN,
10772 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10773 if (PN->use_empty()) return true;
10774 if (!PN->hasOneUse()) return false;
10776 // Remember this node, and if we find the cycle, return.
10777 if (!PotentiallyDeadPHIs.insert(PN))
10780 // Don't scan crazily complex things.
10781 if (PotentiallyDeadPHIs.size() == 16)
10784 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10785 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10790 /// PHIsEqualValue - Return true if this phi node is always equal to
10791 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10792 /// z = some value; x = phi (y, z); y = phi (x, z)
10793 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10794 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10795 // See if we already saw this PHI node.
10796 if (!ValueEqualPHIs.insert(PN))
10799 // Don't scan crazily complex things.
10800 if (ValueEqualPHIs.size() == 16)
10803 // Scan the operands to see if they are either phi nodes or are equal to
10805 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10806 Value *Op = PN->getIncomingValue(i);
10807 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10808 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10810 } else if (Op != NonPhiInVal)
10818 // PHINode simplification
10820 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10821 // If LCSSA is around, don't mess with Phi nodes
10822 if (MustPreserveLCSSA) return 0;
10824 if (Value *V = PN.hasConstantValue())
10825 return ReplaceInstUsesWith(PN, V);
10827 // If all PHI operands are the same operation, pull them through the PHI,
10828 // reducing code size.
10829 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10830 isa<Instruction>(PN.getIncomingValue(1)) &&
10831 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10832 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10833 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10834 // than themselves more than once.
10835 PN.getIncomingValue(0)->hasOneUse())
10836 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10839 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10840 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10841 // PHI)... break the cycle.
10842 if (PN.hasOneUse()) {
10843 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10844 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10845 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10846 PotentiallyDeadPHIs.insert(&PN);
10847 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10848 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10851 // If this phi has a single use, and if that use just computes a value for
10852 // the next iteration of a loop, delete the phi. This occurs with unused
10853 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10854 // common case here is good because the only other things that catch this
10855 // are induction variable analysis (sometimes) and ADCE, which is only run
10857 if (PHIUser->hasOneUse() &&
10858 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10859 PHIUser->use_back() == &PN) {
10860 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10864 // We sometimes end up with phi cycles that non-obviously end up being the
10865 // same value, for example:
10866 // z = some value; x = phi (y, z); y = phi (x, z)
10867 // where the phi nodes don't necessarily need to be in the same block. Do a
10868 // quick check to see if the PHI node only contains a single non-phi value, if
10869 // so, scan to see if the phi cycle is actually equal to that value.
10871 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10872 // Scan for the first non-phi operand.
10873 while (InValNo != NumOperandVals &&
10874 isa<PHINode>(PN.getIncomingValue(InValNo)))
10877 if (InValNo != NumOperandVals) {
10878 Value *NonPhiInVal = PN.getOperand(InValNo);
10880 // Scan the rest of the operands to see if there are any conflicts, if so
10881 // there is no need to recursively scan other phis.
10882 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10883 Value *OpVal = PN.getIncomingValue(InValNo);
10884 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10888 // If we scanned over all operands, then we have one unique value plus
10889 // phi values. Scan PHI nodes to see if they all merge in each other or
10891 if (InValNo == NumOperandVals) {
10892 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10893 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10894 return ReplaceInstUsesWith(PN, NonPhiInVal);
10901 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10902 Value *PtrOp = GEP.getOperand(0);
10903 // Eliminate 'getelementptr %P, i32 0' and 'getelementptr %P', they are noops.
10904 if (GEP.getNumOperands() == 1)
10905 return ReplaceInstUsesWith(GEP, PtrOp);
10907 if (isa<UndefValue>(GEP.getOperand(0)))
10908 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10910 bool HasZeroPointerIndex = false;
10911 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10912 HasZeroPointerIndex = C->isNullValue();
10914 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10915 return ReplaceInstUsesWith(GEP, PtrOp);
10917 // Eliminate unneeded casts for indices.
10919 bool MadeChange = false;
10920 unsigned PtrSize = TD->getPointerSizeInBits();
10922 gep_type_iterator GTI = gep_type_begin(GEP);
10923 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
10924 I != E; ++I, ++GTI) {
10925 if (!isa<SequentialType>(*GTI)) continue;
10927 // If we are using a wider index than needed for this platform, shrink it
10928 // to what we need. If narrower, sign-extend it to what we need. This
10929 // explicit cast can make subsequent optimizations more obvious.
10930 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
10931 if (OpBits == PtrSize)
10934 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
10937 if (MadeChange) return &GEP;
10940 // Combine Indices - If the source pointer to this getelementptr instruction
10941 // is a getelementptr instruction, combine the indices of the two
10942 // getelementptr instructions into a single instruction.
10944 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
10945 // Note that if our source is a gep chain itself that we wait for that
10946 // chain to be resolved before we perform this transformation. This
10947 // avoids us creating a TON of code in some cases.
10949 if (GetElementPtrInst *SrcGEP =
10950 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
10951 if (SrcGEP->getNumOperands() == 2)
10952 return 0; // Wait until our source is folded to completion.
10954 SmallVector<Value*, 8> Indices;
10956 // Find out whether the last index in the source GEP is a sequential idx.
10957 bool EndsWithSequential = false;
10958 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
10960 EndsWithSequential = !isa<StructType>(*I);
10962 // Can we combine the two pointer arithmetics offsets?
10963 if (EndsWithSequential) {
10964 // Replace: gep (gep %P, long B), long A, ...
10965 // With: T = long A+B; gep %P, T, ...
10968 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
10969 Value *GO1 = GEP.getOperand(1);
10970 if (SO1 == Constant::getNullValue(SO1->getType())) {
10972 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10975 // If they aren't the same type, then the input hasn't been processed
10976 // by the loop above yet (which canonicalizes sequential index types to
10977 // intptr_t). Just avoid transforming this until the input has been
10979 if (SO1->getType() != GO1->getType())
10981 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10984 // Update the GEP in place if possible.
10985 if (Src->getNumOperands() == 2) {
10986 GEP.setOperand(0, Src->getOperand(0));
10987 GEP.setOperand(1, Sum);
10990 Indices.append(Src->op_begin()+1, Src->op_end()-1);
10991 Indices.push_back(Sum);
10992 Indices.append(GEP.op_begin()+2, GEP.op_end());
10993 } else if (isa<Constant>(*GEP.idx_begin()) &&
10994 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10995 Src->getNumOperands() != 1) {
10996 // Otherwise we can do the fold if the first index of the GEP is a zero
10997 Indices.append(Src->op_begin()+1, Src->op_end());
10998 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
11001 if (!Indices.empty())
11002 return (cast<GEPOperator>(&GEP)->isInBounds() &&
11003 Src->isInBounds()) ?
11004 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
11005 Indices.end(), GEP.getName()) :
11006 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
11007 Indices.end(), GEP.getName());
11010 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
11011 if (Value *X = getBitCastOperand(PtrOp)) {
11012 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
11014 // If the input bitcast is actually "bitcast(bitcast(x))", then we don't
11015 // want to change the gep until the bitcasts are eliminated.
11016 if (getBitCastOperand(X)) {
11017 Worklist.AddValue(PtrOp);
11021 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11022 // into : GEP [10 x i8]* X, i32 0, ...
11024 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11025 // into : GEP i8* X, ...
11027 // This occurs when the program declares an array extern like "int X[];"
11028 if (HasZeroPointerIndex) {
11029 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11030 const PointerType *XTy = cast<PointerType>(X->getType());
11031 if (const ArrayType *CATy =
11032 dyn_cast<ArrayType>(CPTy->getElementType())) {
11033 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11034 if (CATy->getElementType() == XTy->getElementType()) {
11035 // -> GEP i8* X, ...
11036 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11037 return cast<GEPOperator>(&GEP)->isInBounds() ?
11038 GetElementPtrInst::CreateInBounds(X, Indices.begin(), Indices.end(),
11040 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11044 if (const ArrayType *XATy = dyn_cast<ArrayType>(XTy->getElementType())){
11045 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11046 if (CATy->getElementType() == XATy->getElementType()) {
11047 // -> GEP [10 x i8]* X, i32 0, ...
11048 // At this point, we know that the cast source type is a pointer
11049 // to an array of the same type as the destination pointer
11050 // array. Because the array type is never stepped over (there
11051 // is a leading zero) we can fold the cast into this GEP.
11052 GEP.setOperand(0, X);
11057 } else if (GEP.getNumOperands() == 2) {
11058 // Transform things like:
11059 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11060 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11061 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11062 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11063 if (TD && isa<ArrayType>(SrcElTy) &&
11064 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11065 TD->getTypeAllocSize(ResElTy)) {
11067 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11068 Idx[1] = GEP.getOperand(1);
11069 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11070 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11071 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11072 // V and GEP are both pointer types --> BitCast
11073 return new BitCastInst(NewGEP, GEP.getType());
11076 // Transform things like:
11077 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11078 // (where tmp = 8*tmp2) into:
11079 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11081 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
11082 uint64_t ArrayEltSize =
11083 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11085 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11086 // allow either a mul, shift, or constant here.
11088 ConstantInt *Scale = 0;
11089 if (ArrayEltSize == 1) {
11090 NewIdx = GEP.getOperand(1);
11091 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11092 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11093 NewIdx = ConstantInt::get(CI->getType(), 1);
11095 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11096 if (Inst->getOpcode() == Instruction::Shl &&
11097 isa<ConstantInt>(Inst->getOperand(1))) {
11098 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11099 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11100 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11102 NewIdx = Inst->getOperand(0);
11103 } else if (Inst->getOpcode() == Instruction::Mul &&
11104 isa<ConstantInt>(Inst->getOperand(1))) {
11105 Scale = cast<ConstantInt>(Inst->getOperand(1));
11106 NewIdx = Inst->getOperand(0);
11110 // If the index will be to exactly the right offset with the scale taken
11111 // out, perform the transformation. Note, we don't know whether Scale is
11112 // signed or not. We'll use unsigned version of division/modulo
11113 // operation after making sure Scale doesn't have the sign bit set.
11114 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11115 Scale->getZExtValue() % ArrayEltSize == 0) {
11116 Scale = ConstantInt::get(Scale->getType(),
11117 Scale->getZExtValue() / ArrayEltSize);
11118 if (Scale->getZExtValue() != 1) {
11119 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11121 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
11124 // Insert the new GEP instruction.
11126 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11128 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11129 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
11130 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11131 // The NewGEP must be pointer typed, so must the old one -> BitCast
11132 return new BitCastInst(NewGEP, GEP.getType());
11138 /// See if we can simplify:
11139 /// X = bitcast A* to B*
11140 /// Y = gep X, <...constant indices...>
11141 /// into a gep of the original struct. This is important for SROA and alias
11142 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11143 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11145 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11146 // Determine how much the GEP moves the pointer. We are guaranteed to get
11147 // a constant back from EmitGEPOffset.
11148 ConstantInt *OffsetV =
11149 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11150 int64_t Offset = OffsetV->getSExtValue();
11152 // If this GEP instruction doesn't move the pointer, just replace the GEP
11153 // with a bitcast of the real input to the dest type.
11155 // If the bitcast is of an allocation, and the allocation will be
11156 // converted to match the type of the cast, don't touch this.
11157 if (isa<AllocationInst>(BCI->getOperand(0)) ||
11158 isMalloc(BCI->getOperand(0))) {
11159 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11160 if (Instruction *I = visitBitCast(*BCI)) {
11163 BCI->getParent()->getInstList().insert(BCI, I);
11164 ReplaceInstUsesWith(*BCI, I);
11169 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11172 // Otherwise, if the offset is non-zero, we need to find out if there is a
11173 // field at Offset in 'A's type. If so, we can pull the cast through the
11175 SmallVector<Value*, 8> NewIndices;
11177 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11178 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11179 Value *NGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
11180 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
11181 NewIndices.end()) :
11182 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
11185 if (NGEP->getType() == GEP.getType())
11186 return ReplaceInstUsesWith(GEP, NGEP);
11187 NGEP->takeName(&GEP);
11188 return new BitCastInst(NGEP, GEP.getType());
11196 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11197 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11198 if (AI.isArrayAllocation()) { // Check C != 1
11199 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11200 const Type *NewTy =
11201 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11202 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11203 AllocationInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
11204 New->setAlignment(AI.getAlignment());
11206 // Scan to the end of the allocation instructions, to skip over a block of
11207 // allocas if possible...also skip interleaved debug info
11209 BasicBlock::iterator It = New;
11210 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11212 // Now that I is pointing to the first non-allocation-inst in the block,
11213 // insert our getelementptr instruction...
11215 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
11219 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
11220 New->getName()+".sub", It);
11222 // Now make everything use the getelementptr instead of the original
11224 return ReplaceInstUsesWith(AI, V);
11225 } else if (isa<UndefValue>(AI.getArraySize())) {
11226 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11230 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11231 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11232 // Note that we only do this for alloca's, because malloc should allocate
11233 // and return a unique pointer, even for a zero byte allocation.
11234 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11235 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11237 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11238 if (AI.getAlignment() == 0)
11239 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11245 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11246 Value *Op = FI.getOperand(0);
11248 // free undef -> unreachable.
11249 if (isa<UndefValue>(Op)) {
11250 // Insert a new store to null because we cannot modify the CFG here.
11251 new StoreInst(ConstantInt::getTrue(*Context),
11252 UndefValue::get(Type::getInt1PtrTy(*Context)), &FI);
11253 return EraseInstFromFunction(FI);
11256 // If we have 'free null' delete the instruction. This can happen in stl code
11257 // when lots of inlining happens.
11258 if (isa<ConstantPointerNull>(Op))
11259 return EraseInstFromFunction(FI);
11261 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11262 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11263 FI.setOperand(0, CI->getOperand(0));
11267 // Change free (gep X, 0,0,0,0) into free(X)
11268 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11269 if (GEPI->hasAllZeroIndices()) {
11270 Worklist.Add(GEPI);
11271 FI.setOperand(0, GEPI->getOperand(0));
11276 if (isMalloc(Op)) {
11277 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
11278 if (Op->hasOneUse() && CI->hasOneUse()) {
11279 EraseInstFromFunction(FI);
11280 EraseInstFromFunction(*CI);
11281 return EraseInstFromFunction(*cast<Instruction>(Op));
11284 // Op is a call to malloc
11285 if (Op->hasOneUse()) {
11286 EraseInstFromFunction(FI);
11287 return EraseInstFromFunction(*cast<Instruction>(Op));
11296 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11297 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11298 const TargetData *TD) {
11299 User *CI = cast<User>(LI.getOperand(0));
11300 Value *CastOp = CI->getOperand(0);
11301 LLVMContext *Context = IC.getContext();
11304 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11305 // Instead of loading constant c string, use corresponding integer value
11306 // directly if string length is small enough.
11308 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11309 unsigned len = Str.length();
11310 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11311 unsigned numBits = Ty->getPrimitiveSizeInBits();
11312 // Replace LI with immediate integer store.
11313 if ((numBits >> 3) == len + 1) {
11314 APInt StrVal(numBits, 0);
11315 APInt SingleChar(numBits, 0);
11316 if (TD->isLittleEndian()) {
11317 for (signed i = len-1; i >= 0; i--) {
11318 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11319 StrVal = (StrVal << 8) | SingleChar;
11322 for (unsigned i = 0; i < len; i++) {
11323 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11324 StrVal = (StrVal << 8) | SingleChar;
11326 // Append NULL at the end.
11328 StrVal = (StrVal << 8) | SingleChar;
11330 Value *NL = ConstantInt::get(*Context, StrVal);
11331 return IC.ReplaceInstUsesWith(LI, NL);
11337 const PointerType *DestTy = cast<PointerType>(CI->getType());
11338 const Type *DestPTy = DestTy->getElementType();
11339 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11341 // If the address spaces don't match, don't eliminate the cast.
11342 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11345 const Type *SrcPTy = SrcTy->getElementType();
11347 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11348 isa<VectorType>(DestPTy)) {
11349 // If the source is an array, the code below will not succeed. Check to
11350 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11352 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11353 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11354 if (ASrcTy->getNumElements() != 0) {
11356 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::getInt32Ty(*Context));
11357 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11358 SrcTy = cast<PointerType>(CastOp->getType());
11359 SrcPTy = SrcTy->getElementType();
11362 if (IC.getTargetData() &&
11363 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11364 isa<VectorType>(SrcPTy)) &&
11365 // Do not allow turning this into a load of an integer, which is then
11366 // casted to a pointer, this pessimizes pointer analysis a lot.
11367 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11368 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11369 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11371 // Okay, we are casting from one integer or pointer type to another of
11372 // the same size. Instead of casting the pointer before the load, cast
11373 // the result of the loaded value.
11375 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
11376 // Now cast the result of the load.
11377 return new BitCastInst(NewLoad, LI.getType());
11384 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11385 Value *Op = LI.getOperand(0);
11387 // Attempt to improve the alignment.
11389 unsigned KnownAlign =
11390 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11392 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11393 LI.getAlignment()))
11394 LI.setAlignment(KnownAlign);
11397 // load (cast X) --> cast (load X) iff safe.
11398 if (isa<CastInst>(Op))
11399 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11402 // None of the following transforms are legal for volatile loads.
11403 if (LI.isVolatile()) return 0;
11405 // Do really simple store-to-load forwarding and load CSE, to catch cases
11406 // where there are several consequtive memory accesses to the same location,
11407 // separated by a few arithmetic operations.
11408 BasicBlock::iterator BBI = &LI;
11409 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11410 return ReplaceInstUsesWith(LI, AvailableVal);
11412 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11413 const Value *GEPI0 = GEPI->getOperand(0);
11414 // TODO: Consider a target hook for valid address spaces for this xform.
11415 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
11416 // Insert a new store to null instruction before the load to indicate
11417 // that this code is not reachable. We do this instead of inserting
11418 // an unreachable instruction directly because we cannot modify the
11420 new StoreInst(UndefValue::get(LI.getType()),
11421 Constant::getNullValue(Op->getType()), &LI);
11422 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11426 if (Constant *C = dyn_cast<Constant>(Op)) {
11427 // load null/undef -> undef
11428 // TODO: Consider a target hook for valid address spaces for this xform.
11429 if (isa<UndefValue>(C) ||
11430 (C->isNullValue() && LI.getPointerAddressSpace() == 0)) {
11431 // Insert a new store to null instruction before the load to indicate that
11432 // this code is not reachable. We do this instead of inserting an
11433 // unreachable instruction directly because we cannot modify the CFG.
11434 new StoreInst(UndefValue::get(LI.getType()),
11435 Constant::getNullValue(Op->getType()), &LI);
11436 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11439 // Instcombine load (constant global) into the value loaded.
11440 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11441 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11442 return ReplaceInstUsesWith(LI, GV->getInitializer());
11444 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11445 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11446 if (CE->getOpcode() == Instruction::GetElementPtr) {
11447 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11448 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11450 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
11451 return ReplaceInstUsesWith(LI, V);
11452 if (CE->getOperand(0)->isNullValue()) {
11453 // Insert a new store to null instruction before the load to indicate
11454 // that this code is not reachable. We do this instead of inserting
11455 // an unreachable instruction directly because we cannot modify the
11457 new StoreInst(UndefValue::get(LI.getType()),
11458 Constant::getNullValue(Op->getType()), &LI);
11459 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11462 } else if (CE->isCast()) {
11463 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11469 // If this load comes from anywhere in a constant global, and if the global
11470 // is all undef or zero, we know what it loads.
11471 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11472 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11473 if (GV->getInitializer()->isNullValue())
11474 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11475 else if (isa<UndefValue>(GV->getInitializer()))
11476 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11480 if (Op->hasOneUse()) {
11481 // Change select and PHI nodes to select values instead of addresses: this
11482 // helps alias analysis out a lot, allows many others simplifications, and
11483 // exposes redundancy in the code.
11485 // Note that we cannot do the transformation unless we know that the
11486 // introduced loads cannot trap! Something like this is valid as long as
11487 // the condition is always false: load (select bool %C, int* null, int* %G),
11488 // but it would not be valid if we transformed it to load from null
11489 // unconditionally.
11491 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11492 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11493 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11494 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11495 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
11496 SI->getOperand(1)->getName()+".val");
11497 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
11498 SI->getOperand(2)->getName()+".val");
11499 return SelectInst::Create(SI->getCondition(), V1, V2);
11502 // load (select (cond, null, P)) -> load P
11503 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11504 if (C->isNullValue()) {
11505 LI.setOperand(0, SI->getOperand(2));
11509 // load (select (cond, P, null)) -> load P
11510 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11511 if (C->isNullValue()) {
11512 LI.setOperand(0, SI->getOperand(1));
11520 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11521 /// when possible. This makes it generally easy to do alias analysis and/or
11522 /// SROA/mem2reg of the memory object.
11523 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11524 User *CI = cast<User>(SI.getOperand(1));
11525 Value *CastOp = CI->getOperand(0);
11527 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11528 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11529 if (SrcTy == 0) return 0;
11531 const Type *SrcPTy = SrcTy->getElementType();
11533 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11536 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11537 /// to its first element. This allows us to handle things like:
11538 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11539 /// on 32-bit hosts.
11540 SmallVector<Value*, 4> NewGEPIndices;
11542 // If the source is an array, the code below will not succeed. Check to
11543 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11545 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11546 // Index through pointer.
11547 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
11548 NewGEPIndices.push_back(Zero);
11551 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11552 if (!STy->getNumElements()) /* Struct can be empty {} */
11554 NewGEPIndices.push_back(Zero);
11555 SrcPTy = STy->getElementType(0);
11556 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11557 NewGEPIndices.push_back(Zero);
11558 SrcPTy = ATy->getElementType();
11564 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11567 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11570 // If the pointers point into different address spaces or if they point to
11571 // values with different sizes, we can't do the transformation.
11572 if (!IC.getTargetData() ||
11573 SrcTy->getAddressSpace() !=
11574 cast<PointerType>(CI->getType())->getAddressSpace() ||
11575 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11576 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11579 // Okay, we are casting from one integer or pointer type to another of
11580 // the same size. Instead of casting the pointer before
11581 // the store, cast the value to be stored.
11583 Value *SIOp0 = SI.getOperand(0);
11584 Instruction::CastOps opcode = Instruction::BitCast;
11585 const Type* CastSrcTy = SIOp0->getType();
11586 const Type* CastDstTy = SrcPTy;
11587 if (isa<PointerType>(CastDstTy)) {
11588 if (CastSrcTy->isInteger())
11589 opcode = Instruction::IntToPtr;
11590 } else if (isa<IntegerType>(CastDstTy)) {
11591 if (isa<PointerType>(SIOp0->getType()))
11592 opcode = Instruction::PtrToInt;
11595 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11596 // emit a GEP to index into its first field.
11597 if (!NewGEPIndices.empty())
11598 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
11599 NewGEPIndices.end());
11601 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
11602 SIOp0->getName()+".c");
11603 return new StoreInst(NewCast, CastOp);
11606 /// equivalentAddressValues - Test if A and B will obviously have the same
11607 /// value. This includes recognizing that %t0 and %t1 will have the same
11608 /// value in code like this:
11609 /// %t0 = getelementptr \@a, 0, 3
11610 /// store i32 0, i32* %t0
11611 /// %t1 = getelementptr \@a, 0, 3
11612 /// %t2 = load i32* %t1
11614 static bool equivalentAddressValues(Value *A, Value *B) {
11615 // Test if the values are trivially equivalent.
11616 if (A == B) return true;
11618 // Test if the values come form identical arithmetic instructions.
11619 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
11620 // its only used to compare two uses within the same basic block, which
11621 // means that they'll always either have the same value or one of them
11622 // will have an undefined value.
11623 if (isa<BinaryOperator>(A) ||
11624 isa<CastInst>(A) ||
11626 isa<GetElementPtrInst>(A))
11627 if (Instruction *BI = dyn_cast<Instruction>(B))
11628 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
11631 // Otherwise they may not be equivalent.
11635 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11636 // return the llvm.dbg.declare.
11637 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11638 if (!V->hasNUses(2))
11640 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11642 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11644 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11645 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11652 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11653 Value *Val = SI.getOperand(0);
11654 Value *Ptr = SI.getOperand(1);
11656 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11657 EraseInstFromFunction(SI);
11662 // If the RHS is an alloca with a single use, zapify the store, making the
11664 // If the RHS is an alloca with a two uses, the other one being a
11665 // llvm.dbg.declare, zapify the store and the declare, making the
11666 // alloca dead. We must do this to prevent declare's from affecting
11668 if (!SI.isVolatile()) {
11669 if (Ptr->hasOneUse()) {
11670 if (isa<AllocaInst>(Ptr)) {
11671 EraseInstFromFunction(SI);
11675 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11676 if (isa<AllocaInst>(GEP->getOperand(0))) {
11677 if (GEP->getOperand(0)->hasOneUse()) {
11678 EraseInstFromFunction(SI);
11682 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11683 EraseInstFromFunction(*DI);
11684 EraseInstFromFunction(SI);
11691 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11692 EraseInstFromFunction(*DI);
11693 EraseInstFromFunction(SI);
11699 // Attempt to improve the alignment.
11701 unsigned KnownAlign =
11702 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11704 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11705 SI.getAlignment()))
11706 SI.setAlignment(KnownAlign);
11709 // Do really simple DSE, to catch cases where there are several consecutive
11710 // stores to the same location, separated by a few arithmetic operations. This
11711 // situation often occurs with bitfield accesses.
11712 BasicBlock::iterator BBI = &SI;
11713 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11716 // Don't count debug info directives, lest they affect codegen,
11717 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11718 // It is necessary for correctness to skip those that feed into a
11719 // llvm.dbg.declare, as these are not present when debugging is off.
11720 if (isa<DbgInfoIntrinsic>(BBI) ||
11721 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11726 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11727 // Prev store isn't volatile, and stores to the same location?
11728 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11729 SI.getOperand(1))) {
11732 EraseInstFromFunction(*PrevSI);
11738 // If this is a load, we have to stop. However, if the loaded value is from
11739 // the pointer we're loading and is producing the pointer we're storing,
11740 // then *this* store is dead (X = load P; store X -> P).
11741 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11742 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11743 !SI.isVolatile()) {
11744 EraseInstFromFunction(SI);
11748 // Otherwise, this is a load from some other location. Stores before it
11749 // may not be dead.
11753 // Don't skip over loads or things that can modify memory.
11754 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11759 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11761 // store X, null -> turns into 'unreachable' in SimplifyCFG
11762 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
11763 if (!isa<UndefValue>(Val)) {
11764 SI.setOperand(0, UndefValue::get(Val->getType()));
11765 if (Instruction *U = dyn_cast<Instruction>(Val))
11766 Worklist.Add(U); // Dropped a use.
11769 return 0; // Do not modify these!
11772 // store undef, Ptr -> noop
11773 if (isa<UndefValue>(Val)) {
11774 EraseInstFromFunction(SI);
11779 // If the pointer destination is a cast, see if we can fold the cast into the
11781 if (isa<CastInst>(Ptr))
11782 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11784 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11786 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11790 // If this store is the last instruction in the basic block (possibly
11791 // excepting debug info instructions and the pointer bitcasts that feed
11792 // into them), and if the block ends with an unconditional branch, try
11793 // to move it to the successor block.
11797 } while (isa<DbgInfoIntrinsic>(BBI) ||
11798 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11799 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11800 if (BI->isUnconditional())
11801 if (SimplifyStoreAtEndOfBlock(SI))
11802 return 0; // xform done!
11807 /// SimplifyStoreAtEndOfBlock - Turn things like:
11808 /// if () { *P = v1; } else { *P = v2 }
11809 /// into a phi node with a store in the successor.
11811 /// Simplify things like:
11812 /// *P = v1; if () { *P = v2; }
11813 /// into a phi node with a store in the successor.
11815 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11816 BasicBlock *StoreBB = SI.getParent();
11818 // Check to see if the successor block has exactly two incoming edges. If
11819 // so, see if the other predecessor contains a store to the same location.
11820 // if so, insert a PHI node (if needed) and move the stores down.
11821 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11823 // Determine whether Dest has exactly two predecessors and, if so, compute
11824 // the other predecessor.
11825 pred_iterator PI = pred_begin(DestBB);
11826 BasicBlock *OtherBB = 0;
11827 if (*PI != StoreBB)
11830 if (PI == pred_end(DestBB))
11833 if (*PI != StoreBB) {
11838 if (++PI != pred_end(DestBB))
11841 // Bail out if all the relevant blocks aren't distinct (this can happen,
11842 // for example, if SI is in an infinite loop)
11843 if (StoreBB == DestBB || OtherBB == DestBB)
11846 // Verify that the other block ends in a branch and is not otherwise empty.
11847 BasicBlock::iterator BBI = OtherBB->getTerminator();
11848 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11849 if (!OtherBr || BBI == OtherBB->begin())
11852 // If the other block ends in an unconditional branch, check for the 'if then
11853 // else' case. there is an instruction before the branch.
11854 StoreInst *OtherStore = 0;
11855 if (OtherBr->isUnconditional()) {
11857 // Skip over debugging info.
11858 while (isa<DbgInfoIntrinsic>(BBI) ||
11859 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11860 if (BBI==OtherBB->begin())
11864 // If this isn't a store, or isn't a store to the same location, bail out.
11865 OtherStore = dyn_cast<StoreInst>(BBI);
11866 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11869 // Otherwise, the other block ended with a conditional branch. If one of the
11870 // destinations is StoreBB, then we have the if/then case.
11871 if (OtherBr->getSuccessor(0) != StoreBB &&
11872 OtherBr->getSuccessor(1) != StoreBB)
11875 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11876 // if/then triangle. See if there is a store to the same ptr as SI that
11877 // lives in OtherBB.
11879 // Check to see if we find the matching store.
11880 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11881 if (OtherStore->getOperand(1) != SI.getOperand(1))
11885 // If we find something that may be using or overwriting the stored
11886 // value, or if we run out of instructions, we can't do the xform.
11887 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11888 BBI == OtherBB->begin())
11892 // In order to eliminate the store in OtherBr, we have to
11893 // make sure nothing reads or overwrites the stored value in
11895 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11896 // FIXME: This should really be AA driven.
11897 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11902 // Insert a PHI node now if we need it.
11903 Value *MergedVal = OtherStore->getOperand(0);
11904 if (MergedVal != SI.getOperand(0)) {
11905 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11906 PN->reserveOperandSpace(2);
11907 PN->addIncoming(SI.getOperand(0), SI.getParent());
11908 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11909 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11912 // Advance to a place where it is safe to insert the new store and
11914 BBI = DestBB->getFirstNonPHI();
11915 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11916 OtherStore->isVolatile()), *BBI);
11918 // Nuke the old stores.
11919 EraseInstFromFunction(SI);
11920 EraseInstFromFunction(*OtherStore);
11926 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11927 // Change br (not X), label True, label False to: br X, label False, True
11929 BasicBlock *TrueDest;
11930 BasicBlock *FalseDest;
11931 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11932 !isa<Constant>(X)) {
11933 // Swap Destinations and condition...
11934 BI.setCondition(X);
11935 BI.setSuccessor(0, FalseDest);
11936 BI.setSuccessor(1, TrueDest);
11940 // Cannonicalize fcmp_one -> fcmp_oeq
11941 FCmpInst::Predicate FPred; Value *Y;
11942 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11943 TrueDest, FalseDest)) &&
11944 BI.getCondition()->hasOneUse())
11945 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11946 FPred == FCmpInst::FCMP_OGE) {
11947 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
11948 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
11950 // Swap Destinations and condition.
11951 BI.setSuccessor(0, FalseDest);
11952 BI.setSuccessor(1, TrueDest);
11953 Worklist.Add(Cond);
11957 // Cannonicalize icmp_ne -> icmp_eq
11958 ICmpInst::Predicate IPred;
11959 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11960 TrueDest, FalseDest)) &&
11961 BI.getCondition()->hasOneUse())
11962 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11963 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11964 IPred == ICmpInst::ICMP_SGE) {
11965 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
11966 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
11967 // Swap Destinations and condition.
11968 BI.setSuccessor(0, FalseDest);
11969 BI.setSuccessor(1, TrueDest);
11970 Worklist.Add(Cond);
11977 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11978 Value *Cond = SI.getCondition();
11979 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11980 if (I->getOpcode() == Instruction::Add)
11981 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11982 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11983 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11985 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11987 SI.setOperand(0, I->getOperand(0));
11995 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11996 Value *Agg = EV.getAggregateOperand();
11998 if (!EV.hasIndices())
11999 return ReplaceInstUsesWith(EV, Agg);
12001 if (Constant *C = dyn_cast<Constant>(Agg)) {
12002 if (isa<UndefValue>(C))
12003 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12005 if (isa<ConstantAggregateZero>(C))
12006 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12008 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12009 // Extract the element indexed by the first index out of the constant
12010 Value *V = C->getOperand(*EV.idx_begin());
12011 if (EV.getNumIndices() > 1)
12012 // Extract the remaining indices out of the constant indexed by the
12014 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12016 return ReplaceInstUsesWith(EV, V);
12018 return 0; // Can't handle other constants
12020 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12021 // We're extracting from an insertvalue instruction, compare the indices
12022 const unsigned *exti, *exte, *insi, *inse;
12023 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12024 exte = EV.idx_end(), inse = IV->idx_end();
12025 exti != exte && insi != inse;
12027 if (*insi != *exti)
12028 // The insert and extract both reference distinctly different elements.
12029 // This means the extract is not influenced by the insert, and we can
12030 // replace the aggregate operand of the extract with the aggregate
12031 // operand of the insert. i.e., replace
12032 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12033 // %E = extractvalue { i32, { i32 } } %I, 0
12035 // %E = extractvalue { i32, { i32 } } %A, 0
12036 return ExtractValueInst::Create(IV->getAggregateOperand(),
12037 EV.idx_begin(), EV.idx_end());
12039 if (exti == exte && insi == inse)
12040 // Both iterators are at the end: Index lists are identical. Replace
12041 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12042 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12044 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12045 if (exti == exte) {
12046 // The extract list is a prefix of the insert list. i.e. replace
12047 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12048 // %E = extractvalue { i32, { i32 } } %I, 1
12050 // %X = extractvalue { i32, { i32 } } %A, 1
12051 // %E = insertvalue { i32 } %X, i32 42, 0
12052 // by switching the order of the insert and extract (though the
12053 // insertvalue should be left in, since it may have other uses).
12054 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
12055 EV.idx_begin(), EV.idx_end());
12056 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12060 // The insert list is a prefix of the extract list
12061 // We can simply remove the common indices from the extract and make it
12062 // operate on the inserted value instead of the insertvalue result.
12064 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12065 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12067 // %E extractvalue { i32 } { i32 42 }, 0
12068 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12071 // Can't simplify extracts from other values. Note that nested extracts are
12072 // already simplified implicitely by the above (extract ( extract (insert) )
12073 // will be translated into extract ( insert ( extract ) ) first and then just
12074 // the value inserted, if appropriate).
12078 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12079 /// is to leave as a vector operation.
12080 static bool CheapToScalarize(Value *V, bool isConstant) {
12081 if (isa<ConstantAggregateZero>(V))
12083 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12084 if (isConstant) return true;
12085 // If all elts are the same, we can extract.
12086 Constant *Op0 = C->getOperand(0);
12087 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12088 if (C->getOperand(i) != Op0)
12092 Instruction *I = dyn_cast<Instruction>(V);
12093 if (!I) return false;
12095 // Insert element gets simplified to the inserted element or is deleted if
12096 // this is constant idx extract element and its a constant idx insertelt.
12097 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12098 isa<ConstantInt>(I->getOperand(2)))
12100 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12102 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12103 if (BO->hasOneUse() &&
12104 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12105 CheapToScalarize(BO->getOperand(1), isConstant)))
12107 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12108 if (CI->hasOneUse() &&
12109 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12110 CheapToScalarize(CI->getOperand(1), isConstant)))
12116 /// Read and decode a shufflevector mask.
12118 /// It turns undef elements into values that are larger than the number of
12119 /// elements in the input.
12120 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12121 unsigned NElts = SVI->getType()->getNumElements();
12122 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12123 return std::vector<unsigned>(NElts, 0);
12124 if (isa<UndefValue>(SVI->getOperand(2)))
12125 return std::vector<unsigned>(NElts, 2*NElts);
12127 std::vector<unsigned> Result;
12128 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12129 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12130 if (isa<UndefValue>(*i))
12131 Result.push_back(NElts*2); // undef -> 8
12133 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12137 /// FindScalarElement - Given a vector and an element number, see if the scalar
12138 /// value is already around as a register, for example if it were inserted then
12139 /// extracted from the vector.
12140 static Value *FindScalarElement(Value *V, unsigned EltNo,
12141 LLVMContext *Context) {
12142 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12143 const VectorType *PTy = cast<VectorType>(V->getType());
12144 unsigned Width = PTy->getNumElements();
12145 if (EltNo >= Width) // Out of range access.
12146 return UndefValue::get(PTy->getElementType());
12148 if (isa<UndefValue>(V))
12149 return UndefValue::get(PTy->getElementType());
12150 else if (isa<ConstantAggregateZero>(V))
12151 return Constant::getNullValue(PTy->getElementType());
12152 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12153 return CP->getOperand(EltNo);
12154 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12155 // If this is an insert to a variable element, we don't know what it is.
12156 if (!isa<ConstantInt>(III->getOperand(2)))
12158 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12160 // If this is an insert to the element we are looking for, return the
12162 if (EltNo == IIElt)
12163 return III->getOperand(1);
12165 // Otherwise, the insertelement doesn't modify the value, recurse on its
12167 return FindScalarElement(III->getOperand(0), EltNo, Context);
12168 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12169 unsigned LHSWidth =
12170 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12171 unsigned InEl = getShuffleMask(SVI)[EltNo];
12172 if (InEl < LHSWidth)
12173 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12174 else if (InEl < LHSWidth*2)
12175 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12177 return UndefValue::get(PTy->getElementType());
12180 // Otherwise, we don't know.
12184 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12185 // If vector val is undef, replace extract with scalar undef.
12186 if (isa<UndefValue>(EI.getOperand(0)))
12187 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12189 // If vector val is constant 0, replace extract with scalar 0.
12190 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12191 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12193 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12194 // If vector val is constant with all elements the same, replace EI with
12195 // that element. When the elements are not identical, we cannot replace yet
12196 // (we do that below, but only when the index is constant).
12197 Constant *op0 = C->getOperand(0);
12198 for (unsigned i = 1; i != C->getNumOperands(); ++i)
12199 if (C->getOperand(i) != op0) {
12204 return ReplaceInstUsesWith(EI, op0);
12207 // If extracting a specified index from the vector, see if we can recursively
12208 // find a previously computed scalar that was inserted into the vector.
12209 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12210 unsigned IndexVal = IdxC->getZExtValue();
12211 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
12213 // If this is extracting an invalid index, turn this into undef, to avoid
12214 // crashing the code below.
12215 if (IndexVal >= VectorWidth)
12216 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12218 // This instruction only demands the single element from the input vector.
12219 // If the input vector has a single use, simplify it based on this use
12221 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12222 APInt UndefElts(VectorWidth, 0);
12223 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12224 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12225 DemandedMask, UndefElts)) {
12226 EI.setOperand(0, V);
12231 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12232 return ReplaceInstUsesWith(EI, Elt);
12234 // If the this extractelement is directly using a bitcast from a vector of
12235 // the same number of elements, see if we can find the source element from
12236 // it. In this case, we will end up needing to bitcast the scalars.
12237 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12238 if (const VectorType *VT =
12239 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12240 if (VT->getNumElements() == VectorWidth)
12241 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12242 IndexVal, Context))
12243 return new BitCastInst(Elt, EI.getType());
12247 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12248 // Push extractelement into predecessor operation if legal and
12249 // profitable to do so
12250 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12251 if (I->hasOneUse() &&
12252 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
12254 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
12255 EI.getName()+".lhs");
12257 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
12258 EI.getName()+".rhs");
12259 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12261 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12262 // Extracting the inserted element?
12263 if (IE->getOperand(2) == EI.getOperand(1))
12264 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12265 // If the inserted and extracted elements are constants, they must not
12266 // be the same value, extract from the pre-inserted value instead.
12267 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
12268 Worklist.AddValue(EI.getOperand(0));
12269 EI.setOperand(0, IE->getOperand(0));
12272 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12273 // If this is extracting an element from a shufflevector, figure out where
12274 // it came from and extract from the appropriate input element instead.
12275 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12276 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12278 unsigned LHSWidth =
12279 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12281 if (SrcIdx < LHSWidth)
12282 Src = SVI->getOperand(0);
12283 else if (SrcIdx < LHSWidth*2) {
12284 SrcIdx -= LHSWidth;
12285 Src = SVI->getOperand(1);
12287 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12289 return ExtractElementInst::Create(Src,
12290 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx,
12294 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12299 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12300 /// elements from either LHS or RHS, return the shuffle mask and true.
12301 /// Otherwise, return false.
12302 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12303 std::vector<Constant*> &Mask,
12304 LLVMContext *Context) {
12305 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12306 "Invalid CollectSingleShuffleElements");
12307 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12309 if (isa<UndefValue>(V)) {
12310 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12312 } else if (V == LHS) {
12313 for (unsigned i = 0; i != NumElts; ++i)
12314 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12316 } else if (V == RHS) {
12317 for (unsigned i = 0; i != NumElts; ++i)
12318 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
12320 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12321 // If this is an insert of an extract from some other vector, include it.
12322 Value *VecOp = IEI->getOperand(0);
12323 Value *ScalarOp = IEI->getOperand(1);
12324 Value *IdxOp = IEI->getOperand(2);
12326 if (!isa<ConstantInt>(IdxOp))
12328 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12330 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12331 // Okay, we can handle this if the vector we are insertinting into is
12332 // transitively ok.
12333 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12334 // If so, update the mask to reflect the inserted undef.
12335 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
12338 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12339 if (isa<ConstantInt>(EI->getOperand(1)) &&
12340 EI->getOperand(0)->getType() == V->getType()) {
12341 unsigned ExtractedIdx =
12342 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12344 // This must be extracting from either LHS or RHS.
12345 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12346 // Okay, we can handle this if the vector we are insertinting into is
12347 // transitively ok.
12348 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12349 // If so, update the mask to reflect the inserted value.
12350 if (EI->getOperand(0) == LHS) {
12351 Mask[InsertedIdx % NumElts] =
12352 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12354 assert(EI->getOperand(0) == RHS);
12355 Mask[InsertedIdx % NumElts] =
12356 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
12365 // TODO: Handle shufflevector here!
12370 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12371 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12372 /// that computes V and the LHS value of the shuffle.
12373 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12374 Value *&RHS, LLVMContext *Context) {
12375 assert(isa<VectorType>(V->getType()) &&
12376 (RHS == 0 || V->getType() == RHS->getType()) &&
12377 "Invalid shuffle!");
12378 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12380 if (isa<UndefValue>(V)) {
12381 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12383 } else if (isa<ConstantAggregateZero>(V)) {
12384 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
12386 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12387 // If this is an insert of an extract from some other vector, include it.
12388 Value *VecOp = IEI->getOperand(0);
12389 Value *ScalarOp = IEI->getOperand(1);
12390 Value *IdxOp = IEI->getOperand(2);
12392 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12393 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12394 EI->getOperand(0)->getType() == V->getType()) {
12395 unsigned ExtractedIdx =
12396 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12397 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12399 // Either the extracted from or inserted into vector must be RHSVec,
12400 // otherwise we'd end up with a shuffle of three inputs.
12401 if (EI->getOperand(0) == RHS || RHS == 0) {
12402 RHS = EI->getOperand(0);
12403 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12404 Mask[InsertedIdx % NumElts] =
12405 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
12409 if (VecOp == RHS) {
12410 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12412 // Everything but the extracted element is replaced with the RHS.
12413 for (unsigned i = 0; i != NumElts; ++i) {
12414 if (i != InsertedIdx)
12415 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
12420 // If this insertelement is a chain that comes from exactly these two
12421 // vectors, return the vector and the effective shuffle.
12422 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12424 return EI->getOperand(0);
12429 // TODO: Handle shufflevector here!
12431 // Otherwise, can't do anything fancy. Return an identity vector.
12432 for (unsigned i = 0; i != NumElts; ++i)
12433 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12437 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12438 Value *VecOp = IE.getOperand(0);
12439 Value *ScalarOp = IE.getOperand(1);
12440 Value *IdxOp = IE.getOperand(2);
12442 // Inserting an undef or into an undefined place, remove this.
12443 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12444 ReplaceInstUsesWith(IE, VecOp);
12446 // If the inserted element was extracted from some other vector, and if the
12447 // indexes are constant, try to turn this into a shufflevector operation.
12448 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12449 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12450 EI->getOperand(0)->getType() == IE.getType()) {
12451 unsigned NumVectorElts = IE.getType()->getNumElements();
12452 unsigned ExtractedIdx =
12453 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12454 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12456 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12457 return ReplaceInstUsesWith(IE, VecOp);
12459 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12460 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12462 // If we are extracting a value from a vector, then inserting it right
12463 // back into the same place, just use the input vector.
12464 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12465 return ReplaceInstUsesWith(IE, VecOp);
12467 // If this insertelement isn't used by some other insertelement, turn it
12468 // (and any insertelements it points to), into one big shuffle.
12469 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12470 std::vector<Constant*> Mask;
12472 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12473 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12474 // We now have a shuffle of LHS, RHS, Mask.
12475 return new ShuffleVectorInst(LHS, RHS,
12476 ConstantVector::get(Mask));
12481 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12482 APInt UndefElts(VWidth, 0);
12483 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12484 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12491 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12492 Value *LHS = SVI.getOperand(0);
12493 Value *RHS = SVI.getOperand(1);
12494 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12496 bool MadeChange = false;
12498 // Undefined shuffle mask -> undefined value.
12499 if (isa<UndefValue>(SVI.getOperand(2)))
12500 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12502 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12504 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12507 APInt UndefElts(VWidth, 0);
12508 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12509 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12510 LHS = SVI.getOperand(0);
12511 RHS = SVI.getOperand(1);
12515 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12516 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12517 if (LHS == RHS || isa<UndefValue>(LHS)) {
12518 if (isa<UndefValue>(LHS) && LHS == RHS) {
12519 // shuffle(undef,undef,mask) -> undef.
12520 return ReplaceInstUsesWith(SVI, LHS);
12523 // Remap any references to RHS to use LHS.
12524 std::vector<Constant*> Elts;
12525 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12526 if (Mask[i] >= 2*e)
12527 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12529 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12530 (Mask[i] < e && isa<UndefValue>(LHS))) {
12531 Mask[i] = 2*e; // Turn into undef.
12532 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12534 Mask[i] = Mask[i] % e; // Force to LHS.
12535 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
12539 SVI.setOperand(0, SVI.getOperand(1));
12540 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12541 SVI.setOperand(2, ConstantVector::get(Elts));
12542 LHS = SVI.getOperand(0);
12543 RHS = SVI.getOperand(1);
12547 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12548 bool isLHSID = true, isRHSID = true;
12550 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12551 if (Mask[i] >= e*2) continue; // Ignore undef values.
12552 // Is this an identity shuffle of the LHS value?
12553 isLHSID &= (Mask[i] == i);
12555 // Is this an identity shuffle of the RHS value?
12556 isRHSID &= (Mask[i]-e == i);
12559 // Eliminate identity shuffles.
12560 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12561 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12563 // If the LHS is a shufflevector itself, see if we can combine it with this
12564 // one without producing an unusual shuffle. Here we are really conservative:
12565 // we are absolutely afraid of producing a shuffle mask not in the input
12566 // program, because the code gen may not be smart enough to turn a merged
12567 // shuffle into two specific shuffles: it may produce worse code. As such,
12568 // we only merge two shuffles if the result is one of the two input shuffle
12569 // masks. In this case, merging the shuffles just removes one instruction,
12570 // which we know is safe. This is good for things like turning:
12571 // (splat(splat)) -> splat.
12572 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12573 if (isa<UndefValue>(RHS)) {
12574 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12576 std::vector<unsigned> NewMask;
12577 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12578 if (Mask[i] >= 2*e)
12579 NewMask.push_back(2*e);
12581 NewMask.push_back(LHSMask[Mask[i]]);
12583 // If the result mask is equal to the src shuffle or this shuffle mask, do
12584 // the replacement.
12585 if (NewMask == LHSMask || NewMask == Mask) {
12586 unsigned LHSInNElts =
12587 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12588 std::vector<Constant*> Elts;
12589 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12590 if (NewMask[i] >= LHSInNElts*2) {
12591 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12593 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), NewMask[i]));
12596 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12597 LHSSVI->getOperand(1),
12598 ConstantVector::get(Elts));
12603 return MadeChange ? &SVI : 0;
12609 /// TryToSinkInstruction - Try to move the specified instruction from its
12610 /// current block into the beginning of DestBlock, which can only happen if it's
12611 /// safe to move the instruction past all of the instructions between it and the
12612 /// end of its block.
12613 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12614 assert(I->hasOneUse() && "Invariants didn't hold!");
12616 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12617 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12620 // Do not sink alloca instructions out of the entry block.
12621 if (isa<AllocaInst>(I) && I->getParent() ==
12622 &DestBlock->getParent()->getEntryBlock())
12625 // We can only sink load instructions if there is nothing between the load and
12626 // the end of block that could change the value.
12627 if (I->mayReadFromMemory()) {
12628 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12630 if (Scan->mayWriteToMemory())
12634 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12636 CopyPrecedingStopPoint(I, InsertPos);
12637 I->moveBefore(InsertPos);
12643 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12644 /// all reachable code to the worklist.
12646 /// This has a couple of tricks to make the code faster and more powerful. In
12647 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12648 /// them to the worklist (this significantly speeds up instcombine on code where
12649 /// many instructions are dead or constant). Additionally, if we find a branch
12650 /// whose condition is a known constant, we only visit the reachable successors.
12652 static bool AddReachableCodeToWorklist(BasicBlock *BB,
12653 SmallPtrSet<BasicBlock*, 64> &Visited,
12655 const TargetData *TD) {
12656 bool MadeIRChange = false;
12657 SmallVector<BasicBlock*, 256> Worklist;
12658 Worklist.push_back(BB);
12660 std::vector<Instruction*> InstrsForInstCombineWorklist;
12661 InstrsForInstCombineWorklist.reserve(128);
12663 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
12665 while (!Worklist.empty()) {
12666 BB = Worklist.back();
12667 Worklist.pop_back();
12669 // We have now visited this block! If we've already been here, ignore it.
12670 if (!Visited.insert(BB)) continue;
12672 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12673 Instruction *Inst = BBI++;
12675 // DCE instruction if trivially dead.
12676 if (isInstructionTriviallyDead(Inst)) {
12678 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
12679 Inst->eraseFromParent();
12683 // ConstantProp instruction if trivially constant.
12684 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
12685 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12686 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
12688 Inst->replaceAllUsesWith(C);
12690 Inst->eraseFromParent();
12697 // See if we can constant fold its operands.
12698 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
12700 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
12701 if (CE == 0) continue;
12703 // If we already folded this constant, don't try again.
12704 if (!FoldedConstants.insert(CE))
12708 ConstantFoldConstantExpression(CE, BB->getContext(), TD);
12709 if (NewC && NewC != CE) {
12711 MadeIRChange = true;
12717 InstrsForInstCombineWorklist.push_back(Inst);
12720 // Recursively visit successors. If this is a branch or switch on a
12721 // constant, only visit the reachable successor.
12722 TerminatorInst *TI = BB->getTerminator();
12723 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12724 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12725 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12726 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12727 Worklist.push_back(ReachableBB);
12730 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12731 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12732 // See if this is an explicit destination.
12733 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12734 if (SI->getCaseValue(i) == Cond) {
12735 BasicBlock *ReachableBB = SI->getSuccessor(i);
12736 Worklist.push_back(ReachableBB);
12740 // Otherwise it is the default destination.
12741 Worklist.push_back(SI->getSuccessor(0));
12746 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12747 Worklist.push_back(TI->getSuccessor(i));
12750 // Once we've found all of the instructions to add to instcombine's worklist,
12751 // add them in reverse order. This way instcombine will visit from the top
12752 // of the function down. This jives well with the way that it adds all uses
12753 // of instructions to the worklist after doing a transformation, thus avoiding
12754 // some N^2 behavior in pathological cases.
12755 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
12756 InstrsForInstCombineWorklist.size());
12758 return MadeIRChange;
12761 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12762 MadeIRChange = false;
12764 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12765 << F.getNameStr() << "\n");
12768 // Do a depth-first traversal of the function, populate the worklist with
12769 // the reachable instructions. Ignore blocks that are not reachable. Keep
12770 // track of which blocks we visit.
12771 SmallPtrSet<BasicBlock*, 64> Visited;
12772 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12774 // Do a quick scan over the function. If we find any blocks that are
12775 // unreachable, remove any instructions inside of them. This prevents
12776 // the instcombine code from having to deal with some bad special cases.
12777 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12778 if (!Visited.count(BB)) {
12779 Instruction *Term = BB->getTerminator();
12780 while (Term != BB->begin()) { // Remove instrs bottom-up
12781 BasicBlock::iterator I = Term; --I;
12783 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12784 // A debug intrinsic shouldn't force another iteration if we weren't
12785 // going to do one without it.
12786 if (!isa<DbgInfoIntrinsic>(I)) {
12788 MadeIRChange = true;
12791 // If I is not void type then replaceAllUsesWith undef.
12792 // This allows ValueHandlers and custom metadata to adjust itself.
12793 if (!I->getType()->isVoidTy())
12794 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12795 I->eraseFromParent();
12800 while (!Worklist.isEmpty()) {
12801 Instruction *I = Worklist.RemoveOne();
12802 if (I == 0) continue; // skip null values.
12804 // Check to see if we can DCE the instruction.
12805 if (isInstructionTriviallyDead(I)) {
12806 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12807 EraseInstFromFunction(*I);
12809 MadeIRChange = true;
12813 // Instruction isn't dead, see if we can constant propagate it.
12814 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
12815 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
12816 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
12818 // Add operands to the worklist.
12819 ReplaceInstUsesWith(*I, C);
12821 EraseInstFromFunction(*I);
12822 MadeIRChange = true;
12826 // See if we can trivially sink this instruction to a successor basic block.
12827 if (I->hasOneUse()) {
12828 BasicBlock *BB = I->getParent();
12829 Instruction *UserInst = cast<Instruction>(I->use_back());
12830 BasicBlock *UserParent;
12832 // Get the block the use occurs in.
12833 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
12834 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
12836 UserParent = UserInst->getParent();
12838 if (UserParent != BB) {
12839 bool UserIsSuccessor = false;
12840 // See if the user is one of our successors.
12841 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12842 if (*SI == UserParent) {
12843 UserIsSuccessor = true;
12847 // If the user is one of our immediate successors, and if that successor
12848 // only has us as a predecessors (we'd have to split the critical edge
12849 // otherwise), we can keep going.
12850 if (UserIsSuccessor && UserParent->getSinglePredecessor())
12851 // Okay, the CFG is simple enough, try to sink this instruction.
12852 MadeIRChange |= TryToSinkInstruction(I, UserParent);
12856 // Now that we have an instruction, try combining it to simplify it.
12857 Builder->SetInsertPoint(I->getParent(), I);
12862 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
12863 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
12865 if (Instruction *Result = visit(*I)) {
12867 // Should we replace the old instruction with a new one?
12869 DEBUG(errs() << "IC: Old = " << *I << '\n'
12870 << " New = " << *Result << '\n');
12872 // Everything uses the new instruction now.
12873 I->replaceAllUsesWith(Result);
12875 // Push the new instruction and any users onto the worklist.
12876 Worklist.Add(Result);
12877 Worklist.AddUsersToWorkList(*Result);
12879 // Move the name to the new instruction first.
12880 Result->takeName(I);
12882 // Insert the new instruction into the basic block...
12883 BasicBlock *InstParent = I->getParent();
12884 BasicBlock::iterator InsertPos = I;
12886 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12887 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12890 InstParent->getInstList().insert(InsertPos, Result);
12892 EraseInstFromFunction(*I);
12895 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
12896 << " New = " << *I << '\n');
12899 // If the instruction was modified, it's possible that it is now dead.
12900 // if so, remove it.
12901 if (isInstructionTriviallyDead(I)) {
12902 EraseInstFromFunction(*I);
12905 Worklist.AddUsersToWorkList(*I);
12908 MadeIRChange = true;
12913 return MadeIRChange;
12917 bool InstCombiner::runOnFunction(Function &F) {
12918 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12919 Context = &F.getContext();
12920 TD = getAnalysisIfAvailable<TargetData>();
12923 /// Builder - This is an IRBuilder that automatically inserts new
12924 /// instructions into the worklist when they are created.
12925 IRBuilder<true, TargetFolder, InstCombineIRInserter>
12926 TheBuilder(F.getContext(), TargetFolder(TD, F.getContext()),
12927 InstCombineIRInserter(Worklist));
12928 Builder = &TheBuilder;
12930 bool EverMadeChange = false;
12932 // Iterate while there is work to do.
12933 unsigned Iteration = 0;
12934 while (DoOneIteration(F, Iteration++))
12935 EverMadeChange = true;
12938 return EverMadeChange;
12941 FunctionPass *llvm::createInstructionCombiningPass() {
12942 return new InstCombiner();